feat: ddddocr-rs 完成 core/ort/tract2 规范整改与发布

准备

  - core：内置官方字符集（OLD/BETA）与 ModelMetadata::from_builtin_*
  构造器
  - ort：导出 Session、实现 Info、修正 cuda feature 接线、共享推理工具
  - tract2：包名更名（原 ddddocr-tract 已被占用）并完成规范整改
  - 集成测试按领域拆分（ocr / det / slide / common / api_surface）
  - 发布准备：Cargo.toml 元数据、workspace 版本 0.2.4、LICENSE/
  NOTICE、README 模型下载说明

Cargo.toml

@@ -1,11 +1,29 @@
-[package]
-name = "ddddocr-rs"
-version = "0.1.0"
+[workspace]
+resolver = "2"
+members = [
+    "ddddocr-core", "ddddocr-ort",
+    "ddddocr-tract2",
+]
+
+[workspace.package]
+version = "0.2.4"
 edition = "2024"
 license = "MIT OR Apache-2.0"
 
-[dependencies]
-tract-onnx = { version = "0.21.10" }
-anyhow = "1.0.102"
+
+[workspace.dependencies]
+tract-onnx = "0.23.4"
+tract-linalg = { version = "0.23.4",features = ["multithread-mm"]}
+ort = "2.0.0-rc.12"
+ndarray = "0.17.2"
+
 image = "0.25.10"
 base64 = "0.22.1"
+imageproc = { version = "0.26.2", default-features = true }
+
+serde = { version = "1.0.228", features = ["derive"] }
+serde_json = "1.0.150"
+
+anyhow = "1.0.102"
+thiserror = "1.0" # 刚好可以开始接入你需要的标准库错误处理
+tracing = "0.1.44" # 埋入日志打点(后续需要继续优化，现在只是简单尝试)

NOTICE

@@ -3,4 +3,6 @@ Copyright 2026 CNWei
 
 This product includes software developed by:
 - sml2h3 (ddddocr) - Model weights and original logic.
-- Sonos (tract) - ONNX inference engine.
+- Sonos (tract) - ONNX inference engine.
+- pyke.io (ort) - Rust bindings for ONNX Runtime.
+- Microsoft (ONNX Runtime contributors) - ONNX inference engine.

README.md

@@ -1,10 +1,135 @@
 # ddddocr-rs
 
-带带弟弟 OCR (ddddocr) 的 Rust 移植版。高性能、低占用，支持多种验证码识别与检测。
+[带带弟弟 OCR（ddddocr）](https://github.com/sml2h3/ddddocr) 的 Rust 移植版：提供验证码 OCR 识别、目标检测与滑块匹配能力。核心库与推理引擎解耦，同一套 API 可切换 [Tract](https://github.com/sonos/tract) 或 [ONNX Runtime](https://onnxruntime.ai/) 后端。
 
+## 架构
 
+项目为 Cargo workspace，包含三个 crate：
 
-鸣谢 (Credits)
+| crate | 说明 |
+|---|---|
+| `ddddocr-core` | 引擎无关的核心库：OCR / 目标检测 / 滑块匹配、模型元数据与图像预处理 |
+| `ddddocr-ort` | ONNX Runtime 推理后端（可选 `cuda` feature 启用 GPU 加速） |
+| `ddddocr-tract2` | Tract（纯 Rust）推理后端 |
 
-- 本项目是 [ddddocr](https://github.com/sml2h3/ddddocr) 的 Rust 移植版本，原作者为 sml2h3。衷心感谢原作者对 OCR 社区做出的杰出贡献。
-- 推理引擎基于 [tract (Sonos)](https://github.com/sonos/tract)。感谢其为 Rust 生态提供的轻量级推理方案。
+`ddddocr-core` 通过 `traits::InferenceEngine` / `OcrEngine` / `DetEngine` 抽象推理能力，由 `ddddocr-ort` / `ddddocr-tract2` 实现，业务代码只依赖核心库接口即可。
+
+## 快速开始
+
+在 Cargo.toml 中添加：
+
+```toml
+[dependencies]
+ddddocr-core = "0.2"
+ddddocr-ort = "0.2"   # 或 ddddocr-tract2 = "0.2"
+```
+
+核心库自带一个不依赖真实模型的演示示例（用假引擎演示完整调用链）：
+
+```bash
+cargo run -p ddddocr-core --example quick_start
+```
+
+真实推理需要自行准备 ONNX 模型文件（crate 包内不包含模型）。以 ORT 后端为例：
+
+```rust
+use ddddocr_core::traits::{Info, Loader};
+use ddddocr_core::{ModelMetadata, Normalization, Ocr, Resize};
+
+// 1. 构建会话（等价写法：ddddocr_tract2::loader::ModelLoader）
+let session = ddddocr_ort::loader::ModelLoader::default()
+    .build_for_path("models/common.onnx")?;
+
+// 2. 组装 OCR 运行时
+let metadata = ModelMetadata::from_static_slice(
+    &["a", "b"],              // 字符集
+    false,                    // 非单字模型
+    Resize::DynamicWidth(64), // 高度固定 64、宽度等比缩放
+    1,                        // 灰度单通道
+    Normalization::MinusOneToOne,
+);
+let ocr = ddddocr_ort::OcrRuntime::new(session, metadata);
+
+// 3. 识别图片
+let result = Ocr::builder().build_with(&ocr).predict(&image)?;
+println!("识别结果: {}", result);
+```
+
+目标检测与滑块匹配类似：`Detector::new(&engine).predict(&image)` 返回检测框列表；`Slider::new().slide_match(target, background, simple_target)` 返回匹配坐标与置信度。
+
+## 模型下载
+
+OCR / 目标检测需要 ONNX 模型文件。官方模型位于 [ddddocr 仓库 `ddddocr/` 目录](https://github.com/sml2h3/ddddocr/blob/master/ddddocr/)，下载后放入仓库根目录的 `models/` 文件夹：
+
+| 文件 | 大小 | 用途 | 直链 |
+|---|---|---|---|
+| `common.onnx` | 约 51.6 MB | 新版 OCR 模型（默认） | [下载](https://raw.githubusercontent.com/sml2h3/ddddocr/master/ddddocr/common.onnx) |
+| `common_det.onnx` | 约 19.2 MB | 目标检测模型 | [下载](https://raw.githubusercontent.com/sml2h3/ddddocr/master/ddddocr/common_det.onnx) |
+| `common_old.onnx` | 约 13 MB | 旧版 OCR 模型 | [下载](https://raw.githubusercontent.com/sml2h3/ddddocr/master/ddddocr/common_old.onnx) |
+
+> 注意：`ddddocr-tract2` 不支持旧版模型 `common_old.onnx`，请使用 `common.onnx`。
+>
+> 字符集配对：`common.onnx` 对应内置 Beta 字符集（`Charset` / `ModelMetadata::from_builtin_beta`，归一化 `MinusOneToOne`）；`common_old.onnx` 对应旧版字符集（`from_builtin_old`，归一化 `ZeroToOne`）。
+>
+> 仓库测试使用的 `common_sml2h3_f32.onnx`、`common_huashi666_i64.onnx` 为社区转换的 f32 / i64 变体模型（与 `common.onnx` 同源），不在上述官方目录中，需自行获取。
+
+## 字符集与内置默认值
+
+OCR 模型的字符集（token 列表）是模型的一部分：字符 `i` 对应模型输出 logits 的第 `i` 列，必须与模型训练时一致，否则识别结果会错位。
+
+字符集有两种提供方式：
+
+- 通过模型元数据 JSON 的 `charset` 字段（`Metadata::from_json_str` / `from_json_bytes` 自动解析）；
+- 代码内显式指定：`ModelMetadata::from_static_slice(&["", "a", "b"], ...)` 或 `Charset::new(...)`。
+
+`ddddocr-core` 内置官方模型的默认字符集（旧版 `CHARSET_OLD` 与 Beta `CHARSET_BETA`），作为免配置的快速通道。字符集数据独立存放在 `ddddocr-core/src/ocr/builtin.rs` 私有模块中（与业务逻辑分离），通过 `ModelMetadata` 的公开构造方法使用：
+
+```rust
+use ddddocr_core::{ModelMetadata, Normalization, Resize};
+
+// 官方旧版模型配套
+let meta = ModelMetadata::from_builtin_old(
+    false,
+    Resize::DynamicWidth(64),
+    1,
+    Normalization::ZeroToOne,
+);
+
+// 官方 Beta 模型配套
+let meta = ModelMetadata::from_builtin_beta(
+    false,
+    Resize::DynamicWidth(64),
+    1,
+    Normalization::MinusOneToOne,
+);
+```
+
+数据内置于代码，不依赖外部文件。若使用自有模型，请仍以元数据 JSON 或 `from_static_slice` 指定匹配的字符集。
+
+## 滑块匹配核心知识点
+
+项目实现两种匹配模式，底层逻辑与 OpenCV 对齐：
+
+| 模式 | 算法原理 | 适用场景 | 备注 |
+|---|---|---|---|
+| 边缘模式（Edge-based） | 基于 Canny 边缘检测提取轮廓后匹配 | **推荐方案**，适用于绝大多数拼图滑块 | 天然免疫拼图周边透明/黑色留白干扰，坐标最精准 |
+| 简单模式（Simple/Gray） | 基于灰度像素值做归一化互相关（NCC） | 无明显边缘、靠颜色差异识别的场景 | 对背景和透明边框敏感，可能存在重心偏移 |
+
+关键点：
+
+- 简单模式采用归一化互相关（NCC），与 OpenCV 的 `TM_CCORR_NORMED` 完全一致。
+- 若拼图原图四周带透明留白，本项目会将留白计入整张图片框中心；OpenCV 默认的 CCOEFF 会做均值中心化削弱留白影响。
+- 图像预处理与 Python 链路保持一致：灰度权重采用 OpenCV 标准感光公式 `0.299R + 0.587G + 0.114B`；PNG 转 RGB 时透明区域填充为黑色；返回坐标为匹配区域的几何中心 `(x + w/2, y + h/2)`。
+
+> 若识别结果在 X 轴上有约 10px 固定误差，通常是滑块原图自带透明边距所致。此时请确保 `simple_target = false`，边缘模式会自动锁定拼图实体并忽略留白干扰。
+
+## 致谢
+
+- [sml2h3 (ddddocr)](https://github.com/sml2h3/ddddocr) - 模型权重与原始逻辑。
+- [Sonos (tract)](https://github.com/sonos/tract) - 纯 Rust ONNX 推理引擎。
+- [pyke.io (ort)](https://github.com/pykeio/ort) - ONNX Runtime 的 Rust 绑定。
+- [Microsoft (ONNX Runtime)](https://onnxruntime.ai/) - ONNX 推理引擎。
+
+## License
+
+`MIT OR Apache-2.0`，许可证文本见仓库根目录的 `LICENSE-MIT` 与 `LICENSE-APACHE`。

ddddocr-core/Cargo.toml

@@ -0,0 +1,18 @@
+[package]
+name = "ddddocr-core"
+version = { workspace = true }
+edition = { workspace = true }
+license = { workspace = true }
+description = "ddddocr-rs 的核心库：引擎无关的 OCR、目标检测与滑块匹配实现"
+keywords = ["ocr", "captcha", "ddddocr", "onnx", "image"]
+categories = ["multimedia::images", "computer-vision"]
+# repository = "https://github.com/<用户名>/<仓库名>"  # 发布前请补充
+readme = "../README.md"
+
+[dependencies]
+ndarray = { workspace = true }
+base64 = { workspace = true }
+image = { workspace = true }
+imageproc = { workspace = true }
+thiserror = { workspace = true }
+tracing = { workspace = true }

ddddocr-core/examples/quick_start.rs

@@ -0,0 +1,63 @@
+//! 快速开始示例：演示 ddddocr-core 与引擎 crate 的解耦用法。
+//!
+//! 运行：`cargo run -p ddddocr-core --example quick_start`
+
+use ddddocr_core::error::{Result, TensorError};
+use ddddocr_core::traits::{InferenceEngine, Info, OcrEngine};
+use ddddocr_core::types::{ModelInfo, TensorInfo};
+use ddddocr_core::{ModelMetadata, Normalization, OcrBuilder, OcrOutput, Resize};
+
+/// 演示引擎：只实现接口，不接入真实 ONNX 运行时。
+struct DemoEngine {
+    meta: ModelMetadata,
+}
+
+impl Info for DemoEngine {
+    fn input_info(&self) -> Result<Vec<TensorInfo>> {
+        Ok(vec![])
+    }
+    fn output_info(&self) -> Result<Vec<TensorInfo>> {
+        Ok(vec![])
+    }
+    fn model_info(&self) -> Result<ModelInfo> {
+        Ok(ModelInfo {
+            inputs: vec![],
+            outputs: vec![],
+            providers: None,
+        })
+    }
+}
+
+impl InferenceEngine for DemoEngine {
+    type Output = OcrOutput;
+
+    fn inference(&self, input: ndarray::Array4<f32>) -> Result<Self::Output, TensorError> {
+        // 用全零 logits 模拟推理输出：[Steps, Classes]
+        let steps = input.shape()[2];
+        let classes = self.meta.charset.size();
+        Ok(OcrOutput::Logits(ndarray::Array2::zeros((steps, classes))))
+    }
+}
+
+impl OcrEngine for DemoEngine {
+    fn metadata(&self) -> &ModelMetadata {
+        &self.meta
+    }
+}
+
+fn main() {
+    let engine = DemoEngine {
+        meta: ModelMetadata::from_static_slice(
+            &["", "a", "b"],
+            false,
+            Resize::Fixed(64, 64),
+            1,
+            Normalization::ZeroToOne,
+        ),
+    };
+
+    let ocr = OcrBuilder::new().probability(true).build_with(&engine);
+    let image = image::DynamicImage::new_luma8(64, 64);
+    let result = ocr.predict(&image).expect("识别失败");
+    println!("识别结果: {result}");
+}

ddddocr-core/src/det.rs

@@ -0,0 +1,7 @@
+//! 目标检测模块：检测器构建器与执行入口。
+
+mod builder;
+mod executor;
+
+pub use builder::DetBuilder;
+pub use executor::{DetectionResult, Detector};

ddddocr-core/src/det/builder.rs

@@ -0,0 +1,15 @@
+//! 检测器构建器。
+
+use crate::det::executor::Detector;
+use crate::traits::DetEngine;
+
+/// 检测器构建器，通过 [`crate::Detector::builder`] 创建。
+#[derive(Default)]
+pub struct DetBuilder;
+
+impl DetBuilder {
+    /// 绑定检测引擎会话并构建 Detector。
+    pub fn build_with<E: DetEngine>(self, runtime: &E) -> Detector<'_> {
+        Detector { runtime }
+    }
+}

ddddocr-core/src/det/executor.rs

@@ -0,0 +1,304 @@
+//! 目标检测执行器：检测后处理与预测入口。
+
+use crate::error::{Result, TensorError};
+use image::{DynamicImage, GenericImageView, imageops::FilterType};
+use ndarray::{Array2, Array3, Array4, Axis, prelude::*, s};
+use std::fmt;
+
+use crate::traits::DetEngine;
+use crate::{DetBuilder, DetOutput};
+
+/// 目标检测结果：原图像素坐标系下的边界框、置信度与类别 ID。
+#[derive(Debug, Clone, Copy)]
+pub struct DetectionResult {
+    /// 左上角 x 坐标。
+    pub x1: i32,
+    /// 左上角 y 坐标。
+    pub y1: i32,
+    /// 右下角 x 坐标。
+    pub x2: i32,
+    /// 右下角 y 坐标。
+    pub y2: i32,
+    /// 置信度。
+    pub score: f32,
+    /// 类别 ID。
+    pub class_id: u32,
+}
+
+impl fmt::Display for DetectionResult {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        // 结构体只管自己这一行怎么显示，不用管外部的索引 [i]
+        write!(
+            f,
+            "x1={}, y1={}, x2={}, y2={}, 分数={:.4}, 类别ID={}",
+            self.x1, self.y1, self.x2, self.y2, self.score, self.class_id
+        )
+    }
+}
+
+/// 目标检测器：对输入图像执行检测并返回结果。
+pub struct Detector<'a> {
+    pub(crate) runtime: &'a dyn DetEngine,
+}
+
+impl<'a> Detector<'a> {
+    /// 绑定检测引擎会话创建检测器。
+    pub fn new(runtime: &'a dyn DetEngine) -> Self {
+        Detector { runtime }
+    }
+    /// 创建检测器构建器。
+    pub fn builder() -> DetBuilder {
+        DetBuilder
+    }
+}
+impl<'a> Detector<'a> {
+    /// 对输入图像执行目标检测，返回检测框列表。
+    pub fn predict(&self, image: &DynamicImage) -> Result<Vec<DetectionResult>> {
+        // Rust 中通常在调用层处理文件/PIL转换，这里直接进入核心逻辑
+        Ok(self.get_bbox(image)?)
+    }
+    fn preproc(&self, image: &DynamicImage, input_size: (u32, u32)) -> (Array4<f32>, f32) {
+        let (target_h, target_w) = input_size;
+        let (img_w, img_h) = image.dimensions();
+
+        // 计算缩放比例 (Letterbox)
+        let r = (target_h as f32 / img_h as f32).min(target_w as f32 / img_w as f32);
+        let new_h = (img_h as f32 * r) as u32;
+        let new_w = (img_w as f32 * r) as u32;
+
+        // Resize 图像
+        let resized = image.resize_exact(new_w, new_h, FilterType::Triangle);
+        // 2. 关键：将 DynamicImage 显式转换为 RgbImage (Rgb<u8>)
+        let resized_rgb = resized.to_rgb8();
+        // 创建 114 灰度填充的背景
+        let mut base_img =
+            image::ImageBuffer::from_pixel(target_w, target_h, image::Rgb([114u8, 114, 114]));
+
+        // 将 resize 后的图像覆盖到左上角 (类似于原始代码中的 padded_img[:h, :w])
+        image::imageops::overlay(&mut base_img, &resized_rgb, 0, 0);
+
+        // 优化：直接获取底层的扁平 raw buffer，比 enumerate_pixels() 快得多
+        let raw_samples = base_img.as_flat_samples();
+        let slice = raw_samples.as_slice();
+
+        // 构造 NCHW Tensor
+        let mut array = Array4::<f32>::zeros((1, 3, target_h as usize, target_w as usize));
+
+        // 用连续的 stride 步长进行写入，提高 CPU 缓存利用率
+        for y in 0..target_h as usize {
+            for x in 0..target_w as usize {
+                let idx = (y * target_w as usize + x) * 3;
+                // BGR 赋值
+                array[[0, 0, y, x]] = slice[idx + 2] as f32; // B
+                array[[0, 1, y, x]] = slice[idx + 1] as f32; // G
+                array[[0, 2, y, x]] = slice[idx] as f32; // R
+            }
+        }
+
+        (array, r)
+    }
+
+    fn demo_postprocess(&self, mut outputs: Array3<f32>, img_size: (i32, i32)) -> Array3<f32> {
+        let strides = [8, 16, 32];
+
+        // 遍历每一个 Batch（支持动态 Batch 推理）
+        for mut batch in outputs.axis_iter_mut(Axis(0)) {
+            let mut offset = 0;
+
+            for &stride in &strides {
+                // 计算当前特征图的尺寸
+                let h = img_size.0 / stride;
+                let w = img_size.1 / stride;
+                let f_stride = stride as f32;
+
+                for y in 0..h {
+                    for x in 0..w {
+                        // 计算当前格子在 25200 个锚点中的线性索引
+                        let idx = offset + (y * w + x) as usize;
+                        // 1. 还原中心点坐标 (cx, cy)
+                        // 公式: (output + grid_offset) * stride
+                        batch[[idx, 0]] = (batch[[idx, 0]] + x as f32) * f_stride;
+                        batch[[idx, 1]] = (batch[[idx, 1]] + y as f32) * f_stride;
+
+                        // 2. 还原宽高 (w, h)
+                        // 公式: exp(output) * stride
+                        batch[[idx, 2]] = batch[[idx, 2]].exp() * f_stride;
+                        batch[[idx, 3]] = batch[[idx, 3]].exp() * f_stride;
+                    }
+                }
+                // 移动到下一个步长的起始位置
+                offset += (h * w) as usize;
+            }
+        }
+        outputs
+    }
+
+    fn nms(&self, boxes: &Array2<f32>, scores: &Array1<f32>, nms_thr: f32) -> Vec<usize> {
+        let mut keep = Vec::new();
+        let x1 = boxes.column(0);
+        let y1 = boxes.column(1);
+        let x2 = boxes.column(2);
+        let y2 = boxes.column(3);
+        // 在每一项前加上 &，并确保括号内的计算顺序
+        // 注意：ndarray 的 View 运算需要 &view1 - &view2
+        let areas = (&x2 - &x1 + 1.0) * (&y2 - &y1 + 1.0);
+
+        // 初始排序索引
+        let mut v: Vec<usize> = (0..scores.len()).collect();
+        v.sort_unstable_by(|&i, &j| {
+            scores[j]
+                .partial_cmp(&scores[i])
+                .unwrap_or(std::cmp::Ordering::Equal)
+        });
+        // 我们不使用 v.remove(0)，而是直接通过索引池操作
+        let mut active_indices = v;
+
+        while !active_indices.is_empty() {
+            // 取出当前池子中得分最高的框（即第一个元素）
+            let i = active_indices[0];
+            keep.push(i);
+
+            // 如果池子里只剩一个了，直接结束
+            if active_indices.len() == 1 {
+                break;
+            }
+
+            // 5. 核心逻辑：使用 retain 一次性过滤掉：
+            // (a) 当前框自己 (idx == i)
+            // (b) 与当前框重叠度过高的框 (iou > nms_thr)
+            active_indices.retain(|&idx| {
+                // 如果是当前正在处理的框，不保留（因为它已经进入 keep 了）
+                if idx == i {
+                    return false;
+                }
+
+                // 计算 IoU
+                let xx1 = x1[i].max(x1[idx]);
+                let yy1 = y1[i].max(y1[idx]);
+                let xx2 = x2[i].min(x2[idx]);
+                let yy2 = y2[i].min(y2[idx]);
+
+                let w = (xx2 - xx1 + 1.0).max(0.0);
+                let h = (yy2 - yy1 + 1.0).max(0.0);
+                let inter = w * h;
+
+                let iou = inter / (areas[i] + areas[idx] - inter);
+
+                // 只保留 IoU 小于阈值的框
+                iou <= nms_thr
+            });
+        }
+
+        keep
+    }
+
+    /// 多类别 NMS 后处理：按分数阈值筛选候选框，并用 NMS 阈值去重。
+    pub fn multiclass_nms(
+        &self,
+        boxes: &Array2<f32>,  // [25200, 4] -> xyxy 格式
+        scores: &Array2<f32>, // [25200, 80] -> 已经乘以 objectness 的得分
+        nms_thr: f32,
+        score_thr: f32,
+    ) -> Vec<[f32; 6]> {
+        let mut candidates = Vec::new();
+
+        // 1. 筛选高分框 (单次遍历完成 Argmax 和 Threshold 过滤)
+        for i in 0..scores.nrows() {
+            let row = scores.row(i);
+
+            // 找到当前行（即当前锚点）得分最高的类别
+            let mut max_score = 0.0;
+            let mut cls_id = 0;
+            for (j, &s) in row.iter().enumerate() {
+                if s > max_score {
+                    max_score = s;
+                    cls_id = j;
+                }
+            }
+
+            // 仅保留超过阈值的候选框
+            if max_score > score_thr {
+                // 暂时存储索引和元数据，避免频繁创建大数组
+                candidates.push((i, max_score, cls_id));
+            }
+        }
+
+        if candidates.is_empty() {
+            return vec![];
+        }
+
+        // 2. 准备 NMS 输入
+        // 构造 NMS 需要的子集数组
+        let mut b_subset = Array2::<f32>::zeros((candidates.len(), 4));
+        let mut s_subset = Array1::<f32>::zeros(candidates.len());
+
+        for (new_idx, &(orig_idx, score, _)) in candidates.iter().enumerate() {
+            b_subset.row_mut(new_idx).assign(&boxes.row(orig_idx));
+            s_subset[new_idx] = score;
+        }
+
+        // 3. 执行 NMS (返回保留下来的子集索引)
+        let keep = self.nms(&b_subset, &s_subset, nms_thr);
+
+        // 4. 组装最终结果 [x1, y1, x2, y2, score, class_id]
+        keep.into_iter()
+            .map(|k_idx| {
+                let (orig_idx, score, cls_id) = candidates[k_idx];
+                let b = boxes.row(orig_idx);
+                [b[0], b[1], b[2], b[3], score, cls_id as f32]
+            })
+            .collect()
+    }
+    /// 对图像执行完整检测流程（预处理、推理、后处理），返回像素坐标系下的检测框。
+    pub fn get_bbox(
+        &self,
+        dynamic_img: &DynamicImage,
+    ) -> Result<Vec<DetectionResult>, TensorError> {
+        // 使用 utils crate 解码
+        let (orig_w, orig_h) = dynamic_img.dimensions();
+
+        let (input_tensor, ratio) = self.preproc(dynamic_img, (416, 416));
+
+        // tract 推理
+        let outputs = self.runtime.inference(input_tensor)?;
+        // 2. 无缝、安全地解包出标准 3维 矩阵
+        let DetOutput::Detection(output_array) = outputs;
+
+        let predictions = self.demo_postprocess(output_array, (416, 416));
+        let pred = predictions.slice(s![0, .., ..]);
+
+        let boxes = pred.slice(s![.., 0..4]);
+        let obj_conf = pred.slice(s![.., 4..5]);
+        let cls_conf = pred.slice(s![.., 5..]);
+        let obj_broadcast =
+            obj_conf
+                .broadcast(cls_conf.dim())
+                .ok_or_else(|| TensorError::DimensionMismatch {
+                    expected: format!("可广播至 cls_conf 形状 {:?}", cls_conf.shape()),
+                    actual: obj_conf.shape().to_vec(),
+                })?;
+        let scores = &obj_broadcast * &cls_conf;
+
+        let mut boxes_xyxy = Array2::<f32>::zeros(boxes.raw_dim());
+        for i in 0..boxes.nrows() {
+            boxes_xyxy[[i, 0]] = (boxes[[i, 0]] - boxes[[i, 2]] / 2.0) / ratio;
+            boxes_xyxy[[i, 1]] = (boxes[[i, 1]] - boxes[[i, 3]] / 2.0) / ratio;
+            boxes_xyxy[[i, 2]] = (boxes[[i, 0]] + boxes[[i, 2]] / 2.0) / ratio;
+            boxes_xyxy[[i, 3]] = (boxes[[i, 1]] + boxes[[i, 3]] / 2.0) / ratio;
+        }
+
+        let detections = self.multiclass_nms(&boxes_xyxy, &scores, 0.45, 0.1);
+        let final_results = detections
+            .into_iter()
+            .map(|d| DetectionResult {
+                x1: (d[0] as i32).max(0).min(orig_w as i32),
+                y1: (d[1] as i32).max(0).min(orig_h as i32),
+                x2: (d[2] as i32).max(0).min(orig_w as i32),
+                y2: (d[3] as i32).max(0).min(orig_h as i32),
+                score: d[4],
+                class_id: d[5] as u32,
+            })
+            .collect();
+        Ok(final_results)
+    }
+}

ddddocr-core/src/error.rs

@@ -0,0 +1,191 @@
+//! 分层错误类型：预处理、推理、解码三阶段的强类型错误。
+
+use thiserror::Error;
+
+/// 全局统一的 `Result` 别名，默认错误类型为 [`DdddError`]。
+pub type Result<T, E = DdddError> = std::result::Result<T, E>;
+
+/// 顶层错误类型，聚合本库各阶段错误。
+#[derive(Error, Debug)]
+pub enum DdddError {
+    /// 图像预处理阶段异常。
+    #[error("图像预处理失败: {0}")]
+    Preprocess(#[from] ImagePreprocessError),
+
+    /// 推理与张量操作阶段异常。
+    #[error("推理与模型输入/输出张量异常: {0}")]
+    Inference(#[from] TensorError),
+
+    /// 后处理解码阶段异常。
+    #[error("后处理解码错误: {0}")]
+    Decode(#[from] DecodeError),
+
+    /// 框架内部不可恢复的逻辑断言错误（如解析节点 Fact 失败）。
+    #[error("内部严重逻辑错误: {0}")]
+    Internal(String),
+
+    /// 用户自定义扩展错误，用于包装第三方插件产生的错误。
+    #[error("用户自定义扩展错误: {0}")]
+    Other(#[source] Box<dyn std::error::Error + Send + Sync>),
+}
+
+/// 图像预处理阶段错误类型。
+#[derive(Error, Debug)]
+pub enum ImagePreprocessError {
+    /// ndarray 基础操作失败。
+    #[error("图片转矩阵(ndarray)基础操作失败: {0}")]
+    Ndarray(#[from] ndarray::ShapeError),
+
+    /// 图像矩阵维度不合规。
+    #[error("图像矩阵维度不合规！预期: {expected}，实际图像形状: {actual:?}")]
+    InvalidDimensions {
+        /// 期望的维度描述。
+        expected: String,
+        /// 实际的图像形状。
+        actual: Vec<usize>,
+    },
+
+    /// 图像缓冲区长度与分辨率/通道数不匹配。
+    #[error(
+        "图像缓冲区长度不匹配！预期大小: {expected}，实际大小: {actual} (分辨率: {width}x{height}, 通道数: {channels})"
+    )]
+    BufferLengthMismatch {
+        /// 期望的缓冲区长度。
+        expected: usize,
+        /// 实际的缓冲区长度。
+        actual: usize,
+        /// 图像宽度。
+        width: u32,
+        /// 图像高度。
+        height: u32,
+        /// 图像通道数。
+        channels: usize,
+    },
+
+    /// 不支持的图像通道数。
+    #[error("不支持的图像通道数: {0} (仅支持单通道灰度L、3通道RGB、4通道RGBA)")]
+    UnsupportedChannels(usize),
+
+    /// HSV 颜色区间参数非法。
+    #[error("HSV 颜色区间参数非法: {0}")]
+    InvalidHsvRange(String),
+
+    /// 未知的颜色预设名称。
+    #[error("不支持的颜色预设名称: {0}")]
+    UnknownColorPreset(String),
+
+    /// 颜色过滤器配置无效。
+    #[error("颜色过滤器配置无效或初始化失败: {0}")]
+    FilterConfigInvalid(String),
+
+    /// 图像维度不匹配。
+    #[error("图像维度不匹配！{0}")]
+    MismatchDimensions(String),
+
+    /// 滑块模板尺寸大于背景图。
+    #[error("滑块模板尺寸 [{target_w}x{target_h}] 大于背景图 [{bg_w}x{bg_h}]")]
+    TargetExceedsBackground {
+        /// 滑块模板宽度。
+        target_w: usize,
+        /// 滑块模板高度。
+        target_h: usize,
+        /// 背景图宽度。
+        bg_w: usize,
+        /// 背景图高度。
+        bg_h: usize,
+    },
+}
+
+/// 推理与张量操作阶段错误类型。
+#[derive(Error, Debug)]
+pub enum TensorError {
+    /// 推理引擎内部异常。
+    #[error("推理引擎内部发生异常: {0}")]
+    Engine(String),
+
+    /// 模型张量维度不匹配。
+    #[error("模型张量维度不匹配！预期: {expected}，实际 Tensor 形状: {actual:?}")]
+    DimensionMismatch {
+        /// 期望的维度描述。
+        expected: String,
+        /// 实际的 Tensor 形状。
+        actual: Vec<usize>,
+    },
+
+    /// OCR Logits 矩阵变形失败。
+    #[error("OCR Logits 矩阵变形失败: {0}")]
+    LogitsDimensionMismatch(#[from] ndarray::ShapeError),
+
+    /// 张量内存不连续。
+    #[error("内存不连续，无法执行零拷贝操作")]
+    NonContiguousMemory,
+
+    /// 未知的模型输出格式。
+    #[error("未知的模型输出格式")]
+    UnknownOutputFormat,
+}
+
+/// 算法解码阶段错误类型。
+#[derive(Error, Debug)]
+pub enum DecodeError {
+    /// CTC 解码异常。
+    #[error("CTC 解码异常: {0}")]
+    Ctc(String),
+}
+
+impl DdddError {
+    /// 手动包装任意第三方错误为 [`DdddError::Other`]。
+    pub fn new<E>(error: E) -> Self
+    where
+        E: Into<Box<dyn std::error::Error + Send + Sync>>,
+    {
+        DdddError::Other(error.into())
+    }
+
+    /// 是否为图片维度不合规错误。
+    pub fn is_invalid_dimensions(&self) -> bool {
+        matches!(
+            self,
+            DdddError::Preprocess(ImagePreprocessError::InvalidDimensions { .. })
+        )
+    }
+
+    /// 是否因通道数不合规而失败。
+    pub fn is_unsupported_channels(&self) -> bool {
+        matches!(
+            self,
+            DdddError::Preprocess(ImagePreprocessError::UnsupportedChannels(_))
+        )
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn new_wraps_third_party_error() {
+        let io_err = std::io::Error::other("boom");
+        let err = DdddError::new(io_err);
+        assert!(matches!(err, DdddError::Other(_)));
+    }
+
+    #[test]
+    fn preprocess_conversion_and_predicates() {
+        let e: DdddError = ImagePreprocessError::UnsupportedChannels(2).into();
+        assert!(e.is_unsupported_channels());
+
+        let e2: DdddError = ImagePreprocessError::InvalidDimensions {
+            expected: "x".into(),
+            actual: vec![0],
+        }
+        .into();
+        assert!(e2.is_invalid_dimensions());
+    }
+
+    #[test]
+    fn decode_conversion() {
+        let e: DdddError = DecodeError::Ctc("bad".into()).into();
+        assert!(matches!(e, DdddError::Decode(_)));
+    }
+}

ddddocr-core/src/lib.rs

@@ -0,0 +1,43 @@
+//! # ddddocr-core
+//!
+//! `ddddocr-rs` 的核心库：提供与具体推理引擎解耦的 OCR 识别、目标检测（det）与滑块匹配（slide）能力。
+//! 推理能力由 [`traits::InferenceEngine`]、[`traits::OcrEngine`]、[`traits::DetEngine`] 抽象，
+//! 由 `ddddocr-tract2`、`ddddocr-ort` 等引擎 crate 实现。
+//!
+//! 完整可运行示例见 `ddddocr-core/examples/quick_start.rs`。
+
+#![warn(missing_docs)]
+
+mod det;
+/// 分层错误类型。
+pub mod error;
+mod ocr;
+mod slide;
+/// 推理引擎统一抽象接口。
+pub mod traits;
+/// 模型输入输出信息等共享类型。
+pub mod types;
+/// 图像加载、转换与处理工具。
+pub mod utils;
+
+pub use crate::det::{DetBuilder, DetectionResult, Detector};
+pub use crate::ocr::{
+    CharRestrict, Charset, ColorFilter, ColorPreset, HsvRange, IdRestrict, ModelMetadata,
+    MultiOrColorRestrict, MultiOrRestrict, Normalization, Ocr, OcrBuilder, OcrResult, PixelCtx,
+    Resize, TokenFilter, ValidationCtx,
+};
+pub use crate::slide::{SlideResult, Slider};
+
+/// OCR 模型的统一输出枚举，由推理引擎产出，供 [`Ocr`] 后处理。
+pub enum OcrOutput {
+    /// 索引序列输出（CTC 解码输入）。
+    Indices(ndarray::Array1<i64>),
+    /// Logits 矩阵输出 `[Steps, Classes]`。
+    Logits(ndarray::Array2<f32>),
+}
+
+/// 目标检测模型的统一输出枚举，由推理引擎产出，供 [`Detector`] 后处理。
+pub enum DetOutput {
+    /// 原始检测输出张量。
+    Detection(ndarray::Array3<f32>),
+}

ddddocr-core/src/ocr.rs

@@ -0,0 +1,16 @@
+//! OCR 模块：识别器构建器、执行入口及元数据、字符集等类型。
+
+mod builder;
+mod builtin;
+mod charset;
+mod color_filter;
+mod executor;
+mod metadata;
+mod token_filter;
+
+pub use builder::OcrBuilder;
+pub use charset::Charset;
+pub use color_filter::{ColorFilter, ColorPreset, HsvRange, MultiOrColorRestrict, PixelCtx};
+pub use executor::{Ocr, OcrResult};
+pub use metadata::{ModelMetadata, Normalization, Resize};
+pub use token_filter::{CharRestrict, IdRestrict, MultiOrRestrict, TokenFilter, ValidationCtx};

ddddocr-core/src/ocr/builder.rs

@@ -0,0 +1,82 @@
+//! OCR 构建器。
+
+use crate::ocr::color_filter::ColorFilter;
+use crate::ocr::executor::Ocr;
+use crate::ocr::token_filter::TokenFilter;
+use crate::traits::OcrEngine;
+
+/// OCR 构建器：配置识别选项后绑定引擎会话构建 [`crate::Ocr`]。
+#[derive(Default)]
+pub struct OcrBuilder {
+    /// 是否修复PNG格式问题
+    png_fix: bool,
+    /// 是否返回概率信息
+    probability: bool,
+    /// 颜色过滤：保留的颜色列表
+    color_filter: Option<Box<dyn ColorFilter + Send + Sync>>,
+    /// 字符集范围
+    charset_restrict: Option<Box<dyn TokenFilter + Send + Sync>>,
+}
+
+impl OcrBuilder {
+    // 初始化任务，设置默认参数
+    /// 创建默认配置的构建器。
+    pub fn new() -> Self {
+        Self {
+            png_fix: false, // 默认值
+            probability: false,
+            color_filter: None,
+            charset_restrict: None,
+        }
+    }
+    /// 设置是否修复 PNG 透明背景问题。
+    pub fn png_fix(mut self, value: bool) -> Self {
+        self.png_fix = value;
+        self
+    }
+    /// 设置是否返回概率信息。
+    pub fn probability(mut self, value: bool) -> Self {
+        self.probability = value;
+        self
+    }
+
+    /// 设置颜色过滤约束。
+    pub fn color_filter<T>(mut self, filter: T) -> Self
+    where
+        T: ColorFilter + Send + Sync + 'static,
+    {
+        self.color_filter = Some(Box::new(filter));
+        self
+    }
+
+    /// 设置字符集限制。
+    pub fn charset_restrict<T>(mut self, restrict: T) -> Self
+    where
+        T: TokenFilter + Send + Sync + 'static,
+    {
+        self.charset_restrict = Some(Box::new(restrict));
+        self
+    }
+    /// 绑定引擎会话并构建 OCR 识别器。
+    pub fn build_with<E: OcrEngine>(self, runtime: &E) -> Ocr<'_> {
+        // 1. 原地解析颜色过滤器
+        let final_color_ranges = match &self.color_filter {
+            Some(filter) => filter.collect_to_vec(),
+            None => Ok(None),
+        };
+        // 2. 原地解析字符集过滤
+        let tokens = &runtime.metadata().charset.tokens;
+        let final_charset_indices = match &self.charset_restrict {
+            Some(restrict) => restrict.apply_to_charset(tokens),
+            None => None,
+        };
+
+        Ocr {
+            runtime,
+            png_fix: self.png_fix, // 原地解构出来
+            probability: self.probability,
+            final_color_ranges,
+            final_charset_indices,
+        }
+    }
+}

ddddocr-core/src/ocr/charset.rs

@@ -0,0 +1,103 @@
+//! 字符集：token 列表与索引的双向映射。
+
+use std::borrow::Cow;
+use std::collections::HashMap;
+
+// ==========================================
+// 3. 字符集核心结构体 (重命名为 Charset)
+// ==========================================
+/// 字符集：token 列表与索引的双向映射。
+#[derive(Debug, Clone)]
+pub struct Charset {
+    /// 字符集 token 列表。
+    // 使用 Cow 统一静态切片和动态读取的 Vec<String>，内部实现真正的零拷贝
+    pub tokens: Vec<Cow<'static, str>>,
+    /// 字符到索引的反查表。
+    // 反向查找表，保证字符转索引为 O(1)
+    pub char_to_idx: HashMap<Cow<'static, str>, usize>,
+}
+
+impl Charset {
+    // 内部底层统一收拢构造
+    /// 从 token 列表构建字符集。
+    pub fn new(tokens: Vec<Cow<'static, str>>) -> Self {
+        let mut char_to_idx = HashMap::with_capacity(tokens.len());
+        for (idx, token) in tokens.iter().enumerate() {
+            char_to_idx.entry(token.clone()).or_insert(idx);
+            // 如果字符集有重复，保留第一个遇到的索引 (符合 Python .index 逻辑)
+        }
+
+        Self {
+            tokens,
+            char_to_idx,
+        }
+    }
+
+    // --- 业务策略方法 ---
+
+    /// 字符转索引，不存在时返回 -1（与 Python 版行为一致）。
+    pub fn char_to_index(&self, char_str: &str) -> i32 {
+        if let Some(&idx) = self.char_to_idx.get(char_str) {
+            idx as i32
+        } else {
+            -1
+        }
+    }
+
+    /// 索引转字符引用，越界时返回 `None`。
+    pub fn index_to_char_ref(&self, index: usize) -> Option<&str> {
+        self.tokens.get(index).map(|cow| cow.as_ref())
+    }
+
+    /// 判断字符是否在字符集中。
+    pub fn is_valid_char(&self, char_str: &str) -> bool {
+        self.char_to_idx.contains_key(char_str)
+    }
+    /// 返回字符集大小。
+    pub fn size(&self) -> usize {
+        self.tokens.len()
+    }
+}
+
+// ==========================================
+// 4. 标准 Display 接口实现 (对应 __str__)
+// ==========================================
+impl std::fmt::Display for Charset {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "Charset [Total Size: {}", self.size(),)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn sample_tokens() -> Vec<Cow<'static, str>> {
+        vec![Cow::Borrowed(""), Cow::Borrowed("a"), Cow::Borrowed("b")]
+    }
+
+    #[test]
+    fn char_to_index_roundtrip() {
+        let cs = Charset::new(sample_tokens());
+        assert_eq!(cs.char_to_index("a"), 1);
+        assert_eq!(cs.char_to_index("z"), -1);
+        assert_eq!(cs.index_to_char_ref(2), Some("b"));
+        assert_eq!(cs.index_to_char_ref(99), None);
+    }
+
+    #[test]
+    fn duplicate_tokens_keep_first_index() {
+        let cs = Charset::new(vec![Cow::Borrowed("x"), Cow::Borrowed("x")]);
+        assert_eq!(cs.char_to_index("x"), 0);
+        assert_eq!(cs.size(), 2);
+    }
+
+    #[test]
+    fn is_valid_char_and_size() {
+        let cs = Charset::new(sample_tokens());
+        assert!(cs.is_valid_char(""));
+        assert!(cs.is_valid_char("a"));
+        assert!(!cs.is_valid_char("A"));
+        assert_eq!(cs.size(), 3);
+    }
+}

ddddocr-core/src/ocr/color_filter.rs

@@ -0,0 +1,370 @@
+//! 颜色过滤：HSV 区间匹配与颜色预设。
+
+use crate::error::{ImagePreprocessError, Result};
+use crate::utils::image_processor::rgb_to_opencv_hsv;
+use image::{DynamicImage, ImageBuffer, Rgb};
+use std::str::FromStr;
+
+#[inline(always)]
+fn is_pixel_matched(ranges: &[HsvRange], h: u8, s: u8, v: u8) -> bool {
+    ranges.iter().any(|range| {
+        h >= range.lower.0
+            && h <= range.upper.0
+            && s >= range.lower.1
+            && s <= range.upper.1
+            && v >= range.lower.2
+            && v <= range.upper.2
+    })
+}
+/// 按 HSV 区间过滤图像：未命中任一区间的像素刷白。
+pub fn apply_to_image(
+    image: &DynamicImage,
+    hsv_ranges: &[HsvRange],
+) -> Result<DynamicImage, ImagePreprocessError> {
+    // 1. 统一转换为连续内存的 RGB8 缓冲区 (对应 Python 的 Image 到 RGB/BGR 数组转换)
+    let rgb_img = image.to_rgb8();
+    let (width, height) = rgb_img.dimensions();
+    let mut raw_pixels = rgb_img.into_raw();
+    let actual_len = raw_pixels.len();
+    let expected_len = (width as usize) * (height as usize) * 3;
+    // 2. 密集计算核心：原地流式迭代修改
+    // 每次取出 3 个 u8 字节，分别代表 [R, G, B]，无多余掩膜矩阵内存分配
+    for chunk in raw_pixels.chunks_exact_mut(3) {
+        let r = chunk[0];
+        let g = chunk[1];
+        let b = chunk[2];
+
+        // 像素级转换为 OpenCV 标准的 HSV
+        let (h, s, v) = rgb_to_opencv_hsv(r, g, b);
+
+        // 模拟 Python 的多范围 mask bitwise_or 并在 mask == 0 处刷白
+        // 如果该像素没有命中任何一个配置的颜色区间，立刻原地刷白 [255, 255, 255]
+        if !is_pixel_matched(hsv_ranges, h, s, v) {
+            chunk[0] = 255;
+            chunk[1] = 255;
+            chunk[2] = 255;
+        }
+    }
+
+    // 3. 将扁平字节数组重新打包回 DynamicImage 容器
+    let filtered_buffer = ImageBuffer::<Rgb<u8>, Vec<u8>>::from_raw(width, height, raw_pixels)
+        .ok_or(ImagePreprocessError::BufferLengthMismatch {
+            expected: expected_len,
+            actual: actual_len,
+            width,
+            height,
+            channels: 3,
+        })?;
+    Ok(DynamicImage::ImageRgb8(filtered_buffer))
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord)]
+/// HSV 颜色区间，下界与上界各为 `(H, S, V)`。
+pub struct HsvRange {
+    /// 区间下界 `(H, S, V)`。
+    pub lower: (u8, u8, u8), // (H, S, V)
+    /// 区间上界 `(H, S, V)`。
+    pub upper: (u8, u8, u8), // (H, S, V)
+}
+
+impl HsvRange {
+    /// 创建 HSV 区间。
+    pub const fn new(lower: (u8, u8, u8), upper: (u8, u8, u8)) -> Self {
+        Self { lower, upper }
+    }
+}
+impl HsvRange {
+    /// 校验区间是否合法（H 0-180，S/V 0-255，且下界 <= 上界）。
+    pub fn validate(&self) -> Result<(), ImagePreprocessError> {
+        // 1. 校验 H 通道边界 (OpenCV 中 H 范围是 0-180)
+        if self.lower.0 > 180 || self.upper.0 > 180 {
+            return Err(ImagePreprocessError::InvalidHsvRange(
+                "H通道值必须在 0-180 范围内".to_string(),
+            ));
+        }
+
+        // 2. 校验下界不能大于上界
+        if self.lower.0 > self.upper.0 || self.lower.1 > self.upper.1 || self.lower.2 > self.upper.2
+        {
+            return Err(ImagePreprocessError::InvalidHsvRange(
+                "HSV范围下界不能大于上界".to_string(),
+            ));
+        }
+
+        Ok(())
+    }
+}
+#[derive(Debug, Clone, PartialEq, Eq)]
+/// 颜色预设：常见颜色对应的 HSV 区间集合。
+pub enum ColorPreset {
+    /// 红色。
+    Red,
+    /// 蓝色。
+    Blue,
+    /// 绿色。
+    Green,
+    /// 黄色。
+    Yellow,
+    /// 橙色。
+    Orange,
+    /// 紫色。
+    Purple,
+    /// 青色。
+    Cyan,
+    /// 黑色。
+    Black,
+    /// 白色。
+    White,
+    /// 灰色。
+    Gray,
+    /// 自定义区间列表。
+    Custom(Vec<HsvRange>),
+}
+
+impl ColorPreset {
+    /// 返回预设对应的 HSV 区间列表。
+    pub fn matches(&self) -> &[HsvRange] {
+        match self {
+            ColorPreset::Red => &[
+                HsvRange {
+                    lower: (0, 50, 50),
+                    upper: (10, 255, 255),
+                },
+                HsvRange {
+                    lower: (170, 50, 50),
+                    upper: (180, 255, 255),
+                },
+            ],
+            ColorPreset::Blue => &[HsvRange {
+                lower: (100, 50, 50),
+                upper: (130, 255, 255),
+            }],
+            ColorPreset::Green => &[HsvRange {
+                lower: (40, 50, 50),
+                upper: (80, 255, 255),
+            }],
+            ColorPreset::Yellow => &[HsvRange {
+                lower: (20, 50, 50),
+                upper: (40, 255, 255),
+            }],
+            ColorPreset::Orange => &[HsvRange {
+                lower: (10, 50, 50),
+                upper: (20, 255, 255),
+            }],
+            ColorPreset::Purple => &[HsvRange {
+                lower: (130, 50, 50),
+                upper: (170, 255, 255),
+            }],
+            ColorPreset::Cyan => &[HsvRange {
+                lower: (80, 50, 50),
+                upper: (100, 255, 255),
+            }],
+            ColorPreset::Black => &[HsvRange {
+                lower: (0, 0, 0),
+                upper: (180, 255, 50),
+            }],
+            ColorPreset::White => &[HsvRange {
+                lower: (0, 0, 200),
+                upper: (180, 30, 255),
+            }],
+            ColorPreset::Gray => &[HsvRange {
+                lower: (0, 0, 50),
+                upper: (180, 30, 200),
+            }],
+            ColorPreset::Custom(ranges) => ranges,
+        }
+    }
+    /// 校验预设的 HSV 区间是否合法。
+    pub fn validate(&self) -> Result<(), ImagePreprocessError> {
+        match self {
+            // 1. 快捷变体：完全绕过，根本不校验，0 运行时开销放行！
+            ColorPreset::Custom(ranges) => {
+                // 2. 只有 Custom 变体需要接受严格的参数政审
+                for r in ranges {
+                    r.validate()?;
+                }
+                Ok(())
+            }
+            _ => Ok(()),
+        }
+    }
+}
+
+impl FromStr for ColorPreset {
+    type Err = ImagePreprocessError;
+    fn from_str(s: &str) -> Result<Self, Self::Err> {
+        match s.to_lowercase().as_str() {
+            "red" => Ok(ColorPreset::Red),
+            "blue" => Ok(ColorPreset::Blue),
+            "green" => Ok(ColorPreset::Green),
+            "yellow" => Ok(ColorPreset::Yellow),
+            "orange" => Ok(ColorPreset::Orange),
+            "purple" => Ok(ColorPreset::Purple),
+            "cyan" => Ok(ColorPreset::Cyan),
+            "black" => Ok(ColorPreset::Black),
+            "white" => Ok(ColorPreset::White),
+            "gray" => Ok(ColorPreset::Gray),
+            _ => Err(ImagePreprocessError::UnknownColorPreset(s.to_string())),
+        }
+    }
+}
+
+// =====================================================================
+// 3. 颜色约束特征（Trait）与组合子设计模式
+// =====================================================================
+
+/// 颜色匹配上下文：当前像素的 HSV 值。
+pub struct PixelCtx {
+    /// 当前像素的 HSV 值。
+    pub hsv: (u8, u8, u8),
+}
+
+/// 颜色过滤约束接口：提供一组 HSV 区间。
+pub trait ColorFilter {
+    /// 将有效区间追加到目标容器。
+    fn append_ranges(&self, target: &mut Vec<HsvRange>);
+    /// 预估有效区间数量。
+    fn estimated_count(&self) -> usize;
+    /// 验证过滤器配置是否合法，默认直接放行。
+    fn validate_self(&self) -> Result<(), ImagePreprocessError> {
+        Ok(())
+    }
+    /// 收集全部有效区间并排序去重；无有效区间时返回 `None`。
+    fn collect_to_vec(&self) -> Result<Option<Vec<HsvRange>>, ImagePreprocessError> {
+        // 1. 触发自检
+        self.validate_self()?;
+
+        let total_capacity = self.estimated_count();
+        if total_capacity == 0 {
+            return Ok(None);
+        }
+
+        // 2. 永远一击必中分配精准内存，不需要再考虑追加和扩容！
+        let mut v = Vec::with_capacity(total_capacity.max(16));
+
+        // 2. 倒入数据
+        self.append_ranges(&mut v);
+
+        // 3. 原地完成排序与去重
+        v.sort_unstable();
+        v.dedup();
+
+        Ok(Some(v))
+    }
+}
+
+impl ColorFilter for ColorPreset {
+    fn append_ranges(&self, target: &mut Vec<HsvRange>) {
+        // 直接利用我们第一步写好的 matches() 拿到切片，整块高速拷贝倒入目标容器
+        target.extend_from_slice(self.matches());
+    }
+    fn estimated_count(&self) -> usize {
+        // 直接获取切片长度
+        self.matches().len()
+    }
+    fn validate_self(&self) -> Result<(), ImagePreprocessError> {
+        // 直接调用我们在第一步中为 ColorPreset 实现的精细化分流校验
+        // 快捷变体在这里会直接返回 Ok(()), 只有 Custom 才会去真正校验
+        self.validate()
+    }
+}
+
+/// 多路颜色“或”逻辑组合子（并集网络）
+pub struct MultiOrColorRestrict<'a> {
+    /// 参与「或」组合的过滤器列表。
+    pub filters: Vec<&'a dyn ColorFilter>,
+}
+
+impl<'a> ColorFilter for MultiOrColorRestrict<'a> {
+    fn append_ranges(&self, target: &mut Vec<HsvRange>) {
+        // 管道递延：依次指挥内部每一个子过滤器把数据倒进目标容器
+        for f in &self.filters {
+            f.append_ranges(target);
+        }
+    }
+    fn estimated_count(&self) -> usize {
+        // 数量累加：$O(1)$ 地把所有子过滤器的预估容量加起来
+        self.filters.iter().map(|f| f.estimated_count()).sum()
+    }
+
+    fn validate_self(&self) -> Result<(), ImagePreprocessError> {
+        // 递归政审：只要其中一个子过滤器校验失败（比如某个 Custom 变体非法），立刻熔断
+        for f in &self.filters {
+            f.validate_self()?;
+        }
+        Ok(())
+    }
+}
+
+// =====================================================================
+// 4. 声明式宏：一语定乾坤
+// =====================================================================
+
+/// 组合多个颜色过滤器为「或」关系的快捷宏。
+#[macro_export]
+macro_rules! color_any_of {
+    ($only:expr) => {
+        &$only as &dyn $crate::ColorFilter
+    };
+    ($($filter:expr),+ $(,)?) => {
+        &$crate::MultiOrColorRestrict {
+            filters: vec![ $( &$filter as &dyn $crate::ColorFilter ),+ ]
+        }
+    };
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn hsv_range_validate() {
+        assert!(HsvRange::new((0, 0, 0), (180, 255, 255)).validate().is_ok());
+        assert!(
+            HsvRange::new((181, 0, 0), (255, 255, 255))
+                .validate()
+                .is_err()
+        );
+        assert!(
+            HsvRange::new((20, 50, 50), (10, 255, 255))
+                .validate()
+                .is_err()
+        );
+    }
+
+    #[test]
+    fn color_preset_matches_counts() {
+        assert_eq!(ColorPreset::Red.matches().len(), 2);
+        assert_eq!(ColorPreset::Blue.matches().len(), 1);
+        let custom = ColorPreset::Custom(vec![HsvRange::new((0, 0, 0), (1, 1, 1))]);
+        assert_eq!(custom.matches().len(), 1);
+    }
+
+    #[test]
+    fn color_preset_from_str() {
+        assert_eq!("red".parse::<ColorPreset>().unwrap(), ColorPreset::Red);
+        assert!("pink".parse::<ColorPreset>().is_err());
+    }
+
+    #[test]
+    fn color_any_of_collects() {
+        let ranges = crate::color_any_of!(ColorPreset::Red, ColorPreset::Blue)
+            .collect_to_vec()
+            .unwrap()
+            .unwrap();
+        assert_eq!(ranges.len(), 3);
+    }
+
+    #[test]
+    fn apply_to_image_whitens_non_matching() {
+        let img = DynamicImage::ImageRgb8(image::ImageBuffer::from_pixel(
+            2,
+            2,
+            image::Rgb([255, 0, 0]),
+        ));
+        let filtered = apply_to_image(&img, ColorPreset::Blue.matches()).unwrap();
+        for p in filtered.to_rgb8().pixels() {
+            assert_eq!(p.0, [255, 255, 255]);
+        }
+    }
+}

ddddocr-core/src/ocr/executor.rs

@@ -0,0 +1,421 @@
+//! OCR 执行器：预测入口与结果类型。
+
+use crate::ocr::metadata::Resize;
+
+use crate::error::{ImagePreprocessError, Result, TensorError};
+use crate::ocr::color_filter::{HsvRange, apply_to_image};
+use crate::traits::OcrEngine;
+use crate::utils::image_convert::png_rgba_white_preprocess;
+use crate::utils::image_processor::{convert_to_grayscale, resize_image};
+use crate::{OcrBuilder, OcrOutput};
+use image::DynamicImage;
+use ndarray::ArrayView2;
+use std::borrow::Cow;
+use std::fmt;
+use tracing::{debug, warn};
+/// OCR 识别结果：纯文本或携带概率的文本。
+#[derive(Debug, Clone)]
+pub enum OcrResult {
+    /// 纯文本结果（`probability = false` 时返回）。
+    Text(String),
+    /// 携带概率的结果（`probability = true` 时返回）。
+    Probability {
+        /// 识别出的文本。
+        text: String,
+        /// 全量概率矩阵 `[Steps, Classes]`。
+        probabilities: Vec<Vec<f32>>,
+        /// 全局平均置信度。
+        confidence: f64,
+    },
+    /// 不支持的模型或未知输出。
+    Unsupported {
+        /// 不支持原因说明。
+        message: String,
+    },
+}
+impl OcrResult {
+    /// 消费自身并提取最终文本。
+    pub fn into_text(self) -> String {
+        match self {
+            OcrResult::Text(text) => text,
+            OcrResult::Probability { text, .. } => text,
+            OcrResult::Unsupported { message } => {
+                // 作为库，这里可以返回空，或者直接携带错误信息，取决于你的设计
+                format!("Error: {}", message)
+            }
+        }
+    }
+}
+impl fmt::Display for OcrResult {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        match self {
+            OcrResult::Text(text) => {
+                // 纯文本分支，直接输出文本内容
+                write!(f, "{}", text)
+            }
+            OcrResult::Probability {
+                text,
+                probabilities,
+                confidence,
+            } => {
+                // 概率分支，友好地展示文本以及百分比形式的置信度
+                // 1. 基本信息
+                write!(f, "{} (置信度: {:.2}%)", text, confidence * 100.0)?;
+
+                // 2. 概率矩阵流式安全打印
+                write!(f, " [概率矩阵预览: ")?;
+
+                let max_steps_to_show = 10;
+                let take_steps = probabilities.iter().take(max_steps_to_show);
+
+                for (i, step_probs) in take_steps.enumerate() {
+                    if i > 0 {
+                        write!(f, ", ")?;
+                    }
+
+                    // 为了防止单行内部数据过长，单行也做一下截断保护（比如每行最多显示前 3 个概率）
+                    let max_classes_to_show = 3;
+                    write!(f, "[")?;
+                    for (j, prob) in step_probs.iter().take(max_classes_to_show).enumerate() {
+                        if j > 0 {
+                            write!(f, ", ")?;
+                        }
+                        write!(f, "{:.4}", prob)?;
+                    }
+                    if step_probs.len() > max_classes_to_show {
+                        write!(f, ", ..")?;
+                    }
+                    write!(f, "]")?;
+                }
+
+                // 如果总 Step 数量超过 10，末尾追加 .. 表示截断
+                if probabilities.len() > max_steps_to_show {
+                    write!(f, ", ..")?;
+                }
+                write!(f, "]")
+            }
+            OcrResult::Unsupported { message } => {
+                // 错误分支，直观输出异常原因
+                write!(f, "未识别成功: {}", message)
+            }
+        }
+    }
+}
+
+/// OCR 识别器：预处理、推理调度与后处理解码。
+pub struct Ocr<'a> {
+    pub(crate) runtime: &'a dyn OcrEngine,
+    pub(crate) png_fix: bool,
+    pub(crate) probability: bool,
+    /// 颜色过滤：保留的颜色列表
+    pub(crate) final_color_ranges: Result<Option<Vec<HsvRange>>, ImagePreprocessError>,
+
+    /// 字符集范围
+    pub(crate) final_charset_indices: Option<Vec<usize>>,
+}
+
+impl<'a> Ocr<'a> {
+    // 初始化任务，设置默认参数
+
+    /// 绑定引擎会话创建 OCR 识别器。
+    pub fn new(runtime: &'a dyn OcrEngine) -> Self {
+        Ocr {
+            runtime,
+            png_fix: false, // 默认值
+            probability: false,
+            final_color_ranges: Ok(None),
+            final_charset_indices: None,
+        }
+    }
+    /// 创建 OCR 构建器。
+    pub fn builder() -> OcrBuilder {
+        OcrBuilder::default()
+    }
+}
+impl<'a> Ocr<'a> {
+    /// 对输入图像执行 OCR 识别并返回结果。
+    pub fn predict(&self, image: &DynamicImage) -> Result<OcrResult> {
+        debug!("当前颜色过滤器状态: {:?}", self.final_color_ranges);
+
+        // =====================================================================
+        // 管道节点 1: 颜色过滤流水线
+        // 使用 Cow (Copy-On-Write) 智能指针。
+        // 如果未开启过滤，img_cow 内部只是持有原图的【只读借用】，发生【零内存分配】！
+        // =====================================================================
+        let img_cow = match &self.final_color_ranges {
+            Err(err_msg) => {
+                return Err(ImagePreprocessError::FilterConfigInvalid(
+                    err_msg.to_string(),
+                ))?;
+            }
+            Ok(None) => {
+                // 核心优化点：直接借用原图，不发生任何克隆
+                Cow::Borrowed(image)
+            }
+            Ok(Some(ranges)) => {
+                // 只有真正需要过滤时，才在内部提取像素并生成清洗后的 Owned 新图
+                let filtered_img = apply_to_image(image, ranges)?;
+                Cow::Owned(filtered_img)
+            }
+        };
+        let tensor = self.preprocess_image(&img_cow)?;
+
+        let raw_tensor = self.runtime.inference(tensor)?;
+
+        // 3. 后处理分流：直接返回 OcrResult
+        let ocr_output = self.process_model_output(raw_tensor)?;
+        Ok(ocr_output)
+    }
+    fn preprocess_image(
+        &self,
+        img: &DynamicImage,
+    ) -> Result<ndarray::Array4<f32>, ImagePreprocessError> {
+        // 1. 获取模型元数据配置
+        let meta = self.runtime.metadata();
+        let norm = &meta.normalization; // 获取归一化器
+
+        // A. 修复 PNG 透明背景 (内部逻辑你之前已实现)
+        let current_img = if self.png_fix && img.color().has_alpha() {
+            // 只有满足条件才去触发分配，生成新图
+            Cow::Owned(png_rgba_white_preprocess(img))
+        } else {
+            // 正常情况下，仅仅是再次安全借用，无开销
+            Cow::Borrowed(img)
+        };
+
+        // 3. 管道节点 2: 根据 Resize 策略计算目标宽高并进行缩放
+        let (target_w, target_h) = match meta.resize {
+            Resize::Fixed(w, h) => (w, h),
+            Resize::DynamicWidth(h) => {
+                // 高度固定，宽度根据原始比例动态计算：W_target = W_orig * (H_target / H_orig)
+                let w =
+                    (current_img.width() as f32 * (h as f32 / current_img.height() as f32)) as u32;
+                (w, h)
+            }
+            Resize::Square(size) => {
+                // 单字识别模型，直接缩放为正方形
+                (size, size)
+            }
+        };
+        // 执行缩放
+        let resized_img = resize_image(&current_img, target_w, target_h);
+
+        // 4. 管道节点 3: 颜色通道转换（单通道灰度 vs 三通道 RGB）与 4D 张量填充
+        let array4 = match meta.channel {
+            // --- 情况 A: 单通道（灰度图），对应 Python 的 len(shape) == 2 展开 ---
+            1 => {
+                let gray_img = convert_to_grayscale(&resized_img);
+
+                ndarray::Array4::from_shape_fn(
+                    (1, 1, target_h as usize, target_w as usize),
+                    |(_, _, y, x)| {
+                        let pixel = gray_img.get_pixel(x as u32, y as u32)[0] as f32;
+                        // pixel / 255.0 // 严格对齐 Python 归一化 [0.0, 1.0]
+                        // (pixel / 255.0 - 0.5) / 0.5
+                        norm.normalize(pixel)
+                    },
+                )
+            }
+
+            // --- 情况 B: 三通道（RGB），对应 Python 的 transpose(2, 0, 1) 的 CHW 布局 ---
+            3 => {
+                let rgb_img = resized_img.to_rgb8();
+
+                ndarray::Array4::from_shape_fn(
+                    (1, 3, target_h as usize, target_w as usize),
+                    |(_, c, y, x)| {
+                        let pixel = rgb_img.get_pixel(x as u32, y as u32)[c] as f32;
+                        // pixel / 255.0 // 严格对齐 Python 归一化 [0.0, 1.0]
+                        // (pixel / 255.0 - 0.5) / 0.5
+                        norm.normalize(pixel)
+                    },
+                )
+            }
+
+            // _ => return Err(anyhow::anyhow!("不支持的通道数配置: {}", meta.channel)),
+            _ => {
+                return Err(ImagePreprocessError::UnsupportedChannels(
+                    meta.channel as usize,
+                ));
+            }
+        };
+        Ok(array4)
+    }
+
+    fn process_model_output(&self, output: OcrOutput) -> Result<OcrResult, TensorError> {
+        match output {
+            OcrOutput::Indices(array1) => {
+                // 对应原来的 process_i64_tensor
+                let slice = array1.as_slice().ok_or(TensorError::NonContiguousMemory)?;
+                let final_text = self.ctc_decode_to_string(slice);
+
+                if self.probability {
+                    Ok(OcrResult::Probability {
+                        text: final_text,
+                        probabilities: vec![],
+                        confidence: 1.0,
+                    })
+                } else {
+                    Ok(OcrResult::Text(final_text))
+                }
+            }
+            OcrOutput::Logits(matrix_view) => {
+                // 对应原来的 process_f32_tensor
+                // 注意：此时的 matrix_view 已经是干净的标准的 ndarray::Array2<f32>，且保证是 [Steps, Classes] 2D 形状
+                if self.probability {
+                    let (probabilities_list, confidence, predicted_indices) =
+                        self.compute_f32_full_probability(matrix_view.view());
+                    let final_text = self.ctc_decode_to_string(&predicted_indices);
+                    Ok(OcrResult::Probability {
+                        text: final_text,
+                        probabilities: probabilities_list,
+                        confidence: confidence as f64,
+                    })
+                } else {
+                    let predicted_indices: Vec<i64> = matrix_view
+                        .outer_iter()
+                        .map(|row| {
+                            row.iter()
+                                .enumerate()
+                                .max_by(|(_, a), (_, b)| a.total_cmp(b))
+                                .map(|(idx, _)| idx as i64)
+                                .unwrap_or(0)
+                        })
+                        .collect();
+
+                    let final_text = self.ctc_decode_to_string(&predicted_indices);
+                    Ok(OcrResult::Text(final_text))
+                }
+            }
+        }
+    }
+}
+impl<'a> Ocr<'a> {
+    /// 判断索引是否为当前限制下的有效字符索引。
+    pub fn is_valid_indices(&self, idx: usize) -> bool {
+        if idx >= self.runtime.metadata().charset.size() {
+            return false;
+        }
+
+        match &self.final_charset_indices {
+            Some(v) => v.binary_search(&idx).is_ok(),
+            None => true,
+        }
+    }
+    /// 返回当前生效的可用 token 列表。
+    pub fn valid_tokens(&self) -> Vec<&str> {
+        let charset = &self.runtime.metadata().charset;
+        let tokens = &charset.tokens;
+        match &self.final_charset_indices {
+            Some(indices) => indices
+                .iter()
+                .filter_map(|&idx| tokens.get(idx).map(|cow| cow.as_ref()))
+                .collect(),
+            // 如果是 None，现场映射出全量 Token 视图给外部
+            None => tokens.iter().map(|cow| cow.as_ref()).collect(),
+        }
+    }
+    /// 返回当前生效的可用 token 数量。
+    pub fn valid_size(&self) -> usize {
+        match &self.final_charset_indices {
+            Some(indices) => indices.len(),
+            None => self.runtime.metadata().charset.tokens.len(),
+        }
+    }
+    fn compute_f32_full_probability(
+        &self,
+        matrix_view: ArrayView2<f32>,
+    ) -> (Vec<Vec<f32>>, f32, Vec<i64>) {
+        let steps = matrix_view.nrows();
+        let classes = matrix_view.ncols();
+
+        // 1. 预分配满额概率矩阵内存
+        let mut prob_matrix = ndarray::Array2::<f32>::zeros((steps, classes));
+        let mut predicted_indices = Vec::with_capacity(steps);
+        let mut confidence_sum = 0.0f32;
+
+        // 2. 融合单次遍历
+        for (step_idx, row) in matrix_view.outer_iter().enumerate() {
+            // 寻找当前 Step 的最大值和最大值索引 (Argmax)
+            let (row_max_idx, max_logit) = row
+                .iter()
+                .enumerate()
+                .max_by(|(_, a), (_, b)| a.total_cmp(b))
+                .map(|(idx, &val)| (idx, val))
+                .unwrap_or((0, 0.0));
+
+            predicted_indices.push(row_max_idx as i64);
+
+            // 计算单行 exp 溢出防范和
+            let mut exp_sum = 0.0f32;
+            for &val in row.iter() {
+                exp_sum += (val - max_logit).exp();
+            }
+
+            // 归一化 Softmax 顺序写入
+            for (class_idx, &val) in row.iter().enumerate() {
+                prob_matrix[[step_idx, class_idx]] = (val - max_logit).exp() / exp_sum;
+            }
+
+            // 当前 Step 最大概率在线累加
+            confidence_sum += 1.0f32 / exp_sum;
+        }
+
+        // 3. 统计全局平均置信度
+        let confidence = if steps > 0 {
+            confidence_sum / steps as f32
+        } else {
+            1.0
+        };
+
+        // 4. 将矩阵转化为标准安全序列化格式 [Steps, Classes]
+        let probabilities_list: Vec<Vec<f32>> =
+            prob_matrix.outer_iter().map(|row| row.to_vec()).collect();
+
+        (probabilities_list, confidence, predicted_indices)
+    }
+    fn ctc_decode_to_string(&self, predicted_indices: &[i64]) -> String {
+        debug!("indices模型输出原始数据: {:?}", predicted_indices);
+        let charset = &self.runtime.metadata().charset;
+        let tokens = &charset.tokens;
+
+        // 对应 _ctc_decode_indices 的逻辑：去重、去 blank (0)
+        let mut res = String::new();
+        let mut prev_idx: i64 = -1;
+
+        for &idx in predicted_indices {
+            // 1. CTC 去重：如果是连续重复的，直接跳过
+            if idx == prev_idx {
+                continue;
+            }
+            // 【关键核心】只要不是连续重复，立刻更新 prev_idx 状态，绝对不能被后续的过滤短路！
+            prev_idx = idx;
+
+            // 2. CTC 过滤 Blank (0)
+            if idx == 0 {
+                continue;
+            }
+            // 3. 类型安全转换
+            let u_idx = match usize::try_from(idx) {
+                Ok(u) => u,
+                Err(_) => continue,
+            };
+
+            // 史诗级加速点：如果是 None，说明没限制，根本不进入分支，直接放行！
+            // 只有当有具体限制（Some）时，才去跑 4-5 次 CPU 寄存器级别的二分查找
+            if let Some(ref indices) = self.final_charset_indices
+                && indices.binary_search(&u_idx).is_err()
+            {
+                continue;
+            }
+
+            // 5. 字符映射
+            if let Some(char_str) = tokens.get(u_idx) {
+                res.push_str(char_str);
+            } else {
+                warn!("警告: 预测索引 {} 超出字符集范围", u_idx);
+            }
+        }
+        res
+    }
+}

ddddocr-core/src/ocr/metadata.rs

@@ -0,0 +1,156 @@
+//! OCR 模型元数据：归一化、缩放策略与模型信息。
+
+// =====================================================================
+// 1. 辅助定义的枚举与结构体
+// =====================================================================
+
+use crate::ocr::Charset;
+use std::borrow::Cow;
+
+/// 像素归一化策略。
+#[derive(Debug, Clone, Copy)]
+pub enum Normalization {
+    /// 映射到 [0.0, 1.0] -> pixel / 255.0
+    ZeroToOne,
+    /// 映射到 [-1.0, 1.0] -> (pixel / 255.0 - 0.5) / 0.5
+    MinusOneToOne,
+}
+
+impl Normalization {
+    /// 对像素值执行归一化。
+    #[inline(always)]
+    pub fn normalize(&self, pixel: f32) -> f32 {
+        match self {
+            Normalization::ZeroToOne => pixel / 255.0,
+            Normalization::MinusOneToOne => (pixel / 255.0 - 0.5) / 0.5,
+        }
+    }
+}
+
+/// 图像缩放策略枚举
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum Resize {
+    /// 固定宽高，例如 (64, 64)
+    Fixed(u32, u32),
+    /// 高度固定，宽度根据原始比例动态计算（对应 Python 的 [-1, H]）
+    DynamicWidth(u32),
+    /// 单字识别的正方形切图（对应 Python 的 word 为 True 且 [-1, H]）
+    Square(u32),
+}
+
+/// OCR 模型元数据：字符集、缩放策略、通道数与归一化配置。
+#[derive(Debug, Clone)]
+pub struct ModelMetadata {
+    /// 字符集。
+    pub charset: Charset,
+    /// 是否为单字识别模型
+    pub word: bool,
+    /// 缩放策略。
+    pub resize: Resize,
+    /// 图像通道数（1 或 3）
+    pub channel: u8,
+    /// 像素归一化配置。
+    pub normalization: Normalization,
+}
+
+impl ModelMetadata {
+    /// 创建模型元数据。
+    pub fn new(
+        charset: Charset,
+        word: bool,
+        resize: Resize,
+        channel: u8,
+        normalization: Normalization,
+    ) -> Self {
+        Self {
+            charset,
+            word,
+            resize,
+            channel,
+            normalization,
+        }
+    }
+    // --- 优雅的工厂模式构造器 ---
+    /// 从静态字符切片构建元数据并自动生成字符集。
+    pub fn from_static_slice(
+        slice: &[&'static str],
+        word: bool,
+        resize: Resize,
+        channel: u8,
+        normalization: Normalization,
+    ) -> Self {
+        let tokens: Vec<Cow<'static, str>> = slice.iter().map(|&s| Cow::Borrowed(s)).collect();
+        Self {
+            charset: Charset::new(tokens),
+            word,
+            resize,
+            channel,
+            normalization,
+        }
+    }
+    /// 使用内置旧版字符集（与官方旧版模型配套）。
+    pub fn from_builtin_old(
+        word: bool,
+        resize: Resize,
+        channel: u8,
+        normalization: Normalization,
+    ) -> Self {
+        Self::from_static_slice(
+            super::builtin::CHARSET_OLD,
+            word,
+            resize,
+            channel,
+            normalization,
+        )
+    }
+
+    /// 使用内置 Beta 字符集（与官方 Beta 模型配套）。
+    pub fn from_builtin_beta(
+        word: bool,
+        resize: Resize,
+        channel: u8,
+        normalization: Normalization,
+    ) -> Self {
+        Self::from_static_slice(
+            super::builtin::CHARSET_BETA,
+            word,
+            resize,
+            channel,
+            normalization,
+        )
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn normalization_zero_to_one() {
+        let n = Normalization::ZeroToOne;
+        assert_eq!(n.normalize(0.0), 0.0);
+        assert_eq!(n.normalize(255.0), 1.0);
+    }
+
+    #[test]
+    fn normalization_minus_one_to_one() {
+        let n = Normalization::MinusOneToOne;
+        assert_eq!(n.normalize(0.0), -1.0);
+        assert_eq!(n.normalize(255.0), 1.0);
+    }
+
+    #[test]
+    fn from_static_slice_builds_charset() {
+        let meta = ModelMetadata::from_static_slice(
+            &["", "a"],
+            false,
+            Resize::Fixed(64, 64),
+            1,
+            Normalization::ZeroToOne,
+        );
+        assert_eq!(meta.charset.size(), 2);
+        assert_eq!(meta.charset.char_to_index("a"), 1);
+        assert_eq!(meta.channel, 1);
+        assert!(!meta.word);
+    }
+}

ddddocr-core/src/ocr/token_filter.rs

@@ -0,0 +1,232 @@
+//! 字符集限制：按字符属性或索引过滤识别范围。
+
+use std::borrow::Cow;
+use tracing::warn;
+
+/// 字符集校验上下文：当前 token 的文本与索引。
+pub struct ValidationCtx<'a> {
+    /// 当前 token 的文本内容。
+    pub text: &'a str, // 当前 Token 的文本内容
+    /// 当前 token 的 ID 索引。
+    pub token_id: usize, // 当前 Token 的 ID 索引
+}
+
+/// 字符集限制接口：决定某个 token 是否放行。
+pub trait TokenFilter {
+    /// 判断 token 是否放行。
+    fn matches(&self, ctx: &ValidationCtx) -> bool;
+    /// 预估匹配数量的容量提示。
+    fn estimated_capacity(&self) -> usize {
+        128
+    }
+
+    /// 遍历全量字符集筛选可用索引（放行 CTC blank、排序去重、空交集返回 `None`）。
+    fn apply_to_charset(&self, tokens: &[Cow<str>]) -> Option<Vec<usize>> {
+        let mut has_any_match = false;
+        let estimated_capacity = self.estimated_capacity();
+
+        // 1. 精准开辟内存，完美利用容量提示，避免动态乱涨
+        let mut temp_indices = Vec::with_capacity(estimated_capacity.max(16));
+
+        // 2. 高性能原地单次流式迭代
+        for (idx, token) in tokens.iter().enumerate() {
+            let token_str = token.as_ref();
+
+            // 规则 A: CTC Blank 空字符串或 0 号索引无条件放行
+            if token_str.is_empty() || idx == 0 {
+                temp_indices.push(idx);
+                continue; // 关键：直接跳过，防止后续 matches 匹配成功导致重复 push 产生 Bug
+            }
+
+            // 规则 B: 组装无拷贝上下文
+            let ctx = ValidationCtx {
+                text: token_str,
+                token_id: idx,
+            };
+
+            // 规则 C: 路由到各自具体实现的特异性匹配中（如 Digit 判定、TopN 判定、组合子判定等）
+            if self.matches(&ctx) {
+                temp_indices.push(idx);
+                has_any_match = true;
+            }
+        }
+
+        // 3. 终极防御：如果整个模型字符集除了 Blank，一个都没对上，直接退化为 None（全量识别）
+        if !has_any_match {
+            warn!("当前限制策略与模型字符集完全没有交集，已自动恢复全量识别");
+            None
+        } else {
+            // 4. 排序并去重，为 Ocr 引擎后续进行极其高频的『二分查找』筑起绝对安全的底层保障
+            temp_indices.sort_unstable();
+            temp_indices.dedup();
+            Some(temp_indices)
+        }
+    }
+}
+
+/// 按字符属性限制：数字、大小写字母或自定义列表。
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum CharRestrict {
+    /// 仅数字。
+    Digit,
+    /// 仅小写字母。
+    Lowercase,
+    /// 仅大写字母。
+    Uppercase,
+    /// 自定义字符列表。
+    CustomList(Vec<String>),
+}
+
+impl TokenFilter for CharRestrict {
+    fn matches(&self, ctx: &ValidationCtx) -> bool {
+        match self {
+            Self::Digit => ctx.text.len() == 1 && ctx.text.as_bytes()[0].is_ascii_digit(),
+            Self::Lowercase => ctx.text.len() == 1 && ctx.text.as_bytes()[0].is_ascii_lowercase(),
+            Self::Uppercase => ctx.text.len() == 1 && ctx.text.as_bytes()[0].is_ascii_uppercase(),
+            Self::CustomList(vec) => vec.iter().any(|t| t == ctx.text),
+        }
+    }
+    fn estimated_capacity(&self) -> usize {
+        match self {
+            Self::Digit => 16,
+            Self::Lowercase | Self::Uppercase => 32,
+            Self::CustomList(vec) => vec.len() + 1,
+        }
+    }
+}
+
+/// 按索引限制：前 N 个、索引范围或索引列表。
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum IdRestrict {
+    /// 前 N 个索引。
+    TopN(usize),
+    /// 指定索引范围。
+    IdRange(std::ops::Range<usize>),
+    /// 指定索引列表。
+    IdList(Vec<usize>),
+}
+
+impl TokenFilter for IdRestrict {
+    fn matches(&self, ctx: &ValidationCtx) -> bool {
+        match self {
+            Self::TopN(n) => ctx.token_id < *n,
+            Self::IdRange(range) => range.contains(&ctx.token_id),
+            Self::IdList(vec) => vec.contains(&ctx.token_id),
+        }
+    }
+    fn estimated_capacity(&self) -> usize {
+        match self {
+            Self::TopN(n) => *n + 1,
+            // 2. IdRange：标准标准库 Range 的长度
+            // 注意：因为范围可能是 1000..2000，它的 len() 返回的是 usize
+            Self::IdRange(range) => range.len() + 1,
+            // 3. IdList：Vec 里的元素个数
+            Self::IdList(vec) => vec.len() + 1,
+        }
+    }
+}
+
+/// 多路“或”逻辑组合子（支持 N 个规则无缝并集）
+pub struct MultiOrRestrict<'a> {
+    /// 参与「或」组合的过滤器列表。
+    pub filters: Vec<&'a dyn TokenFilter>,
+}
+
+impl<'a> TokenFilter for MultiOrRestrict<'a> {
+    fn matches(&self, ctx: &ValidationCtx) -> bool {
+        // 核心高阶函数：只要有一个过滤器命中，该 Token 即可放行
+        self.filters.iter().any(|f| f.matches(ctx))
+    }
+
+    fn estimated_capacity(&self) -> usize {
+        // 将所有过滤器的预估容量累加，作为最终容量参考
+        self.filters.iter().map(|f| f.estimated_capacity()).sum()
+    }
+}
+
+/// 组合多个字符限制规则为「或」关系的快捷宏。
+#[macro_export]
+macro_rules! any_of {
+    // 场景 A：如果用户只传了一个规则，免去构建 Vec 的开销，直接返回其引用
+    ($only:expr) => {
+        &$only as &dyn $crate::TokenFilter
+    };
+
+    // 场景 B：如果用户传入了多个规则，自动组合
+    ($($filter:expr),+ $(,)?) => {
+        &$crate::MultiOrRestrict {
+            filters: vec![ $( &$filter as &dyn $crate::TokenFilter ),+ ]
+        }
+    };
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn tokens() -> Vec<Cow<'static, str>> {
+        vec![
+            Cow::Borrowed(""),
+            Cow::Borrowed("1"),
+            Cow::Borrowed("a"),
+            Cow::Borrowed("A"),
+            Cow::Borrowed("!"),
+        ]
+    }
+
+    #[test]
+    fn char_restrict_digit() {
+        let indices = CharRestrict::Digit.apply_to_charset(&tokens()).unwrap();
+        assert_eq!(indices, vec![0, 1]);
+    }
+
+    #[test]
+    fn char_restrict_case() {
+        assert_eq!(
+            CharRestrict::Lowercase.apply_to_charset(&tokens()).unwrap(),
+            vec![0, 2]
+        );
+        assert_eq!(
+            CharRestrict::Uppercase.apply_to_charset(&tokens()).unwrap(),
+            vec![0, 3]
+        );
+    }
+
+    #[test]
+    fn custom_list() {
+        let r = CharRestrict::CustomList(vec!["!".into(), "a".into()]);
+        assert_eq!(r.apply_to_charset(&tokens()).unwrap(), vec![0, 2, 4]);
+    }
+
+    #[test]
+    fn no_intersection_falls_back_to_none() {
+        let r = CharRestrict::CustomList(vec!["z".into()]);
+        assert_eq!(r.apply_to_charset(&tokens()), None);
+    }
+
+    #[test]
+    fn id_restrict_variants() {
+        assert_eq!(
+            IdRestrict::TopN(3).apply_to_charset(&tokens()).unwrap(),
+            vec![0, 1, 2]
+        );
+        assert_eq!(
+            IdRestrict::IdRange(1..3)
+                .apply_to_charset(&tokens())
+                .unwrap(),
+            vec![0, 1, 2]
+        );
+        assert_eq!(
+            IdRestrict::IdList(vec![0, 4])
+                .apply_to_charset(&tokens())
+                .unwrap(),
+            vec![0, 4]
+        );
+    }
+
+    #[test]
+    fn multi_or_restrict_unions() {
+        let combined = crate::any_of!(CharRestrict::Digit, CharRestrict::Uppercase);
+        assert_eq!(combined.apply_to_charset(&tokens()).unwrap(), vec![0, 1, 3]);
+    }
+}

ddddocr-core/src/slide.rs

@@ -0,0 +1,307 @@
+//! 滑块验证码匹配：模板匹配与差异比较两种模式。
+
+use crate::error::{ImagePreprocessError, Result};
+use crate::utils::image_convert::{ColorMode, image_to_ndarray};
+use crate::utils::image_processor;
+use crate::utils::image_processor::{abs_diff, min_max_loc, ndarray_to_luma8, rgb_to_gray};
+use image::DynamicImage;
+use image::Luma;
+use imageproc::contrast::{ThresholdType, threshold};
+use imageproc::distance_transform::Norm;
+use imageproc::edges::canny;
+use imageproc::morphology::{close, open};
+use imageproc::region_labelling::{Connectivity, connected_components};
+use imageproc::template_matching::{MatchTemplateMethod, match_template};
+use ndarray::{ArrayView2, ArrayView3};
+use std::fmt;
+use tracing::debug;
+
+/// 滑块匹配结果：检测中心坐标与置信度。
+#[derive(Debug)]
+pub struct SlideResult {
+    /// 检测中心坐标 `[x, y]`。
+    pub target: [i32; 2],
+    /// 检测中心的 x 坐标。
+    pub target_x: i32,
+    /// 检测中心的 y 坐标。
+    pub target_y: i32,
+    /// 匹配置信度。
+    pub confidence: f64,
+}
+impl fmt::Display for SlideResult {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        writeln!(f, "滑块匹配测试结果:")?;
+        writeln!(f, "检测坐标: [x: {}, y: {}]", self.target_x, self.target_y)?;
+        // 注意：这里保留 4 位小数，如果想让外部控制，也可以直接写 {:.4}
+        write!(f, "置信度: {:.4}", self.confidence)?;
+        Ok(())
+    }
+}
+
+/// 滑块匹配服务：提供模板匹配与差异比较两种识别模式。
+#[derive(Default)]
+pub struct Slider;
+
+impl Slider {
+    /// 创建滑块匹配服务。
+    pub fn new() -> Self {
+        Self
+    }
+    /// 模板匹配滑块：在背景图中定位滑块中心（对应 Python 的 `slide_match`）。
+    pub fn slide_match(
+        &self,
+        target_image: &DynamicImage,
+        background_image: &DynamicImage,
+        simple_target: bool,
+    ) -> Result<SlideResult> {
+        let target_array = image_to_ndarray(target_image, ColorMode::RGB)?;
+        let background_array = image_to_ndarray(background_image, ColorMode::RGB)?;
+
+        self.perform_slide_match(target_array.view(), background_array.view(), simple_target)
+            .map_err(Into::into)
+    }
+    /// 差异比较滑块：对比带坑位的图与原始背景图，定位差异中心（对应 Python 的 `slide_comparison`）。
+    pub fn slide_comparison(
+        &self,
+        target_image: &DynamicImage,
+        background_image: &DynamicImage,
+    ) -> Result<SlideResult> {
+        // 1. 转换为 ndarray (HWC RGB)
+        let target_array = image_to_ndarray(target_image, ColorMode::RGB)?;
+        let background_array = image_to_ndarray(background_image, ColorMode::RGB)?;
+
+        // 2. 执行比较逻辑 (对应 _perform_slide_comparison)
+        self.perform_slide_comparison(target_array.view(), background_array.view())
+            .map_err(Into::into)
+    }
+    /// 差异比较核心实现（对应 Python 的 `_perform_slide_comparison`）。
+    pub fn perform_slide_comparison(
+        &self,
+        target: ArrayView3<u8>,
+        background: ArrayView3<u8>,
+    ) -> Result<SlideResult, ImagePreprocessError> {
+        // 1. 计算差异数组 (复用 cv2::absdiff)
+        let (th, tw, tc) = target.dim();
+        let (bh, bw, bc) = background.dim();
+
+        // 1. 比较模式下的严格尺寸校验
+        if th != bh || tw != bw || tc != bc {
+            return Err(ImagePreprocessError::MismatchDimensions(format!(
+                "比较模式要求两张图分辨率与通道数完全一致！Target: [{}x{}x{}], Background: [{}x{}x{}]",
+                tw, th, tc, bw, bh, bc
+            )));
+        }
+        if th == 0 || tw == 0 {
+            return Err(ImagePreprocessError::InvalidDimensions {
+                expected: "输入图像尺寸不能为0".to_string(),
+                actual: vec![th, tw],
+            });
+        }
+
+        let diff_array = abs_diff(&target, &background);
+
+        // 2. 转换为灰度数组 (复用你的 cv2.cvtColor)
+        let gray_array = rgb_to_gray(diff_array.view());
+        // 3. 转为 ImageBuffer 以使用 imageproc 的高级功能
+        let gray_buffer = ndarray_to_luma8(gray_array.view());
+
+        // 2. 二值化 (对应 cv2.threshold(..., 30, 255, cv2.THRESH_BINARY))
+        let binary = threshold(&gray_buffer, 30, ThresholdType::Binary);
+        // 3. 形态学操作去噪 (对应 cv2.morphologyEx)
+        // 闭运算 (Close): 先膨胀后腐蚀，用于填补缺口内的细小黑色空洞
+        // 开运算 (Open): 先腐蚀后膨胀，用于消除背景中的白色噪点点
+        let norm = Norm::LInf; // 对应 3x3 的矩形内核
+        let radius = 1u8; // 1 表示 3x3 的范围，2 表示 5x5 的范围
+        let closed = close(&binary, norm, radius);
+        let cleaned = open(&closed, norm, radius);
+
+        // connected_components 会给每个独立的白色区域打上不同的标签 (ID)
+        let background_label = Luma([0u8]);
+        let labelled = connected_components(&cleaned, Connectivity::Eight, background_label);
+
+        // 4. 寻找最大连通区域 (对应 findContours + max area)
+        if let Some(max_label) = image_processor::find_contours_and_max(&labelled) {
+            // 5. 计算最大区域的边界框 (对应 cv2.boundingRect)
+            let (x, y, w, h) = image_processor::bounding_rect(&labelled, max_label);
+            // 6. 计算中心点 (调用之前封装的 calculate_center)
+            let (center_x, center_y) =
+                image_processor::calculate_center((x, y), w as usize, h as usize);
+
+            Ok(SlideResult {
+                target: [center_x, center_y],
+                target_x: center_x,
+                target_y: center_y,
+                confidence: 1.0, // Comparison 模式下通常认为找到即为 1.0
+            })
+        } else {
+            Ok(SlideResult {
+                target: [0, 0],
+                target_x: 0,
+                target_y: 0,
+                confidence: 0.0,
+            })
+        }
+    }
+
+    /// 模板匹配核心实现（对应 Python 的 `_perform_slide_match`）。
+    fn perform_slide_match(
+        &self,
+        target: ArrayView3<u8>,
+        background: ArrayView3<u8>,
+        simple_target: bool, // 增加这个参数
+    ) -> Result<SlideResult, ImagePreprocessError> {
+        let (th, tw, tc) = target.dim();
+        let (bh, bw, bc) = background.dim();
+
+        // 1. 严格的鲁棒性校验（防止底层的 imageproc 算子崩溃）
+        if th == 0 || tw == 0 || bh == 0 || bw == 0 {
+            return Err(ImagePreprocessError::InvalidDimensions {
+                expected: "输入图像的宽度或高度不能为0".to_string(),
+                actual: vec![th, tw, tc],
+            });
+        }
+        if th > bh || tw > bw {
+            return Err(ImagePreprocessError::TargetExceedsBackground {
+                target_w: tw,
+                target_h: th,
+                bg_w: bw,
+                bg_h: bh,
+            });
+        }
+        if tc != bc {
+            return Err(ImagePreprocessError::MismatchDimensions(format!(
+                "目标图与背景图的通道数不一致 (target: {}, bg: {})",
+                tc, bc
+            )));
+        }
+
+        // 1. 统一灰度化
+        let target_gray = rgb_to_gray(target);
+        let background_gray = rgb_to_gray(background);
+
+        if simple_target {
+            // 2a. 简单模式：直接在灰度图上匹配
+            Ok(self.simple_template_match(target_gray.view(), background_gray.view()))
+        } else {
+            // 2b. 复杂模式：先提取边缘，再匹配
+
+            Ok(self.edge_based_match(target_gray.view(), background_gray.view()))
+        }
+    }
+    /// 简单模式模板匹配：直接对灰度图做归一化互相关（对应 Python 的 `_simple_template_match`）。
+    fn simple_template_match(
+        &self,
+        target: ArrayView2<u8>,
+        background: ArrayView2<u8>,
+    ) -> SlideResult {
+        // 1. 将 ndarray 转换为 imageproc 需要的 ImageBuffer (无拷贝或轻量转换)
+        // 转换逻辑 (假设你已经有方法转回 ImageBuffer)
+        let t_buf = ndarray_to_luma8(target);
+        let b_buf = ndarray_to_luma8(background);
+
+        // 2. 调用 imageproc 的 NCC 算法 (等价于 cv2.TM_CCOEFF_NORMED)
+        // 模板匹配 (完全对齐 cv2.matchTemplate(..., cv2.TM_CCOEFF_NORMED))
+        let result = match_template(
+            &b_buf,
+            &t_buf,
+            MatchTemplateMethod::CrossCorrelationNormalized,
+        );
+        // 3. 寻找最大值 (等价于 cv2.minMaxLoc)
+        let (max_val, max_loc) = min_max_loc(&result);
+
+        // 4. 计算中心点 (与 Python 逻辑完全一致)
+        let (th, tw) = target.dim();
+
+        let (center_x, center_y) = image_processor::calculate_center(max_loc, tw, th);
+        SlideResult {
+            target: [center_x, center_y],
+            target_x: center_x,
+            target_y: center_y,
+            confidence: max_val as f64,
+        }
+    }
+
+    /// 边缘模式模板匹配：基于 Canny 边缘检测后再匹配（对应 Python 的 `_edge_based_match`）。
+    pub fn edge_based_match(
+        &self,
+        target: ArrayView2<u8>,
+        background: ArrayView2<u8>,
+    ) -> SlideResult {
+        // 1. 将 ndarray 转换为 ImageBuffer
+        // 注意：Canny 和 match_template 需要 ImageBuffer 格式
+        let t_buf = ndarray_to_luma8(target);
+        let b_buf = ndarray_to_luma8(background);
+
+        // 2. 边缘检测 (完全对齐 cv2.Canny(50, 150))
+        // 这步会生成黑底白线的二值化边缘图
+        let target_edges = canny(&t_buf, 50.0, 150.0);
+        let background_edges = canny(&b_buf, 50.0, 150.0);
+
+        // 3. 模板匹配 (完全对齐 cv2.matchTemplate(..., cv2.TM_CCOEFF_NORMED))
+        // 在边缘图上计算归一化互相关系数
+        let result = match_template(
+            &background_edges,
+            &target_edges,
+            MatchTemplateMethod::CrossCorrelationNormalized,
+        );
+
+        // 4. 找到最佳匹配位置 (对齐 cv2.minMaxLoc)
+        let (max_val, max_loc) = min_max_loc(&result);
+        // 5. 计算中心位置 (对齐 Python 逻辑)
+        // target_w, target_h 来自输入数组的维度
+        let (th, tw) = target.dim();
+        let (center_x, center_y) = image_processor::calculate_center(max_loc, tw, th);
+
+        // 打印调试信息，方便与 Python 对比
+        debug!("-Rust Target Width (tw): {}", tw);
+        debug!("-Rust Best Max Loc X: {}", max_loc.0);
+        debug!("-Rust Final Center X: {}", center_x);
+        SlideResult {
+            target: [center_x, center_y],
+            target_x: center_x,
+            target_y: center_y,
+            confidence: max_val as f64,
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn paint_block(img: &mut DynamicImage, x: u32, y: u32, w: u32, h: u32) {
+        let mut luma = img.to_luma8();
+        for yy in y..y + h {
+            for xx in x..x + w {
+                luma.put_pixel(xx, yy, Luma([255u8]));
+            }
+        }
+        *img = DynamicImage::ImageLuma8(luma);
+    }
+
+    #[test]
+    fn slide_match_finds_block_center() {
+        let slider = Slider::new();
+        let mut target = DynamicImage::new_luma8(10, 10);
+        paint_block(&mut target, 3, 3, 4, 4);
+        let mut background = DynamicImage::new_luma8(30, 30);
+        paint_block(&mut background, 8, 8, 4, 4);
+
+        let res = slider.slide_match(&target, &background, true).unwrap();
+        // 模板内白块位于 (3,3)，因此最佳匹配原点为 (8-3, 8-3)=(5,5)，中心为 (5+5, 5+5)
+        assert_eq!(res.target_x, 10);
+        assert_eq!(res.target_y, 10);
+        assert!((res.confidence - 1.0).abs() < 1e-3);
+    }
+
+    #[test]
+    fn slide_comparison_identical_returns_zero() {
+        let slider = Slider::new();
+        let mut img = DynamicImage::new_luma8(16, 16);
+        paint_block(&mut img, 4, 4, 4, 4);
+
+        let res = slider.slide_comparison(&img, &img).unwrap();
+        assert_eq!(res.target, [0, 0]);
+        assert_eq!(res.confidence, 0.0);
+    }
+}

ddddocr-core/src/traits.rs

@@ -0,0 +1,55 @@
+//! 推理引擎统一抽象接口。
+
+use crate::error::TensorError;
+use crate::types::{ModelInfo, TensorInfo};
+use crate::{DetOutput, ModelMetadata, OcrOutput};
+use std::path::Path;
+
+/// 查询模型输入/输出信息的接口。
+pub trait Info {
+    /// 获取输入张量信息列表。
+    fn input_info(&self) -> crate::error::Result<Vec<TensorInfo>>;
+    /// 获取输出张量信息列表。
+    fn output_info(&self) -> crate::error::Result<Vec<TensorInfo>>;
+    /// 获取模型完整输入/输出信息。
+    fn model_info(&self) -> crate::error::Result<ModelInfo>;
+}
+
+/// 推理引擎统一抽象接口，由 ddddocr-tract2、ddddocr-ort 等引擎 crate 实现。
+pub trait InferenceEngine {
+    /// 引擎产出的输出枚举（OCR 为 [`crate::OcrOutput`]，检测为 [`crate::DetOutput`]）。
+    type Output;
+    /// 对输入张量执行推理并返回引擎定义的输出。
+    fn inference(
+        &self,
+        input_array: ndarray::Array4<f32>,
+    ) -> crate::error::Result<Self::Output, TensorError>;
+}
+
+/// OCR 引擎接口：输出 [`crate::OcrOutput`]，并提供模型元数据。
+pub trait OcrEngine: InferenceEngine<Output = OcrOutput> + Info {
+    /// 获取模型元数据。
+    fn metadata(&self) -> &ModelMetadata;
+}
+
+/// 目标检测引擎接口：输出 [`crate::DetOutput`]。
+pub trait DetEngine: InferenceEngine<Output = DetOutput> {}
+
+/// 模型加载器：从本地路径或字节流构建引擎会话。
+pub trait Loader {
+    /// 构建出的引擎会话类型。
+    type Session;
+    /// 构建过程中的错误类型。
+    type Error;
+    /// 从本地模型路径构建会话。
+    fn build_for_path<P: AsRef<Path>>(
+        &self,
+        model_path: P,
+    ) -> crate::error::Result<Self::Session, Self::Error>;
+
+    /// 从模型字节流构建会话。
+    fn build_from_bytes(
+        &self,
+        model_bytes: &[u8],
+    ) -> crate::error::Result<Self::Session, Self::Error>;
+}

ddddocr-core/src/types.rs

@@ -0,0 +1,60 @@
+//! 模型输入输出信息等共享类型，由 [`crate::traits::Info`] 接口返回。
+
+/// 张量元素的数据类型标记。
+#[derive(Debug, Clone)]
+pub enum TensorType {
+    /// 32 位浮点类型。
+    F32,
+    /// 64 位整数类型。
+    I64,
+    /// 其他类型。
+    Other,
+}
+
+/// 模型某个轴的维度特征：静态数值或动态符号。
+#[derive(Clone, PartialEq, Eq)]
+pub enum AxisDim {
+    /// 静态固定维度。
+    Static(usize),
+    /// 动态符号维度。
+    Dynamic(String),
+}
+
+impl AxisDim {
+    /// 是否为动态维度。
+    pub fn is_dynamic(&self) -> bool {
+        matches!(self, AxisDim::Dynamic(_))
+    }
+}
+
+/// 自定义 `Debug` 输出：静态维度显示数值，动态维度显示符号名。
+impl std::fmt::Debug for AxisDim {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            AxisDim::Static(size) => write!(f, "{}", size),
+            AxisDim::Dynamic(expr) => write!(f, "Dynamic(\"{}\")", expr),
+        }
+    }
+}
+
+/// 单个张量（输入或输出）的名称、形状与数据类型描述。
+#[derive(Debug, Clone)]
+pub struct TensorInfo {
+    /// 张量名称。
+    pub name: String,
+    /// 各轴形状描述。
+    pub shape: Vec<AxisDim>,
+    /// 元素数据类型。
+    pub tensor_type: TensorType,
+}
+
+/// 模型的完整输入/输出信息。
+#[derive(Debug, Clone)]
+pub struct ModelInfo {
+    /// 输入张量列表。
+    pub inputs: Vec<TensorInfo>,
+    /// 输出张量列表。
+    pub outputs: Vec<TensorInfo>,
+    /// 硬件执行提供者（`None` 表示使用引擎默认后端）。
+    pub providers: Option<Vec<String>>,
+}

ddddocr-core/src/utils.rs

@@ -0,0 +1,10 @@
+//! 图像与张量工具：加载、格式转换、图像处理与归一化。
+
+pub mod image_convert;
+mod image_helper;
+pub mod image_processor;
+mod tensor_transform;
+
+// 对外统一暴露干净的 API 语义层
+pub use image_convert::ColorMode;
+pub use tensor_transform::normalize_ocr_logits;

ddddocr-core/src/utils/image_convert.rs

@@ -0,0 +1,217 @@
+//! 图像格式转换：DynamicImage 与 ndarray 数组互转。
+
+use crate::error::{ImagePreprocessError, Result};
+use image::{DynamicImage, GenericImageView, ImageBuffer, Luma, Rgb, Rgba};
+use ndarray::{Array3, ArrayViewD};
+
+/// 图像通道模式。
+#[derive(Debug)]
+pub enum ColorMode {
+    /// RGB 三通道。
+    RGB,
+    /// RGBA 四通道。
+    RGBA,
+    /// 灰度单通道。
+    L,
+}
+
+/// 将 ndarray 数组转换为图像（自动识别 HWC 通道数）。
+// 对应 Python 版 _numpy_to_pil_image
+pub fn ndarray_to_hwc_image(array: ArrayViewD<u8>) -> Result<DynamicImage, ImagePreprocessError> {
+    let shape = array.shape();
+    let dim = shape.len();
+
+    // 1. 确保数据在内存中是连续的 (C order / Standard Layout)
+    // 如果 arr 是经过切片或转置的，这一步会进行必要的内存拷贝
+
+    let color_mode = match dim {
+        // 对应 Python: len(array.shape) == 2 (灰度图 H, W)
+        2 => ColorMode::L,
+
+        // 对应 Python: len(array.shape) == 3 (H, W, C)
+        3 => {
+            let (_h, _w, c) = (shape[0], shape[1], shape[2]);
+            match c {
+                // 对应 Python: array.shape[2] == 1 (单通道 H, W, 1)
+                1 => ColorMode::L,
+                // 对应 Python: array.shape[2] == 3 (RGB H, W, 3)
+                3 => ColorMode::RGB,
+                // 对应 Python: array.shape[2] == 4 (RGBA H, W, 4)
+                4 => ColorMode::RGBA,
+                _ => {
+                    return Err(ImagePreprocessError::UnsupportedChannels(c));
+                }
+            }
+        }
+        _ => {
+            return Err(ImagePreprocessError::InvalidDimensions {
+                expected: "2D (H,W) 或 3D (H,W,C)".to_string(),
+                actual: shape.to_vec(),
+            });
+        }
+    };
+    from_ndarray(array, color_mode)
+}
+
+/// 处理PNG图片的RGBA透明背景，将透明部分设置为白色背景
+// 对应 Python 的 png_rgba_black_preprocess
+pub fn png_rgba_white_preprocess(img: &DynamicImage) -> DynamicImage {
+    // 1. 检查是否包含透明通道，如果没有，直接克隆并返回
+    if !img.color().has_alpha() {
+        return DynamicImage::ImageRgb8(img.to_rgb8());
+    }
+
+    let (width, height) = img.dimensions();
+
+    // 2. 创建一个新的 RGB 图像缓冲，默认填充为白色 (255, 255, 255)
+    let mut background = ImageBuffer::from_pixel(width, height, Rgb([255u8, 255u8, 255u8]));
+
+    // 3. 获取原图的 RGBA 视图
+    let rgba_img = img.to_rgba8();
+
+    // 4. 遍历像素并手动进行 Alpha 混合
+    // 对应 Python 的 utils.paste(img, ..., mask=img)
+    // 使用 enumerate_pixels_mut 同时获取坐标和背景像素的可变引用，减少查找开销
+    for (x, y, bg_pixel) in background.enumerate_pixels_mut() {
+        // 安全性说明：x, y 源自 background 尺寸，与 rgba_img 一致，get_pixel 是安全的
+        let src_pixel = rgba_img.get_pixel(x, y);
+        let alpha_u8 = src_pixel[3];
+
+        match alpha_u8 {
+            // 情况 A：完全不透明，直接覆盖背景色
+            255 => {
+                bg_pixel.0 = [src_pixel[0], src_pixel[1], src_pixel[2]];
+            }
+            // 情况 B：完全透明，保持背景色（白色），无需操作
+            0 => {
+                continue;
+            }
+            // 情况 C：半透明，进行 Alpha 混合计算
+            _ => {
+                let alpha = alpha_u8 as f32 / 255.0;
+                let inv_alpha = 1.0 - alpha;
+
+                bg_pixel[0] = (src_pixel[0] as f32 * alpha + 255.0 * inv_alpha).round() as u8;
+                bg_pixel[1] = (src_pixel[1] as f32 * alpha + 255.0 * inv_alpha).round() as u8;
+                bg_pixel[2] = (src_pixel[2] as f32 * alpha + 255.0 * inv_alpha).round() as u8;
+            }
+        }
+    }
+
+    DynamicImage::ImageRgb8(background)
+}
+/// 将 DynamicImage 转换为 array 数组
+pub fn image_to_ndarray(
+    image: &DynamicImage,
+    mode: ColorMode,
+) -> Result<Array3<u8>, ImagePreprocessError> {
+    // Rust utils 库通过 to_rgb8, to_luma8 等方法实现转换
+    let (width, height) = image.dimensions();
+
+    let (channels, raw) = match mode {
+        ColorMode::L => (1, image.to_luma8().into_raw()),
+        ColorMode::RGB => (3, image.to_rgb8().into_raw()),
+        ColorMode::RGBA => (4, image.to_rgba8().into_raw()),
+    };
+
+    let array = Array3::from_shape_vec((height as usize, width as usize, channels), raw)
+        .map_err(ImagePreprocessError::from)?;
+    Ok(array)
+}
+/// 将 array 数组转换为 DynamicImage
+pub fn ndarray_to_image(
+    array: ArrayViewD<u8>,
+    mode: ColorMode,
+) -> Result<DynamicImage, ImagePreprocessError> {
+    let shape = array.shape();
+
+    // 基础边界检查：至少要有 H 和 W 两个维度
+    if shape.len() < 2 {
+        return Err(ImagePreprocessError::InvalidDimensions {
+            expected: "至少为 2D array [H, W]".to_string(),
+            actual: shape.to_vec(),
+        });
+    }
+    from_ndarray(array, mode)
+}
+
+fn from_ndarray(
+    array: ArrayViewD<u8>,
+    mode: ColorMode,
+) -> Result<DynamicImage, ImagePreprocessError> {
+    let shape = array.shape();
+
+    // 映射：ndarray 的 shape 默认是 [Height, Width, (Channels)]
+    // image 库的 from_raw 接收 (width, height)
+    let height = shape[0] as u32;
+    let width = shape[1] as u32;
+
+    // 1. 确保数据在内存中是连续的 (C order)
+    let standard = array.as_standard_layout();
+    let (raw_data, _) = standard.to_owned().into_raw_vec_and_offset();
+    let raw_len = raw_data.len();
+
+    // 获取当前模式对应的通道数
+    let channels = match mode {
+        ColorMode::L => 1,
+        ColorMode::RGB => 3,
+        ColorMode::RGBA => 4,
+    };
+
+    let expected_len = (width * height) as usize * channels;
+
+    // 构造通用错误闭包，避免 match 分支中重复编写冗长的错误对象
+    let make_err = || ImagePreprocessError::BufferLengthMismatch {
+        expected: expected_len,
+        actual: raw_len,
+        width,
+        height,
+        channels,
+    };
+
+    // 2. 重新解释内存并构建 ImageBuffer
+    match mode {
+        ColorMode::L => ImageBuffer::<Luma<u8>, _>::from_raw(width, height, raw_data)
+            .map(DynamicImage::ImageLuma8)
+            .ok_or_else(make_err),
+        ColorMode::RGB => ImageBuffer::<Rgb<u8>, _>::from_raw(width, height, raw_data)
+            .map(DynamicImage::ImageRgb8)
+            .ok_or_else(make_err),
+        ColorMode::RGBA => ImageBuffer::<Rgba<u8>, _>::from_raw(width, height, raw_data)
+            .map(DynamicImage::ImageRgba8)
+            .ok_or_else(make_err),
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn image_to_ndarray_rgb_dims() {
+        let img = DynamicImage::new_rgb8(4, 3);
+        let arr = image_to_ndarray(&img, ColorMode::RGB).unwrap();
+        assert_eq!(arr.dim(), (3, 4, 3));
+    }
+
+    #[test]
+    fn ndarray_roundtrip() {
+        let img = DynamicImage::new_rgb8(2, 2);
+        let arr = image_to_ndarray(&img, ColorMode::RGB).unwrap();
+        let back = ndarray_to_image(arr.view().into_dyn(), ColorMode::RGB).unwrap();
+        assert_eq!(back.to_rgb8().dimensions(), (2, 2));
+    }
+
+    #[test]
+    fn png_white_preprocess_fills_transparent() {
+        let img = DynamicImage::ImageRgba8(image::RgbaImage::from_pixel(
+            2,
+            2,
+            image::Rgba([255, 0, 0, 0]),
+        ));
+        let out = png_rgba_white_preprocess(&img);
+        for p in out.to_rgb8().pixels() {
+            assert_eq!(p.0, [255, 255, 255]);
+        }
+    }
+}

ddddocr-core/src/utils/image_helper.rs

@@ -0,0 +1,156 @@
+//! 图像输入源：从路径、Base64、字节等统一加载图片。
+
+use crate::error::{DdddError, Result};
+use crate::utils::image_convert::ndarray_to_hwc_image;
+use base64::{Engine as _, engine::general_purpose};
+use image::DynamicImage;
+use ndarray::ArrayViewD;
+use std::fmt;
+use std::fmt::{Debug, Formatter};
+use std::fs;
+use std::path::Path;
+use std::path::PathBuf;
+/// Base64 字符串包装。
+pub struct Base64<'a>(pub &'a str);
+/// 统一图像输入源，包装加载完成的图像。
+pub struct ImageSource {
+    inner: DynamicImage,
+}
+// 1. 将 into_inner 优化为 into_image，符合 Rust 官方命名规范
+impl ImageSource {
+    /// 消耗当前包装器，获取最终的 DynamicImage
+    pub fn into_image(self) -> DynamicImage {
+        self.inner
+    }
+}
+
+/// 从多种输入类型转换为 [`ImageSource`] 的转换接口。
+pub trait TryFromImage<T>: Sized {
+    // 唯一的转换入口，通过目标类型来调用
+    fn try_from_image(value: T) -> Result<Self>;
+}
+
+// 1. 本身是 DynamicImage
+impl TryFromImage<DynamicImage> for ImageSource {
+    fn try_from_image(img: DynamicImage) -> Result<Self> {
+        Ok(Self { inner: img })
+    }
+}
+#[derive(Debug)]
+enum Base64ProcessError {
+    InvalidBase64Header,
+}
+impl fmt::Display for Base64ProcessError {
+    fn fmt(&self, f: &mut Formatter<'_>) -> fmt::Result {
+        match self {
+            Base64ProcessError::InvalidBase64Header => {
+                write!(f, "Base64 头部格式不正确，缺少 ';base64,' 分隔符")
+            }
+        }
+    }
+}
+
+impl std::error::Error for Base64ProcessError {}
+// 2.2 路径类型 A: &str (最常用)
+impl<'a> TryFromImage<&'a str> for ImageSource {
+    fn try_from_image(path_or_b64: &'a str) -> Result<Self> {
+        // 1. 嗅探：如果包含 Base64 特征
+        if path_or_b64.starts_with("data:image/") && path_or_b64.contains(";base64,") {
+            // 提取出真正的 base64 数据部分
+            let (_, clean_b64) = path_or_b64.split_once(";base64,").ok_or_else(|| {
+                // 返回一个明确的、可读性极佳的格式错误
+                DdddError::new(Base64ProcessError::InvalidBase64Header)
+            })?;
+            // 转换为 Base64 包装器，并复用其 TryFromImage 实现
+            Self::try_from_image(Base64(clean_b64))
+        } else {
+            // 2. 否则，老老实实当作本地路径打开
+            let img = image::open(path_or_b64).map_err(DdddError::new)?;
+            Ok(Self { inner: img })
+        }
+    }
+}
+
+// 2.3 路径类型 B: &Path (标准借用)
+impl<'a> TryFromImage<&'a Path> for ImageSource {
+    fn try_from_image(path: &'a Path) -> Result<Self> {
+        let img = image::open(path).map_err(DdddError::new)?;
+        Ok(Self { inner: img })
+    }
+}
+// 2.4 路径类型 C: PathBuf / String (拥有所有权，透传给借用)
+impl TryFromImage<PathBuf> for ImageSource {
+    fn try_from_image(path: PathBuf) -> Result<Self> {
+        Self::try_from_image(path.as_path())
+    }
+}
+
+impl TryFromImage<String> for ImageSource {
+    fn try_from_image(path_or_b64: String) -> Result<Self> {
+        Self::try_from_image(path_or_b64.as_str())
+    }
+}
+// 2. 支持带有生命周期的借用：直接支持 &[u8]（不强绑生命周期到 ImageSource 结构体上！）
+impl<'a> TryFromImage<&'a [u8]> for ImageSource {
+    fn try_from_image(bytes: &'a [u8]) -> Result<Self> {
+        let img = image::load_from_memory(bytes).map_err(DdddError::new)?;
+        Ok(Self { inner: img })
+    }
+}
+
+// 4. 完美支持 ndarray 的借用 ArrayViewD
+impl<'a> TryFromImage<ArrayViewD<'a, u8>> for ImageSource {
+    fn try_from_image(array: ArrayViewD<'a, u8>) -> Result<Self> {
+        let img = ndarray_to_hwc_image(array)?;
+        Ok(Self { inner: img })
+    }
+}
+
+impl<'a> TryFromImage<Base64<'a>> for ImageSource {
+    fn try_from_image(b64_str: Base64<'a>) -> Result<Self> {
+        let img = base64_to_image(b64_str.0)?;
+        Ok(Self { inner: img })
+    }
+}
+
+/// 从任意受支持的输入类型加载图像。
+#[allow(dead_code)]
+pub fn load_image_from_input<I>(input: I) -> Result<DynamicImage>
+where
+    ImageSource: TryFromImage<I>,
+{
+    let img = ImageSource::try_from_image(input)?.into_image();
+    Ok(img)
+}
+
+/// 将 Base64 编码的图片转换为 DynamicImage。
+pub fn base64_to_image(b64_str: &str) -> Result<DynamicImage> {
+    // 过滤掉可能存在的 base64 前缀，例如 "data:utils/png;base64,"
+    let clean_b64 = if let Some(pos) = b64_str.find(",") {
+        &b64_str[pos + 1..]
+    } else {
+        &b64_str
+    };
+
+    let bytes = general_purpose::STANDARD
+        .decode(clean_b64.trim())
+        // .map_err(|e| DdddError::new(e))?;
+        .map_err(DdddError::new)?;
+
+    let img = image::load_from_memory(&bytes).map_err(DdddError::new)?;
+    Ok(img)
+}
+
+/// 读取图片文件并转换为 base64 编码字符串
+// 对应 Python 版 get_img_base64
+pub fn img_to_base64<P: AsRef<Path>>(image_path: P) -> Result<String> {
+    // 1. 读取文件原始字节流
+    // 使用 AsRef<Path> 泛型可以让函数同时支持 String, &str, PathBuf 等类型
+    let image_data = fs::read(&image_path).map_err(DdddError::new)?;
+
+    // 2. 进行 Base64 编码
+    // 使用 STANDARD 引擎对齐 Python 的 base64.b64encode
+    let b64_string = general_purpose::STANDARD.encode(image_data);
+
+    Ok(b64_string)
+}

ddddocr-core/src/utils/image_processor.rs

@@ -0,0 +1,232 @@
+//! 图像处理算法：OpenCV 风格的常用函数封装。
+
+use image::{DynamicImage, GrayImage, ImageBuffer, Luma, imageops::FilterType};
+
+use ndarray::{Array2, Array3, ArrayView2, ArrayView3, azip};
+use std::cmp::{max, min};
+
+// 模拟openCV
+/// 计算两个 HWC 数组的绝对差值（对应 cv2.absdiff）。
+pub fn abs_diff(a: &ArrayView3<u8>, b: &ArrayView3<u8>) -> Array3<u8> {
+    // 利用 ndarray 的 map_collect，生成差值的绝对值数组
+    // 或者直接使用 zip_mut_with 处理以减少内存分配
+    let mut diff = Array3::zeros(a.dim());
+    azip!((res in &mut diff, &va in a, &vb in b) {
+        *res = va.abs_diff(vb);
+    });
+    diff
+}
+
+/// RGB 到灰度转换
+pub fn rgb_to_gray(rgb: ArrayView3<u8>) -> Array2<u8> {
+    let (h, w, _) = rgb.dim();
+    Array2::from_shape_fn((h, w), |(y, x)| {
+        let r = rgb[[y, x, 0]] as f32;
+        let g = rgb[[y, x, 1]] as f32;
+        let b = rgb[[y, x, 2]] as f32;
+        // 完全忽略 a，只按权重计算
+        (0.299 * r + 0.587 * g + 0.114 * b) as u8
+    })
+}
+
+/// 查找匹配结果图中的最大值及其坐标（对应 cv2.minMaxLoc）。
+pub fn min_max_loc(result_map: &ImageBuffer<Luma<f32>, Vec<f32>>) -> (f32, (u32, u32)) {
+    // 4. 找到最佳匹配位置 (对齐 cv2.minMaxLoc)
+    let mut max_val: f32 = -1.0;
+    let mut max_loc = (0, 0);
+
+    // 遍历匹配得分图
+    for (x, y, score) in result_map.enumerate_pixels() {
+        let s = score.0[0];
+
+        // 可以在此处加入你之前验证过的起始位过滤
+        // if x < 15 { continue; }
+
+        if s > max_val {
+            max_val = s;
+            max_loc = (x, y);
+        }
+    }
+    (max_val, max_loc)
+}
+
+/// 模拟 findContours：返回面积最大的连通域标签，找不到时返回 `None`。
+pub fn find_contours_and_max(labelled: &ImageBuffer<Luma<u32>, Vec<u32>>) -> Option<u32> {
+    // 统计每个标签出现的频率（即面积）
+    let mut max_label = 0;
+    let mut max_area = 0;
+    let mut areas = std::collections::HashMap::new();
+
+    for pixel in labelled.pixels() {
+        let label = pixel.0[0];
+        if label == 0 {
+            continue;
+        } // 跳过背景
+        let count = areas.entry(label).or_insert(0);
+        *count += 1;
+        if *count > max_area {
+            max_area = *count;
+            max_label = label;
+        }
+    }
+    if max_label == 0 {
+        None
+    } else {
+        Some(max_label)
+    }
+}
+/// 计算指定连通域标签的外接矩形（对应 cv2.boundingRect），返回 `(min_x, min_y, width, height)`。
+pub fn bounding_rect(
+    labelled: &ImageBuffer<Luma<u32>, Vec<u32>>,
+    max_label: u32,
+) -> (u32, u32, u32, u32) {
+    // 5. 计算最大区域的边界框 (对应 cv2.boundingRect)
+    let mut min_x = labelled.width();
+    let mut max_x = 0;
+    let mut min_y = labelled.height();
+    let mut max_y = 0;
+
+    for (x, y, pixel) in labelled.enumerate_pixels() {
+        if pixel.0[0] == max_label {
+            min_x = min(min_x, x);
+            max_x = max(max_x, x);
+            min_y = min(min_y, y);
+            max_y = max(max_y, y);
+        }
+    }
+
+    let w = max_x - min_x;
+    let h = max_y - min_y;
+    (min_x, min_y, w, h)
+}
+
+/// 根据左上角坐标与矩形长宽，计算其中央核心点坐标
+#[inline]
+pub fn calculate_center(top_left: (u32, u32), width: usize, height: usize) -> (i32, i32) {
+    let center_x = top_left.0 as i32 + (width as i32 / 2);
+    let center_y = top_left.1 as i32 + (height as i32 / 2);
+    (center_x, center_y)
+}
+
+/// 将 2D 灰度 ndarray 视图转换为灰度 ImageBuffer。
+pub fn ndarray_to_luma8(array: ArrayView2<u8>) -> ImageBuffer<Luma<u8>, Vec<u8>> {
+    let (height, width) = array.dim();
+    // 技巧：直接将已有的规整连续内存打平转换，或用 from_raw 包装
+    // 此处保留安全的一步转换，但用更内聚的迭代器或切片拷贝进行速度优化
+    let mut buffer = ImageBuffer::new(width as u32, height as u32);
+    for y in 0..height {
+        for x in 0..width {
+            buffer.put_pixel(x as u32, y as u32, Luma([array[[y, x]]]));
+        }
+    }
+    buffer
+}
+// =====================================================================
+// 5. 核心高性能图像转换算法 (纯 Rust 编写)
+// =====================================================================
+
+/// RGB 像素转换为 OpenCV 风格的 HSV 值。
+#[inline(always)]
+pub fn rgb_to_opencv_hsv(r: u8, g: u8, b: u8) -> (u8, u8, u8) {
+    // 1. 规避高昂的除法，直接转为 f32 进行比对
+    let r_f = r as f32;
+    let g_f = g as f32;
+    let b_f = b as f32;
+
+    let max = r_f.max(g_f).max(b_f);
+    let min = r_f.min(g_f).min(b_f);
+    let delta = max - min;
+
+    // 2. 计算 H (色调) - 移除负数取余陷阱，改用平铺分支
+    let h = if delta == 0.0 {
+        0.0
+    } else if max == r_f {
+        let mut diff = (g_f - b_f) / delta;
+        if diff < 0.0 {
+            diff += 6.0; // 规避 Rust f32 % 负数的行为
+        }
+        60.0 * diff
+    } else if max == g_f {
+        60.0 * (((b_f - r_f) / delta) + 2.0)
+    } else {
+        60.0 * (((r_f - g_f) / delta) + 4.0)
+    };
+
+    // OpenCV 的 H 量化：H / 2
+    // 注意：OpenCV 底层使用截断还是四舍五入与特定版本有关，
+    // 标准的 cvtColor 内部实现通常是： h * (180.0 / 360.0) -> h * 0.5
+    // 这里使用强转(截断)，若单测对齐发现差1，可改为 (h * 0.5 + 0.5) 或 round()
+    let h_opencv = (h * 0.5) as u8;
+
+    // 3. 计算 S (饱和度)
+    // OpenCV 公式: S = max == 0 ? 0 : 255 * delta / max
+    let s_opencv = if max == 0.0 {
+        0
+    } else {
+        ((255.0 * delta) / max) as u8
+    };
+
+    // 4. 计算 V (明度)
+    let v_opencv = max as u8;
+
+    (h_opencv, s_opencv, v_opencv)
+}
+
+/// 将图像转换为灰度图（L 模式）。
+pub fn convert_to_grayscale(image: &DynamicImage) -> GrayImage {
+    // Rust utils 库的 to_luma8 会根据标准的亮度公式进行转换
+    image.to_luma8()
+}
+
+/// 按指定宽高调整图像尺寸。
+pub fn resize_image(
+    image: &DynamicImage,
+    target_width: u32,
+    target_height: u32,
+    // resample 参数我们直接使用 FilterType，Lanczos3 是最接近 Python LANCZOS 的
+) -> DynamicImage {
+    // image::imageops::resize 的最高层封装
+    // FilterType::Lanczos3 与 Python Pillow 的 Image.LANCZOS 算法完全对齐，缩放质量最高
+    image.resize_exact(target_width, target_height, FilterType::Lanczos3)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn abs_diff_absolutes() {
+        let a = Array3::from_shape_vec((1, 1, 3), vec![10, 200, 5]).unwrap();
+        let b = Array3::from_shape_vec((1, 1, 3), vec![200, 100, 5]).unwrap();
+        let d = abs_diff(&a.view(), &b.view());
+        assert_eq!(d[[0, 0, 0]], 190);
+        assert_eq!(d[[0, 0, 1]], 100);
+        assert_eq!(d[[0, 0, 2]], 0);
+    }
+
+    #[test]
+    fn rgb_to_gray_white_is_255() {
+        let rgb = Array3::from_shape_vec((1, 1, 3), vec![255, 255, 255]).unwrap();
+        assert_eq!(rgb_to_gray(rgb.view())[[0, 0]], 255);
+    }
+
+    #[test]
+    fn min_max_loc_finds_max() {
+        let buf = ImageBuffer::<Luma<f32>, Vec<f32>>::from_fn(3, 2, |x, y| {
+            Luma([if x == 2 && y == 1 { 0.9 } else { 0.1 }])
+        });
+        let (val, loc) = min_max_loc(&buf);
+        assert_eq!(val, 0.9);
+        assert_eq!(loc, (2, 1));
+    }
+
+    #[test]
+    fn calculate_center_midpoint() {
+        assert_eq!(calculate_center((10, 20), 6, 4), (13, 22));
+    }
+
+    #[test]
+    fn rgb_to_opencv_hsv_red() {
+        assert_eq!(rgb_to_opencv_hsv(255, 0, 0), (0, 255, 255));
+    }
+}

ddddocr-core/src/utils/tensor_transform.rs

@@ -0,0 +1,84 @@
+//! 张量变换：将异构形状的模型输出规整为统一格式。
+
+use crate::OcrOutput;
+use crate::error::{Result, TensorError};
+use ndarray::s;
+
+/// 将异构形状的模型输出规整为标准 `[Steps, Classes]` Logits 矩阵。
+pub fn normalize_ocr_logits(
+    array: ndarray::ArrayViewD<f32>,
+    shape: &[usize],
+) -> Result<OcrOutput, TensorError> {
+    let (steps, classes, data_dyn_view) = match shape.len() {
+        3 => {
+            if shape[1] == 1 {
+                // 形状: [Steps, 1, Classes]
+                (shape[0], shape[2], array)
+            } else if shape[0] == 1 {
+                // 形状: [1, Steps, Classes]
+                (shape[1], shape[2], array)
+            } else {
+                // 默认取第一个 batch: [Batch, Steps, Classes]
+                // 使用 ndarray 的 s! 宏，对应 Python 的 output[0, :, :]
+                let sliced = array.slice_move(s![0, .., ..]);
+                (shape[1], shape[2], sliced.into_dyn())
+            }
+        }
+        // 形状: [Steps, Classes]
+        2 => (shape[0], shape[1], array),
+        // 形状: [Classes] -> 单字符输出（对应 Python 的 ndim == 0 保护逻辑）
+        // 我们把它虚构成一个 [1, Classes] 的 2D 矩阵来复用后面的 argmax 逻辑
+        1 => (1, shape[0], array),
+        _ => {
+            return Err(TensorError::DimensionMismatch {
+                expected: "1D, 2D, or 3D OCR Logits".to_string(),
+                actual: shape.to_vec(),
+            });
+        }
+    };
+
+    // 转换为标准的 2D 静态矩阵 [Steps, Classes]
+    let matrix_cow = data_dyn_view
+        .to_shape(ndarray::Ix2(steps, classes))
+        .map_err(|shape_err| {
+            // 如果是因为切片导致不连续且无法进行零拷贝变换，抛出 NonContiguousMemory
+            if !data_dyn_view.is_standard_layout() {
+                TensorError::NonContiguousMemory
+            } else {
+                // 否则，说明是纯粹的数据元素数量不对（Shape 不匹配），抛出专属的强类型错误
+                TensorError::LogitsDimensionMismatch(shape_err)
+            }
+        })?
+        .to_owned();
+
+    Ok(OcrOutput::Logits(matrix_cow))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn normalize_2d_logits() {
+        let array = ndarray::Array2::<f32>::zeros((2, 3));
+        match normalize_ocr_logits(array.view().into_dyn(), &[2, 3]).unwrap() {
+            crate::OcrOutput::Logits(m) => assert_eq!(m.dim(), (2, 3)),
+            _ => panic!("应为 Logits 输出"),
+        }
+    }
+
+    #[test]
+    fn normalize_3d_batch_first() {
+        let array = ndarray::Array3::<f32>::zeros((1, 2, 3));
+        match normalize_ocr_logits(array.view().into_dyn(), &[1, 2, 3]).unwrap() {
+            crate::OcrOutput::Logits(m) => assert_eq!(m.dim(), (2, 3)),
+            _ => panic!("应为 Logits 输出"),
+        }
+    }
+
+    #[test]
+    fn unsupported_dim_errors() {
+        let array = ndarray::Array4::<f32>::zeros((1, 1, 1, 1));
+        assert!(normalize_ocr_logits(array.view().into_dyn(), &[1, 1, 1, 1]).is_err());
+    }
+}

ddddocr-core/tests/api_surface.rs

@@ -0,0 +1,72 @@
+//! 外部视角 API 测试：验证颜色过滤与字符集限制扩展点对外可用。
+
+use std::borrow::Cow;
+
+use ddddocr_core::color_any_of;
+use ddddocr_core::{
+    CharRestrict, ColorFilter, ColorPreset, HsvRange, IdRestrict, OcrBuilder, TokenFilter, any_of,
+};
+
+#[test]
+fn color_filter_setter_accepts_owned_preset() {
+    let builder = OcrBuilder::new().color_filter(ColorPreset::Red);
+    let _ = builder;
+}
+
+#[test]
+fn color_any_of_macro_expands_and_collects() {
+    let ranges = color_any_of!(ColorPreset::Red, ColorPreset::Blue)
+        .collect_to_vec()
+        .expect("收集颜色区间失败")
+        .expect("应存在有效颜色区间");
+    // Red 两段 + Blue 一段
+    assert_eq!(ranges.len(), 3);
+}
+
+#[test]
+fn charset_restrict_setter_accepts_owned_restrict() {
+    let builder = OcrBuilder::new().charset_restrict(CharRestrict::Digit);
+    let _ = builder;
+}
+
+#[test]
+fn any_of_macro_expands_and_filters() {
+    let tokens: Vec<Cow<'static, str>> = vec![
+        Cow::Borrowed(""),
+        Cow::Borrowed("1"),
+        Cow::Borrowed("a"),
+        Cow::Borrowed("A"),
+    ];
+
+    let indices = any_of!(CharRestrict::Digit, CharRestrict::Lowercase)
+        .apply_to_charset(&tokens)
+        .expect("字符集应存在交集");
+    // 0 号 blank 放行 + 数字 '1' + 小写 'a'
+    assert_eq!(indices, vec![0, 1, 2]);
+}
+
+#[test]
+fn id_restrict_filters_by_index() {
+    let tokens: Vec<Cow<'static, str>> = vec![
+        Cow::Borrowed(""),
+        Cow::Borrowed("1"),
+        Cow::Borrowed("2"),
+        Cow::Borrowed("3"),
+    ];
+
+    let indices = IdRestrict::TopN(2)
+        .apply_to_charset(&tokens)
+        .expect("字符集应存在交集");
+    assert_eq!(indices, vec![0, 1]);
+}
+
+#[test]
+fn hsv_range_validate() {
+    let range = HsvRange::new((0, 50, 50), (10, 255, 255));
+    assert!(range.validate().is_ok());
+    assert!(
+        HsvRange::new((200, 0, 0), (255, 255, 255))
+            .validate()
+            .is_err()
+    );
+}

ddddocr-ort/Cargo.toml

@@ -0,0 +1,25 @@
+[package]
+name = "ddddocr-ort"
+version = { workspace = true }
+edition = { workspace = true }
+license = { workspace = true }
+description = "ddddocr-rs 的 ONNX Runtime 推理引擎实现"
+keywords = ["ocr", "captcha", "ddddocr", "onnxruntime", "ort"]
+categories = ["multimedia::images", "computer-vision"]
+# repository = "https://github.com/<用户名>/<仓库名>"  # 发布前请补充
+readme = "../README.md"
+
+[dependencies]
+ddddocr-core = { path = "../ddddocr-core", version = "0.2.4" }
+ort = { workspace = true }
+ndarray = { workspace = true }
+image = { workspace = true }
+serde = { workspace = true }
+serde_json = { workspace = true }
+anyhow = { workspace = true }
+thiserror = { workspace = true }
+tracing = { workspace = true }
+
+[features]
+# 开启后启用底层 ort 库的 CUDA 执行提供者支持。
+cuda = ["ort/cuda"]

ddddocr-ort/src/det.rs

@@ -0,0 +1 @@
+pub mod session;

ddddocr-ort/src/det/session.rs

@@ -0,0 +1,41 @@
+use crate::runtime::{extract_tensor, lock_session, run_session};
+use crate::types::Session;
+use ddddocr_core::DetOutput;
+use ddddocr_core::error::{Result, TensorError};
+use ddddocr_core::traits::{DetEngine, InferenceEngine};
+use ndarray::Ix3;
+use tracing::debug;
+
+/// 目标检测推理运行时：持有 ORT 会话，输出 [`DetOutput`]。
+#[derive(Debug)]
+pub struct DetRuntime {
+    /// ORT 会话句柄。
+    pub session: Session,
+}
+
+impl DetRuntime {
+    /// 基于已构建的会话创建检测运行时。
+    pub fn new(session: Session) -> Self {
+        Self { session }
+    }
+}
+
+impl InferenceEngine for DetRuntime {
+    type Output = DetOutput;
+    fn inference(&self, input_array: ndarray::Array4<f32>) -> Result<Self::Output, TensorError> {
+        let mut session_guard = lock_session(&self.session)?;
+        let result = run_session(&mut session_guard, &input_array)?;
+        debug!("模型输出原始数据: {:?}", result);
+
+        let (view, shape_vec) = extract_tensor::<f32>(&result[0])?;
+        let array3 = view.to_owned().into_dimensionality::<Ix3>().map_err(|_| {
+            TensorError::DimensionMismatch {
+                expected: "3D 检测矩阵 [Batch, Box_Count, Box_Attributes]".to_string(),
+                actual: shape_vec,
+            }
+        })?;
+        Ok(DetOutput::Detection(array3))
+    }
+}
+
+impl DetEngine for DetRuntime {}

ddddocr-ort/src/lib.rs

@@ -0,0 +1,20 @@
+//! # ddddocr-ort
+//!
+//! 基于 ONNX Runtime（ORT）的推理引擎实现：通过 [`loader::ModelLoader`] 构建会话，
+//! 再由 [`OcrRuntime`] / [`DetRuntime`] 实现 core 的
+//! [`ddddocr_core::traits::InferenceEngine`] 接口。
+//! 同时重导出 core 的 [`OcrBuilder`]、[`Slider`]、[`SlideResult`] 等便捷 API。
+
+#![warn(missing_docs)]
+
+mod det;
+/// 模型加载器：从路径或字节流构建 ORT 会话。
+pub mod loader;
+mod ocr;
+mod runtime;
+mod types;
+
+pub use ddddocr_core::{OcrBuilder, SlideResult, Slider};
+pub use det::session::DetRuntime;
+pub use ocr::session::OcrRuntime;
+pub use types::Session;

ddddocr-ort/src/loader.rs

@@ -0,0 +1,7 @@
+mod error;
+mod metadata;
+mod model;
+
+pub use error::{Error, ParseError, Result};
+pub use metadata::{Metadata, ModelMetadataDto, NormalizationDto};
+pub use model::ModelLoader;

ddddocr-ort/src/loader/error.rs

@@ -0,0 +1,101 @@
+use ort::Error as OrtError;
+/// 模型加载与解析的通用结果类型。
+pub type Result<T> = std::result::Result<T, Error>;
+#[derive(thiserror::Error, Debug)]
+/// 模型加载、解析与 Session 构建阶段的错误。
+pub enum Error {
+    /// 底层构建器配置失败。
+    #[error("builder构建失败")]
+    Build(#[from] BuildError),
+    /// 解析 ONNX 模型/路径失败（如文件损坏、算子不支持、路径非法）
+    #[error("解析 ONNX 模型结构失败: {0}")]
+    ModelParse(#[from] ParseError),
+
+    /// 模型计算图优化失败（如常量折叠、形状推导失败）
+    #[error("优化 ORT 模型图失败: {0}")]
+    OptimizationFailed(#[source] OrtError),
+
+    /// 构建可执行 Session 失败（如输入输出 Tensor 类型/形状未确定）
+    #[error("构建可运行的 ORT Session 失败: {0}")]
+    RunnableBuildFailed(#[source] OrtError),
+
+    /// JSON 反序列化失败（自动透传 serde_json 报错）
+    #[error("模型 Metadata JSON 解析失败: {0}")]
+    JsonParse(#[from] serde_json::Error),
+
+    /// 字节流非合法 UTF-8 编码（自动透传 Utf8Error）
+    #[error("Metadata 字节流不是合法的 UTF-8 编码: {0}")]
+    InvalidUtf8(#[from] std::str::Utf8Error),
+
+    /// 模型元数据内容解析失败。
+    #[error("模型元数据解析失败: {0}")]
+    MetadataParse(String),
+
+    /// 承载任何第三方扩展、解密、特定预处理插件在执行时产生的自定义错误
+    #[error("{0}: {1}")]
+    Other(String, #[source] Box<dyn std::error::Error + Send + Sync>),
+}
+
+impl Error {
+    /// 方便将任何第三方 Error 包装为 Error::Other
+    pub fn new<E>(msg: impl Into<String>, err: E) -> Self
+    where
+        E: Into<Box<dyn std::error::Error + Send + Sync>>,
+    {
+        Self::Other(msg.into(), err.into())
+    }
+}
+#[derive(thiserror::Error, Debug)]
+/// 从路径或字节流解析 ONNX 模型失败的错误。
+pub enum ParseError {
+    /// 策略 A：从文件路径加载失败（附带路径上下文信息，方便排查是找不到文件还是格式不对）
+    #[error("从路径 '{0}' 加载 ONNX 模型失败: {1}")]
+    Path(String, #[source] OrtError),
+
+    /// 策略 B：从内存字节流加载失败（如 include_bytes! 传入的字节流损坏）
+    #[error("从内存字节流解析 ONNX 模型失败: {0}")]
+    Bytes(#[source] OrtError),
+}
+#[derive(thiserror::Error, Debug)]
+pub enum BuildError {
+    /// 底层 SessionBuilder 构建失败。
+    #[error("构建 ORT SessionBuilder 失败: {0}")]
+    BuildFailed(#[from] OrtError),
+    /// 线程数配置失败。
+    #[error("{0}")]
+    Threads(String),
+    /// 启用 CUDA 执行提供者失败。
+    #[error("{0}")]
+    EnabledCudaFailed(String),
+    /// 未编译 CUDA 支持时尝试启用 GPU。
+    #[error("{0}")]
+    NotEnabledCuda(String),
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn wraps_external_error_via_new() {
+        let io_err = std::io::Error::other("boom");
+        let err = Error::new("自定义错误", io_err);
+        assert!(matches!(err, Error::Other(_, _)));
+        assert!(err.to_string().contains("自定义错误"));
+    }
+
+    #[test]
+    fn serde_json_error_converts_to_error() {
+        let json_err = serde_json::from_str::<serde_json::Value>("{").unwrap_err();
+        let err: Error = json_err.into();
+        assert!(matches!(err, Error::JsonParse(_)));
+    }
+
+    #[test]
+    fn ut8_error_converts_to_error() {
+        let bytes = vec![0xffu8];
+        let utf8_err = std::str::from_utf8(&bytes).unwrap_err();
+        let err: Error = utf8_err.into();
+        assert!(matches!(err, Error::InvalidUtf8(_)));
+    }
+}

ddddocr-ort/src/loader/metadata.rs

@@ -0,0 +1,193 @@
+use crate::loader::error::{Error, Result};
+
+use ddddocr_core::{Charset, ModelMetadata, Normalization, Resize};
+use serde::Deserialize;
+use std::borrow::Cow;
+
+#[derive(Deserialize)]
+#[serde(rename_all = "snake_case")] // 支持 json 中写 "zero_to_one" 或 "minus_one_to_one"
+/// 归一化策略的 JSON 反序列化中间表示。
+pub enum NormalizationDto {
+    /// 映射到 [0.0, 1.0] -> pixel / 255.0
+    ZeroToOne,
+    /// 映射到 [-1.0, 1.0] -> (pixel / 255.0 - 0.5) / 0.5
+    MinusOneToOne,
+}
+
+impl From<NormalizationDto> for Normalization {
+    fn from(dto: NormalizationDto) -> Self {
+        match dto {
+            NormalizationDto::ZeroToOne => Normalization::ZeroToOne,
+            NormalizationDto::MinusOneToOne => Normalization::MinusOneToOne,
+        }
+    }
+}
+
+/// 仅用于反序列化 JSON 的中间临时结构体（DTO）
+#[derive(Deserialize)]
+pub struct ModelMetadataDto {
+    charset: Vec<String>,
+    word: bool,
+    #[serde(alias = "image")]
+    resize: Vec<i32>,
+    channel: u8,
+    /// 新增：允许在配置文件中指定归一化策略。
+    /// 使用 serde(default) 可以在不配置时提供一个默认值（比如默认 ZeroToOne）
+    #[serde(default = "default_normalization")]
+    normalization: NormalizationDto,
+}
+fn default_normalization() -> NormalizationDto {
+    NormalizationDto::ZeroToOne
+}
+
+/// 从 JSON 字符串或字节流解析模型元数据的扩展接口。
+pub trait Metadata: Sized {
+    /// 从 JSON 字符串解析模型元数据。
+    fn from_json_str(json_str: &str) -> Result<Self>;
+    /// 机制 2：从内存字节流加载（极大地方便 include_bytes! 或网络下载）
+    fn from_json_bytes(bytes: &[u8]) -> Result<Self> {
+        let json_str = std::str::from_utf8(bytes)?;
+        Self::from_json_str(json_str)
+    }
+}
+impl Metadata for ModelMetadata {
+    // --- 优雅的工厂模式构造器 ---
+    fn from_json_str(json_str: &str) -> Result<ModelMetadata> {
+        let dto: ModelMetadataDto = serde_json::from_str(json_str)?;
+
+        // 1. 将 DTO 的字符串数组转化为强类型的 Charset
+        let tokens: Vec<Cow<'static, str>> = dto.charset.into_iter().map(Cow::Owned).collect();
+        let charset = Charset::new(tokens);
+
+        // 2. 解析 resize 策略（重现 Python 的复杂条件判断）
+        if dto.resize.len() != 2 {
+            return Err(Error::MetadataParse(
+                "'resize (or image)' 字段必须是包含两个元素的数组，例如 [-1, 64]".to_string(),
+            ));
+        }
+        let r0 = dto.resize[0];
+        let r1 = dto.resize[1];
+
+        let resize = if r0 == -1 {
+            if dto.word {
+                // 如果 word 为 true，且包含 -1，Python 里是 resize 为 (r1, r1) 的正方形
+                Resize::Square(r1 as u32)
+            } else {
+                // 如果 word 为 false，且包含 -1，Python 里是高度固定为 r1，宽度按原图比例缩放
+                Resize::DynamicWidth(r1 as u32)
+            }
+        } else {
+            // 正常的固定宽高
+            Resize::Fixed(r0 as u32, r1 as u32)
+        };
+
+        Ok(ModelMetadata::new(
+            charset,
+            dto.word,
+            resize,
+            dto.channel,
+            dto.normalization.into(),
+        ))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    const MINIMAL_JSON: &str = r#"{
+        "charset": ["a", "b", "c"],
+        "word": false,
+        "resize": [-1, 64],
+        "channel": 1
+    }"#;
+
+    #[test]
+    fn parses_minimal_json_with_default_normalization() {
+        let meta = ModelMetadata::from_json_str(MINIMAL_JSON).unwrap();
+        assert_eq!(meta.charset.size(), 3);
+        assert!(!meta.word);
+        assert_eq!(meta.channel, 1);
+        assert!(matches!(meta.resize, Resize::DynamicWidth(64)));
+        assert!(matches!(meta.normalization, Normalization::ZeroToOne));
+    }
+
+    #[test]
+    fn parses_minus_one_to_one_normalization() {
+        let json = r#"{
+            "charset": ["a"],
+            "word": false,
+            "resize": [-1, 64],
+            "channel": 1,
+            "normalization": "minus_one_to_one"
+        }"#;
+        let meta = ModelMetadata::from_json_str(json).unwrap();
+        assert!(matches!(meta.normalization, Normalization::MinusOneToOne));
+    }
+
+    #[test]
+    fn parses_word_model_as_square_resize() {
+        let json = r#"{
+            "charset": ["a"],
+            "word": true,
+            "resize": [-1, 64],
+            "channel": 1
+        }"#;
+        let meta = ModelMetadata::from_json_str(json).unwrap();
+        assert!(matches!(meta.resize, Resize::Square(64)));
+    }
+
+    #[test]
+    fn parses_fixed_resize() {
+        let json = r#"{
+            "charset": ["a"],
+            "word": false,
+            "resize": [100, 64],
+            "channel": 1
+        }"#;
+        let meta = ModelMetadata::from_json_str(json).unwrap();
+        assert!(matches!(meta.resize, Resize::Fixed(100, 64)));
+    }
+
+    #[test]
+    fn parses_image_alias() {
+        let json = r#"{
+            "charset": ["a"],
+            "word": false,
+            "image": [200, 64],
+            "channel": 1
+        }"#;
+        let meta = ModelMetadata::from_json_str(json).unwrap();
+        assert!(matches!(meta.resize, Resize::Fixed(200, 64)));
+    }
+
+    #[test]
+    fn rejects_resize_with_wrong_length() {
+        let json = r#"{
+            "charset": ["a"],
+            "word": false,
+            "resize": [1, 2, 3],
+            "channel": 1
+        }"#;
+        assert!(matches!(
+            ModelMetadata::from_json_str(json),
+            Err(Error::MetadataParse(_))
+        ));
+    }
+
+    #[test]
+    fn rejects_invalid_json() {
+        assert!(matches!(
+            ModelMetadata::from_json_str("not json"),
+            Err(Error::JsonParse(_))
+        ));
+    }
+
+    #[test]
+    fn rejects_invalid_utf8_bytes() {
+        assert!(matches!(
+            ModelMetadata::from_json_bytes(&[0xff, 0xfe]),
+            Err(Error::InvalidUtf8(_))
+        ));
+    }
+}

ddddocr-ort/src/loader/model.rs

@@ -0,0 +1,131 @@
+use crate::loader::Error;
+use crate::loader::error::{BuildError, ParseError, Result};
+use crate::types::Session;
+use ddddocr_core::traits::Loader;
+use ort::session::Session as OrtSession;
+use ort::session::builder::SessionBuilder;
+use std::sync::{Arc, Mutex};
+
+/// ORT 专用的链式构建器
+#[derive(Debug, Clone, Default)]
+pub struct ModelLoader {
+    use_gpu: bool,
+    device_id: i32,
+    intra_threads: Option<usize>,
+}
+impl ModelLoader {
+    /// 开启或关闭 GPU 加速
+    pub fn use_gpu(mut self, enable: bool) -> Self {
+        self.use_gpu = enable;
+        self
+    }
+
+    /// 指定 GPU 设备 ID
+    pub fn device_id(mut self, id: i32) -> Self {
+        self.device_id = id;
+        self
+    }
+    /// 指定 ORT 内部线程数。
+    pub fn num_threads(mut self, threads: usize) -> Self {
+        self.intra_threads = Some(threads);
+        self
+    }
+
+    /// 内部辅助方法：根据当前的配置构建 ORT 底层的 SessionBuilder
+    fn create_session_builder(&self) -> Result<SessionBuilder> {
+        let mut builder = OrtSession::builder().map_err(BuildError::BuildFailed)?;
+        // 如果用户显式设置了线程数，则配置给 ORT
+        if let Some(threads) = self.intra_threads {
+            builder = builder
+                .with_intra_threads(threads)
+                .map_err(|e| BuildError::Threads(format!("设置线程数失败: {e}")))?;
+        }
+
+        if self.use_gpu {
+            // 根据 ort 库版本配置 CUDA 执行提供者 (Execution Provider)
+            #[cfg(feature = "cuda")]
+            {
+                use ort::ep::CUDAExecutionProvider;
+                let cuda_ep = CUDAExecutionProvider::default().with_device_id(self.device_id);
+                builder = builder
+                    .with_execution_providers([cuda_ep.build()])
+                    .map_err(|e| {
+                        BuildError::EnabledCudaFailed(format!("配置 CUDA 硬件加速失败: {e}"))
+                    })?
+            }
+            #[cfg(not(feature = "cuda"))]
+            {
+                // 如果用户明确开启了 GPU，但 Feature 没编译进去，明确抛错提醒
+                return Err(BuildError::NotEnabledCuda(
+                    "未启用 CUDA 支持：请在 Cargo.toml 中为 ddddocr-ort 开启 `cuda` feature"
+                        .to_string(),
+                ))?;
+            }
+        }
+
+        Ok(builder)
+    }
+}
+
+impl Loader for ModelLoader {
+    type Session = Session;
+    type Error = Error;
+
+    fn build_for_path<P>(&self, model_path: P) -> Result<Session>
+    where
+        P: AsRef<std::path::Path>,
+    {
+        let path_ref = model_path.as_ref();
+        let mut builder = self.create_session_builder()?;
+        // Session::builder() 会返回 Result<SessionBuilder, ort::Error>
+
+        let session = builder
+            .commit_from_file(path_ref)
+            .map_err(|e| ParseError::Path(path_ref.display().to_string(), e))?;
+
+        Ok(Arc::new(Mutex::new(session)))
+    }
+    /// 策略 B：从内存字节流加载模型（配合 include_bytes! 使用）
+    fn build_from_bytes(&self, model_bytes: &[u8]) -> Result<Session> {
+        let mut builder = self.create_session_builder()?;
+
+        let session = builder
+            .commit_from_memory(model_bytes)
+            .map_err(ParseError::Bytes)?;
+
+        Ok(Arc::new(Mutex::new(session)))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn default_config() {
+        let loader = ModelLoader::default();
+        assert!(!loader.use_gpu);
+        assert_eq!(loader.device_id, 0);
+        assert!(loader.intra_threads.is_none());
+    }
+
+    #[test]
+    fn builder_configures_fields() {
+        let loader = ModelLoader::default()
+            .use_gpu(true)
+            .device_id(2)
+            .num_threads(4);
+        assert!(loader.use_gpu);
+        assert_eq!(loader.device_id, 2);
+        assert_eq!(loader.intra_threads, Some(4));
+    }
+
+    #[test]
+    fn builder_is_chainable_and_immutable() {
+        let base = ModelLoader::default();
+        let _configured = base.clone().use_gpu(true).device_id(1).num_threads(8);
+        assert!(!base.use_gpu);
+        assert_eq!(base.device_id, 0);
+        assert!(base.intra_threads.is_none());
+    }
+}

ddddocr-ort/src/ocr.rs

@@ -0,0 +1 @@
+pub mod session;

ddddocr-ort/src/ocr/session.rs

@@ -0,0 +1,188 @@
+use crate::runtime::{extract_tensor, lock_session, run_session};
+use crate::types::Session;
+use ddddocr_core::ModelMetadata;
+use ddddocr_core::OcrOutput;
+use ddddocr_core::error::{Result, TensorError};
+use ddddocr_core::traits::{InferenceEngine, Info, OcrEngine};
+use ddddocr_core::types::{AxisDim, ModelInfo, TensorInfo, TensorType};
+use ddddocr_core::utils::normalize_ocr_logits;
+use ort::value::{Outlet, TensorElementType, ValueType};
+use tracing::debug;
+
+/// OCR 推理运行时：持有 ORT 会话与模型元数据，输出 [`OcrOutput`]。
+pub struct OcrRuntime {
+    /// ORT 会话句柄。
+    pub session: Session,
+    /// 模型元数据（字符集、归一化策略等）。
+    pub metadata: ModelMetadata,
+}
+
+impl OcrRuntime {
+    /// 基于已构建的会话与元数据创建 OCR 运行时。
+    pub fn new(session: Session, metadata: ModelMetadata) -> Self {
+        Self { session, metadata }
+    }
+
+    /// 将 ORT 输入/输出出口解析为 core 的 [`TensorInfo`] 列表。
+    fn resolve_outlets(&self, outlets: &[Outlet]) -> Result<Vec<TensorInfo>> {
+        Ok(outlets
+            .iter()
+            .map(|outlet| TensorInfo {
+                name: outlet.name().to_string(),
+                shape: resolve_shape(outlet.dtype()),
+                tensor_type: tensor_type_from_ort(outlet.dtype()),
+            })
+            .collect())
+    }
+}
+
+/// 将 ORT 值类型映射为 core 的 [`TensorType`]。
+fn tensor_type_from_ort(dtype: &ValueType) -> TensorType {
+    match dtype.tensor_type() {
+        Some(TensorElementType::Float32) => TensorType::F32,
+        Some(TensorElementType::Int64) => TensorType::I64,
+        _ => TensorType::Other,
+    }
+}
+
+/// 将 ORT 张量形状解析为 [`AxisDim`] 列表，动态维度（`-1`）标记为符号维度。
+fn resolve_shape(dtype: &ValueType) -> Vec<AxisDim> {
+    dtype
+        .tensor_shape()
+        .map(|shape| {
+            shape
+                .iter()
+                .map(|&dim| {
+                    if dim >= 0 {
+                        AxisDim::Static(dim as usize)
+                    } else {
+                        AxisDim::Dynamic("dynamic".to_string())
+                    }
+                })
+                .collect()
+        })
+        .unwrap_or_default()
+}
+
+impl OcrEngine for OcrRuntime {
+    fn metadata(&self) -> &ModelMetadata {
+        &self.metadata
+    }
+}
+
+impl InferenceEngine for OcrRuntime {
+    type Output = OcrOutput;
+
+    fn inference(&self, input_array: ndarray::Array4<f32>) -> Result<Self::Output, TensorError> {
+        let mut session_guard = lock_session(&self.session)?;
+        let result = run_session(&mut session_guard, &input_array)?;
+        debug!("模型输出原始数据: {:?}", result);
+        let raw_value = &result[0];
+
+        match raw_value.dtype().tensor_type() {
+            Some(TensorElementType::Int64) => {
+                let (view, actual_shape) = extract_tensor::<i64>(raw_value)?;
+                let array1 = view
+                    .to_owned()
+                    .into_dimensionality::<ndarray::Ix1>()
+                    .map_err(|_| TensorError::DimensionMismatch {
+                        expected: "1D 字符索引静态矩阵".to_string(),
+                        actual: actual_shape,
+                    })?;
+                Ok(OcrOutput::Indices(array1))
+            }
+            Some(TensorElementType::Float32) => {
+                let shape = raw_value.shape();
+                debug!("模型输出 shape 数据: {:?}", shape);
+                let (view, shape_vec) = extract_tensor::<f32>(raw_value)?;
+                normalize_ocr_logits(view, shape_vec.as_slice())
+            }
+            _ => Err(TensorError::UnknownOutputFormat),
+        }
+    }
+}
+
+impl Info for OcrRuntime {
+    fn input_info(&self) -> Result<Vec<TensorInfo>> {
+        let session_guard = lock_session(&self.session)?;
+        self.resolve_outlets(session_guard.inputs())
+    }
+
+    fn output_info(&self) -> Result<Vec<TensorInfo>> {
+        let session_guard = lock_session(&self.session)?;
+        self.resolve_outlets(session_guard.outputs())
+    }
+
+    fn model_info(&self) -> Result<ModelInfo> {
+        Ok(ModelInfo {
+            inputs: self.input_info()?,
+            outputs: self.output_info()?,
+            providers: None,
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use ort::value::{Shape, SymbolicDimensions, TensorElementType};
+
+    fn tensor_dtype(ty: TensorElementType) -> ValueType {
+        ValueType::Tensor {
+            ty,
+            shape: Shape::new([1, 64]),
+            dimension_symbols: SymbolicDimensions::new([String::new(), String::new()]),
+        }
+    }
+
+    #[test]
+    fn maps_tensor_element_types() {
+        assert!(matches!(
+            tensor_type_from_ort(&tensor_dtype(TensorElementType::Float32)),
+            TensorType::F32
+        ));
+        assert!(matches!(
+            tensor_type_from_ort(&tensor_dtype(TensorElementType::Int64)),
+            TensorType::I64
+        ));
+        assert!(matches!(
+            tensor_type_from_ort(&tensor_dtype(TensorElementType::Uint8)),
+            TensorType::Other
+        ));
+    }
+
+    #[test]
+    fn resolves_static_shape() {
+        let shape = resolve_shape(&tensor_dtype(TensorElementType::Float32));
+        assert_eq!(shape, vec![AxisDim::Static(1), AxisDim::Static(64)]);
+    }
+
+    #[test]
+    fn resolves_dynamic_dimension_as_symbol() {
+        let dtype = ValueType::Tensor {
+            ty: TensorElementType::Float32,
+            shape: Shape::new([1, 64, -1]),
+            dimension_symbols: SymbolicDimensions::new([
+                String::new(),
+                String::new(),
+                String::from("width"),
+            ]),
+        };
+        let shape = resolve_shape(&dtype);
+        assert_eq!(
+            shape,
+            vec![
+                AxisDim::Static(1),
+                AxisDim::Static(64),
+                AxisDim::Dynamic("dynamic".to_string())
+            ]
+        );
+    }
+
+    #[test]
+    fn non_tensor_dtype_yields_empty_shape() {
+        let dtype = ValueType::Sequence(Box::new(tensor_dtype(TensorElementType::Float32)));
+        assert!(resolve_shape(&dtype).is_empty());
+        assert!(matches!(tensor_type_from_ort(&dtype), TensorType::Other));
+    }
+}

ddddocr-ort/src/runtime.rs

@@ -0,0 +1,41 @@
+//! 会话执行相关的共享工具函数。
+
+use crate::types::Session;
+use ddddocr_core::error::TensorError;
+use ort::inputs;
+use ort::session::Session as OrtSession;
+use ort::session::SessionOutputs;
+use ort::value::{PrimitiveTensorElementType, TensorRef, Value};
+use std::sync::MutexGuard;
+
+/// 获取会话互斥锁，失败时转换为引擎错误。
+pub(crate) fn lock_session(session: &Session) -> Result<MutexGuard<'_, OrtSession>, TensorError> {
+    session
+        .lock()
+        .map_err(|_| TensorError::Engine("获取 Session 锁失败 (Poisoned)".to_string()))
+}
+
+/// 对输入张量执行一次 ORT 推理，返回原始输出值列表。
+pub(crate) fn run_session<'a>(
+    session: &'a mut OrtSession,
+    input_array: &ndarray::Array4<f32>,
+) -> Result<SessionOutputs<'a>, TensorError> {
+    session
+        .run(inputs![TensorRef::from_array_view(input_array).map_err(
+            |e| { TensorError::Engine(format!("构建输入失败: {e}")) }
+        )?])
+        .map_err(|e| TensorError::Engine(format!("执行模型推理失败: {e}")))
+}
+
+/// 从输出值中提取类型化张量视图，同时返回其真实维度。
+pub(crate) fn extract_tensor<T: PrimitiveTensorElementType>(
+    value: &Value,
+) -> Result<(ndarray::ArrayViewD<'_, T>, Vec<usize>), TensorError> {
+    let (shape_ref, slice) = value
+        .try_extract_tensor::<T>()
+        .map_err(|_| TensorError::Engine("无法获取张量内存视图".to_string()))?;
+    let shape: Vec<usize> = shape_ref.iter().map(|v| *v as usize).collect();
+    let view = ndarray::ArrayViewD::from_shape(shape.as_slice(), slice)
+        .map_err(|_| TensorError::Engine("构建 ndarray ArrayViewD 失败".to_string()))?;
+    Ok((view, shape))
+}

ddddocr-ort/src/types.rs

@@ -0,0 +1,7 @@
+//! ORT 会话共享类型。
+
+use ort::session::Session as OrtSession;
+use std::sync::{Arc, Mutex};
+
+/// ORT 会话句柄：由 [`crate::loader::ModelLoader`] 构建，推理时通过互斥锁串行访问。
+pub type Session = Arc<Mutex<OrtSession>>;

ddddocr-ort/tests/api_surface.rs

@@ -0,0 +1,76 @@
+//! 外部视角 API 测试：验证 `ddddocr-ort` 的公开类型与方法在外部 crate 中可正常使用。
+
+use ddddocr_core::traits::{Info, Loader};
+use ddddocr_core::types::AxisDim;
+use ddddocr_core::{ModelMetadata, Normalization, OcrBuilder, Resize, SlideResult, Slider};
+use ddddocr_ort::loader::ModelLoader;
+use ddddocr_ort::{DetRuntime, OcrRuntime, Session};
+use std::path::{Path, PathBuf};
+
+fn model_path(name: &str) -> PathBuf {
+    Path::new(env!("CARGO_MANIFEST_DIR"))
+        .join("..")
+        .join("models")
+        .join(name)
+}
+
+/// 验证对外导出的类型（含 `Session` 与 core 便捷重导出）均可直接命名。
+#[test]
+fn public_types_are_nameable() {
+    let _: Option<Session> = None;
+    let _: Option<OcrBuilder> = None;
+    let _: Option<SlideResult> = None;
+    let _slider = Slider::new();
+}
+
+/// 验证构建器链式 API 可组合使用，且返回的会话类型可显式标注。
+#[test]
+fn loader_chain_builds_session() -> anyhow::Result<()> {
+    let path = model_path("common_sml2h3_f32.onnx");
+    assert!(path.exists(), "缺少测试模型: {}", path.display());
+    let _session: Session = ModelLoader::default()
+        .use_gpu(false)
+        .device_id(0)
+        .num_threads(4)
+        .build_for_path(path)?;
+    Ok(())
+}
+
+/// 验证 `Info` trait 能从真实会话中解析输入/输出信息。
+#[test]
+fn info_trait_returns_model_metadata() -> anyhow::Result<()> {
+    let path = model_path("common_sml2h3_f32.onnx");
+    assert!(path.exists(), "缺少测试模型: {}", path.display());
+    let session: Session = ModelLoader::default().build_for_path(path)?;
+    let metadata = ModelMetadata::from_static_slice(
+        &["a", "b"],
+        false,
+        Resize::DynamicWidth(64),
+        1,
+        Normalization::MinusOneToOne,
+    );
+    let ocr = OcrRuntime::new(session, metadata);
+
+    let inputs = ocr.input_info()?;
+    let outputs = ocr.output_info()?;
+    assert!(!inputs.is_empty());
+    assert!(!outputs.is_empty());
+    assert!(matches!(inputs[0].shape[0], AxisDim::Static(1)));
+    assert!(matches!(inputs[0].shape[2], AxisDim::Static(64)));
+
+    let model_info = ocr.model_info()?;
+    assert_eq!(model_info.inputs.len(), inputs.len());
+    assert_eq!(model_info.outputs.len(), outputs.len());
+    assert!(model_info.providers.is_none());
+    Ok(())
+}
+
+/// 验证检测运行时可以从构建的会话创建。
+#[test]
+fn det_runtime_builds_from_session() -> anyhow::Result<()> {
+    let path = model_path("common_det.onnx");
+    assert!(path.exists(), "缺少测试模型: {}", path.display());
+    let session: Session = ModelLoader::default().build_for_path(path)?;
+    let _det = DetRuntime::new(session);
+    Ok(())
+}

ddddocr-ort/tests/common/mod.rs

@@ -0,0 +1,25 @@
+//! 集成测试共享工具。
+
+use std::path::{Path, PathBuf};
+
+/// 仓库根目录 `models/` 下模型文件的路径。
+pub fn model_path(name: &str) -> PathBuf {
+    Path::new(env!("CARGO_MANIFEST_DIR"))
+        .join("..")
+        .join("models")
+        .join(name)
+}
+
+/// 仓库根目录 `samples/` 下样例图片的路径。
+pub fn sample_path(name: &str) -> PathBuf {
+    Path::new(env!("CARGO_MANIFEST_DIR"))
+        .join("..")
+        .join("samples")
+        .join(name)
+}
+
+/// 加载图片，失败时附带路径上下文。
+pub fn load_image<P: AsRef<Path>>(path: P) -> anyhow::Result<image::DynamicImage> {
+    let path_ref = path.as_ref();
+    image::open(path_ref).map_err(|e| anyhow::anyhow!("无法加载图片 {:?}: {}", path_ref, e))
+}

ddddocr-ort/tests/det.rs

@@ -0,0 +1,33 @@
+//! 目标检测集成测试。
+
+mod common;
+
+use common::{model_path, sample_path};
+use ddddocr_core::Detector;
+use ddddocr_core::traits::Loader;
+use ddddocr_ort::DetRuntime;
+use ddddocr_ort::loader::ModelLoader;
+use image::GenericImageView;
+
+/// 检测模型应能从样例图片中找到至少一个目标，且坐标在图片范围内。
+#[test]
+fn det_model_detects_targets_in_image() -> anyhow::Result<()> {
+    let session = ModelLoader::default()
+        .build_for_path(model_path("common_det.onnx"))
+        .expect("模型加载失败");
+    let det = DetRuntime::new(session);
+    let img = image::open(sample_path("det1.png")).expect("测试图片不存在");
+
+    let bboxes = Detector::new(&det).predict(&img)?;
+    assert!(!bboxes.is_empty(), "应检测到至少一个目标");
+
+    let (width, height) = img.dimensions();
+    for bbox in &bboxes {
+        assert!(bbox.x1 >= 0 && bbox.y1 >= 0, "检测框左上角不应为负");
+        assert!(
+            bbox.x2 <= width as i32 && bbox.y2 <= height as i32,
+            "检测框右下角不应超出图片范围"
+        );
+    }
+    Ok(())
+}

ddddocr-ort/tests/ocr.rs

@@ -0,0 +1,56 @@
+//! OCR 识别与模型信息集成测试。
+
+mod common;
+
+use common::{model_path, sample_path};
+use ddddocr_core::traits::{Info, Loader};
+use ddddocr_core::{ModelMetadata, Normalization, Ocr, Resize};
+use ddddocr_ort::OcrRuntime;
+use ddddocr_ort::loader::ModelLoader;
+
+/// 用官方 sml2h3 f32 模型识别验证码图片，结果不应为空。
+#[test]
+fn ocr_classification_recognizes_code_image() {
+    let session = ModelLoader::default()
+        .use_gpu(false)
+        .build_for_path(model_path("common_sml2h3_f32.onnx"))
+        .expect("模型加载失败");
+    let metadata = ModelMetadata::from_builtin_beta(
+        false,
+        Resize::DynamicWidth(64),
+        1,
+        Normalization::MinusOneToOne,
+    );
+    let ocr = OcrRuntime::new(session, metadata);
+
+    let img = image::open(sample_path("code2.png")).expect("测试图片不存在");
+    let text = Ocr::builder()
+        .build_with(&ocr)
+        .predict(&img)
+        .expect("识别过程出错")
+        .into_text();
+
+    println!("识别结果: {text}");
+    assert!(!text.is_empty(), "识别结果不应为空");
+}
+
+/// 真实模型应能通过 `Info` trait 返回输入/输出张量信息。
+#[test]
+fn model_info_lists_inputs_and_outputs() -> anyhow::Result<()> {
+    let session = ModelLoader::default()
+        .build_for_path(model_path("common_huashi666_i64.onnx"))
+        .expect("建立测试模型图失败");
+    let metadata = ModelMetadata::from_builtin_beta(
+        false,
+        Resize::DynamicWidth(64),
+        1,
+        Normalization::MinusOneToOne,
+    );
+    let ocr = OcrRuntime::new(session, metadata);
+
+    let inputs = ocr.input_info()?;
+    let outputs = ocr.output_info()?;
+    assert!(!inputs.is_empty(), "模型应有输入张量信息");
+    assert!(!outputs.is_empty(), "模型应有输出张量信息");
+    Ok(())
+}

ddddocr-ort/tests/slide.rs

@@ -0,0 +1,52 @@
+//! 滑块匹配集成测试。
+
+mod common;
+
+use common::{load_image, sample_path};
+use ddddocr_core::Slider;
+
+/// 边缘模式匹配应定位到预期坐标。
+#[test]
+fn slide_match_locates_target_position() {
+    let engine = Slider::new();
+    let target = load_image(sample_path("target1.png")).expect("请确保 samples/target1.png 存在");
+    let background = load_image(sample_path("background1.png")).expect("请确保 samples/background1.png 存在");
+
+    let start = std::time::Instant::now();
+    let result = engine
+        .slide_match(&target, &background, false)
+        .expect("Slide match 执行失败");
+    let elapsed = start.elapsed();
+
+    println!("边缘模式匹配: {result}");
+    println!("耗时: {elapsed:?}");
+
+    assert_eq!(result.target_x, 237);
+    assert_eq!(result.target_y, 77);
+    assert!(result.confidence > 0.0);
+}
+
+/// 灰度对比匹配应定位到预期坐标。
+#[test]
+fn slide_comparison_locates_target_position() {
+    let engine = Slider::new();
+    let target = load_image(sample_path("target2.jpg")).expect("请确保 samples/target2.jpg 存在");
+    let background =
+        load_image(sample_path("background2.jpg")).expect("请确保 samples/background2.jpg 存在");
+
+    let start = std::time::Instant::now();
+    let result = engine
+        .slide_comparison(&target, &background)
+        .expect("Slide comparison 执行失败");
+    let elapsed = start.elapsed();
+
+    println!(
+        "灰度对比匹配: 坐标 [x: {}, y: {}], 置信度 {:.4}",
+        result.target_x, result.target_y, result.confidence
+    );
+    println!("耗时: {elapsed:?}");
+
+    assert_eq!(result.target_x, 171);
+    assert_eq!(result.target_y, 90);
+    assert!(result.confidence > 0.0);
+}

ddddocr-tract2/Cargo.toml

@@ -0,0 +1,22 @@
+[package]
+name = "ddddocr-tract2"
+version = { workspace = true }
+edition = { workspace = true }
+license = { workspace = true }
+description = "ddddocr-rs 的 Tract 推理引擎实现"
+keywords = ["ocr", "captcha", "ddddocr", "onnxruntime", "tract"]
+categories = ["multimedia::images", "computer-vision"]
+# repository = "https://github.com/<用户名>/<仓库名>"  # 发布前请补充
+readme = "../README.md"
+
+[dependencies]
+ddddocr-core = { path = "../ddddocr-core", version = "0.2.4" }
+tract-onnx = { workspace = true }
+tract-linalg = { workspace = true }
+ndarray = { workspace = true }
+image = { workspace = true }
+serde = { workspace = true }
+serde_json = { workspace = true }
+anyhow = { workspace = true }
+thiserror = { workspace = true }
+tracing = { workspace = true }

ddddocr-tract2/src/det.rs

@@ -0,0 +1 @@
+pub mod session;

ddddocr-tract2/src/det/session.rs

@@ -0,0 +1,44 @@
+use crate::runtime::{extract_plain_array, run_session};
+use crate::types::Session;
+use ddddocr_core::DetOutput;
+use ddddocr_core::error::{Result, TensorError};
+use ddddocr_core::traits::{DetEngine, InferenceEngine};
+use ndarray::Ix3;
+use tracing::debug;
+use tract_onnx::prelude::*;
+
+#[derive(Debug)]
+/// 目标检测推理运行时：持有 Tract 会话，输出 [`DetOutput`]。
+pub struct DetRuntime {
+    /// Tract 会话句柄。
+    pub session: Session,
+}
+
+impl DetRuntime {
+    /// 基于已构建的会话创建检测运行时。
+    pub fn new(session: Session) -> Self {
+        Self { session }
+    }
+}
+
+impl InferenceEngine for DetRuntime {
+    type Output = DetOutput;
+    fn inference(&self, input_array: ndarray::Array4<f32>) -> Result<Self::Output, TensorError> {
+        let mut result = run_session(&self.session, input_array)?;
+        debug!("模型输出原始数据: {:?}", result);
+        let raw_tensor = result.swap_remove(0).into_tensor();
+        let array_d = extract_plain_array::<f32>(raw_tensor)?;
+        let actual_shape = array_d.shape().to_vec();
+
+        let array3 =
+            array_d
+                .into_dimensionality::<Ix3>()
+                .map_err(|_| TensorError::DimensionMismatch {
+                    expected: "3D 检测矩阵 [Batch, Box_Count, Box_Attributes]".to_string(),
+                    actual: actual_shape,
+                })?;
+        Ok(DetOutput::Detection(array3))
+    }
+}
+
+impl DetEngine for DetRuntime {}

ddddocr-tract2/src/lib.rs

@@ -0,0 +1,22 @@
+//! # ddddocr-tract2
+//!
+//! 基于 Tract（纯 Rust ONNX 推理框架）的推理引擎实现：通过 [`loader::ModelLoader`] 构建会话，
+//! 再由 [`OcrRuntime`] / [`DetRuntime`] 实现 core 的
+//! [`ddddocr_core::traits::InferenceEngine`] 接口。
+//! 同时重导出 core 的 [`OcrBuilder`]、[`Slider`]、[`SlideResult`] 等便捷 API。
+
+#![warn(missing_docs)]
+
+mod det;
+/// 模型加载器：从路径或字节流构建 Tract 会话。
+pub mod loader;
+mod ocr;
+mod runtime;
+mod types;
+
+pub use ddddocr_core::{
+    DetectionResult, Detector, ModelMetadata, Normalization, Ocr, OcrBuilder, SlideResult, Slider,
+};
+pub use det::session::DetRuntime;
+pub use ocr::session::OcrRuntime;
+pub use types::Session;

ddddocr-tract2/src/loader.rs

@@ -0,0 +1,7 @@
+mod error;
+mod metadata;
+mod model;
+
+pub use error::{Error, ParseError, Result};
+pub use metadata::{Metadata, ModelMetadataDto, NormalizationDto};
+pub use model::ModelLoader;

ddddocr-tract2/src/loader/error.rs

@@ -0,0 +1,82 @@
+use tract_onnx::prelude::TractError;
+/// 模型加载与解析的通用结果类型。
+pub type Result<T> = std::result::Result<T, Error>;
+#[derive(thiserror::Error, Debug)]
+/// 模型加载、解析与 Session 构建阶段的错误。
+pub enum Error {
+    /// 解析 ONNX 模型/路径失败（如文件损坏、算子不支持、路径非法）
+    #[error("解析 ONNX 模型结构失败: {0}")]
+    ModelParse(#[from] ParseError),
+
+    /// 模型计算图优化失败（如常量折叠、形状推导失败）
+    #[error("优化 Tract 模型图失败: {0}")]
+    OptimizationFailed(#[source] TractError),
+
+    /// 构建可执行 Session 失败（如输入输出 Tensor 类型/形状未确定）
+    #[error("构建可运行 Tract 实例失败: {0}")]
+    RunnableBuildFailed(#[source] TractError),
+
+    /// JSON 反序列化失败（自动透传 serde_json 报错）
+    #[error("模型 Metadata JSON 解析失败: {0}")]
+    JsonParse(#[from] serde_json::Error),
+
+    /// 字节流非合法 UTF-8 编码（自动透传 Utf8Error）
+    #[error("Metadata 字节流不是合法的 UTF-8 编码: {0}")]
+    InvalidUtf8(#[from] std::str::Utf8Error),
+
+    /// 模型元数据内容解析失败。
+    #[error("模型元数据解析失败: {0}")]
+    MetadataParse(String),
+
+    /// 承载任何第三方扩展、解密、特定预处理插件在执行时产生的自定义错误
+    #[error("{0}: {1}")]
+    Other(String, #[source] Box<dyn std::error::Error + Send + Sync>),
+}
+impl Error {
+    /// 方便将任何第三方 Error 包装为 Error::Other
+    pub fn new<E>(msg: impl Into<String>, err: E) -> Self
+    where
+        E: Into<Box<dyn std::error::Error + Send + Sync>>,
+    {
+        Self::Other(msg.into(), err.into())
+    }
+}
+#[derive(thiserror::Error, Debug)]
+/// 从路径或字节流解析 ONNX 模型失败的错误。
+pub enum ParseError {
+    /// 策略 A：从文件路径加载失败（附带路径上下文信息，方便排查是找不到文件还是格式不对）
+    #[error("从路径 '{0}' 加载 ONNX 模型失败: {1}")]
+    Path(String, #[source] TractError),
+
+    /// 策略 B：从内存字节流加载失败（如 include_bytes! 传入的字节流损坏）
+    #[error("从内存字节流解析 ONNX 模型失败: {0}")]
+    Bytes(#[source] TractError),
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn wraps_external_error_via_new() {
+        let io_err = std::io::Error::other("boom");
+        let err = Error::new("自定义错误", io_err);
+        assert!(matches!(err, Error::Other(_, _)));
+        assert!(err.to_string().contains("自定义错误"));
+    }
+
+    #[test]
+    fn serde_json_error_converts_to_error() {
+        let json_err = serde_json::from_str::<serde_json::Value>("{").unwrap_err();
+        let err: Error = json_err.into();
+        assert!(matches!(err, Error::JsonParse(_)));
+    }
+
+    #[test]
+    fn invalid_utf8_converts_to_error() {
+        let bytes = vec![0xffu8];
+        let utf8_err = std::str::from_utf8(&bytes).unwrap_err();
+        let err: Error = utf8_err.into();
+        assert!(matches!(err, Error::InvalidUtf8(_)));
+    }
+}

ddddocr-tract2/src/loader/metadata.rs

@@ -0,0 +1,195 @@
+use crate::loader::error::{Error, Result};
+
+use ddddocr_core::ModelMetadata;
+use ddddocr_core::Resize;
+use ddddocr_core::{Charset, Normalization};
+use serde::Deserialize;
+use std::borrow::Cow;
+
+#[derive(Deserialize)]
+#[serde(rename_all = "snake_case")] // 支持 json 中写 "zero_to_one" 或 "minus_one_to_one"
+/// 归一化策略的 JSON 反序列化中间表示。
+pub enum NormalizationDto {
+    /// 映射到 [0.0, 1.0] -> pixel / 255.0
+    ZeroToOne,
+    /// 映射到 [-1.0, 1.0] -> (pixel / 255.0 - 0.5) / 0.5
+    MinusOneToOne,
+}
+
+impl From<NormalizationDto> for Normalization {
+    fn from(dto: NormalizationDto) -> Self {
+        match dto {
+            NormalizationDto::ZeroToOne => Normalization::ZeroToOne,
+            NormalizationDto::MinusOneToOne => Normalization::MinusOneToOne,
+        }
+    }
+}
+
+/// 仅用于反序列化 JSON 的中间临时结构体（DTO）
+#[derive(Deserialize)]
+pub struct ModelMetadataDto {
+    charset: Vec<String>,
+    word: bool,
+    #[serde(alias = "image")]
+    resize: Vec<i32>,
+    channel: u8,
+    /// 新增：允许在配置文件中指定归一化策略。
+    /// 使用 serde(default) 可以在不配置时提供一个默认值（比如默认 ZeroToOne）
+    #[serde(default = "default_normalization")]
+    normalization: NormalizationDto,
+}
+fn default_normalization() -> NormalizationDto {
+    NormalizationDto::ZeroToOne
+}
+
+/// 从 JSON 字符串或字节流解析模型元数据的扩展接口。
+pub trait Metadata: Sized {
+    /// 从 JSON 字符串解析模型元数据。
+    fn from_json_str(json_str: &str) -> Result<Self>;
+    /// 机制 2：从内存字节流加载（极大地方便 include_bytes! 或网络下载）
+    fn from_json_bytes(bytes: &[u8]) -> Result<Self> {
+        let json_str = std::str::from_utf8(bytes)?;
+        Self::from_json_str(json_str)
+    }
+}
+impl Metadata for ModelMetadata {
+    // --- 优雅的工厂模式构造器 ---
+    fn from_json_str(json_str: &str) -> Result<ModelMetadata> {
+        let dto: ModelMetadataDto = serde_json::from_str(json_str)?;
+
+        // 1. 将 DTO 的字符串数组转化为强类型的 Charset
+        let tokens: Vec<Cow<'static, str>> = dto.charset.into_iter().map(Cow::Owned).collect();
+        let charset = Charset::new(tokens);
+
+        // 2. 解析 resize 策略（重现 Python 的复杂条件判断）
+        if dto.resize.len() != 2 {
+            return Err(Error::MetadataParse(
+                "'resize (or image)' 字段必须是包含两个元素的数组，例如 [-1, 64]".to_string(),
+            ));
+        }
+        let r0 = dto.resize[0];
+        let r1 = dto.resize[1];
+
+        let resize = if r0 == -1 {
+            if dto.word {
+                // 如果 word 为 true，且包含 -1，Python 里是 resize 为 (r1, r1) 的正方形
+                Resize::Square(r1 as u32)
+            } else {
+                // 如果 word 为 false，且包含 -1，Python 里是高度固定为 r1，宽度按原图比例缩放
+                Resize::DynamicWidth(r1 as u32)
+            }
+        } else {
+            // 正常的固定宽高
+            Resize::Fixed(r0 as u32, r1 as u32)
+        };
+
+        Ok(ModelMetadata::new(
+            charset,
+            dto.word,
+            resize,
+            dto.channel,
+            dto.normalization.into(),
+        ))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    const MINIMAL_JSON: &str = r#"{
+        "charset": ["a", "b", "c"],
+        "word": false,
+        "resize": [-1, 64],
+        "channel": 1
+    }"#;
+
+    #[test]
+    fn parses_minimal_json_with_default_normalization() {
+        let meta = ModelMetadata::from_json_str(MINIMAL_JSON).unwrap();
+        assert_eq!(meta.charset.size(), 3);
+        assert!(!meta.word);
+        assert_eq!(meta.channel, 1);
+        assert!(matches!(meta.resize, Resize::DynamicWidth(64)));
+        assert!(matches!(meta.normalization, Normalization::ZeroToOne));
+    }
+
+    #[test]
+    fn parses_minus_one_to_one_normalization() {
+        let json = r#"{
+            "charset": ["a"],
+            "word": false,
+            "resize": [-1, 64],
+            "channel": 1,
+            "normalization": "minus_one_to_one"
+        }"#;
+        let meta = ModelMetadata::from_json_str(json).unwrap();
+        assert!(matches!(meta.normalization, Normalization::MinusOneToOne));
+    }
+
+    #[test]
+    fn parses_word_model_as_square_resize() {
+        let json = r#"{
+            "charset": ["a"],
+            "word": true,
+            "resize": [-1, 64],
+            "channel": 1
+        }"#;
+        let meta = ModelMetadata::from_json_str(json).unwrap();
+        assert!(matches!(meta.resize, Resize::Square(64)));
+    }
+
+    #[test]
+    fn parses_fixed_resize() {
+        let json = r#"{
+            "charset": ["a"],
+            "word": false,
+            "resize": [100, 64],
+            "channel": 1
+        }"#;
+        let meta = ModelMetadata::from_json_str(json).unwrap();
+        assert!(matches!(meta.resize, Resize::Fixed(100, 64)));
+    }
+
+    #[test]
+    fn parses_image_alias() {
+        let json = r#"{
+            "charset": ["a"],
+            "word": false,
+            "image": [200, 64],
+            "channel": 1
+        }"#;
+        let meta = ModelMetadata::from_json_str(json).unwrap();
+        assert!(matches!(meta.resize, Resize::Fixed(200, 64)));
+    }
+
+    #[test]
+    fn rejects_resize_with_wrong_length() {
+        let json = r#"{
+            "charset": ["a"],
+            "word": false,
+            "resize": [1, 2, 3],
+            "channel": 1
+        }"#;
+        assert!(matches!(
+            ModelMetadata::from_json_str(json),
+            Err(Error::MetadataParse(_))
+        ));
+    }
+
+    #[test]
+    fn rejects_invalid_json() {
+        assert!(matches!(
+            ModelMetadata::from_json_str("not json"),
+            Err(Error::JsonParse(_))
+        ));
+    }
+
+    #[test]
+    fn rejects_invalid_utf8_bytes() {
+        assert!(matches!(
+            ModelMetadata::from_json_bytes(&[0xff, 0xfe]),
+            Err(Error::InvalidUtf8(_))
+        ));
+    }
+}

ddddocr-tract2/src/loader/model.rs

@@ -0,0 +1,95 @@
+use crate::loader::error::{Error, ParseError, Result};
+use crate::types::Session;
+use ddddocr_core::traits::Loader;
+use std::io::Cursor;
+use tract_linalg::multithread::{Executor, set_default_executor};
+use tract_onnx::onnx;
+use tract_onnx::prelude::*;
+
+/// Tract 专用的链式构建器
+#[derive(Debug, Clone, Default)]
+pub struct ModelLoader {
+    num_threads: Option<usize>,
+}
+
+impl ModelLoader {
+    /// 可选扩展：设置 CPU 线程数（不提供任何 GPU 相关的 API）
+    pub fn num_threads(mut self, threads: usize) -> Self {
+        self.num_threads = Some(threads);
+        self
+    }
+    fn setup_tract_threads(&self) {
+        // 💡 1. 如果设置了线程数，可以通过 tract 的 multithread 配置应用给 model
+        if let Some(threads) = self.num_threads {
+            // 在 Tract 中可以通过 set_num_threads 或设置底层环境控制并发
+            // （注：Tract 0.20+ 版本支持全局/局部线程控制）
+            let executor = if threads <= 1 {
+                Executor::SingleThread
+            } else {
+                Executor::multithread(threads)
+            };
+            set_default_executor(executor);
+        }
+    }
+}
+
+impl Loader for ModelLoader {
+    type Session = Session;
+    type Error = Error;
+    fn build_for_path<P>(&self, model_path: P) -> Result<Session>
+    where
+        P: AsRef<std::path::Path>,
+    {
+        self.setup_tract_threads();
+        let path_ref = model_path.as_ref();
+
+        let session = onnx()
+            .model_for_path(path_ref)
+            .map_err(|e| ParseError::Path(path_ref.display().to_string(), e))?
+            .into_optimized()
+            .map_err(Error::OptimizationFailed)?
+            .into_runnable()
+            .map_err(Error::RunnableBuildFailed)?;
+        Ok(session)
+    }
+    /// 策略 B：从内存字节流加载模型（配合 include_bytes! 使用）
+    fn build_from_bytes(&self, model_bytes: &[u8]) -> Result<Session> {
+        self.setup_tract_threads();
+        // 使用 std::io::Cursor 将 &[u8] 包装为可读的流（实现 std::io::Read）
+        let mut cursor = Cursor::new(model_bytes);
+
+        let session = onnx()
+            .model_for_read(&mut cursor)
+            .map_err(ParseError::Bytes)?
+            .into_optimized()
+            .map_err(Error::OptimizationFailed)?
+            .into_runnable()
+            .map_err(Error::RunnableBuildFailed)?;
+
+        Ok(session)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn default_config() {
+        let loader = ModelLoader::default();
+        assert!(loader.num_threads.is_none());
+    }
+
+    #[test]
+    fn builder_configures_threads() {
+        let loader = ModelLoader::default().num_threads(4);
+        assert_eq!(loader.num_threads, Some(4));
+    }
+
+    #[test]
+    fn builder_is_chainable_and_immutable() {
+        let base = ModelLoader::default();
+        let _configured = base.clone().num_threads(8);
+        assert!(base.num_threads.is_none());
+    }
+}

ddddocr-tract2/src/ocr.rs

@@ -0,0 +1 @@
+pub mod session;

ddddocr-tract2/src/ocr/session.rs

@@ -0,0 +1,174 @@
+use crate::runtime::{extract_plain_array, run_session};
+use crate::types::Session;
+use ddddocr_core::ModelMetadata;
+use ddddocr_core::OcrOutput;
+use ddddocr_core::error::{DdddError, Result, TensorError};
+use ddddocr_core::traits::{InferenceEngine, Info, OcrEngine};
+use ddddocr_core::types::{AxisDim, ModelInfo, TensorInfo, TensorType};
+use ddddocr_core::utils::normalize_ocr_logits;
+use tracing::debug;
+use tract_onnx::prelude::{DatumType, IntoTensor, OutletId, ShapeFact, TypedModel};
+
+/// OCR 推理运行时：持有 Tract 会话与模型元数据，输出 [`OcrOutput`]。
+pub struct OcrRuntime {
+    /// Tract 会话句柄。
+    pub session: Session,
+    /// 模型元数据（字符集、归一化策略等）。
+    pub metadata: ModelMetadata,
+}
+impl OcrRuntime {
+    /// 基于已构建的会话与元数据创建 OCR 运行时。
+    pub fn new(session: Session, metadata: ModelMetadata) -> Self {
+        Self { session, metadata }
+    }
+
+    /// 提取出来的公共转换逻辑：将一组 OutletId 解析为 TensorInfo 列表
+    fn resolve_tensors(&self, model: &TypedModel, outlets: &[OutletId]) -> Result<Vec<TensorInfo>> {
+        outlets
+            .iter()
+            .map(|&outlet_id| {
+                let fact = model.outlet_fact(outlet_id).map_err(DdddError::new)?;
+                let shape = resolve_shape(&fact.shape);
+                let node_name = model.node(outlet_id.node).name.clone();
+                let tensor_type = tensor_type_from_datum(fact.datum_type);
+
+                Ok(TensorInfo {
+                    name: node_name,
+                    shape,
+                    tensor_type,
+                })
+            })
+            .collect() // 函数式声明：自动传播第一处发生的错误
+    }
+}
+
+/// 将 Tract 的 DatumType 映射为 core 的 [`TensorType`]。
+fn tensor_type_from_datum(datum: DatumType) -> TensorType {
+    match datum {
+        DatumType::F32 => TensorType::F32,
+        DatumType::I64 => TensorType::I64,
+        _ => TensorType::Other,
+    }
+}
+
+/// 安全还原 Tract 维度至 [`AxisDim`] 列表。
+fn resolve_shape(shape_fact: &ShapeFact) -> Vec<AxisDim> {
+    shape_fact
+        .to_tvec()
+        .iter()
+        .map(|dim| {
+            // 防御性编程：必须同时满足能够转换为 i64 且 大于等于 0
+            if let Ok(size) = dim.to_i64() {
+                if size >= 0 {
+                    AxisDim::Static(size as usize)
+                } else {
+                    // 如果 ONNX 导出时某些动态维度被标记为了 -1，安全地作为动态符号捕获
+                    AxisDim::Dynamic(dim.to_string())
+                }
+            } else {
+                AxisDim::Dynamic(dim.to_string())
+            }
+        })
+        .collect()
+}
+
+impl OcrEngine for OcrRuntime {
+    fn metadata(&self) -> &ModelMetadata {
+        &self.metadata
+    }
+}
+impl InferenceEngine for OcrRuntime {
+    type Output = OcrOutput;
+    fn inference(&self, input_array: ndarray::Array4<f32>) -> Result<Self::Output, TensorError> {
+        let mut result = run_session(&self.session, input_array)?;
+        debug!("模型输出原始数据: {:?}", result);
+        let raw_tensor = result.swap_remove(0).into_tensor();
+        match raw_tensor.datum_type() {
+            DatumType::I64 => {
+                let array_d = extract_plain_array::<i64>(raw_tensor)?;
+                let actual_shape = array_d.shape().to_vec();
+                let array1 = array_d.into_dimensionality::<ndarray::Ix1>().map_err(|_| {
+                    TensorError::DimensionMismatch {
+                        expected: "1D 字符索引静态矩阵".to_string(),
+                        actual: actual_shape,
+                    }
+                })?;
+                Ok(OcrOutput::Indices(array1))
+            }
+            DatumType::F32 => {
+                let shape = raw_tensor.shape().to_vec();
+                debug!("模型输出 shape 数据: {:?}", shape);
+                let array = extract_plain_array::<f32>(raw_tensor)?;
+                normalize_ocr_logits(array.view(), array.shape())
+            }
+            _ => Err(TensorError::UnknownOutputFormat),
+        }
+    }
+}
+impl Info for OcrRuntime {
+    fn input_info(&self) -> Result<Vec<TensorInfo>> {
+        let model = self.session.model();
+        let outlets = model.input_outlets().map_err(DdddError::new)?;
+        self.resolve_tensors(model, outlets)
+    }
+
+    /// 获取模型输出的节点信息列表
+    fn output_info(&self) -> Result<Vec<TensorInfo>> {
+        let model = self.session.model();
+        let outlets = model.output_outlets().map_err(DdddError::new)?;
+        self.resolve_tensors(model, outlets)
+    }
+
+    /// 获取模型详细元数据信息（对标 Python ddddocr 的 get_model_info）
+    /// 完美包容 [1, 1, 64, image_width] 这样的变长图像模型
+    /// 获取模型详细元数据信息（代码更紧凑、优雅）
+    fn model_info(&self) -> Result<ModelInfo> {
+        Ok(ModelInfo {
+            inputs: self.input_info()?,
+            outputs: self.output_info()?,
+            providers: None,
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use tract_onnx::prelude::*;
+
+    #[test]
+    fn maps_datum_types() {
+        assert!(matches!(
+            tensor_type_from_datum(DatumType::F32),
+            TensorType::F32
+        ));
+        assert!(matches!(
+            tensor_type_from_datum(DatumType::I64),
+            TensorType::I64
+        ));
+        assert!(matches!(
+            tensor_type_from_datum(DatumType::U8),
+            TensorType::Other
+        ));
+    }
+
+    #[test]
+    fn resolves_static_shape() {
+        let fact = ShapeFact::from_dims(vec![TDim::from(1), TDim::from(64)]);
+        assert_eq!(
+            resolve_shape(&fact),
+            vec![AxisDim::Static(1), AxisDim::Static(64)]
+        );
+    }
+
+    #[test]
+    fn resolves_dynamic_dimension_as_symbol() {
+        let scope = SymbolScope::default();
+        let dims = vec![TDim::from(1), TDim::from(64), scope.sym("width").into()];
+        let fact = ShapeFact::from_dims(dims);
+        let shape = resolve_shape(&fact);
+        assert_eq!(shape[0], AxisDim::Static(1));
+        assert_eq!(shape[1], AxisDim::Static(64));
+        assert!(matches!(&shape[2], AxisDim::Dynamic(_)));
+    }
+}

ddddocr-tract2/src/runtime.rs

@@ -0,0 +1,25 @@
+//! 会话执行相关的共享工具函数。
+
+use crate::types::Session;
+use ddddocr_core::error::TensorError;
+use tract_onnx::prelude::{Datum, TValue, TVec, Tensor, tvec};
+
+/// 对输入张量执行一次 Tract 推理，返回原始输出值列表。
+pub(crate) fn run_session(
+    session: &Session,
+    input_array: ndarray::Array4<f32>,
+) -> Result<TVec<TValue>, TensorError> {
+    let tensor = Tensor::from(input_array);
+    session
+        .run(tvec!(tensor.into()))
+        .map_err(|e| TensorError::Engine(format!("执行模型推理失败: {e}")))
+}
+
+/// 将 Tensor 转换为 ndarray 的 `ArrayD`。
+pub(crate) fn extract_plain_array<D: Datum>(
+    tensor: Tensor,
+) -> Result<ndarray::ArrayD<D>, TensorError> {
+    tensor
+        .into_plain_array::<D>()
+        .map_err(|_| TensorError::Engine("无法获取张量内存视图".to_string()))
+}

ddddocr-tract2/src/types.rs

@@ -0,0 +1,7 @@
+//! Tract 会话共享类型。
+
+use std::sync::Arc;
+use tract_onnx::prelude::TypedRunnableModel;
+
+/// Tract 会话句柄：由 [`crate::loader::ModelLoader`] 构建，可直接并发执行推理。
+pub type Session = Arc<TypedRunnableModel>;

ddddocr-tract2/tests/api_surface.rs

@@ -0,0 +1,74 @@
+//! 外部视角 API 测试：验证 `ddddocr-tract2` 的公开类型与方法在外部 crate 中可正常使用。
+
+use ddddocr_core::Resize;
+use ddddocr_core::traits::{Info, Loader};
+use ddddocr_core::types::AxisDim;
+use ddddocr_tract2::loader::ModelLoader;
+use ddddocr_tract2::{
+    DetRuntime, ModelMetadata, Normalization, OcrBuilder, OcrRuntime, Session, SlideResult, Slider,
+};
+use std::path::{Path, PathBuf};
+
+fn model_path(name: &str) -> PathBuf {
+    Path::new(env!("CARGO_MANIFEST_DIR"))
+        .join("..")
+        .join("models")
+        .join(name)
+}
+
+/// 验证对外导出的类型（含 `Session` 与 core 便捷重导出）均可直接命名。
+#[test]
+fn public_types_are_nameable() {
+    let _: Option<Session> = None;
+    let _: Option<OcrBuilder> = None;
+    let _: Option<SlideResult> = None;
+    let _slider = Slider::new();
+}
+
+/// 验证构建器链式 API 可组合使用，且返回的会话类型可显式标注。
+#[test]
+fn loader_chain_builds_session() -> anyhow::Result<()> {
+    let path = model_path("common_sml2h3_f32.onnx");
+    assert!(path.exists(), "缺少测试模型: {}", path.display());
+    let _session: Session = ModelLoader::default().num_threads(4).build_for_path(path)?;
+    Ok(())
+}
+
+/// 验证 `Info` trait 能从真实会话中解析输入/输出信息。
+#[test]
+fn info_trait_returns_model_metadata() -> anyhow::Result<()> {
+    let path = model_path("common_sml2h3_f32.onnx");
+    assert!(path.exists(), "缺少测试模型: {}", path.display());
+    let session: Session = ModelLoader::default().build_for_path(path)?;
+    let metadata = ModelMetadata::from_static_slice(
+        &["a", "b"],
+        false,
+        Resize::DynamicWidth(64),
+        1,
+        Normalization::MinusOneToOne,
+    );
+    let ocr = OcrRuntime::new(session, metadata);
+
+    let inputs = ocr.input_info()?;
+    let outputs = ocr.output_info()?;
+    assert!(!inputs.is_empty());
+    assert!(!outputs.is_empty());
+    assert!(matches!(inputs[0].shape[0], AxisDim::Static(1)));
+    assert!(matches!(inputs[0].shape[2], AxisDim::Static(64)));
+
+    let model_info = ocr.model_info()?;
+    assert_eq!(model_info.inputs.len(), inputs.len());
+    assert_eq!(model_info.outputs.len(), outputs.len());
+    assert!(model_info.providers.is_none());
+    Ok(())
+}
+
+/// 验证检测运行时可以从构建的会话创建。
+#[test]
+fn det_runtime_builds_from_session() -> anyhow::Result<()> {
+    let path = model_path("common_det.onnx");
+    assert!(path.exists(), "缺少测试模型: {}", path.display());
+    let session: Session = ModelLoader::default().build_for_path(path)?;
+    let _det = DetRuntime::new(session);
+    Ok(())
+}

ddddocr-tract2/tests/common/mod.rs

@@ -0,0 +1,25 @@
+//! 集成测试共享工具。
+
+use std::path::{Path, PathBuf};
+
+/// 仓库根目录 `models/` 下模型文件的路径。
+pub fn model_path(name: &str) -> PathBuf {
+    Path::new(env!("CARGO_MANIFEST_DIR"))
+        .join("..")
+        .join("models")
+        .join(name)
+}
+
+/// 仓库根目录 `samples/` 下样例图片的路径。
+pub fn sample_path(name: &str) -> PathBuf {
+    Path::new(env!("CARGO_MANIFEST_DIR"))
+        .join("..")
+        .join("samples")
+        .join(name)
+}
+
+/// 加载图片，失败时附带路径上下文。
+pub fn load_image<P: AsRef<Path>>(path: P) -> anyhow::Result<image::DynamicImage> {
+    let path_ref = path.as_ref();
+    image::open(path_ref).map_err(|e| anyhow::anyhow!("无法加载图片 {:?}: {}", path_ref, e))
+}

ddddocr-tract2/tests/det.rs

@@ -0,0 +1,33 @@
+//! 目标检测集成测试。
+
+mod common;
+
+use common::{model_path, sample_path};
+use ddddocr_core::Detector;
+use ddddocr_core::traits::Loader;
+use ddddocr_tract2::DetRuntime;
+use ddddocr_tract2::loader::ModelLoader;
+use image::GenericImageView;
+
+/// 检测模型应能从样例图片中找到至少一个目标，且坐标在图片范围内。
+#[test]
+fn det_model_detects_targets_in_image() -> anyhow::Result<()> {
+    let session = ModelLoader::default()
+        .build_for_path(model_path("common_det.onnx"))
+        .expect("模型加载失败");
+    let det = DetRuntime::new(session);
+    let img = image::open(sample_path("det1.png")).expect("测试图片不存在");
+
+    let bboxes = Detector::new(&det).predict(&img)?;
+    assert!(!bboxes.is_empty(), "应检测到至少一个目标");
+
+    let (width, height) = img.dimensions();
+    for bbox in &bboxes {
+        assert!(bbox.x1 >= 0 && bbox.y1 >= 0, "检测框左上角不应为负");
+        assert!(
+            bbox.x2 <= width as i32 && bbox.y2 <= height as i32,
+            "检测框右下角不应超出图片范围"
+        );
+    }
+    Ok(())
+}

ddddocr-tract2/tests/ocr.rs

@@ -0,0 +1,55 @@
+//! OCR 识别与模型信息集成测试。
+
+mod common;
+
+use common::{model_path, sample_path};
+use ddddocr_core::traits::{Info, Loader};
+use ddddocr_core::{ModelMetadata, Normalization, Ocr, Resize};
+use ddddocr_tract2::OcrRuntime;
+use ddddocr_tract2::loader::ModelLoader;
+
+/// 用官方 sml2h3 f32 模型识别验证码图片，结果不应为空。
+#[test]
+fn ocr_classification_recognizes_code_image() {
+    let session = ModelLoader::default()
+        .build_for_path(model_path("common_sml2h3_f32.onnx"))
+        .expect("模型加载失败");
+    let metadata = ModelMetadata::from_builtin_beta(
+        false,
+        Resize::DynamicWidth(64),
+        1,
+        Normalization::MinusOneToOne,
+    );
+    let ocr = OcrRuntime::new(session, metadata);
+
+    let img = image::open(sample_path("code2.png")).expect("测试图片不存在");
+    let text = Ocr::builder()
+        .build_with(&ocr)
+        .predict(&img)
+        .expect("识别过程出错")
+        .into_text();
+
+    println!("识别结果: {text}");
+    assert!(!text.is_empty(), "识别结果不应为空");
+}
+
+/// 真实模型应能通过 `Info` trait 返回输入/输出张量信息。
+#[test]
+fn model_info_lists_inputs_and_outputs() -> anyhow::Result<()> {
+    let session = ModelLoader::default()
+        .build_for_path(model_path("common_huashi666_i64.onnx"))
+        .expect("建立测试模型图失败");
+    let metadata = ModelMetadata::from_builtin_beta(
+        false,
+        Resize::DynamicWidth(64),
+        1,
+        Normalization::MinusOneToOne,
+    );
+    let ocr = OcrRuntime::new(session, metadata);
+
+    let inputs = ocr.input_info()?;
+    let outputs = ocr.output_info()?;
+    assert!(!inputs.is_empty(), "模型应有输入张量信息");
+    assert!(!outputs.is_empty(), "模型应有输出张量信息");
+    Ok(())
+}

ddddocr-tract2/tests/slide.rs

@@ -0,0 +1,52 @@
+//! 滑块匹配集成测试。
+
+mod common;
+
+use common::{load_image, sample_path};
+use ddddocr_core::Slider;
+
+/// 边缘模式匹配应定位到预期坐标。
+#[test]
+fn slide_match_locates_target_position() {
+    let engine = Slider::new();
+    let target = load_image(sample_path("target1.png")).expect("请确保 samples/target1.png 存在");
+    let background = load_image(sample_path("background1.png")).expect("请确保 samples/background1.png 存在");
+
+    let start = std::time::Instant::now();
+    let result = engine
+        .slide_match(&target, &background, false)
+        .expect("Slide match 执行失败");
+    let elapsed = start.elapsed();
+
+    println!("边缘模式匹配: {result}");
+    println!("耗时: {elapsed:?}");
+
+    assert_eq!(result.target_x, 237);
+    assert_eq!(result.target_y, 77);
+    assert!(result.confidence > 0.0);
+}
+
+/// 灰度对比匹配应定位到预期坐标。
+#[test]
+fn slide_comparison_locates_target_position() {
+    let engine = Slider::new();
+    let target = load_image(sample_path("target2.jpg")).expect("请确保 samples/target2.jpg 存在");
+    let background =
+        load_image(sample_path("background2.jpg")).expect("请确保 samples/background2.jpg 存在");
+
+    let start = std::time::Instant::now();
+    let result = engine
+        .slide_comparison(&target, &background)
+        .expect("Slide comparison 执行失败");
+    let elapsed = start.elapsed();
+
+    println!(
+        "灰度对比匹配: 坐标 [x: {}, y: {}], 置信度 {:.4}",
+        result.target_x, result.target_y, result.confidence
+    );
+    println!("耗时: {elapsed:?}");
+
+    assert_eq!(result.target_x, 171);
+    assert_eq!(result.target_y, 90);
+    assert!(result.confidence > 0.0);
+}

examples/simple_usage.rs

@@ -1,5 +1,5 @@
 fn main() {
-    let ocr = ddddocr_rs::DdddOcr::new("model/common.onnx").unwrap();
-    let img = image::open("samples/code3.png").unwrap();
-    println!("Result: {}", ocr.classification(&img).unwrap());
+    // let ocr = ddddocr_rs::DdddOcrBuilder::new().build().unwrap();
+    // let img = image::open("samples/code3.png").unwrap();
+    // println!("Result: {}", ocr.classification(&img).unwrap());
 }

src/charset.rs

@@ -1,516 +0,0 @@
-pub const CHARSET_BETA: &[&str] = &[
-    "", "笤", "谴", "膀", "荔", "佰", "电", "臁", "矍", "同", "奇", "芄", "吠", "6", "曛", "荇",
-    "砥", "蹅", "晃", "厄", "殣", "ｃ", "辱", "钋", "杻", "價", "眙", "鴿", "⒄", "裙", "训", "涛",
-    "酉", "挞", "忙", "怍", "︰", "镍", "檐", "眯", "茓", "辖", "淩", "啟", "蜀", "芟", "裟", "楝",
-    "彘", "嶪", "费", "亞", "滁", "榉", "朝", "f", "倻", "裎", "谧", "崂", "卑", "助", "触", "氐",
-    "锟", "铢", "膪", "脐", "渲", "荫", "佾", "琯", "钣", "珰", "翦", "膻", "娥", "浥", "淄", "猸",
-    "内", "消", "粞", "反", "苪", "冽", "酵", "玩", "父", "存", "屯", "殷", "俐", "篱", "俛", "塮",
-    "苕", "耲", "输", "壖", "溶", "琤", "氏", "真", "黩", "瑄", "阶", "茔", "眩", "浙", "痄", "噔",
-    "烤", "楯", "²", "铆", "裈", "偬", "盏", "祐", "伯", "庙", "獯", "榀", "裒", "综", "蒸", "架",
-    "蜱", "鹖", "涴", "肌", "廖", "祾", "蔗", "破", "!", "鑫", "瓷", "Ｈ", "宛", "倪", "贝", "酝",
-    "倞", "榼", "菊", "帕", "胍", "淌", "抨", "倕", "味", "独", "à", "庐", "蹲", "肸", "洹", "騳",
-    "绖", "觉", "蝙", "铜", "選", "郚", "奄", "手", "篦", "忮", "潺", "歁", "湖", "貔", "缚", "癫",
-    "捣", "翎", "勇", "徒", "杪", "捃", "纴", "郞", "蛱", "浓", "講", "薇", "汊", "彻", "琖", "觇",
-    "驩", "野", "闺", "彩", "膊", "簸", "瑭", "龏", "栓", "攸", "堕", "鹿", "檠", "锽", "晟", "煊",
-    "衖", "p", "L", "侞", "吹", "岵", "捡", "邃", "曩", "泼", "娌", "磙", "鞮", "號", "苤", "骁",
-    "感", "氙", "榜", "菟", "蠲", "∶", "焌", "漯", "胪", "以", "剜", "=", "衰", "剔", "疏", "韩",
-    "邋", "探", "搌", "握", "舵", "腰", "咐", "郎", "鈺", "赴", "斩", "铥", "棂", "褫", "秾", "城",
-    "葬", "缺", "甓", "沙", "鴨", "恧", "吩", "膈", "俗", "引", "濑", "坛", "蛘", "谍", "飘", "鎚",
-    "貋", "袅", "圃", "肽", "祖", "瑢", "鄣", "卡", "恸", "饷", "撖", "阖", "碌", "墉", "⇋", "抗",
-    "僇", "撩", "狷", "静", "荽", "憯", "虻", "滹", "簧", "Y", "汾", "嫁", "蚴", "岈", "榔", "邶",
-    "挛", "火", "w", "旌", "線", "3", "跏", "F", "楸", "瞬", "證", "現", "符", "鲀", "窸", "朗",
-    "m", "劉", "襄", "鸻", "敖", "憩", "濛", "胶", "雏", "禽", "缜", "鐀", "澄", "泉", "懈", "鹟",
-    "牙", "叟", "镖", "膨", "硷", "钏", "嵌", "冲", "橪", "厨", "溘", "妻", "贺", "耀", "潲", "瞳",
-    "惺", "涑", "鄄", "舡", "战", "钹", "盍", "窀", "凳", "锋", "（", "绤", "翻", "》", "嶷", "戛",
-    "照", "设", "两", "霹", "風", "格", "栖", "椹", "№", "蔌", "达", "悠", "旒", "函", "抔", "逆",
-    "疙", "玫", "箢", "恩", "樘", "远", "考", "荭", "殒", "靡", "蝾", "舅", "䀲", "偾", "灿", "埴",
-    "瀍", "特", "诸", "搦", "恒", "妊", "课", "劂", "殊", "艋", "柚", "硕", "捅", "钍", "芘", "脖",
-    "襁", "募", "卅", "疆", "嫖", "黹", "臀", "豇", "瘗", "憎", "嗯", "Θ", "跺", "喧", "捺", "爪",
-    "鲠", "縠", "屋", "撑", "者", "娶", "喝", "墼", "丶", "茚", "髅", "瓢", "农", "橼", "攰", "折",
-    "诋", "镒", "赂", "捲", "耑", "沐", "窜", "亰", "煌", "阕", "羯", "纤", "滇", "值", "琥", "渴",
-    "祗", "鳌", "蛞", "庄", "鹇", "訄", "犷", "弋", "琐", "佗", "氛", "揄", "旺", "聱", "榨", "湿",
-    "蟥", "湎", "敕", "轷", "耿", "三", "運", "瑶", "困", "勁", "蚱", "泸", "螫", "斜", "蟠", "轨",
-    "镞", "霆", "嶙", "烷", "瘩", "敦", "塾", "僚", "澙", "關", "酩", "殴", "—", "例", "筷", "乘",
-    "颗", "核", "孬", "舣", "糇", "劾", "黧", "镌", "罅", "X", "仅", "哐", "蟊", "呻", "呕", "粟",
-    "配", "伊", "槊", "昌", "宰", "盘", "肫", "鳥", "圾", "恬", "辘", "绿", "時", "丐", "扃", "敓",
-    "摄", "陕", "滿", "鹆", "嗬", "龅", "渣", "釣", "萦", "督", "孑", "∧", "疥", "噱", "蝎", "君",
-    "笮", "泌", "镔", "称", "柘", "鬣", "罨", "潍", "垣", "顔", "褶", "礓", "△", "骐", "湍", "獾",
-    "羖", "戾", "预", "祭", "鹨", "凤", "茕", "珏", "蛊", "毛", "枰", "鄭", "娲", "丕", "蜜", "纟",
-    "蛹", "粮", "嚰", "嚓", "螨", "裰", "G", "集", "删", "郜", "舻", "嵛", "鵰", "腊", "峄", "脸",
-    "鲺", "坢", "寞", "撰", "顸", "枋", "荠", "夡", "豆", "馆", "赭", "傎", "淼", "镪", "許", "礌",
-    "带", "訂", "饧", "锜", "及", "漾", "編", "究", "仍", "糜", "喳", "嗝", "醦", "堅", "企", "烔",
-    "图", "垲", "枥", "畿", "踱", "槟", "◆", "酽", "溢", "酥", "谚", "缨", "死", "镑", "干", "用",
-    "紊", "坑", "副", "枭", "琮", "鸨", "獻", "弈", "伺", "醍", "氓", "宁", "臻", "贾", "啐", "玉",
-    "咀", "孀", "烆", "嘱", "频", "蜥", "楡", "瞩", "委", "锺", "赀", "睬", "旅", "刿", "韪", "抹",
-    "鞘", "x", "钝", "倌", "奖", "蟭", "灌", "肼", "曰", "啊", "属", "唐", "彦", "煦", "鄘", "坦",
-    "鬻", "告", "单", "菽", "匏", "浚", "仺", "怜", "拦", "鸡", "拄", "乍", "燠", "暧", "竭", "⒂",
-    "济", "Ⓡ", "趑", "舴", "体", "拒", "罂", "说", "猎", "闫", "鄗", "妲", "鑙", "葶", "匍", "等",
-    "略", "盲", "唔", "钰", "渠", "镦", "葫", "蹒", "姘", "婵", "夯", "实", "何", "株", "锌", "礁",
-    "桺", "捉", "鮕", "莓", "轶", "辉", "溋", "视", "嗒", "猞", "猁", "杓", "怡", "咴", "巷", "仂",
-    "婉", "睛", "？", "淡", "/", "郛", "绨", "较", "毁", "沓", "瞎", "馏", "蕙", "戏", "i", "董",
-    "臣", "鹫", "栳", "锹", "裕", "蜷", "唣", "９", "缂", "螬", "笥", "惶", "蚰", "徨", "忭", "传",
-    "绛", "离", "锑", "候", "拓", "德", "损", "附", "紡", "徕", "錾", "蕻", "⑧", "構", "镊", "脘",
-    "靸", "涒", "镡", "光", "廒", "尴", "荥", "佳", "弼", "暮", "榆", "鼑", "辅", "钆", "湝", "佤",
-    "瘌", "炽", "筰", "嗈", "電", "飙", "坶", "椿", "俾", "灊", "泵", "像", "咅", "樓", "苒", "烨",
-    "溺", "棽", "戒", "箅", "愔", "缢", "楞", "庤", "塑", "湮", "沽", "蝌", "赌", "薨", "锾", "圜",
-    "骥", "秉", "瞌", "惇", "诊", "圣", "睢", "我", "廰", "苠", "襻", "鲚", "酮", "厮", "评", "沈",
-    "愀", "垞", "习", "敷", "比", "欢", "尚", "钪", "卣", "她", ",", "伴", "赃", "蚬", "喇", "醇",
-    "嚅", "T", "姝", "鍊", "悯", "N", "抚", "颡", "獐", "趟", "洑", "缝", "喁", "帷", "憷", "获",
-    "阉", "镫", "臨", "炮", "奴", "揩", "叩", "恺", "粱", "胁", "憬", "痿", "´", "沘", "彀", "饩",
-    "滋", "竑", "嗑", "鸟", "Ｔ", "璠", "快", "銀", "舶", "羞", "桞", "飛", "茹", "師", "偌", "節",
-    "冁", "叡", "臑", "踔", "酔", "養", "溦", "岐", "豉", "杌", "胩", "仳", "沾", "窑", "曳", "闷",
-    "垢", "垆", "磨", "髹", "態", "丽", "見", "洃", "遣", "场", "铭", "溉", "衬", "橦", "详", "馊",
-    "濂", "瑾", "鲫", "贞", "搴", "胺", " ", "景", "执", "袴", "炖", "箭", "楂", "婚", "镧", "厩",
-    "宣", "缣", "跃", "痘", "亓", "宦", "豪", "侣", "郧", "與", "痳", "挚", "殻", "嘢", "洽", "蹉",
-    "孽", "森", "俚", "八", "砹", "凹", "訇", "屹", "啰", "宴", "廨", "沇", "麇", "泅", "绡", "辛",
-    "鹲", "鋆", "旮", "婀", "幔", "赞", "：", "永", "活", "萸", "霉", "←", "媲", "阚", "鲮", "佝",
-    "獒", "圹", "隋", "征", "蚶", "龊", "搀", "嫫", "鬏", "幞", "學", "然", "沧", "萑", "襜", "撻",
-    "苯", "狳", "鞠", "咋", "壹", "栅", "款", "镘", "阳", "蚜", "荬", "糊", "疳", "糕", "镵", "寐",
-    "褴", "v", "醋", "诀", "汰", "啸", "备", "娘", "氚", "镓", "室", "簟", "硝", "嘌", "釐", "争",
-    "男", "疬", "鹄", "艳", "倒", "忸", "庶", "葒", "岖", "涨", "羝", "诼", "纨", "纰", "扉", "酎",
-    "藉", "疮", "枞", "黜", "戮", "芽", "鳑", "末", "蒿", "茈", "透", "渊", "秷", "〈", "试", "络",
-    "羼", "滪", "奋", "虏", "脁", "沫", "蓰", "襆", "披", "鲌", "艚", "逹", "炳", "泔", "叹", "轳",
-    "锵", "嫜", "佥", "严", "迅", "筠", "逵", "铿", "钇", "拇", "诏", "绯", "吊", "纠", "蟾", "c",
-    "涝", "汩", "盐", "跋", "拤", "邹", "镨", "羚", "龆", "脊", "攉", "傕", "短", "團", "蹬", "嘤",
-    "奎", "熨", "芪", "鸢", "濉", "''", "莴", "義", "赜", "踺", "皂", "努", "偏", "狡", "遭", "吞",
-    "嘿", "婊", "媸", "增", "殿", "刮", "燘", "劫", "娜", "瞄", "寡", "优", "捋", "佴", "菰", "蓠",
-    "笙", "镃", "樇", "瘫", "B", "橇", "逯", "堍", "О", "磐", "腼", "て", "送", "狭", "皓", "亡",
-    "嗉", "菠", "顺", "連", "嶶", "瑪", "辟", "婷", "牛", "笫", "窅", "萁", "戟", "覃", "馍", "建",
-    "謇", "旘", "镣", "燏", "葉", "轺", "倏", "堪", "见", "葛", "钕", "键", "押", "僊", "槐", "戎",
-    "窨", "洙", "鲢", "鞒", "慒", "雁", "圭", "Ｄ", "陌", "肱", "蜿", "洧", "惑", "祛", "樟", "矧",
-    "呵", "峻", "凝", "蕨", "拯", "珮", "塥", "展", "贻", "囐", "弱", "庳", "嫔", "緣", "呈", "策",
-    "漉", "瑗", "鲂", "鹂", "吾", "灶", "并", "挲", "重", "奭", "皙", "侪", "埗", "烬", "纾", "椒",
-    "技", "ɔ", "擀", "恍", "遨", "订", "雨", "卵", "锏", "猗", "癸", "）", "谡", "稷", "枨", "蹽",
-    "荑", "沅", "稽", "間", "冉", "颇", "酺", "份", "瞾", "毯", "藥", "蕞", "狲", "吡", "慷", "卯",
-    "摽", "肿", "嗛", "悒", "丨", "横", "鳡", "仫", "狎", "砗", "聿", "腥", "酡", "飱", "柳", "遽",
-    "汇", "湔", "麋", "垃", "粽", "坷", "鳗", "迫", "丢", "\"", "⒀", "嗲", "肐", "結", "署", "飨",
-    "蠡", "涩", "挈", "浿", "鐾", "姞", "隧", "铘", "呜", "蜕", "鷄", "逼", "哌", "病", "係", "偿",
-    "Ⅲ", "埋", "妤", "赘", "悉", "陷", "沸", "呲", "誓", "舆", "髀", "挫", "羑", "据", "顿", "淝",
-    "抟", "珧", "郑", "仗", "怛", "掠", "稳", "尥", "祙", "找", "郐", "锔", "轹", "钓", "黙", "饸",
-    "谌", "斐", "龙", "噫", "駆", "浼", "峒", "育", "纣", "溠", "铊", "亨", "杮", "呓", "钌", "业",
-    "繻", "溪", "戌", "蓿", "椱", "悱", "仉", "阮", "芈", "濋", "搔", "纽", "琛", "趄", "双", "镯",
-    "☆", "敉", "啬", "讦", "娱", "爾", "遶", "漱", "郗", "锪", "颃", "靰", "醊", "驊", "呢", "術",
-    "妙", "蚣", "溽", "酇", "巾", "舐", "却", "废", "邾", "砣", "乙", "鲜", "蒤", "囍", "璈", "稔",
-    "蘘", "匐", "業", "碟", "渺", "贤", "绋", "畑", "颞", "侥", "盟", "鼍", "阊", "蔽", "标", "吮",
-    "淬", "鏾", "圗", "夜", "乕", "娇", "瞿", "循", "讲", "懒", "熘", "禚", "观", "钷", "万", "n",
-    "未", "藝", "愿", "圈", "浩", "伦", "扛", "暄", "饶", "梧", "欣", "咿", "檔", "吼", "妮", "覆",
-    "辰", "誤", "允", "危", "硗", "惫", "瘪", "李", "焱", "沣", "坯", "穄", "归", "画", "营", ":",
-    "色", "哔", "矢", "巯", "祆", "傍", "享", "悻", "取", "凫", "铒", "唅", "眈", "疹", "败", "晴",
-    "顼", "绶", "剃", "斗", "禾", "４", "誜", "俨", "２", "患", "结", "可", "帇", "抍", "筝", "衢",
-    "鹛", "跸", "颢", "钾", "渡", "棒", "丛", "皱", "梓", "将", "压", "#", "岛", "?", "砺", "过",
-    "党", "挣", "瞋", "谶", "妯", "羡", "化", "淫", "歪", "鼗", "阄", "蔓", "烩", "餘", "猊", "．",
-    "畯", "祧", "狒", "碁", "咛", "鲈", "叨", "哞", "５", "娈", "半", "免", "拿", "畎", "媾", "棚",
-    "丈", "周", "匋", "酯", "奚", "爇", "摇", "搭", "蓇", "陽", "岢", "禤", "藠", "雅", "哲", "弹",
-    "按", "↑", "蹀", "察", "螾", "渎", "褂", "觳", "耍", "皲", "骗", "箫", "蕺", "亚", "保", "棵",
-    "放", "踪", "了", "熣", "亦", "痛", "币", "馐", "夢", "诱", "梱", "鲰", "郕", "璜", "祯", "颦",
-    "走", "踣", "嫚", "旯", "雲", "湟", "墨", "笃", "肇", "撝", "腦", "账", "舞", "⑨", "噻", "幂",
-    "僵", "崦", "’", "牢", "号", "嫡", "囱", "肥", "代", "锶", "掏", "随", "棓", "殉", "嘅", "掰",
-    "功", "垛", "踶", "娠", "霜", "碣", "鲼", "伉", "凄", "骋", "鹞", "洺", "乌", "赧", "瑛", "黎",
-    "曚", "鲴", "髫", "瘴", "藏", "雍", "畐", "蔻", "爼", "蹴", "巨", "贱", "汜", "胡", "虬", "椎",
-    "逸", "魇", "亶", "Φ", "忆", "赉", "塞", "潢", "垌", "简", "鼹", "發", "枢", "麝", "虹", "惭",
-    "唛", "春", "瑟", "郰", "桡", "捩", "堙", "嗨", "驳", "Ｆ", "荪", "忑", "贪", "躅", "步", "揜",
-    "闪", "垟", "晶", "分", "韭", "戴", "泪", "啧", "機", "峙", "和", "鸱", "绎", "屠", "阋", "黍",
-    "淸", "萩", "汉", "吐", "匙", "铗", "蠔", "簠", "鲵", "须", "蛣", "躏", "完", "咻", "釜", "馼",
-    "崤", "欻", "珐", "于", "郅", "焓", "轴", "递", "堰", "嗷", "儇", "壕", "嘟", "酸", "庾", "龂",
-    "妍", "锅", "雳", "桦", "抬", "谄", "气", "六", "诎", "绀", "张", "復", "客", "荞", "鳚", "衔",
-    "亁", "昂", "漤", "鞚", "筘", "绫", "彝", "枪", "苊", "榟", "饺", "苦", "顶", "衷", "聚", "寮",
-    "揆", "轪", "栋", "臂", "葖", "颋", "镐", "愕", "贸", "Q", "琼", "糥", "世", "莪", "龁", "禁",
-    "绲", "陶", "弑", "黢", "铵", "睐", "沄", "紬", "防", "癣", "曾", "钉", "纶", "膘", "句", "莸",
-    "踝", "躐", "酤", "腑", "雄", "堤", "喀", "姣", "孢", "阡", "褐", "胂", "髙", ";", "骖", "膺",
-    "糙", "辢", "⒃", "险", "砻", "缫", "骎", "低", "蚵", "箐", "苞", "劭", "峪", "工", "盈", "腹",
-    "袄", "祉", "癔", "笨", "Ｒ", "乚", "畏", "虍", "臾", "泛", "噙", "杷", "麗", "蹋", "逍", "迓",
-    "摅", "页", "戥", "胞", "艄", "壅", "啶", "趼", "牟", "翙", "蓝", "府", "轿", "砼", "荜", "杆",
-    "惊", "起", "瘅", "墈", "氖", "匀", "麃", "阘", "虮", "蘇", "蚤", "汗", "鳞", "籁", "缲", "畈",
-    "亟", "劬", "課", "蓄", "缅", "楮", "湜", "珩", "斋", "塬", "殁", "魃", "脞", "H", "澼", "钚",
-    "饕", "缕", "Ⅱ", "攮", "卿", "莅", "镆", "熹", "藩", "汁", "順", "趿", "拆", "蟹", "砒", "惴",
-    "㎡", "忖", "寝", "戕", "螭", "酿", "™", "柬", "枧", "凶", "蚁", "島", "殄", "鲊", "忠", "肉",
-    "辕", "叫", "徙", "漆", "缞", "夀", "楦", "佪", "兴", "粉", "裳", "蘧", "國", "旬", "看", "Ⅰ",
-    "剑", "痭", "襟", "恐", "遹", "◎", "窃", "穰", "澎", "敬", "旱", "燚", "坩", "彤", "尜", "猃",
-    "夏", "穈", "媒", "柑", "駡", "孛", "脉", "车", "零", "菩", "痊", "卉", "桔", "距", "吧", "漦",
-    "启", "仁", "滬", "馋", "帅", "鳈", "鄌", "超", "芡", "窘", "刽", "掌", "氤", "梽", "拎", "踏",
-    "勋", "甍", "玑", "稱", "鞍", "浍", "翅", "饟", "鼎", "罩", "加", "虚", "蕰", "簉", "堇", "巢",
-    "疲", "蟑", "狝", "瓮", "潋", "行", "饥", "散", "糌", "牵", "貢", "偉", "咄", "痕", "沃", "苓",
-    "锂", "狻", "褿", "畸", "姿", "煎", "胜", "觅", "烊", "質", "疵", "擢", "椤", "米", "累", "诳",
-    "斡", "K", "恻", "匦", "烫", "湾", "鹎", "吟", "摘", "涞", "恿", "嫉", "炎", "婧", "朽", "铑",
-    "ㆍ", "讧", "曜", "挑", "〇", "搅", "鹐", "丁", "彼", "棠", "饪", "箬", "祎", "魄", "囿", "犬",
-    "市", "髃", "勚", "桶", "辎", "瓞", "财", "缄", "園", "睒", "护", "尿", "融", "围", "水", "糍",
-    "虢", "呦", "越", "棺", "砮", "邓", "鹦", "稣", "呃", "柙", "鎌", "转", "袋", "湉", "亘", "俩",
-    "腆", "谣", "飔", "撂", "鄳", "爲", "盦", "谳", "卸", "W", "嚆", "婕", "卫", "拚", "呀", "汽",
-    "洣", "冻", "鳝", "录", "毋", "閥", "熬", "谜", "齐", "匳", "慧", "猴", "撬", "妳", "諾", "蠼",
-    "瘟", "伐", "颤", "奶", "陧", "麾", "岌", "浇", "邸", "「", "不", "哼", "热", "旳", "慙", "＆",
-    "苔", "郿", "钗", "氡", "纹", "侬", "霓", "靈", "扁", "聢", "疼", "岣", "甥", "恭", "喷", "芫",
-    "骂", "肪", "熥", "揠", "鲷", "遁", "霎", "娆", "圩", "爬", "傲", "贽", "紫", "觑", "琇", "蟆",
-    "怙", "玙", "庼", "筮", "慗", "层", "娓", "蚨", "糟", "璩", "隼", "锧", "疱", "铎", "祠", "绁",
-    "速", "湛", "蝮", "立", "媵", "禇", "撸", "禳", ">", "恋", "⑥", "鹡", "蓑", "樱", "奸", "蝣",
-    "埭", "聪", "慭", "睑", "肢", "焗", "骃", "毵", "潼", "塘", "烧", "劓", "栾", "牯", "〉", "毓",
-    "釉", "庞", "宕", "磚", "夺", "畜", "俏", "筼", "虱", "釆", "计", "陀", "诧", "臱", "牌", "固",
-    "鹉", "凰", "擞", "■", "㛃", "词", "店", "當", "许", "妣", "耋", "硼", "根", "焅", "砸", "霈",
-    "锃", "巳", "誉", "咕", "锊", "P", "云", "乞", "为", "姆", "桕", "丞", "鳤", "楷", "醒", "趔",
-    "怔", "砉", "潏", "肝", "拷", "也", "璐", "厘", "致", "昇", "绥", "抃", "佩", "斥", "⑶", "断",
-    "纵", "翁", "耄", "沭", "洄", "别", "吻", "渍", "r", "愦", "替", "骓", "攻", "旦", "来", "案",
-    "坫", "辞", "班", "锝", "拜", "掎", "穸", "笛", "痂", "笈", "鲶", "僻", "依", "碎", "蕤", "炕",
-    "寥", "拔", "髑", "慨", "眺", "⑿", "珞", "歧", "湲", "錦", "拮", "哭", "爰", "验", "寸", "%",
-    "口", "阔", "积", "篝", "狴", "殂", "痧", "绽", "搪", "祸", "缘", "遯", "祜", "种", "缓", "冒",
-    "庵", "窥", "颂", "哕", "谘", "蜗", "掖", "驮", "郎", "跂", "蝓", "贲", "馔", "琰", "大", "芊",
-    "粝", "啻", "蒉", "谒", "羔", "蘑", "菑", "布", "徂", "l", "炬", "蘩", "弭", "嵎", "補", "岑",
-    "佯", "棍", "没", "志", "裣", "咸", "权", "豺", "韦", "優", "紧", "嚚", "牴", "酚", "沤", "睥",
-    "葆", "乓", "划", "犒", "惚", "埕", "锴", "就", "滓", "汕", "冏", "缵", "囟", "旸", "麂", "接",
-    "餮", "瓤", "偕", "灏", "溻", "揶", "钯", "荸", "宪", "泠", "槠", "抺", "威", "燴", "井", "骣",
-    "耜", "磔", "懜", "淇", "瘘", "蹼", "觏", "圬", "纔", "恤", "铝", "薏", "鞫", "衉", "Ｅ", "球",
-    "姽", "毹", "嵘", "睦", "蛉", "伽", "橥", "痤", "卞", "警", "则", "芬", "磉", "悢", "逾", "吆",
-    "朿", "锒", "卧", "E", "焉", "纥", "髦", "鞯", "牞", "蹿", "琪", "洌", "幛", "淳", "菲", "涔",
-    "噗", "勺", "势", "哓", "毪", "刖", "雜", "浪", "懵", "棁", "秋", "◇", "玡", "埯", "谥", "芑",
-    "拌", "Ｚ", "埂", "寻", "虼", "豫", "蛄", "蜞", "椟", "〔", "哩", "絷", "咱", "郈", "湫", "抡",
-    "矮", "庖", "锰", "鎖", "葸", "棣", "砑", "题", "棋", "晕", "兠", "蓣", "貂", "裤", "昊", "扳",
-    "讴", "弩", "蚀", "尓", "暗", "莘", "黏", "寖", "蛔", "И", "痪", "饳", "妈", "Z", "秏", "矽",
-    "蘭", "艴", "菖", "寿", "撅", "秤", "颓", "廪", "醪", "耢", "螣", "椰", "麈", "听", "敞", "铠",
-    "润", "穑", "涟", "澹", "茨", "昨", "嚯", "钟", "護", "榇", "曦", "蜎", "呷", "邦", "瘵", "鲬",
-    "谖", "概", "九", "飞", "潇", "高", "锫", "钐", "徘", "诟", "缧", "剞", "【", "簋", "噍", "痋",
-    "关", "糁", "鹊", "潤", "邬", "赔", "序", "盤", "拙", "盾", "鸫", "迟", "氇", "塄", "庀", "噩",
-    "臃", "绔", "岔", "珠", "癀", "骡", "隶", "堋", "笆", "影", "昶", "飗", "α", "薄", "碜", "瞅",
-    "奠", "鲽", "牂", "裆", "刀", "遑", "隙", "沔", "訾", "旇", "s", "诮", "炱", "榚", "些", "縯",
-    "提", "甚", "埏", "峣", "意", "出", "郡", "摊", "哦", "轼", "嚒", "杈", "痢", "裥", "袖", "陔",
-    "溷", "裀", "迄", "枇", "筑", "貊", "鲃", "鸪", "Ｖ", "鞭", "粳", "菥", "盒", "毙", "风", "壽",
-    "胝", "章", "泄", "浜", "菇", "釘", "揖", "蝰", "榧", "掊", "瘿", "亍", "欤", "髻", "歃", "朱",
-    "撇", "铁", "兜", "喋", "趵", "瘼", "芤", "邳", "除", "滌", "么", "践", "矜", "曆", "炒", "咔",
-    "促", "姐", "酦", "货", "箔", "锡", "吝", "医", "汀", "精", "馘", "骶", "碉", "啷", "缉", "褪",
-    "阙", "滕", "蓐", "搏", "俅", "癌", "股", "圪", "孕", "鳕", "话", "鳳", "妩", "馬", "鞣", "璁",
-    "测", "螠", "圆", "颥", "卒", "钲", "䨱", "骢", "批", "陨", "婶", "熵", "鸣", "梁", "怅", "滑",
-    "驵", "帑", "鑽", "痈", "鲑", "犍", "茑", "糧", "谫", "咒", "绘", "塆", "耖", "铮", "殳", "角",
-    "齮", "挻", "捎", "2", "З", "廋", "泻", "犊", "儋", "椋", "聍", "鼬", "鳏", "收", "跞", "盥",
-    "信", "嘴", "眉", "黠", "纡", "Ω", "愆", "桹", "昽", "讠", "蚧", "龀", "鳓", "麓", "括", "蘊",
-    "鈇", "写", "焕", "哒", "忌", "咳", "蜮", "衄", "氰", "邘", "缇", "缮", "表", "续", "篇", "旄",
-    "闻", "門", "偃", "茳", "晗", "鋪", "讶", "衙", "闭", "管", "傒", "藔", "鳯", "萼", "泥", "汔",
-    "浆", "姚", "几", "含", "沆", "繇", "灾", "嗦", "饬", "系", "熠", "漶", "潭", "牲", "湘", "稁",
-    "宄", "亳", "杏", "豳", "監", "印", "旁", "辙", "播", "让", "乳", "镚", "腒", "衅", "铅", "殛",
-    "形", "颚", "茸", "绚", "铲", "贬", "溆", "寘", "傫", "解", "事", "旻", "差", "竣", "篃", "渥",
-    "嵩", "@", "鹀", "哈", "痍", "绪", "囔", "渗", "帐", "叙", "⑦", "跖", "貝", "盅", "碶", "哗",
-    "裼", "窝", "滙", "郸", "延", "炼", "芹", "簃", "谭", "莱", "佞", "猝", "褥", "陆", "王", "碴",
-    "粑", "怃", "倭", "貉", "硈", "裴", "溍", "坂", "勐", "踬", "敌", "改", "膠", "罘", "罚", "繁",
-    "幢", "煸", "扩", "黑", "氮", "孓", "饔", "姹", "蠖", "倬", "慝", "沩", "嚏", "雌", "琏", "莹",
-    "觊", "颟", "坤", "筅", "桉", "肈", "幻", "”", "烽", "蓅", "噌", "桎", "矻", "指", "瀚", "羊",
-    "邒", "筒", "匣", "旨", "葑", "团", "偻", "镛", "忉", "帱", "往", "刎", "害", "母", "串", "酢",
-    "舟", "栝", "娣", "⒉", "装", "笊", "篑", "垄", "镳", "牝", "煲", "氵", "遂", "骜", "钜", "屡",
-    "宇", "z", "惨", "潆", "鹃", "豌", "浞", "幄", "麺", "暽", "弁", "菤", "酲", "虑", "样", "访",
-    "睁", "黼", "悞", "禧", "涠", "舂", "駕", "屈", "噀", "邽", "炅", "欸", "隆", "描", "绦", "派",
-    "溲", "颙", "忐", "鹺", "只", "诈", "腭", "亵", "这", "炀", "趣", "喙", "跨", "場", "全", "搢",
-    "舨", "踢", "巿", "糸", "乐", "澈", "羰", "匯", "D", "濒", "莛", "癯", "娴", "姒", "祥", "渖",
-    "庭", "渑", "挹", "狉", "萍", "鹚", "焻", "自", "舸", "岫", "飑", "戍", "淙", "愠", "导", "賣",
-    "进", "赳", "鱿", "硂", "僤", "O", "拟", "钨", "笳", "汴", "挤", "稞", "柏", "阐", "弥", "艽",
-    "爻", "魭", "俶", "聂", "鲎", "齌", "菀", "僜", "煅", "满", "兀", "辔", "舊", "胥", "卢", "额",
-    "膝", "嵬", "昼", "唏", "璨", "苷", "舫", "倨", "耵", "碳", "鄺", "橘", "裔", "裸", "宋", "窋",
-    "悚", "昱", "错", "咣", "檄", "硁", "莆", "肃", "曼", "垍", "绵", "役", "內", "館", "规", "犋",
-    "毅", "锨", "瓘", "丘", "疝", "杀", "袗", "屏", "阎", "衹", "嚣", "冼", "铪", "熏", "ｔ", "储",
-    "傧", "髁", "邙", "窴", "泾", "止", "晾", "涵", "庸", "庠", "靿", "睚", "减", "衲", "思", "簖",
-    "耘", "之", "斑", "扬", "袼", "缗", "菏", "滍", "研", "垚", "艘", "黇", "蜘", "汨", "侦", "疃",
-    "再", "沂", "榛", "励", "嚼", "拴", "钴", "珑", "撕", "鸯", "共", "醐", "揿", "醴", "凭", "菹",
-    "鹬", "捷", "掼", "芝", "缪", "咏", "挺", "蒗", "疖", "伸", "索", "黉", "髋", "習", "姜", "蔺",
-    "扶", "忽", "锉", "戚", "珲", "摸", "黛", "跌", "螈", "冚", "洱", "鼯", "庡", "痒", "哆", "品",
-    "歆", "登", "呔", "追", "鉫", "碾", "祚", "总", "帚", "薹", "趴", "容", "滘", "❋", "迷", "拾",
-    "炯", "析", "佼", "嗾", "-", "针", "滩", "禅", "讯", "織", "饱", "哟", "蹓", "侧", "蹯", "踽",
-    "趙", "漫", "卷", "朴", "卺", "狈", "绉", "浉", "玳", "榕", "畛", "蛾", "帔", "欠", "劙", "矾",
-    "進", "蚺", "骨", "捂", "疾", "璱", "圠", "坚", "弯", "秩", "逧", "烟", "佛", "嫱", "鳢", "裹",
-    "耥", "潮", "巽", "嶝", "廷", "鳆", "五", "西", "壤", "慌", "鴈", "爡", "‧", "儆", "萘", "愁",
-    "昔", "咩", "侃", "炫", "迳", "构", "缶", "皴", "捱", "菜", "秀", "粢", "畵", "辊", "伟", "榄",
-    "嗪", "已", "聖", "枕", "恢", "厍", "涅", "鄢", "饨", "鲨", "谪", "礤", "搁", "疔", "雎", "羿",
-    "犁", "硐", "盹", "若", "操", "昀", "亩", "凑", "睃", "贰", "穷", "荆", "葹", "嘃", "廓", "憔",
-    "制", "胧", "染", "麿", "镴", "砀", "组", "茎", "畫", "哑", "坨", "罱", "刳", "嘲", "瞰", "缏",
-    "侩", "谝", "谮", "还", "惋", "佈", "谏", "酗", "怩", "ｉ", "秽", "抻", "蟪", "狠", "辑", "羟",
-    "娉", "肯", "毕", "枸", "啮", "鳙", "要", "娑", "蜈", "忍", "侗", "Р", "凛", "巉", "拽", "该",
-    "俎", "惠", "薅", "猬", "忲", "●", "哏", "柔", "沉", "磁", "脾", "龃", "鲩", "钞", "厂", "鱽",
-    "濞", "群", "唱", "攝", "刃", "蹻", "剕", "朊", "啜", "语", "蜻", "遠", "呖", "都", "嫄", "戢",
-    "丫", "酞", "修", "户", "膏", "侈", "浣", "捍", "馒", "谱", "點", "滏", "跟", "鹁", "书", "愉",
-    "扭", "票", "耒", "燁", "魚", "拨", "鞑", "匆", "最", "憨", "晨", "癍", "邻", "醅", "骘", "骇",
-    "蔫", "。", "珪", "安", "歉", "邑", "淦", "胱", "漏", "咂", "叮", "瘛", "钺", "聘", "玎", "荻",
-    "涰", "淏", "烈", "囚", "㙟", "矫", "嗵", "澶", "磛", "肖", "蹑", "支", "哪", "久", "嗅", "稍",
-    "抠", "具", "裨", "韨", "虿", "漪", "外", "埸", "ä", "葡", "鲇", "廛", "蝥", "广", "兒", "涧",
-    "阌", "谊", "嵴", "拂", "悸", "哳", "4", "佘", "篁", "基", "赪", "掬", "演", "谋", "酣", "植",
-    "波", "恅", "杼", "胀", "纯", "缩", "盆", "芷", "绺", "施", "礞", "莳", "稻", "狐", "馌", "椑",
-    "恃", "镩", "骼", "孱", "硭", "曈", "焼", "秕", "觥", "茛", "射", "肣", "蟞", "′", "彰", "罔",
-    "麻", "逋", "谓", "革", "藤", "堆", "忿", "茂", "屦", "炟", "碹", "啤", "垤", "文", "佐", "叼",
-    "蹭", "踉", "眍", "龟", "黾", "崋", "氘", "尼", "瞢", "悦", "罢", "瑀", "睇", "贴", "涿", "拍",
-    "庥", "粪", "陳", "闌", "蓍", "元", "磊", "忻", "葴", "买", "诺", "㐂", "沦", "苏", "岂", "浊",
-    "脍", "誠", "嚬", "蔬", "黄", "銘", "钩", "劃", "嫘", "嵖", "笄", "鸰", "熳", "堑", "河", "遮",
-    "尝", "灭", "愬", "娅", "鸽", "锢", "狂", "弓", "侨", "滚", "释", "晏", "徵", "苡", "佚", "切",
-    "惰", "降", "槎", "脑", "磺", "谟", "裉", "弘", "醺", "眇", "顒", "兰", "忧", "础", "－", "茫",
-    "袒", "喤", "慶", "話", "跽", "堵", "能", "武", "壳", "汤", "橐", "圧", "嗡", "簌", "斧", "林",
-    "楊", "净", "嫌", "猛", "炸", "冠", "弦", "洵", "°", "創", "葜", "苁", "顏", "薛", "平", "舜",
-    "崧", "撼", "哮", "窣", "殖", "象", "飒", "瞥", "玥", "斝", "胭", "葱", "擎", "踰", "盛", "稇",
-    "斫", "滴", "硫", "胬", "唁", "拐", "殍", "冱", "陪", "鋈", "烦", "辋", "妇", "轸", "潴", "荤",
-    "碡", "徭", "膜", "秧", "淀", "缤", "琶", "^", "肋", "违", "轭", "$", "箕", "墩", "罗", "摆",
-    "槛", "笠", "塃", "畚", "嘛", "秸", "≥", "剪", "羅", "帜", "庇", "骙", "酾", "康", "陬", "蝶",
-    "闇", "圏", "琴", "鬈", "苶", "黡", "瓠", "酋", "琎", "奏", "科", "搒", "億", "翡", "熊", "性",
-    "痞", "篾", "邴", "篥", "饵", "揸", "狙", "鳂", "否", "煺", "畦", "凖", "鹧", "宵", "庛", "鬯",
-    "袂", "昴", "徽", "奧", "鸥", "锸", "庚", "彪", "槿", "趱", "铡", "讹", "莎", "約", "璞", "┌",
-    "铫", "鸈", "瞀", "鼙", "廙", "估", "屺", "锘", "渝", "汶", "囵", "跻", "酏", "扎", "孜", "腔",
-    "吗", "块", "潽", "“", "焚", "倘", "靥", "默", "笺", "遍", "剅", "桠", "极", "夷", "眸", "泳",
-    "谢", "懦", ")", "泜", "物", "开", "淮", "邵", "卟", "徜", "轧", "苍", "酶", "捌", "耨", "膄",
-    "琦", "痦", "涪", "舰", "迥", "涘", "季", "悫", "初", "贩", "碍", "沼", "腮", "勻", "恣", "他",
-    "院", "７", "蜒", "叉", "昉", "瞽", "疢", "蓼", "相", "臜", "ü", "秘", "蚌", "摉", "莞", "番",
-    "盼", "冀", "缠", "学", "馗", "搞", "苛", "羹", "潟", "嬴", "杶", "妖", "↓", "銭", "斄", "猢",
-    "搂", "芦", "埶", "癜", "铴", "從", "留", "硎", "簇", "臬", "侮", "鳎", "頉", "所", "邝", "哂",
-    "弛", "垂", "暑", "仿", "療", "轾", "熜", "悃", "勘", "鋫", "泞", "閃", "＋", "士", "虐", "盡",
-    "溯", "劣", "1", "锇", "蜇", "碛", "婴", "８", "捶", "藻", "嫰", "跹", "芋", "淠", "疯", "Й",
-    "诨", "醢", "痨", "栌", "摧", "噤", "灸", "馉", "蒇", "洲", "孪", "孵", "亿", "枝", "涕", "孟",
-    "缎", "攒", "湧", "褛", "铺", "廣", "诃", "蕈", "维", "珣", "逖", "瓯", "苘", "榫", "寰", "供",
-    "铽", "暲", "义", "兖", "喆", "祇", "蝽", "膛", "髭", "尹", "乒", "柈", "抓", "镋", "脏", "乩",
-    "挎", "酬", "靳", "躔", "霞", "離", "桌", "ｏ", "　", "滨", "懷", "萭", "望", "仆", "怂", "麸",
-    "槁", "洨", "俬", "孚", "匕", "栩", "t", "被", "邗", "纺", "狄", "国", "苋", "眢", "秣", "畤",
-    "∵", "创", "嵫", "胖", "垮", "镁", "壑", "Ａ", "趾", "荟", "汝", "嗥", "驰", "向", "磴", "讪",
-    "溅", "箍", "驾", "讙", "蛐", "眨", "醯", "筣", "鼫", "叠", "沥", "恕", "埽", "甩", "煜", "端",
-    "蜃", "帏", "丙", "荏", "圻", "夥", "暖", "抛", "岸", "头", "恝", "船", "鞲", "の", "七", "怪",
-    "诅", "茧", "北", "搓", "皖", "綉", "坪", "簪", "贵", "阍", "刑", "偶", "雷", "體", "髡", "岚",
-    "剂", "妾", "茬", "鹣", "鲳", "蜊", "假", "认", "嗍", "傈", "焘", "驷", "賀", "氆", "炝", "聋",
-    "嘡", "悺", "﹐", "僭", "证", "侉", "屐", "吃", "十", "〕", "瀦", "骤", "俳", "舁", "驸", "竿",
-    "坍", "|", "普", "軒", "瘃", "现", "醾", "铈", "榭", "蓁", "唷", "睫", "绰", "叶", "頤", "锤",
-    "浸", "杂", "汹", "璟", "蓖", "迦", "截", "錎", "年", "達", "劁", "码", "乃", "捯", "髪", "埚",
-    "呯", "頭", "镈", "汆", "β", "舷", "菂", "纸", "造", "晁", "卦", "蛎", "枉", "脯", "诛", "生",
-    "令", "桄", "恶", "鬟", "怯", "囡", "充", "省", "控", "螂", "笕", "邺", "扅", "斲", "骏", "鹋",
-    "队", "套", "镢", "牖", "Ｉ", "援", "封", "泚", "艿", "沮", "铙", "揳", "抄", "帨", "近", "唾",
-    "珥", "销", "非", "啱", "驼", "瓅", "遢", "筱", "奥", "善", "氩", "盉", "帖", "杜", "鈉", "蕉",
-    "鳄", "江", "鄂", "矩", "排", "诠", "准", "菼", "刻", "足", "峦", "灰", "僳", "蝢", "孫", "據",
-    "璆", "孳", "徹", "妆", "坭", "渌", "喏", "钦", "朔", "机", "徼", "掾", "其", "愧", "刹", "迹",
-    "惹", "闿", "鹘", "窗", "贳", "箓", "奡", "鸮", "翟", "滠", "辄", "夸", "袱", "】", "阿", "摁",
-    "觽", "垠", "棼", "杉", "甏", "旖", "阜", "猩", "埤", "醛", "馥", "哫", "囝", "晖", "骸", "椭",
-    "军", "哺", "束", "犄", "遛", "涡", "罄", "赇", "幌", "曙", "契", "饹", "圳", "仝", "爝", "示",
-    "敝", "漓", "马", "铐", "掳", "鲱", "蟀", "阵", "栒", "b", "轮", "俯", "嬉", "漩", "佟", "捆",
-    "迩", "液", "炊", "抒", "迭", "换", "菘", "请", "柩", "腓", "妥", "怖", "穆", "骝", "胛", "幺",
-    "燥", "棹", "孖", "洁", "边", "彥", "睨", "桁", "淆", "醤", "奘", "伙", "刘", "o", "過", "鏖",
-    "怄", "镗", "蘖", "卩", "龈", "枣", "联", "圊", "贫", "郭", "梗", "赡", "輸", "拼", "冰", "Ë",
-    "姓", "办", "拉", "富", "对", "肩", "缬", "逻", "瞠", "な", "荚", "宝", "緃", "峋", "鞁", "坉",
-    "译", "银", "猹", "榞", "借", "楠", "殽", "片", "谸", "雩", "琨", "颏", "晷", "；", "莠", "楽",
-    "嬗", "啉", "爍", "視", "岽", "荙", "冗", "『", "鯭", "觌", "喹", "药", "腈", "蹰", "–", "泷",
-    "榷", "卮", "缥", "各", "瘙", "凡", "悍", "劢", "６", "镰", "四", "摈", "驹", "陵", "沟", "麟",
-    "骊", "椀", "★", "滦", "菌", "彊", "疽", "辫", "璎", "郁", "资", "粼", "萄", "鳁", "鸤", "漁",
-    "痃", "绌", "蒴", "尊", "哙", "守", "坠", "胯", "泽", "持", "隅", "耧", "逛", "鄯", "瑜", "召",
-    "眷", "逄", "垦", "窄", "板", "赵", "源", "绹", "總", "适", "轰", "钛", "停", "唝", "燎", "栽",
-    "燃", "搬", "夬", "掐", "簦", "旃", "更", "蹊", "熙", "埃", "偲", "蒙", "舍", "杭", "朋", "霍",
-    "谙", "蓥", "赫", "鲏", "純", "缃", "蓓", "è", "萧", "锆", "诉", "恚", "u", "旷", "⑵", "褲",
-    "鲟", "互", "峂", "绝", "陒", "议", "脽", "态", "岙", "僮", "突", "濯", "犹", "阆", "溧", "宜",
-    "霣", "呆", "鄉", "觖", "濠", "耐", "橱", "敛", "踮", "怦", "鸠", "授", "滫", "叭", "镤", "仵",
-    "圖", "救", "变", "寕", "洋", "泃", "晥", "檎", "缦", "鏠", "悬", "彧", "韓", "籽", "磷", "歙",
-    "÷", "杨", "闯", "罴", "鼠", "陋", "瀣", "揾", "匪", "晚", "丧", "彷", "梏", "诽", "尻", "蹢",
-    "拊", "氕", "芳", "凸", "懊", "型", "媄", "憝", "骆", "纭", "刺", "悼", "藍", "欧", "葳", "艺",
-    "±", "紗", "钡", "钭", "首", "栈", "闩", "勃", "喑", "栉", "栟", "焊", "哀", "竖", "肘", "悲",
-    "鳜", "迨", "怿", "裘", "綮", "蕹", "冕", "无", "鼢", "戽", "鲣", "蜍", "插", "扯", "俣", "袪",
-    "级", "理", "茵", "矣", "障", "禊", "惯", "䄂", "宏", "韫", "栗", "编", "座", "跫", "、", "捞",
-    "谠", "車", "袆", "舖", "暝", "9", "混", "蓊", "韂", "拳", "j", "伶", "啦", "0", "诰", "缑",
-    "郴", "蕲", "谛", "皋", "兢", "件", "隈", "溊", "窈", "檀", "洪", "揭", "油", "颐", "剎", "侠",
-    "或", "腻", "迪", "舔", "牺", "倦", "夭", "蠋", "鹑", "倜", "喱", "豨", "谨", "翥", "馄", "蔚",
-    "类", "鳍", "泯", "岷", "張", "紅", "情", "婪", "睡", "诒", "旋", "舱", "列", "求", "姥", "翯",
-    "锲", "適", "蒡", "㭗", "韵", "撄", "律", "湊", "鲪", "碓", "馓", "汛", "尤", "℃", "泫", "竹",
-    "溃", "诵", "乡", "楹", "肷", "戋", "獭", "撴", "晤", "嘎", "裢", "轱", "歘", "鼱", "舀", "潵",
-    "腩", "挽", "至", "骞", "佻", "蕊", "轫", "岘", "斌", "《", "雇", "枘", "贶", "履", "亭", "绍",
-    "空", "泡", "擒", "蠓", "桼", "舢", "踵", "柢", "躺", "罥", "蔁", "幪", "蒯", "≤", "齉", "況",
-    "锕", "徇", "儿", "帽", "谕", "抽", "獍", "揪", "膳", "酴", "姊", "作", "弍", "郦", "翮", "髌",
-    "勝", "孔", "颁", "泩", "衾", "翊", "噉", "∩", "先", "的", "財", "绢", "崇", "饲", "賓", "绾",
-    "陡", "砾", "纷", "酱", "窎", "浕", "琫", "补", "嗽", "呛", "蚓", "伪", "艨", "搡", "粗", "阱",
-    "商", "郫", "喵", "缯", "炻", "州", "珽", "芒", "脶", "蕖", "拘", "\\", "蕃", "伏", "蛏", "璘",
-    "缡", "伞", "吸", "煞", "嗫", "餪", "━", "8", "漠", "衮", "峭", "筌", "阀", "乇", "杞", "镕",
-    "鹈", "觚", "疋", "猥", "肚", "阃", "地", "鸾", "玚", "镀", "诬", "凇", "惬", "劝", "堠", "睪",
-    "瞪", "涓", "專", "琢", "蠹", "裂", "芎", "轲", "谯", "畋", "长", "攘", "缁", "吵", "汈", "稂",
-    "咤", "甸", "瀑", "讣", "翠", "綦", "笑", "掺", "宫", "唆", "旗", "帘", "遵", "唪", "连", "必",
-    "姬", "枵", "渫", "神", "凱", "虞", "闵", "爭", "迢", "飓", "桴", "冑", "招", "鬶", "赊", "朦",
-    "薜", "逗", "铩", "焖", "狱", "昕", "犀", "捏", "侓", "爆", "箧", "怠", "缐", "苄", "踯", "暹",
-    "絮", "裁", "嘶", "诙", "攥", "¥", "痫", "摹", "朕", "薰", "闼", "樨", "需", "蕾", "扔", "服",
-    "蚩", "漳", "瞻", "陟", "诐", "爽", "耩", "坏", "寶", "瞧", "蘅", "爺", "仇", "瞈", "揉", "隰",
-    "阏", "皤", "鉏", "郇", "啵", "僧", "渐", "骠", "厾", "谦", "榱", "蓮", "乂", "伷", "觐", "遫",
-    "畀", "污", "眭", "蝗", "亢", "腾", "遇", "欺", "狸", "籼", "锯", "赚", "槔", "域", "老", "掩",
-    "锎", "邪", "沬", "袯", "映", "肤", "舒", "箦", "涎", "缔", "孃", "萎", "推", "峧", "嫒", "嵐",
-    "怒", "闾", "阅", "扢", "絡", "剩", "腢", "晌", "铟", "铚", "燔", "鄜", "墠", "擤", "芸", "鼩",
-    "食", "朰", "啥", "贿", "券", "垩", "典", "逭", "成", "帧", "侯", "妒", "顷", "晩", "浦", "汳",
-    "㾄", "坊", "②", "蝼", "倩", "挦", "騠", "蓂", "兹", "渤", "茴", "矶", "煤", "伛", "逃", "堞",
-    "漭", "闽", "埠", "妗", "烙", "傀", "滢", "堂", "骅", "崮", "兿", "螟", "榴", "蒈", "蟛", "员",
-    "嗟", "毗", "鹭", "毖", "籍", "傣", "偈", "澴", "耆", "黔", "薷", "俘", "箸", "蒺", "记", "铣",
-    "拥", "瑰", "抉", "史", "急", "什", "川", "曪", "剖", "罪", "扥", "鱼", "羌", "复", "昵", "朓",
-    "氪", "矅", "ò", "摒", "催", "{", "懂", "莉", "屃", "症", "遒", "璋", "鄒", "玄", "氦", "舳",
-    "納", "锷", "鲥", "羸", "怵", "嬖", "飚", "倔", "辨", "篌", "娄", "洎", "弟", "悝", "婥", "待",
-    "梃", "棕", "撞", "肊", "婺", "桥", "慕", "5", "尖", "监", "囊", "腎", "瘁", "⑴", "伢", "邈",
-    "炷", "籴", "月", "终", "龠", "蹇", "柯", "噜", "多", "玮", "踹", "泣", "圮", "辈", "鄙", "赆",
-    "蠃", "珈", "鬃", "识", "甑", "藓", "宥", "韧", "隗", "跚", "峁", "晒", "戳", "環", "葓", "筐",
-    "瑚", "稆", "巩", "葵", "网", "铨", "仡", "呶", "休", "村", "傩", "浡", "祝", "幅", "晳", "岿",
-    "[", "尃", "响", "咽", "讵", "淹", "糠", "驴", "烹", "甪", "焯", "整", "瘆", "茀", "瑞", "嘁",
-    "赅", "芼", "震", "怼", "膑", "蒟", "钊", "缷", "獨", "俄", "噪", "亏", "芍", "鲲", "黟", "酃",
-    "铻", "→", "鲡", "仑", "有", "扌", "治", "聆", "腙", "冬", "剥", "诓", "祃", "棘", "孥", "濆",
-    "鞋", "赣", "卜", "谆", "投", "琊", "塍", "洗", "津", "肴", "奕", "掴", "螗", "胸", "惦", "赠",
-    "耶", "蛸", "曷", "名", "栏", "崖", "始", "孤", "篆", "椴", "模", "儙", "袁", "颌", "卤", "寂",
-    "呸", "蓺", "噶", "钔", "粹", "潸", "郾", "娡", "啖", "芜", "䧳", "缟", "鞅", "哇", "嚄", "熛",
-    "私", "住", "鉄", "Ｕ", "失", "蜾", "㺃", "莨", "擂", "沕", "铄", "耰", "炔", "藐", "锈", "苇",
-    "芯", "檇", "龍", "珀", "筻", "褚", "缀", "棬", "仲", "杧", "铛", "挖", "缸", "嵇", "剟", "觋",
-    "鬼", "天", "鹪", "豭", "腴", "驿", "粵", "馈", "迸", "夕", "蜚", "禺", "迎", "狞", "嶽", "躯",
-    "暅", "蹾", "甾", "柆", "e", "欃", "無", "羽", "蛛", "僖", "碲", "遴", "厝", "滂", "眦", "鹅",
-    "咖", "潜", "媽", "啴", "寨", "琬", "判", "塌", "与", "剧", "掮", "读", "庆", "愚", "氲", "梾",
-    "稗", "很", "饯", "淘", "灞", "A", "！", "飾", "醚", "壁", "芮", "蔼", "知", "郯", "瞒", "术",
-    "法", "匠", "┅", "尕", "脂", "琅", "孝", "埙", "濟", "袷", "方", "芾", "驽", "线", "咡", "睿",
-    "滥", "鞨", "駹", "馎", "蛲", "申", "涣", "牗", "跄", "鲐", "泐", "铍", "肓", "é", "司", "辣",
-    "慑", "崞", "魏", "髯", "涯", "滃", "薪", "~", "鐵", "町", "芥", "匿", "癞", "嘚", "□", "啪",
-    "俇", "橛", "巅", "鄞", "痱", "回", "着", "魂", "瘀", "埒", "措", "蚪", "锁", "巴", "芏", "另",
-    "忏", "洇", "⑾", "箱", "劳", "榈", "Ｂ", "毒", "泰", "巍", "寇", "斛", "赖", "酒", "密", "左",
-    "莜", "茯", "脆", "饴", "洳", "舉", "酪", "鼒", "職", "桫", "螋", "砌", "戬", "前", "豁", "膲",
-    "鲯", "憋", "珺", "蹂", "亥", "冷", "蜔", "俦", "碇", "慰", "扒", "亮", "弒", "胫", "肠", "弇",
-    "拣", "丸", "涫", "勉", "帝", "玠", "应", "砜", "鬓", "泮", "鳖", "峇", "辗", "吴", "僡", "玊",
-    "钘", "惘", "奁", "翰", "爨", "罽", "莶", "纻", "菪", "興", "钧", "鄹", "缌", "嘉", "玲", "蛋",
-    "倚", "镬", "彳", "氯", "谐", "疗", "魉", "扈", "鏡", "涤", "佑", "勾", "饽", "R", "局", "醉",
-    "燫", "驱", "崽", "榮", "锄", "脎", "枓", "敲", "俑", "镶", "当", "逦", "寅", "捽", "宗", "茼",
-    "稿", "躇", "晋", "东", "渭", "闶", "潦", "慊", "蔟", "疑", "蜡", "嬛", "荒", "扪", "拈", "受",
-    "沱", "螽", "嗣", "瘰", "颖", "百", "楱", "崛", "窿", "蓬", "禨", "螵", "捕", "鳀", "霰", "瘸",
-    "圄", "陂", "敢", "矗", "炉", "篪", "注", "凉", "摔", "衿", "g", "遘", "鳟", "腚", "泱", "荣",
-    "池", "杰", "阻", "里", "峥", "葺", "翔", "柁", "飐", "载", "刨", "瘊", "溏", "嚭", "穝", "站",
-    "笞", "厌", "门", "汞", "雠", "靖", "怨", "葙", "厅", "凉", "荨", "底", "坼", "海", "轩", "秃",
-    "醭", "莒", "坎", "驺", "瀛", "嵝", "溥", "嚋", "寤", "k", "築", "耔", "跎", "小", "弊", "疭",
-    "√", "洞", "利", "次", "迕", "鸿", "刁", "媳", "酹", "磋", "勒", "伎", "浔", "斤", "珊", "鸳",
-    "颊", "塗", "刈", "龛", "縻", "楚", "脓", "南", "伧", "桨", "尸", "退", "亊", "侔", "蓟", "削",
-    "诲", "瑯", "挝", "绩", "嶂", "耷", "太", "著", "讥", "烂", "哥", "菔", "唳", "筫", "嚷", "斯",
-    "蔡", "碧", "慵", "搐", "琵", "逶", "菡", "篙", "驭", "莲", "毫", "犯", "坌", "啕", "疴", "傻",
-    "牧", "肺", "既", "悟", "讳", "练", "钠", "胨", "Ｇ", "迁", "帙", "嫦", "哿", "柝", "迤", "庑",
-    "踟", "粤", "乜", "悭", "霁", "在", "虔", "蓦", "暾", "鲅", "様", "徐", "诫", "料", "蒂", "揎",
-    "溜", "崎", "玦", "隐", "绳", "幼", "钼", "腧", "槌", "镅", "锿", "蹁", "猱", "狁", "蹈", "S",
-    "傢", "}", "珉", "浴", "龚", "吇", "肆", "⑩", "臌", "鏢", "―", "定", "哧", "螳", "爹", "摞",
-    "榍", "苩", "丑", "犴", "璀", "偎", "燹", "洯", "啁", "鱾", "計", "镂", "字", "琉", "壴", "霖",
-    "蟜", "做", "松", "扣", "徊", "叱", "桑", "≠", "寓", "崔", "茌", "`", "殓", "猖", "慎", "坻",
-    "皮", "喔", "耕", "拧", "织", "桂", "垝", "约", "飯", "皿", "移", "微", "桅", "钬", "赑", "宓",
-    "颉", "掂", "希", "沁", "翳", "澝", "崭", "聃", "蔸", "磕", "崚", "奉", "槍", "颠", "任", "飕",
-    "蛴", "歼", "蛭", "￥", "鉵", "刷", "泓", "頫", "诤", "阈", "诖", "漂", "冯", "柠", "本", "狼",
-    "芗", "丹", "柒", "饫", "衫", "摩", "桃", "啼", "譄", "醌", "帡", "少", "螃", "祷", "幸", "喘",
-    "噼", "甲", "刭", "泊", "那", "舄", "畲", "遊", "記", "绠", "蝈", "忒", "耙", "鹌", "搽", "绕",
-    "镠", "凼", "扫", "鸵", "匡", "缈", "铼", "挢", "踊", "噭", "苴", "貌", "啾", "蒜", "茱", "⒅",
-    "邛", "棉", "怀", "筵", "笾", "邢", "濱", "蒲", "嫠", "儡", "益", "萊", "歩", "日", "婢", "楔",
-    "嚎", "髂", "禄", "巫", "懿", "砍", "惆", "袭", "～", "欲", "鳅", "書", "撷", "库", "绻", "⒁",
-    "蒹", "辜", "屉", "奂", "龇", "诡", "醑", "儀", "讷", "聒", "盖", "浅", "長", "蛑", "念", "辒",
-    "蚊", "勖", "揣", "岬", "糅", "刊", "蛆", "硌", "柜", "惗", "激", "颈", "閣", "唬", "捧", "蒽",
-    "爸", "侏", "氧", "蛤", "牡", "瘦", "铉", "撒", "屿", "谔", "钀", "瓦", "恂", "钻", "声", "馨",
-    "贷", "藁", "闰", "人", "瀹", "间", "炜", "煨", "畴", "鞡", "悔", "量", "原", "屁", "杯", "趋",
-    "愎", "瓶", "烘", "韬", "姮", "勰", "眛", "饮", "瘐", "决", "版", "際", "俸", "此", "辽", "愫",
-    "爷", "隘", "螯", "腕", "闟", "厣", "齁", "矿", "女", "扇", "郄", "诚", "焜", "<", "檵", "赙",
-    "郢", "酌", "恳", "侵", "劍", "徳", "脔", "岍", "倍", "隹", "镲", "僬", "○", "啭", "坳", "谰",
-    "璪", "昆", "⒆", "惧", "虺", "篡", "锐", "邮", "嵊", "恹", "茄", "黯", "對", "魑", "禀", "添",
-    "汎", "廉", "戈", "∈", "陈", "肟", "螅", "缰", "啃", "瘾", "辁", "茜", "曝", "Ｃ", "餍", "蜴",
-    "咦", "臺", "扺", "隽", "崴", "遐", "胃", "铕", "楗", "墀", "魈", "淑", "邰", "飧", "勤", "糨",
-    "你", "铂", "蛰", "锻", "塈", "予", "怏", "鷪", "暂", "诞", "會", "鼐", "辏", "矬", "筜", "蟒",
-    "柫", "靴", "垒", "樵", "狨", "魋", "查", "掸", "＇", "每", "浮", "绞", "疍", "於", "饦", "辐",
-    "墟", "窳", "养", "墦", "鲔", "黝", "迈", "砟", "伥", "藜", "墅", "喬", "卬", "赶", "眼", "龌",
-    "主", "嗞", "転", "襕", "，", "杲", "儒", "政", "蹩", "付", "Я", "厢", "点", "如", "茗", "奍",
-    "嗌", "蕴", "问", "膚", "硒", "佃", "居", "赍", "鯨", "槜", "衡", "廐", "鲾", "慈", "莫", "篮",
-    "铱", "园", "范", "靠", "叻", "绸", "众", "抑", "浃", "極", "辀", "千", "蔯", "ｒ", "碱", "應",
-    "夼", "祺", "绣", "澉", "愛", "譬", "刚", "扼", "檫", "时", "荼", "部", "耽", "稀", "琳", "鎮",
-    "鳃", "疌", "挠", "京", "筴", "胳", "橡", "裱", "栊", "蛩", "流", "猄", "葭", "邯", "伤", "秦",
-    "孺", "郓", "筏", "茉", "谈", "黥", "鳇", "魅", "秫", "敫", "肄", "俪", "尺", "度", "廊", "昏",
-    "怎", "钎", "愤", "腿", "Λ", "碘", "囤", "婆", "煋", "桀", "阪", "焒", "砘", "澡", "阒", "臼",
-    "茭", "纛", "唻", "胗", "歌", "吱", "铸", "種", "礳", "舾", "砰", "即", "慥", "缆", "躞", "卖",
-    "眆", "搎", "鲁", "稠", "惝", "毶", "耎", "挡", "异", "咨", "歔", "垫", "素", "鳊", "打", "镏",
-    "耗", "锭", "剀", "讓", "鍋", "歇", "拗", "齿", "聊", "惕", "塔", "况", "汐", "采", "灣", "菁",
-    "峰", "呋", "阴", "衽", "鹏", "挂", "鲞", "第", "浑", "秭", "铯", "胎", "滔", "雕", "渔", "骍",
-    "鴻", "眚", "易", "枫", "麥", "罹", "劈", "苣", "蹄", "是", "锛", "饼", "铳", "器", "秒", "開",
-    "舛", "纳", "庒", "饰", "发", "虎", "乏", "届", "浯", "目", "霪", "氢", "华", "澧", "陴", "誊",
-    "欷", "瀌", "梵", "嗖", "惩", "妞", "螓", "溇", "嬃", "笼", "嬢", "獠", "辩", "骛", "挪", "礅",
-    "銎", "趺", "狍", "阑", "擦", "咙", "+", "鼻", "戆", "宿", "霏", "牁", "耦", "枳", "鲸", "阽",
-    "峯", "瘭", "祁", "镜", "绮", "道", "摭", "﹒", "耠", "限", "轻", "莽", "帛", "唯", "墒", "Ｗ",
-    "荦", "矸", "崃", "洫", "杳", "舌", "涌", "朵", "苑", "瓻", "辂", "凬", "畺", "莩", "奔", "荐",
-    "撵", "嶲", "憧", "沖", "邱", "瞑", "瘢", "腺", "嵯", "诗", "經", "胚", "谎", "央", "稼", "３",
-    "冈", "痣", "故", "皎", "珅", "踅", "硇", "通", "鹱", "草", "粕", "而", "梨", "萜", "搛", "鸺",
-    "托", "倡", "匾", "骷", "桐", "阇", "璃", "抖", "显", "镝", "I", "锣", "辆", "茝", "虫", "谀",
-    "萤", "帼", "滈", "鮼", "乔", "会", "产", "縢", "纱", "跆", "铤", "圯", "旚", "墓", "棰", "苜",
-    "聩", "蚍", "筢", "殇", "槚", "阂", "豊", "翚", "犰", "疚", "浠", "礴", "界", "蛙", "吁", "忪",
-    "讫", "辇", "汧", "澥", "纮", "糺", "拭", "镉", "鄠", "讼", "鎏", "逐", "烝", "扊", "筹", "仪",
-    "段", "杵", "卲", "钤", "曹", "潘", "浏", "陑", "答", "期", "橋", "位", "梅", "榻", "芨", "汭",
-    "贇", "杅", "伋", "馁", "皈", "躁", "鲦", "良", "幤", "曺", "榖", "便", "氍", "窍", "槃", "設",
-    "襦", "菓", "猡", "槲", "觱", "邠", "岩", "骈", "鬲", "睽", "残", "碗", "凯", "览", "伾", "吣",
-    "峡", "箨", "忡", "岁", "醵", "霸", "焐", "谁", "疟", "樾", "竺", "果", "阝", "花", "•", "耱",
-    "嗜", "骰", "萋", "甄", "侍", "厦", "辌", "给", "健", "垧", "娩", "校", "又", "临", "揽", "犟",
-    "蠢", "包", "檬", "穹", "燮", "札", "翏", "领", "深", "诶", "窒", "鸷", "统", "肾", "澍", "畹",
-    "职", "蝟", "眬", "旎", "朘", "眠", "篚", "贼", "貴", "＃", "茆", "磬", "①", "祀", "坞", "盯",
-    "苹", "痼", "耪", "狺", "妓", "湨", "動", "渚", "售", "鸶", "霄", "某", "×", "铀", "桓", "脲",
-    "心", "节", "［", "蜩", "剐", "猜", "盂", "穿", "洮", "坜", "掷", "坐", "烜", "冶", "袍", "穴",
-    "偷", "槽", "面", "媛", "撙", "孩", "询", "苈", "橙", "恙", "吉", "燈", "暇", "杋", "後", "璇",
-    "喟", "妪", "湄", "贡", "兮", "楫", "锗", "肮", "圉", "参", "烃", "葩", "骀", "锖", "犿", "恼",
-    "麯", "牮", "溱", "俵", "祢", "割", "羲", "中", "博", "膦", "7", "钙", "耸", "蚋", "嘬", "逝",
-    "瓿", "且", "苗", "屑", "華", "糯", "吓", "陇", ".", "魆", "d", "拶", "箜", "澳", "蘼", "垸",
-    "钅", "樯", "侑", "娼", "赛", "恪", "击", "濮", "隔", "似", "惜", "滤", "培", "谩", "坝", "煙",
-    "馕", "茡", "颼", "啡", "强", "巡", "糗", "芙", "骒", "常", "鳔", "楣", "嫩", "师", "攀", "旭",
-    "茠", "吏", "剌", "露", "把", "携", "忘", "软", "荩", "弧", "锩", "僦", "掘", "仓", "嵚", "沲",
-    "傅", "蝴", "爵", "纪", "薮", "參", "荀", "珙", "逞", "蹙", "徉", "佶", "庹", "弢", "锞", "想",
-    "窕", "躬", "蚡", "闸", "难", "妃", "唶", "V", "蚂", "酐", "]", "羁", "土", "敏", "蝻", "赎",
-    "闹", "铷", "脬", "攫", "椐", "橹", "(", "暨", "剁", "灵", "咫", "苻", "赈", "萝", "鹮", "颛",
-    "勿", "臆", "福", "蝠", "烺", "跤", "垓", "庋", "鏊", "彭", "碚", "膂", "鐘", "鸹", "择", "疫",
-    "暌", "胙", "翘", "仞", "實", "猫", "嘹", "坟", "泺", "靽", "仟", "鲋", "畅", "箪", "箩", "星",
-    "辍", "餐", "伍", "狃", "蹐", "芴", "龋", "仨", "狩", "诌", "媚", "琚", "❤", "鉰", "仔", "忝",
-    "垅", "嗤", "氅", "填", "詹", "诿", "皑", "椅", "鲭", "繪", "碑", "选", "廠", "嶓", "迮", "唠",
-    "赏", "│", "泍", "姑", "晦", "直", "艾", "邂", "裝", "郊", "疡", "」", "鸲", "兕", "县", "邡",
-    "麤", "旧", "壶", "礼", "梭", "铔", "玕", "硬", "后", "馅", "窠", "俊", "犾", "槭", "姨", "妨",
-    "幡", "沿", "帮", "爿", "褊", "淅", "册", "诔", "使", "唢", "族", "逮", "蟋", "掣", "红", "颔",
-    "闲", "咧", "衍", "矛", "蹜", "抱", "雀", "箴", "雙", "茁", "聲", "芰", "瘤", "落", "峨", "秆",
-    "献", "霭", "胰", "墁", "柽", "跛", "阼", "菝", "谗", "蠊", "懋", "蔷", "拃", "岱", "戡", "逡",
-    "背", "克", "棱", "栲", "乖", "橄", "炭", "萱", "胲", "摺", "殆", "魍", "宾", "拢", "姁", "祲",
-    "·", "歹", "胆", "爱", "旰", "抢", "匜", "硚", "愈", "跳", "▲", "跪", "網", "糈", "維", "痹",
-    "娃", "疸", "痔", "褙", "芭", "鹜", "猷", "檩", "吨", "淖", "竟", "呱", "瘠", "奢", "呐", "扞",
-    "灼", "家", "饿", "玻", "倾", "锚", "遗", "赝", "去", "跣", "纬", "玢", "梆", "陉", "屎", "上",
-    "酆", "搤", "併", "腌", "避", "禹", "绅", "帆", "英", "区", "交", "▪", "巧", "豹", "游", "清",
-    "荧", "娀", "訓", "尾", "觫", "梢", "均", "忤", "继", "變", "喽", "尉", "骚", "趸", "馃", "石",
-    "置", "毡", "竴", "龄", "掭", "骟", "堡", "确", "街", "沌", "恨", "壮", "弄", "鞴", "剿", "貅",
-    "路", "杠", "豚", "澜", "靼", "擐", "租", "媪", "彖", "逢", "撺", "盗", "呑", "锥", "栎", "唧",
-    "翩", "倥", "虽", "郏", "钒", "雉", "捒", "官", "酷", "樽", "房", "屣", "堌", "蒍", "悌", "④",
-    "蛳", "俱", "慢", "筇", "框", "椽", "债", "吭", "柱", "夤", "嘈", "骺", "趹", "刓", "豕", "噢",
-    "J", "论", "俞", "仙", "碰", "经", "铋", "⑤", "顽", "曲", "得", "汪", "浒", "际", "嘏", "扮",
-    "祊", "嵋", "‘", "牒", "禮", "粒", "籀", "闱", "刍", "邕", "美", "苫", "靺", "铹", "莰", "蔑",
-    "敵", "逑", "锬", "宸", "鞥", "匝", "褀", "菅", "逊", "脱", "晔", "罾", "郝", "恰", "丟", "芩",
-    "枷", "鹩", "a", "夙", "述", "钢", "瀵", "铧", "翱", "尧", "醮", "&", "鳐", "懲", "濩", "树",
-    "劵", "虾", "砚", "腄", "罟", "硖", "崩", "埘", "╱", "血", "觯", "巂", "栀", "萏", "雪", "眵",
-    "白", "新", "靶", "猪", "亻", "窬", "Ⅳ", "麽", "遄", "沏", "茅", "笋", "甬", "潔", "悛", "幕",
-    "鹤", "悴", "耻", "觜", "滞", "己", "鏂", "肛", "婤", "嘞", "蘸", "胴", "柿", "穗", "咯", "嵕",
-    "蜣", "僰", "玃", "细", "▏", "戗", "尘", "仄", "子", "历", "翌", "烀", "晞", "桷", "藕", "啄",
-    "腐", "瘥", "诣", "彗", "捭", "懑", "陛", "傥", "沚", "贯", "﹑", "雒", "航", "跬", "晓", "赢",
-    "鸭", "臊", "闳", "稲", "阁", "鼋", "‰", "个", "農", "匮", "辚", "滗", "桢", "詈", "绗", "馇",
-    "蜓", "专", "余", "鳉", "稚", "午", "珍", "癃", "钵", "档", "颍", "厕", "岭", "祓", "诂", "襙",
-    "炙", "枯", "锦", "袤", "桊", "嗓", "䘵", "嫂", "丰", "苎", "梦", "早", "菱", "钱", "灯", "钮",
-    "锓", "豐", "蝤", "协", "h", "鼓", "燕", "覺", "涮", "呗", "専", "篓", "玷", "褓", "从", "鲍",
-    "羧", "俭", "萌", "樣", "颜", "谷", "夹", "囫", "』", "缙", "鹰", "条", "状", "沨", "因", "莺",
-    "楼", "扰", "禢", "砧", "埔", "纩", "臭", "茶", "妹", "莼", "境", "山", "诘", "蒌", "唤", "仃",
-    "眄", "田", "畼", "缱", "祘", "痴", "揍", "猾", "甙", "躲", "返", "砝", "叽", "苟", "肭", "悖",
-    "嫣", "凋", "踞", "勢", "弃", "般", "赤", "佣", "嘘", "智", "寒", "旆", "咎", "戊", "温", "昙",
-    "蝄", "掉", "冤", "凍", "一", "鋼", "茏", "兼", "昃", "峤", "挟", "０", "顾", "颅", "*", "遆",
-    "硅", "烯", "眊", "鳘", "下", "雾", "癖", "贮", "缴", "升", "劲", "婞", "掇", "梯", "壬", "撮",
-    "绑", "氣", "薤", "兔", "赟", "漕", "签", "顇", "砭", "鸼", "明", "佧", "鹗", "哉", "席", "趁",
-    "砖", "澌", "撤", "萨", "熔", "蚝", "狮", "夔", "介", "青", "喈", "郂", "埝", "狗", "哚", "咪",
-    "材", "跑", "炗", "龉", "髓", "衣", "觀", "淋", "寄", "羮", "抵", "恁", "藿", "蜢", "甜", "兽",
-    "洒", "眶", "薯", "绷", "颀", "骑", "鲻", "｜", "q", "繄", "丝", "焙", "瑕", "童", "豸", "讨",
-    "钿", "艮", "踧", "幽", "焰", "宽", "址", "鸩", "堉", "音", "擅", "铞", "埌", "宙", "_", "钥",
-    "承", "率", "噬", "凌", "程", "蚕", "嘣", "饭", "橉", "径", "１", "鄏", "袜", "孙", "樊", "窦",
-    "牦", "歡", "罕", "蓉", "诩", "㧎", "责", "豢", "逅", "甘", "礶", "镥", "穂", "到", "檑", "皇",
-    "泖", "腱", "郤", "竞", "竽", "溴", "木", "蛀", "蛟", "燧", "鹳", "谵", "钳", "魔", "链", "咆",
-    "荡", "婳", "冢", "鲧", "昝", "※", "倫", "才", "簿", "醫", "吕", "姗", "Ｌ", "俺", "隳", "鮮",
-    "衩", "鲆", "崟", "鲹", "岗", "镇", "嗄", "硪", "盎", "氟", "笪", "今", "公", "尽", "佬", "骧",
-    "驻", "翼", "咚", "盱", "镎", "浈", "谅", "洟", "嘭", "跶", "蔹", "咭", "妁", "殚", "砷", "购",
-    "犸", "哄", "笸", "喂", "赋", "倓", "项", "臧", "梳", "笏", "鹝", "昭", "罿", "镄", "П", "湯",
-    "U", "绐", "绂", "鸬", "毐", "台", "鈴", "憾", "痰", "缭", "绒", "赁", "效", "哝", "─", "偆",
-    "泗", "谇", "貘", "搜", "阗", "個", "艟", "嗔", "正", "筛", "鼷", "摛", "馀", "坡", "卓", "厥",
-    "睾", "雯", "篷", "狰", "掀", "夫", "靛", "伻", "茇", "顧", "猕", "占", "俟", "兄", "垡", "兆",
-    "铃", "屙", "轵", "銮", "磲", "桤", "跷", "缳", "粘", "袢", "尬", "獘", "社", "彬", "哨", "煳",
-    "莊", "教", "鸸", "窖", "晰", "捻", "环", "廿", "驶", "椠", "熟", "赓", "挥", "兵", "洚", "桧",
-    "埼", "嘀", "哎", "柞", "悄", "调", "洼", "蛇", "铬", "蝇", "倆", "谂", "姻", "怕", "瑙", "算",
-    "惮", "遥", "惟", "甫", "玺", "袈", "税", "玟", "械", "擗", "淤", "溟", "揞", "纲", "谤", "條",
-    "翕", "荛", "忾", "馿", "糖", "檗", "垱", "┐", "躜", "缒", "您", "蹶", "合", "数", "斟", "粲",
-    "玖", "淞", "孰", "觎", "M", "嫪", "毽", "桩", "蹦", "亹", "胾", "亲", "痉", "莙", "捐", "烁",
-    "瓣", "邀", "但", "误", "饻", "妄", "镱", "民", "嗳", "竦", "柰", "脒", "斓", "瞟", "拱", "霾",
-    "粜", "腽", "咬", "唿", "礻", "y", "褒", "郪", "熰", "梻", "甭", "处", "拖", "蜉", "鄱", "絲",
-    "宠", "瞵", "荷", "靓", "喻", "阢", "伫", "担", "朐", "墙", "儣", "胄", "簏", "皆", "踌", "萃",
-    "命", "滟", "睹", "叁", "蟮", "東", "﹖", "式", "營", "疠", "醳", "尔", "購", "吒", "柴", "途",
-    "力", "疰", "寳", "萬", "挨", "碥", "牍", "嘻", "柃", "剡", "徝", "杖", "螺", "祈", "産", "煽",
-    "讽", "瞭", "由", "牀", "涉", "暴", "枚", "仰", "它", "胤", "驯", "港", "寵", "熄", "蛮", "务",
-    "殃", "亸", "涢", "妫", "粥", "扦", "寺", "抿", "湃", "C", "脚", "涸", "动", "務", "镭", "检",
-    "釂", "耳", "榦", "怆", "氨", "瘳", "钽", "牚", "负", "纂", "獬", "息", "婄", "襞", "囹", "酰",
-    "扑", "臛", "价", "褡", "篼", "媢", "酊", "腋", "愣", "垯", "喃", "牾", "犨", "蜂", "诹", "迂",
-    "槑", "好", "褰", "咝", "③", "劐", "骄", "腉", "仕", "叵", "磅", "金", "皁", "勍", "煮", "鍪",
-    "筚", "吲", "艰", "床", "叛", "﹔", "喜", "齑", "昧", "乎", "谲", "匈", "麒", "身", "唑", "鲤",
-    "栂", "垭", "殪", "沪", "苌", "濡", "蒎", "遏", "®", "质", "纫", "罐", "瞍", "柄", "呼", "锱",
-    "盔", "搋", "烛", "喉", "奈", "赦", "轘", "鉴", "瓒", "铰", "恽", "腠", "韶", "胼", "侄", "珂",
-    "鹕", "䝉", "呤", "椁", "岳", "汲", "疤", "绊", "涂", "乾", "噎", "兑", "絜", "嘧", "蚯", "们",
-    "蝉", "玛", "魁", "黻", "振", "冥", "报", "窟", "笱", "喊", "瓴", "喾", "説", "雹", "恫", "ａ",
-    "颧", "給", "宅", "鞧", "嬷", "鼾", "笔", "鲉", "匹", "蠕", "愍", "氽", "赐", "畔", "絪", "洴",
-    "审", "窂", "褔", "狯", "谑", "毳", "右", "疣", "鲙", "堃", "瓜", "…", "崆", "钫", "樭", "鲛",
-    "砂", "筋", "犏", "乱", "入", "浐", "膙", "厚", "獗", "覌", "焦", "罃", "宬", "隍", "荃", "醜",
-    "厉", "艇", "娟", "樂", "筲", "洛", "唼", "诜", "馑", "香", "缛", "运", "祟", "踩", "陲", "叔",
-    "潞", "二", "搠", "祼", "瀘", "蒋", "樗", "麦", "鸦", "棻", "古", "鹠", "怊", "裾", "够", "璧",
-    "晡", "擘", "毂", "御", "葚", "忱", "觞", "瑁", "唇", "罡", "剽", "殡", "沛", "帻", "举", "瞓",
-    "谬", "溝", "言", "哽", "婿", "猿", "跗", "獴", "俜", "呙", "弗", "凿", "窭", "铌", "友", "唉",
-    "怫", "荘",
-];

src/image_io.rs

@@ -1,62 +0,0 @@
-use anyhow::{Context, Result};
-use base64::{Engine as _, engine::general_purpose};
-use image::{DynamicImage, GenericImageView, ImageBuffer, Rgb, RgbImage};
-use std::path::{Path, PathBuf};
-use tract_onnx::prelude::tract_ndarray::Array3;
-
-/// 定义支持的输入类型枚举
-pub enum ImageInput {
-    Bytes(Vec<u8>),
-    Array(Array3<u8>),
-    Path(PathBuf),
-    Base64(String),
-    DynamicImage(DynamicImage),
-}
-
-/// 模拟 Python 的 load_image_from_input
-#[allow(dead_code)]
-pub fn load_image_from_input(input: ImageInput) -> Result<DynamicImage> {
-    match input {
-        ImageInput::DynamicImage(img) => Ok(img),
-        _ => todo!("后续补充"),
-    }
-}
-
-/// 对应 Python 的 png_rgba_black_preprocess
-/// 将带有透明通道的图片转换为白色背景的 RGB 图片
-#[allow(dead_code)]
-pub fn png_rgba_white_preprocess(img: &DynamicImage) -> DynamicImage {
-    // 1. 检查是否包含透明通道，如果没有，直接克隆并返回
-    if !img.color().has_alpha() {
-        return img.clone();
-    }
-
-    let (width, height) = img.dimensions();
-
-    // 2. 创建一个新的 RGB 图像缓冲，默认填充为白色 (255, 255, 255)
-    let mut background = ImageBuffer::from_pixel(width, height, Rgb([255u8, 255u8, 255u8]));
-
-    // 3. 获取原图的 RGBA 视图
-    let rgba_img = img.to_rgba8();
-
-    // 4. 遍历像素并手动进行 Alpha 混合
-    // 对应 Python 的 image.paste(img, ..., mask=img)
-    for (x, y, pixel) in rgba_img.enumerate_pixels() {
-        let alpha = pixel[3] as f32 / 255.0;
-
-        if alpha >= 1.0 {
-            // 完全不透明，直接覆盖
-            background.put_pixel(x, y, Rgb([pixel[0], pixel[1], pixel[2]]));
-        } else if alpha > 0.0 {
-            // 半透明，执行 Alpha 混合公式： (src * alpha) + (dst * (1 - alpha))
-            let bg_pixel = background.get_pixel(x, y);
-            let r = (pixel[0] as f32 * alpha + bg_pixel[0] as f32 * (1.0 - alpha)) as u8;
-            let g = (pixel[1] as f32 * alpha + bg_pixel[1] as f32 * (1.0 - alpha)) as u8;
-            let b = (pixel[2] as f32 * alpha + bg_pixel[2] as f32 * (1.0 - alpha)) as u8;
-            background.put_pixel(x, y, Rgb([r, g, b]));
-        }
-        // alpha == 0 的情况不需要处理，因为背景已经是白色了
-    }
-
-    DynamicImage::ImageRgb8(background)
-}

src/image_processor.rs

@@ -1,27 +0,0 @@
-use image::{DynamicImage, GrayImage, imageops::FilterType};
-use anyhow::Result;
-
-/// 对应 Python 的 convert_to_grayscale
-/// 将图像转换为灰度图 (L模式)
-pub fn convert_to_grayscale(image: &DynamicImage) -> GrayImage {
-    // Rust image 库的 to_luma8 会根据标准的亮度公式进行转换
-    image.to_luma8()
-}
-
-/// 对应 Python 的 resize_image
-/// 调整图像尺寸。当前版本仅实现 keep_aspect_ratio=false
-pub fn resize_image(
-    image: &GrayImage,
-    target_width: u32,
-    target_height: u32,
-    // resample 参数我们直接使用 FilterType，Lanczos3 是最接近 Python LANCZOS 的
-) -> GrayImage {
-    // 使用 resize 算法进行精确缩放
-    image::imageops::resize(
-        image,
-        target_width,
-        target_height,
-        FilterType::Lanczos3
-    )
-}
-

src/lib.rs

@@ -1,182 +0,0 @@
-mod charset;
-mod image_io;
-mod image_processor;
-mod model;
-mod utils;
-
-use crate::image_io::png_rgba_white_preprocess;
-use crate::image_processor::{convert_to_grayscale, resize_image};
-use anyhow::{Context, Result};
-use image::{DynamicImage, imageops::FilterType};
-use tract_onnx::prelude::*;
-// 关键点：直接使用 tract 重导出的 ndarray
-use tract_onnx::prelude::tract_ndarray::s;
-pub struct DdddOcr {
-    session: RunnableModel<TypedFact, Box<dyn TypedOp>, Graph<TypedFact, Box<dyn TypedOp>>>,
-}
-
-impl DdddOcr {
-    pub fn new<P>(model_path: P) -> Result<Self>
-    where
-        P: AsRef<std::path::Path>,
-    {
-        let session = onnx()
-            .model_for_path(model_path)
-            .with_context(|| "加载 ONNX 模型失败，请检查路径是否正确")?
-            .into_optimized()?
-            .into_runnable()?;
-        Ok(Self { session })
-    }
-
-    pub fn classification(&self, img: &DynamicImage) -> Result<String> {
-        let tensor = self.preprocess_image(img, false)?;
-
-        // let result = self.session.run(tvec!(tensor.into()))?;
-        // 3. 解析结果
-        // let output = result[0].to_array_view::<i64>()?;
-        let output = self.inference(tensor)?;
-        let output2 = self.process_text_output(&output)?;
-        Ok(Self::ctc_decode_indices(&output2))
-    }
-    /// 对应 Python 的 _preprocess_image
-    /// 负责：透明背景修复 -> 灰度化 -> 按比例 Resize -> 归一化 -> 4维张量转换
-    fn preprocess_image(&self, img: &DynamicImage, png_fix: bool) -> Result<Tensor> {
-        // A. 修复 PNG 透明背景 (内部逻辑你之前已实现)
-        let _ = if png_fix && img.color().has_alpha() {
-            png_rgba_white_preprocess(img)
-        } else {
-            img.clone()
-        };
-
-        let h = 64u32;
-        let w = (img.width() as f32 * (h as f32 / img.height() as f32)) as u32;
-        let gray_img = convert_to_grayscale(img);
-        let resized = resize_image(&gray_img, w, h);
-        // resized.save("debug_preprocessed.png").unwrap();
-        // 1. 预处理：转灰度 -> Resize -> 归一化
-        // let resized = img.resize_exact(w, h, FilterType::Lanczos3).to_luma8();
-
-        // 使用 tract_ndarray 构造，避免版本冲突
-        let array =
-            tract_ndarray::Array4::from_shape_fn((1, 1, h as usize, w as usize), |(_, _, y, x)| {
-                let pixel = resized.get_pixel(x as u32, y as u32)[0] as f32;
-                (pixel / 255.0 - 0.5) / 0.5
-            });
-
-        let tensor = Tensor::from(array);
-
-        Ok(tensor)
-    }
-    /// 对应 Python 的 _inference
-    fn inference(&self, tensor: Tensor) -> Result<Tensor> {
-        // tract 的 run 会返回一个 Vec<TValue>，我们通常只需要第一个输出
-        // let result = self.session.run(tvec!(tensor.into()))?;
-        let mut result = self
-            .session
-            .run(tvec!(tensor.into()))
-            .context("执行模型推理失败")?;
-        println!("模型输出原始数据: {:?}", result);
-        Ok(result.remove(0).into_tensor())
-    }
-    /// 核心解析逻辑：将模型输出的各种维度/类型的 Tensor 转为字符索引序列
-    fn process_text_output(&self, raw_tensor: &Tensor) -> Result<Vec<i64>> {
-        let shape = raw_tensor.shape();
-        println!("模型输出shape数据: {:?}", shape);
-        let datum_type = raw_tensor.datum_type();
-        println!("模型输出datum_type数据: {:?}", datum_type);
-
-        match raw_tensor.datum_type() {
-            // 情况 1: huashi666 式模型，直接输出 i64 索引 (通常是模型内部做好了 Argmax)
-            DatumType::I64 => {
-                let view = raw_tensor.to_array_view::<i64>()?;
-                Ok(view.iter().cloned().collect())
-            }
-
-            // 情况 2: sml2h3 原版模型，输出 F32 概率矩阵
-            DatumType::F32 => {
-                let view = raw_tensor.to_array_view::<f32>()?;
-                let (steps, classes, data_view) = match shape.len() {
-                    3 => {
-                        if shape[1] == 1 {
-                            // 形状: [Steps, 1, Classes] -> 你的原有逻辑
-                            (shape[0], shape[2], view.into_dyn())
-                        } else if shape[0] == 1 {
-                            // 形状: [1, Steps, Classes] -> 另一种常见导出格式
-                            (shape[1], shape[2], view.into_dyn())
-                        } else {
-                            // 默认取第一个 batch: [Batch, Steps, Classes]
-                            // 使用 slice 对应 Python 的 output[0, :, :]
-                            let sliced = view.slice(s![0, .., ..]);
-                            (shape[1], shape[2], sliced.into_dyn())
-                        }
-                    }
-                    2 => {
-                        // 形状: [Steps, Classes] -> 已经剥离了 Batch 维度
-                        (shape[0], shape[1], view.into_dyn())
-                    }
-                    _ => return Err(anyhow::anyhow!("不支持的输出维度: {:?}", shape)),
-                };
-                let array_2d = data_view.to_shape((steps, classes))?;
-                //
-                // 对每一行执行 Argmax (寻找概率最大的字符索引)
-                let indices = array_2d
-                    .outer_iter()
-                    .map(|row| {
-                        row.iter()
-                            .enumerate()
-                            .max_by(|(_, a), (_, b)| {
-                                a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal)
-                            })
-                            .map(|(idx, _)| idx as i64)
-                            .unwrap_or(0)
-                    })
-                    .collect();
-                Ok(indices)
-            }
-            _ => Err(anyhow::anyhow!(
-                "不支持的模型输出数据类型: {:?}",
-                raw_tensor.datum_type()
-            )),
-        }
-    }
-    fn ctc_decode_indices(predicted_indices: &[i64]) -> String {
-        println!("indices模型输出原始数据: {:?}", predicted_indices);
-
-        use crate::charset::CHARSET_BETA;
-        // 对应 _ctc_decode_indices 的逻辑：去重、去 blank (0)
-        let mut res = String::new();
-        let mut prev_idx: i64 = -1;
-
-        for &idx in predicted_indices {
-            // 1. 跳过连续重复的索引
-            // 2. 跳过 blank 字符 (假设索引 0 是 blank)
-            if idx != prev_idx && idx != 0 {
-                if let Ok(u_idx) = usize::try_from(idx) {
-                    if let Some(&char_str) = CHARSET_BETA.get(u_idx) {
-                        res.push_str(char_str);
-                    }
-                }
-            }
-            prev_idx = idx;
-        }
-        println!("最终识别出的验证码是: {}", res);
-        res
-    }
-}
-
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-    #[test]
-    fn test_ctc_decode_indices() {
-        // 模拟一个 DdddOcr 实例（如果 decode 不依赖 session，可以设为相关函数）
-        // 这里假设你的 decode_ctc 是公开或内部可访问的
-        let input = vec![1, 1, 0, 1, 2, 2, 0, 2];
-        // 逻辑：[1, 1] -> 1, [0] -> 跳过, [1] -> 1, [2, 2] -> 2, [0] -> 跳过, [2] -> 2
-        // 预期结果索引应该是 [1, 1, 2, 2] 对应的字符
-        // 具体的断言取决于你的 CHARSET_BETA
-        // let result = dddd.ctc_decode_indices(&input);
-        // assert_eq!(result, "AABB");
-    }
-}

tests/ocr_test.rs

@@ -1,16 +0,0 @@
-use ddddocr_rs::DdddOcr; // 假设你的包名是这个
-
-#[test]
-fn test_full_classification() {
-    // 1. 初始化模型
-    let ocr = DdddOcr::new("model/common.onnx").expect("模型加载失败");
-
-    // 2. 加载测试图片
-    let img = image::open("samples/code3.png").expect("测试图片不存在");
-
-    // 3. 执行识别
-    let result = ocr.classification(&img).expect("识别过程出错");
-
-    println!("识别结果: {}", result);
-    assert!(!result.is_empty());
-}