refactor: 抽象解耦推理引擎并重构为多Crate工作空间架构

- 移除 核心层与 tract/Tensor 的强耦合，前/后处理全线转用标准 ndarray
- 针对 OCR 与目标检测（Det）分别设计独立的强类型输出小枚举（OcrOutput/DetOutput）
- 利用 Trait 关联类型（Associated Type）InferenceEngine,OcrEngine,DetEngine 统一接口，实现多后端解耦
- 引入 thiserror 库，建立完备的强类型错误处理机制（DdddError/Result）
- 完成项目结构初拆，剥离为 ddddocr-core 和 ddddocr-tract

Cargo.toml

@@ -1,10 +1,17 @@
-[package]
+[workspace]
-name = "ddddocr-rs"
+resolver = "2"
+members = [
+    "ddddocr-core",
+    "ddddocr-tract",
+]
+
+[workspace.package]
 version = "0.1.0"
 edition = "2024"
 license = "MIT OR Apache-2.0"
 
-[dependencies]
+
+[workspace.dependencies]
 tract-onnx = { version = "0.21.10" }
 anyhow = "1.0.102"
 image = "0.25.10"
@@ -12,3 +19,5 @@ base64 = "0.22.1"
 imageproc = { version = "0.26.2", default-features = true }
 serde = { version = "1.0.228", features = ["derive"] }
 serde_json = "1.0.150"
+ndarray="0.16.1"
+thiserror = "1.0" # 刚好可以开始接入你需要的标准库错误处理

ddddocr-core/Cargo.toml

@@ -0,0 +1,17 @@
+[package]
+name = "ddddocr-core"
+version = { workspace = true }
+edition = { workspace = true }
+license = { workspace = true }
+
+[dependencies]
+anyhow = "1.0.102"
+image = "0.25.10"
+base64 = "0.22.1"
+imageproc = { version = "0.26.2", default-features = true }
+serde = { workspace = true }
+serde_json = "1.0.150"
+ndarray = { workspace = true } # 继承自工作空间
+thiserror = { workspace = true } # 刚好可以开始接入你需要的标准库错误处理
+
+#serde = { workspace = true, features = ["derive"] }

ddddocr-core/src/algo/mod.rs

@@ -0,0 +1,3 @@
+mod slide;
+
+pub use slide::{SlideResult, Slider};

ddddocr-core/src/algo/slide.rs

@@ -1,32 +1,40 @@
-use crate::utils::cv_ops;
+use crate::utils::image_proc;
-use crate::utils::cv_ops::{abs_diff, min_max_loc, ndarray_to_luma8, rgb_to_gray};
+use crate::utils::image_proc::{abs_diff, min_max_loc, ndarray_to_luma8, rgb_to_gray};
 use crate::utils::image_io::image_to_ndarray;
-use anyhow::{Context, Result, anyhow};
+use anyhow::{Result, anyhow};
-use image::{DynamicImage, GenericImageView};
+use image::DynamicImage;
-use image::{ImageBuffer, Luma};
+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 std::cmp::{max, min};
+use std::fmt;
-use tract_onnx::prelude::tract_ndarray::{Array2, Array3, ArrayView2, ArrayView3, Axis, s};
+use ndarray::{ArrayView2, ArrayView3};
-
+#[derive(Debug)]
 pub struct SlideResult {
     pub target: [i32; 2],
     pub target_x: i32,
     pub target_y: i32,
     pub confidence: f64,
 }
-
+impl fmt::Display for SlideResult {
-pub struct Slide;
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-
+        writeln!(f, "滑块匹配测试结果:")?;
-impl Slide {
+        writeln!(f, "检测坐标: [x: {}, y: {}]", self.target_x, self.target_y)?;
-    pub fn new() -> Self {
+        // 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}
-        Self
+        write!(f, "置信度: {:.4}", self.confidence)?;
+        Ok(())
     }
+}
 
+pub struct Slider;
+
+impl Slider {
+    pub fn new() -> Result<Self, anyhow::Error> {
+        Ok(Self)
+    }
     /// 对应 Python: slide_match 滑块匹配接口
     pub fn slide_match(
         &self,
@@ -59,23 +67,8 @@ impl Slide {
         target: ArrayView3<u8>,
         background: ArrayView3<u8>,
     ) -> Result<SlideResult> {
-        // let (h, w, _) = target.dim();
 
-        // 1. 计算图像差异并灰度化 (对应 cv2.absdiff + cv2.cvtColor)
-        // 使用 OpenCV 标准权重公式：0.299R + 0.587G + 0.114B
-        // let mut diff_buffer = ImageBuffer::new(w as u32, h as u32);
-        // for y in 0..h {
-        //     for x in 0..w {
-        //         let r_diff = (target[[y, x, 0]] as i16 - background[[y, x, 0]] as i16).abs() as f32;
-        //         let g_diff = (target[[y, x, 1]] as i16 - background[[y, x, 1]] as i16).abs() as f32;
-        //         let b_diff = (target[[y, x, 2]] as i16 - background[[y, x, 2]] as i16).abs() as f32;
-        //
-        //         let gray_diff = (0.299 * r_diff + 0.587 * g_diff + 0.114 * b_diff) as u8;
-        //         diff_buffer.put_pixel(x as u32, y as u32, Luma([gray_diff]));
-        //     }
-        // }
         // 1. 计算差异数组 (复用 cv2::absdiff)
-
         let (th, tw, tc) = target.dim();
         let (bh, bw, bc) = background.dim();
 
@@ -118,11 +111,11 @@ impl Slide {
 
         // // 统计每个标签出现的频率（即面积）
         // 4. 寻找最大连通区域 (对应 findContours + max area)
-        if let Some(max_label) = cv_ops::find_contours_and_max(&labelled) {
+        if let Some(max_label) = image_proc::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) = cv_ops::bounding_rect(&labelled, max_label);
+            let (x, y, w, h) = image_proc::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) = cv_ops::calculate_center((x, y), w as usize, h as usize);
+            let (center_x, center_y) = image_proc::calculate_center((x, y), w as usize, h as usize);
 
             Ok(SlideResult {
                 target: [center_x, center_y],
@@ -194,9 +187,6 @@ impl Slide {
         background: ArrayView2<u8>,
     ) -> Result<SlideResult> {
         // 1. 将 ndarray 转换为 imageproc 需要的 ImageBuffer (无拷贝或轻量转换)
-
-        // let (bh, bw) = background.dim();
-
         // 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);
@@ -216,7 +206,7 @@ impl Slide {
         // 4. 计算中心点 (与 Python 逻辑完全一致)
         let (th, tw) = target.dim();
 
-        let (center_x, center_y) = cv_ops::calculate_center(max_loc, tw as usize, th as usize);
+        let (center_x, center_y) = image_proc::calculate_center(max_loc, tw as usize, th as usize);
         // println!("Rust Target Width (tw): {}", tw);
         // println!("Rust Best Max Loc X: {}", max_loc.0);
         // println!("Rust Final Center X: {}", center_x);
@@ -261,7 +251,7 @@ impl Slide {
         // 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) = cv_ops::calculate_center(max_loc, tw as usize, th as usize);
+        let (center_x, center_y) = image_proc::calculate_center(max_loc, tw as usize, th as usize);
 
         // 打印调试信息，方便与 Python 对比
         // println!("Edge Match: max_val: {}, max_loc: {:?}", max_val, max_loc);

ddddocr-core/src/error.rs

@@ -0,0 +1,48 @@
+pub(crate) const MODEL_DOWNLOAD_HELP: &str = "\
+================================================================================
+[ddddocr-rust] 错误：未找到默认的模型文件！
+--------------------------------------------------------------------------------
+由于打包体积限制，本库未内置 ONNX 模型。请按照以下步骤操作：
+
+1. 前往官方 GitHub 下载对应的模型权重：
+   - OCR 模型:  https://github.com/sml2h3/ddddocr/raw/master/ddddocr/common_sml2h3_f32.onnx
+   - DET 模型:  https://github.com/sml2h3/ddddocr/raw/master/ddddocr/common_det.onnx
+
+2. 配置加载方式（二选一）：
+   A. 【推荐】设置环境变量指向您下载的文件：
+      Linux/macOS: export DDDD_OCR_MODEL=\"/path/to/common_sml2h3_f32.onnx\"
+      Windows (CMD): set DDDD_OCR_MODEL=C:\\path\\to\\common_sml2h3_f32.onnx
+      Windows (PowerShell): $env:DDDD_OCR_MODEL=\"C:\\path\\to\\common_sml2h3_f32.onnx\"
+
+   B. 或者直接将模型文件重命名并放置在您运行程序的“当前工作目录”或“可执行文件同级目录”下。
+================================================================================";
+
+
+use thiserror::Error;
+
+#[derive(Error, Debug)]
+pub enum DdddError {
+    #[error("图像预处理失败: {0}")]
+    PreprocessError(String),
+
+    #[error("模型推理引擎内部发生异常: {0}")]
+    EngineError(#[from] anyhow::Error),
+
+    #[error("CTC 解码错误: {0}")]
+    DecodeError(String),
+
+    #[error("维度转换失败，预期维度 {expected}，实际形状为 {actual:?}")]
+    DimensionMismatch {
+        expected: String,
+        actual: Vec<usize>,
+    },
+
+    #[error("内存不连续，无法执行零拷贝操作")]
+    NonContiguousMemory,
+
+    #[error("未知的模型输出格式")]
+    UnknownOutputFormat,
+}
+
+/// 统一用我们自己的 DdddError 包装 Result
+pub type Result<T> = std::result::Result<T, DdddError>;

ddddocr-core/src/lib.rs

@@ -0,0 +1,37 @@
+mod algo;
+pub mod error;
+pub mod models;
+pub mod utils;
+
+pub use crate::algo::{SlideResult, Slider};
+use crate::error::Result;
+pub use crate::models::det::{DetBuilder, DetectionResult, Detector};
+pub use crate::models::ocr::{Ocr, OcrBuilder, OcrResult};
+pub use models::ocr::metadata::ModelMetadata;
+// DetSession
+
+pub enum OcrOutput {
+    Indices(ndarray::Array1<i64>), // 拥有完整所有权的 1维数组，可任意传递和返回
+    Logits(ndarray::Array2<f32>),
+}
+/// 2. 目标检测专属的、编译期安全的输出枚举
+pub enum DetOutput {
+    Detection(ndarray::Array3<f32>), // 拥有完整所有权的 2维矩阵，可任意传递和返回
+}
+
+/// 核心层定义的统一推理引擎接口。
+/// 未来的 ddddocr-tract 和 ddddocr-ort 都必须实现这个 Trait
+
+pub trait InferenceEngine {
+    /// 关联类型：具体的 Session 需要声明自己到底产出什么枚举
+    type Output;
+    fn inference(&self, input_array: ndarray::Array4<f32>) -> Result<Self::Output>;
+}
+
+pub trait OcrEngine: InferenceEngine<Output = OcrOutput> {
+    fn metadata(&self) -> &ModelMetadata;
+}
+
+pub trait DetEngine: InferenceEngine<Output = DetOutput> {}
+
+

ddddocr-core/src/models/det/builder.rs

@@ -0,0 +1,26 @@
+use crate::models::det::executor::Detector;
+// use ddddocr_tract::det::session::DetSession;
+use crate::DetEngine;
+
+pub struct DetBuilder {
+    use_gpu: bool,
+    device_id: u8,
+}
+
+impl DetBuilder {
+    fn use_gpu(mut self) -> Self {
+        self.use_gpu = true;
+        self
+    }
+    fn device_id(mut self, device_id: u8) -> Self {
+        self.device_id = device_id;
+        self
+    }
+    fn build(self, session: &dyn DetEngine) -> Detector<'_> {
+        Detector {
+            session,
+            use_gpu: self.use_gpu,
+            device_id: self.device_id,
+        }
+    }
+}

ddddocr-core/src/models/det/executor.rs

@@ -1,11 +1,12 @@
-use crate::models::loader::{ModelLoader, ModelSession, ModelType};
 use anyhow::{Context, Result};
-use image::{DynamicImage, GenericImageView, imageops::FilterType};
+use image::{imageops::FilterType, DynamicImage, GenericImageView};
-use tract_onnx::prelude::tract_ndarray::{Array2, Array3, Array4, Axis, prelude::*, s};
+use std::fmt;
-use tract_onnx::prelude::{Graph, RunnableModel, Tensor, TypedFact, TypedOp, tvec};
+use ndarray::{prelude::*, s, Array2, Array3, Array4, Axis};
-
+// use tract_onnx::prelude::{Tensor};
 
 
+// use ddddocr_tract::det::session::DetSession;
+use crate::{DetEngine, DetOutput};
 #[derive(Debug, Clone, Copy)]
 pub struct DetectionResult {
     pub x1: i32,
@@ -16,30 +17,41 @@ pub struct DetectionResult {
     pub class_id: u32,
 }
 
-
+impl fmt::Display for DetectionResult {
-
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-pub struct Det {
+        // 结构体只管自己这一行怎么显示，不用管外部的索引 [i]
-    session: RunnableModel<TypedFact, Box<dyn TypedOp>, Graph<TypedFact, Box<dyn TypedOp>>>,
+        write!(
-}
+            f,
-impl ModelSession for Det {
+            "x1={}, y1={}, x2={}, y2={}, 分数={:.4}, 类别ID={}",
-    fn get_model_type(&self) -> ModelType {
+            self.x1, self.y1, self.x2, self.y2, self.score, self.class_id
-        todo!()
+        )
-    }
-    fn desc(&self) -> String {
-        "Detection Model 加载成功".to_string()
     }
 }
-impl Det {
+
-    pub fn new(model_path: String) -> Result<Self, anyhow::Error> {
+
-        let session = ModelLoader::load_model(&model_path)?.session;
+pub struct Detector<'a> {
-        Ok(Self { session })
+    pub(crate) session: &'a dyn DetEngine,
+    #[allow(dead_code)]
+    pub(crate) use_gpu: bool,
+    #[allow(dead_code)]
+    pub(crate) device_id: u8,
+}
+
+impl<'a> Detector<'a> {
+    pub fn new(session: &'a dyn DetEngine) -> Self {
+        Detector {
+            session,
+            use_gpu: false,
+            device_id: 0,
+        }
     }
+
     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
         self.get_bbox(image)
     }
     /// 2. preproc: 纯 Rust 实现 (替代 OpenCV)
-    fn preproc(&self, image: &DynamicImage, input_size: (u32, u32)) -> Result<(Tensor, f32)> {
+    fn preproc(&self, image: &DynamicImage, input_size: (u32, u32)) -> Result<(Array4<f32>, f32)> {
         let (target_h, target_w) = input_size;
         let (img_w, img_h) = image.dimensions();
 
@@ -73,12 +85,11 @@ impl Det {
                 // 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[[0, 2, y, x]] = slice[idx] as f32; // R
             }
         }
 
-
+        Ok((array, r))
-        Ok((array.into(), r))
     }
 
     /// 3. demo_postprocess (逻辑与 Python 一致)
@@ -244,11 +255,11 @@ impl Det {
         let (input_tensor, ratio) = self.preproc(dynamic_img, (416, 416))?;
 
         // tract 推理
-        let outputs = self.session.run(tvec!(input_tensor.into()))?;
+        // let outputs = self.session.session.run(tvec!(input_tensor.into()))?;
-        let output_array = outputs[0]
+        let outputs = self.session.inference(input_tensor)?;
-            .to_array_view::<f32>()?
+        // let output_array = outputs[0]
-            .to_owned()
+        // 2. 无缝、安全地解包出标准 3维 矩阵
-            .into_dimensionality::<Ix3>()?;
+        let DetOutput::Detection(output_array) = outputs;
 
         let predictions = self.demo_postprocess(output_array, (416, 416));
         let pred = predictions.slice(s![0, .., ..]);
@@ -273,17 +284,15 @@ impl Det {
         let detections = self.multiclass_nms(&boxes_xyxy, &scores, 0.45, 0.1);
         let final_results = detections
             .into_iter()
-            .map(|d| {
+            .map(|d| DetectionResult {
-                DetectionResult{
+                x1: (d[0] as i32).max(0).min(orig_w as i32),
-                    x1: (d[0] as i32).max(0).min(orig_w as i32),
+                y1: (d[1] as i32).max(0).min(orig_h 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),
-                    x2: (d[2] as i32).max(0).min(orig_w as i32),
+                y2: (d[3] as i32).max(0).min(orig_h as i32),
-                    y2: (d[3] as i32).max(0).min(orig_h as i32),
+                score: d[4],
-                    score: d[4],
+                class_id: d[5] as u32,
-                    class_id: d[5] as u32,
-                }
             })
             .collect();
-        Ok(final_results )
+        Ok(final_results)
     }
 }

ddddocr-core/src/models/det/mod.rs

@@ -0,0 +1,6 @@
+mod builder;
+mod executor;
+
+pub use builder::DetBuilder;
+pub use executor::{DetectionResult, Detector};
+// pub use ddddocr_tract::det::session::DetSession;

ddddocr-core/src/models/mod.rs

@@ -0,0 +1,2 @@
+pub mod ocr;
+pub mod det;

ddddocr-core/src/models/ocr/builder.rs

@@ -0,0 +1,74 @@
+use crate::models::ocr::executor::Ocr;
+// use ddddocr_tract::session::OcrSession;
+use crate::models::ocr::color_filter::ColorFilter;
+use crate::models::ocr::token_filter::TokenFilter;
+use crate::OcrEngine;
+
+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
+    pub fn new() -> Self {
+        Self {
+            png_fix: false, // 默认值
+            probability: false,
+            color_filter: None,
+            charset_restrict: None,
+        }
+    }
+    pub fn png_fix(mut self, value: bool) -> Self {
+        self.png_fix = value;
+        self
+    }
+    pub fn probability(mut self, value: bool) -> Self {
+        self.probability = value;
+        self
+    }
+
+    pub fn color_filter<T>(mut self, filter: T) -> Self
+    where
+        T: ColorFilter + Send + Sync + 'static,
+    {
+        self.color_filter = Some(Box::new(filter));
+        self
+    }
+
+    pub fn charset_restrict<T>(mut self, restrict: T) -> Self
+    where
+        T: TokenFilter + Send + Sync + 'static,
+    {
+        self.charset_restrict = Some(Box::new(restrict));
+        self
+    }
+    pub fn build(self, session: &dyn OcrEngine) -> 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 = &session.metadata().charset.tokens;
+        let final_charset_indices = match &self.charset_restrict {
+            Some(restrict) => restrict.apply_to_charset(tokens),
+            None => None,
+        };
+
+        // Ocr::new(session, self)
+        Ocr {
+            session,
+            png_fix: self.png_fix, // 原地解构出来
+            probability: self.probability,
+            final_color_ranges,
+            final_charset_indices,
+        }
+    }
+}

ddddocr-core/src/models/ocr/color_filter.rs

@@ -1,25 +1,24 @@
-use std::str::FromStr;
+use crate::utils::image_proc::rgb_to_opencv_hsv;
 use anyhow::anyhow;
 use image::{DynamicImage, ImageBuffer, Rgb};
-use crate::utils::cv_ops::rgb_to_opencv_hsv;
+use std::str::FromStr;
-
-#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord)]
-pub struct HsvRange {
-    pub lower: (u8, u8, u8), // (H, S, V)
-    pub upper: (u8, u8, u8), // (H, S, V)
-}
-
 
 /// 核心区间判定辅助函数
 #[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 &&
+        h >= range.lower.0
-            s >= range.lower.1 && s <= range.upper.1 &&
+            && h <= range.upper.0
-            v >= range.lower.2 && v <= range.upper.2
+            && s >= range.lower.1
+            && s <= range.upper.1
+            && v >= range.lower.2
+            && v <= range.upper.2
     })
 }
-pub fn filter_image(image: &DynamicImage, hsv_ranges: &[HsvRange]) -> anyhow::Result<DynamicImage> {
+pub fn apply_to_image(
+    image: &DynamicImage,
+    hsv_ranges: &[HsvRange],
+) -> anyhow::Result<DynamicImage> {
     // 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();
@@ -50,6 +49,13 @@ pub fn filter_image(image: &DynamicImage, hsv_ranges: &[HsvRange]) -> anyhow::Re
 
     Ok(DynamicImage::ImageRgb8(filtered_buffer))
 }
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord)]
+pub struct HsvRange {
+    pub lower: (u8, u8, u8), // (H, S, V)
+    pub upper: (u8, u8, u8), // (H, S, V)
+}
+
 impl HsvRange {
     pub const fn new(lower: (u8, u8, u8), upper: (u8, u8, u8)) -> Self {
         Self { lower, upper }
@@ -65,7 +71,8 @@ impl HsvRange {
         }
 
         // 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 {
+        if self.lower.0 > self.upper.0 || self.lower.1 > self.upper.1 || self.lower.2 > self.upper.2
+        {
             return Err("HSV范围下界不能大于上界".to_string());
         }
 
@@ -87,25 +94,58 @@ pub enum ColorPreset {
     Custom(Vec<HsvRange>),
 }
 
-impl  ColorPreset {
+impl ColorPreset {
     /// 纯裸数据定义，没有任何结构体包装，干净利落
     /// 返回值：(范围数量, 范围数组)
     /// 完美的零成本抽象：利用常量提升将数据直接打入只读数据段 (.rodata)
     pub fn matches(&self) -> &[HsvRange] {
         match self {
             ColorPreset::Red => &[
-                HsvRange { lower: (0, 50, 50), upper: (10, 255, 255) },
+                HsvRange {
-                HsvRange { lower: (170, 50, 50), upper: (180, 255, 255) },
+                    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::Blue => &[HsvRange {
-            ColorPreset::Green  => &[HsvRange { lower: (40, 50, 50), upper: (80, 255, 255) }],
+                lower: (100, 50, 50),
-            ColorPreset::Yellow => &[HsvRange { lower: (20, 50, 50), upper: (40, 255, 255) }],
+                upper: (130, 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::Green => &[HsvRange {
-            ColorPreset::Cyan   => &[HsvRange { lower: (80, 50, 50), upper: (100, 255, 255) }],
+                lower: (40, 50, 50),
-            ColorPreset::Black  => &[HsvRange { lower: (0, 0, 0), upper: (180, 255, 50) }],
+                upper: (80, 255, 255),
-            ColorPreset::White  => &[HsvRange { lower: (0, 0, 200), upper: (180, 30, 255) }],
+            }],
-            ColorPreset::Gray   => &[HsvRange { lower: (0, 0, 50), upper: (180, 30, 200) }],
+            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,
         }
     }
@@ -204,7 +244,6 @@ impl ColorFilter for ColorPreset {
     }
 }
 
-
 /// 多路颜色“或”逻辑组合子（并集网络）
 pub struct MultiOrColorRestrict<'a> {
     pub filters: Vec<&'a dyn ColorFilter>,
@@ -246,4 +285,3 @@ macro_rules! color_any_of {
         }
     };
 }
-

ddddocr-core/src/models/ocr/executor.rs

@@ -1,25 +1,23 @@
-use crate::charset::{TokenFilter, ValidationCtx};
+use crate::models::ocr::metadata::Resize;
-use crate::model_metadata::{ModelMetadata, Resize};
+
-use crate::models::base::ModelArgs;
+use crate::models::ocr::color_filter::{HsvRange, apply_to_image};
-use crate::models::loader::{ModelLoader, ModelSession, ModelType};
+// use ddddocr_tract::session::{ModelOutput, OcrSession};
-use crate::utils::color_filter::{ColorFilter, HsvRange, filter_image};
 use crate::utils::image_io::png_rgba_white_preprocess;
 use crate::utils::image_processor::{convert_to_grayscale, resize_image};
-use anyhow::Context;
+use anyhow::Result;
-use anyhow::{Result, anyhow};
+use image::DynamicImage;
-use image::{DynamicImage, ImageBuffer, Rgb};
 use serde::Serialize;
 use std::borrow::Cow;
-use std::collections::HashSet;
 use std::fmt;
-use tract_onnx::prelude::tract_ndarray::{ArrayView2, Ix2, s};
+// use tract_onnx::prelude::tract_ndarray::{ Ix2, s};
-use tract_onnx::prelude::{
+// use tract_onnx::prelude::{DatumType, Tensor, tract_ndarray};
-    DatumType, Graph, IntoTensor, RunnableModel, Tensor, TypedFact, TypedOp, tract_ndarray, tvec,
+// ！！！【核心纠正】：彻底弃用 tract_ndarray，全线转用标准 ndarray
-};
+use ndarray::ArrayView2;
-// 引入 cv_ops 模块中的 OpenCV HSV 转换算子
+// pub enum ModelOutput {
-use crate::utils::cv_ops::rgb_to_opencv_hsv;
+//     Indices(ndarray::Array1<i64>), // 拥有完整所有权的 1维数组，可任意传递和返回
-
+//     Logits(ndarray::Array2<f32>),  // 拥有完整所有权的 2维矩阵，可任意传递和返回
-/// 推理最终输出的强类型外壳（完全 Owned，无任何生命周期，可直接转 JSON）
+// }
+use crate::{OcrEngine, OcrOutput};
 #[derive(Debug, Clone, Serialize)]
 pub enum OcrResult {
     /// 纯文本分支（对应 probability = false）
@@ -104,103 +102,40 @@ impl fmt::Display for OcrResult {
     }
 }
 
-pub struct Ocr {
+pub struct Ocr<'a> {
-    pub session: RunnableModel<TypedFact, Box<dyn TypedOp>, Graph<TypedFact, Box<dyn TypedOp>>>,
+    pub(crate) session: &'a dyn OcrEngine,
-    pub model_metadata: ModelMetadata,
+    pub(crate) png_fix: bool,
-}
+    pub(crate) probability: bool,
-impl ModelSession for Ocr {
-    fn get_model_type(&self) -> ModelType {
-        todo!("使用thiserror作为错误处理的库,thiserror 专门用于开发库（Library）");
-    }
-    fn desc(&self) -> String {
-        "Ocr Model 加载成功".to_string()
-    }
-}
-impl Ocr {
-    pub fn new(model_path: String, model_metadata: ModelMetadata) -> Result<Self, anyhow::Error> {
-        
-        let session = ModelLoader::load_model(&model_path)?.session;
-        Ok(Self {
-            session,
-            model_metadata,
-        })
-    }
-    /// 对应 Python 的 _inference
-    fn inference(&self, tensor: Tensor) -> anyhow::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.swap_remove(0).into_tensor())
-    }
-
-    pub fn predictor(&'_ self) -> OcrPredictor<'_> {
-        OcrPredictor::new(self)
-    }
-}
-
-pub struct OcrPredictor<'a> {
-    ocr: &'a Ocr,
-    /// 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: Result<Option<Vec<HsvRange>>, String>,
+    pub(crate) final_color_ranges: Result<Option<Vec<HsvRange>>, String>,
 
     /// 字符集范围
-    charset_restrict: Option<Vec<usize>>,
+    pub(crate) final_charset_indices: Option<Vec<usize>>,
 }
 
-impl<'a> OcrPredictor<'a> {
+impl<'a> Ocr<'a> {
     // 初始化任务，设置默认参数
-    pub fn new(ocr: &'a Ocr) -> Self {
+
-        Self {
+    pub fn new(session: &'a dyn OcrEngine) -> Self {
-            ocr,
+        Ocr {
+            session,
             png_fix: false, // 默认值
             probability: false,
-            color_filter: Ok(None),
+            final_color_ranges: Ok(None),
-            charset_restrict: None,
+            final_charset_indices: None,
         }
     }
-    pub fn png_fix(mut self, value: bool) -> Self {
-        self.png_fix = value;
-        self
-    }
-    pub fn probability(mut self, value: bool) -> Self {
-        self.probability = value;
-        self
-    }
-
-    pub fn color_filter(mut self, filter: &dyn ColorFilter) -> Self {
-        // 一句话把活全包了！错误信息无缝传递，完美熔断
-        match filter.collect_to_vec() {
-            Ok(new_ranges) => self.color_filter = Ok(new_ranges),
-            Err(err_msg) => self.color_filter = Err(err_msg), // 校验失败，Builder 正式中毒
-        }
-
-        self
-    }
-
-    pub fn charset_restrict(mut self, restrict: &dyn TokenFilter) -> Self {
-        let charset = &self.ocr.model_metadata.charset;
-        let tokens = &charset.tokens;
-        self.charset_restrict = restrict.apply_to_charset(tokens);
-        self
-    }
 }
-impl<'a> OcrPredictor<'a> {
+impl<'a> Ocr<'a> {
-    pub fn predict(self, image: &DynamicImage) -> anyhow::Result<OcrResult> {
+    pub fn predict(&self, image: &DynamicImage) -> anyhow::Result<OcrResult> {
-        println!("当前颜色过滤器状态: {:?}", self.color_filter);
+        println!("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.color_filter {
+        let img_cow = match &self.final_color_ranges {
             Err(err_msg) => {
                 return Err(anyhow::anyhow!(
                     "颜色过滤器初始化失败，全链路短路: {}",
@@ -213,34 +148,34 @@ impl<'a> OcrPredictor<'a> {
             }
             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 = filter_image(image, ranges)?;
+                let filtered_img = apply_to_image(image, ranges)?;
                 Cow::Owned(filtered_img)
             }
         };
         let tensor = self.preprocess_image(&img_cow)?;
 
-        let raw_tensor = self.ocr.inference(tensor)?;
+        let raw_tensor = self.session.inference(tensor)?;
 
         // 3. 后处理分流：直接返回 OcrResult
-        let ocr_output = match raw_tensor.datum_type() {
+        // let ocr_output = match raw_tensor.datum_type() {
-            DatumType::I64 => self.process_i64_tensor(raw_tensor)?,
+        //     DatumType::I64 => self.process_i64_tensor(raw_tensor)?,
-            DatumType::F32 => self.process_f32_tensor(raw_tensor)?,
+        //     DatumType::F32 => self.process_f32_tensor(raw_tensor)?,
-            _ => OcrResult::Unsupported {
+        //     _ => OcrResult::Unsupported {
-                message: format!("不支持的模型输出数据类型: {:?}", raw_tensor.datum_type()),
+        //         message: format!("不支持的模型输出数据类型: {:?}", raw_tensor.datum_type()),
-            },
+        //     },
-        };
+        // };
 
         // let raw_indices = self.ocr.extract_indices_from_tensor(&raw_tensor)?;
         // // 步骤 2: 将索引切片 `&[i64]` 传给解码器进行 CTC 去重和字符映射
         // let final_text = self.ctc_decode_to_string(&raw_indices);
-
+        let ocr_output = self.process_model_output(raw_tensor);
-        Ok(ocr_output)
+        ocr_output
     }
     /// 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) -> anyhow::Result<Tensor> {
+    fn preprocess_image(&self, img: &DynamicImage) -> anyhow::Result<ndarray::Array4<f32>> {
         // 1. 获取模型元数据配置
-        let meta = &self.ocr.model_metadata;
+        let meta = self.session.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)
@@ -270,12 +205,12 @@ impl<'a> OcrPredictor<'a> {
         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 tensor = match meta.channel {
+        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);
 
-                let array = tract_ndarray::Array4::from_shape_fn(
+                let array = 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;
@@ -284,14 +219,14 @@ impl<'a> OcrPredictor<'a> {
                         norm.normalize(pixel)
                     },
                 );
-                Tensor::from(array)
+                array
             }
 
             // --- 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();
 
-                let array = tract_ndarray::Array4::from_shape_fn(
+                let array = 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;
@@ -300,13 +235,14 @@ impl<'a> OcrPredictor<'a> {
                         norm.normalize(pixel)
                     },
                 );
-                Tensor::from(array)
+                // Tensor::from(array)
+                array
             }
 
             _ => return Err(anyhow::anyhow!("不支持的通道数配置: {}", meta.channel)),
         };
-
+        Ok(array4)
-        Ok(tensor)
+        // Ok(tensor)
 
         // let h = 64u32;
         // let w = (current_img.width() as f32 * (h as f32 / current_img.height() as f32)) as u32;
@@ -327,14 +263,65 @@ impl<'a> OcrPredictor<'a> {
         //
         // Ok(tensor)
     }
+
+    // 这段代码未来直接放入 ddddocr-core
+    fn process_model_output(&self, output: OcrOutput) -> anyhow::Result<OcrResult> {
+        match output {
+            OcrOutput::Indices(array1) => {
+                // 对应你原来的 process_i64_tensor
+                let slice = array1
+                    .as_slice()
+                    .ok_or_else(|| anyhow::anyhow!("内存不连续，无法执行零拷贝解码"))?;
+                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) => {
+                // 对应你原来的 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> OcrPredictor<'a> {
+impl<'a> Ocr<'a> {
     fn is_valid_indices(&self, idx: usize) -> bool {
-        if idx >= self.ocr.model_metadata.charset.size() {
+        if idx >= self.session.metadata().charset.size() {
             return false;
         }
 
-        match &self.charset_restrict {
+        match &self.final_charset_indices {
             Some(v) => v.binary_search(&idx).is_ok(),
             None => true,
         }
@@ -342,9 +329,9 @@ impl<'a> OcrPredictor<'a> {
     /// 【按需延迟打印】：当用户真的需要“知道当前有哪些限制字符”时，一秒反查并打印
     /// 这里的 &str 完美借用了自 tokens，依然是彻底的零拷贝！
     pub fn valid_tokens(&self) -> Vec<&str> {
-        let charset = &self.ocr.model_metadata.charset;
+        let charset = &self.session.metadata().charset;
         let tokens = &charset.tokens;
-        match &self.charset_restrict {
+        match &self.final_charset_indices {
             Some(indices) => indices
                 .iter()
                 .filter_map(|&idx| tokens.get(idx).map(|cow| cow.as_ref()))
@@ -354,9 +341,9 @@ impl<'a> OcrPredictor<'a> {
         }
     }
     pub fn valid_size(&self) -> usize {
-        match &self.charset_restrict {
+        match &self.final_charset_indices {
             Some(indices) => indices.len(),
-            None => self.ocr.model_metadata.charset.tokens.len(),
+            None => self.session.metadata().charset.tokens.len(),
         }
     }
     /// 变体 B 核心处理器：单次遍历 2D 视图，融合计算 Softmax、Argmax、置信度并输出概率大包
@@ -368,7 +355,7 @@ impl<'a> OcrPredictor<'a> {
         let classes = matrix_view.ncols();
 
         // 1. 预分配满额概率矩阵内存
-        let mut prob_matrix = tract_ndarray::Array2::<f32>::zeros((steps, classes));
+        let mut prob_matrix = ndarray::Array2::<f32>::zeros((steps, classes));
         let mut predicted_indices = Vec::with_capacity(steps);
         let mut confidence_sum = 0.0f32;
 
@@ -413,98 +400,98 @@ impl<'a> OcrPredictor<'a> {
         (probabilities_list, confidence, predicted_indices)
     }
     /// 变体 A 专属提取器：直接从 I64 Tensor 零拷贝提取 CTC 文本与初始概率包
-    fn process_i64_tensor(&self, raw_tensor: Tensor) -> anyhow::Result<OcrResult> {
+    // fn process_i64_tensor(&self, raw_tensor: Tensor) -> anyhow::Result<OcrResult> {
-        // 1. 拿到底层的动态维度只读视图
+    //     // 1. 拿到底层的动态维度只读视图
-        let view = raw_tensor.to_array_view::<i64>()?;
+    //     let view = raw_tensor.to_array_view::<i64>()?;
-
+    //
-        // 2. 索要底层连续的只读切片引用
+    //     // 2. 索要底层连续的只读切片引用
-        let slice = view
+    //     let slice = view
-            .as_slice()
+    //         .as_slice()
-            .ok_or_else(|| anyhow::anyhow!("I64 模型输出内存不连续，无法执行零拷贝解码"))?;
+    //         .ok_or_else(|| anyhow::anyhow!("I64 模型输出内存不连续，无法执行零拷贝解码"))?;
-
+    //
-        // 3. 直接喂给 CTC 解码器（无任何物理克隆开销）
+    //     // 3. 直接喂给 CTC 解码器（无任何物理克隆开销）
-        let final_text = self.ctc_decode_to_string(slice);
+    //     let final_text = self.ctc_decode_to_string(slice);
-
+    //
-        // 4. 组装返回
+    //     // 4. 组装返回
-        if self.probability {
+    //     if self.probability {
-            Ok(OcrResult::Probability {
+    //         Ok(OcrResult::Probability {
-                text: final_text,
+    //             text: final_text,
-                probabilities: vec![], // I64 模型物理上丢失了全量 Logits 分值网，降级处理
+    //             probabilities: vec![], // I64 模型物理上丢失了全量 Logits 分值网，降级处理
-                confidence: 1.0,       // 判定即百分之百置信
+    //             confidence: 1.0,       // 判定即百分之百置信
-            })
+    //         })
-        } else {
+    //     } else {
-            Ok(OcrResult::Text(final_text))
+    //         Ok(OcrResult::Text(final_text))
-        }
+    //     }
-    }
+    // }
-    /// 变体二（F32）的总体管线：负责降维，并分流文本和概率
+    // /// 变体二（F32）的总体管线：负责降维，并分流文本和概率
-    fn process_f32_tensor(&self, raw_tensor: Tensor) -> anyhow::Result<OcrResult> {
+    // fn process_f32_tensor(&self, raw_tensor: Tensor) -> anyhow::Result<OcrResult> {
-        let shape = raw_tensor.shape();
+    //     let shape = raw_tensor.shape();
-        println!("模型输出shape数据: {:?}", shape);
+    //     println!("模型输出shape数据: {:?}", shape);
-        let view = raw_tensor.to_array_view::<f32>()?;
+    //     let view = raw_tensor.to_array_view::<f32>()?;
-
+    //
-        // 1. 极其纯粹的、无拷贝的多维 Shape 压扁清洗
+    //     // 1. 极其纯粹的、无拷贝的多维 Shape 压扁清洗
-        let (steps, classes, data_dyn_view) = match shape.len() {
+    //     let (steps, classes, data_dyn_view) = match shape.len() {
-            3 => {
+    //         3 => {
-                if shape[1] == 1 {
+    //             if shape[1] == 1 {
-                    // 形状: [Steps, 1, Classes] -> 你的原有逻辑
+    //                 // 形状: [Steps, 1, Classes] -> 你的原有逻辑
-                    (shape[0], shape[2], view.into_dyn())
+    //                 (shape[0], shape[2], view.into_dyn())
-                } else if shape[0] == 1 {
+    //             } else if shape[0] == 1 {
-                    // 形状: [1, Steps, Classes] -> 另一种常见导出格式
+    //                 // 形状: [1, Steps, Classes] -> 另一种常见导出格式
-                    (shape[1], shape[2], view.into_dyn())
+    //                 (shape[1], shape[2], view.into_dyn())
-                } else {
+    //             } else {
-                    // 默认取第一个 batch: [Batch, Steps, Classes]
+    //                 // 默认取第一个 batch: [Batch, Steps, Classes]
-                    // 使用 slice 对应 Python 的 output[0, :, :]
+    //                 // 使用 slice 对应 Python 的 output[0, :, :]
-                    let sliced = view.slice(s![0, .., ..]);
+    //                 let sliced = view.slice(s![0, .., ..]);
-                    (shape[1], shape[2], sliced.into_dyn())
+    //                 (shape[1], shape[2], sliced.into_dyn())
-                }
+    //             }
-            }
+    //         }
-            // 形状: [Steps, Classes] -> 已经剥离了 Batch 维度
+    //         // 形状: [Steps, Classes] -> 已经剥离了 Batch 维度
-            2 => (shape[0], shape[1], view.into_dyn()),
+    //         2 => (shape[0], shape[1], view.into_dyn()),
-            // 形状: [Classes] -> 单字符输出（对应 Python 的 ndim == 0 保护逻辑）
+    //         // 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
+    //         // 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], view.into_dyn()),
+    //         1 => (1, shape[0], view.into_dyn()),
-            _ => return Err(anyhow::anyhow!("不支持的输出维度: {:?}", shape)),
+    //         _ => return Err(anyhow::anyhow!("不支持的输出维度: {:?}", shape)),
-        };
+    //     };
-        let matrix_cow = data_dyn_view
+    //     let matrix_cow = data_dyn_view
-            .to_shape(Ix2(steps, classes))
+    //         .to_shape(Ix2(steps, classes))
-            .map_err(|e| anyhow::anyhow!("转换为2D静态矩阵失败: {:?}", e))?;
+    //         .map_err(|e| anyhow::anyhow!("转换为2D静态矩阵失败: {:?}", e))?;
-
+    //
-        let matrix_view: ArrayView2<f32> = matrix_cow.view();
+    //     let matrix_view: ArrayView2<f32> = matrix_cow.view();
-
+    //
-        // 2. 根据业务参数明确分流
+    //     // 2. 根据业务参数明确分流
-        if self.probability {
+    //     if self.probability {
-            // 走向 B1：调用刚刚拆分出来的“全量概率计算器”
+    //         // 走向 B1：调用刚刚拆分出来的“全量概率计算器”
-            let (probabilities_list, confidence, predicted_indices) =
+    //         let (probabilities_list, confidence, predicted_indices) =
-                self.compute_f32_full_probability(matrix_view);
+    //             self.compute_f32_full_probability(matrix_view);
-            // 5. 执行 CTC 解码
+    //         // 5. 执行 CTC 解码
-            let final_text = self.ctc_decode_to_string(&predicted_indices);
+    //         let final_text = self.ctc_decode_to_string(&predicted_indices);
-
+    //
-            Ok(OcrResult::Probability {
+    //         Ok(OcrResult::Probability {
-                text: final_text,
+    //             text: final_text,
-                probabilities: probabilities_list,
+    //             probabilities: probabilities_list,
-                confidence: confidence as f64,
+    //             confidence: confidence as f64,
-            })
+    //         })
-        } else {
+    //     } else {
-            // 走向 B2：极速免 Softmax 提取纯文本（代码保持原地提取，简单短小不需要再拆）
+    //         // 走向 B2：极速免 Softmax 提取纯文本（代码保持原地提取，简单短小不需要再拆）
-            let predicted_indices: Vec<i64> = matrix_view
+    //         let predicted_indices: Vec<i64> = matrix_view
-                .outer_iter()
+    //             .outer_iter()
-                .map(|row| {
+    //             .map(|row| {
-                    row.iter()
+    //                 row.iter()
-                        .enumerate()
+    //                     .enumerate()
-                        .max_by(|(_, a), (_, b)| a.total_cmp(b))
+    //                     .max_by(|(_, a), (_, b)| a.total_cmp(b))
-                        .map(|(idx, _)| idx as i64)
+    //                     .map(|(idx, _)| idx as i64)
-                        .unwrap_or(0)
+    //                     .unwrap_or(0)
-                })
+    //             })
-                .collect();
+    //             .collect();
-
+    //
-            let final_text = self.ctc_decode_to_string(&predicted_indices);
+    //         let final_text = self.ctc_decode_to_string(&predicted_indices);
-            Ok(OcrResult::Text(final_text))
+    //         Ok(OcrResult::Text(final_text))
-        }
+    //     }
-    }
+    // }
     /// 获取有效字符索引列表 (用于外部验证或过滤)
     fn ctc_decode_to_string(&self, predicted_indices: &[i64]) -> String {
         println!("indices模型输出原始数据: {:?}", predicted_indices);
-        let charset = &self.ocr.model_metadata.charset;
+        let charset = &self.session.metadata().charset;
         let tokens = &charset.tokens;
         // let valid_indices = &charset.valid_indices;
 
@@ -532,7 +519,7 @@ impl<'a> OcrPredictor<'a> {
 
             // 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.charset_restrict {
+            if let Some(ref indices) = self.final_charset_indices {
                 if indices.binary_search(&u_idx).is_err() {
                     continue;
                 }

ddddocr-core/src/models/ocr/metadata.rs

@@ -1,11 +1,72 @@
-use crate::charset::{CHARSET_BETA, CHARSET_OLD, Charset};
+use anyhow::{anyhow, Result};
-use anyhow::{Result, anyhow};
 use serde::Deserialize;
 use std::borrow::Cow;
-use std::collections::{HashMap, HashSet};
+use std::collections::HashMap;
-use std::fs::File;
+
-use std::io::Read;
+// ==========================================
-use std::path::Path;
+// 3. 字符集核心结构体 (重命名为 Charset)
+// ==========================================
+#[derive(Debug, Clone)]
+pub struct Charset {
+    // 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>>,
+    // 反向查找表，保证字符转索引为 O(1)
+    pub char_to_idx: HashMap<Cow<'static, str>, usize>,
+    // 当前处于激活状态的有效索引缓存 (用于 CTC 解码前的过滤加速)
+    // pub valid_indices: HashSet<usize>,
+}
+
+impl Charset {
+    // 内部底层统一收拢构造
+    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)
+            // char_to_idx.entry(token.to_string()).or_insert(idx);
+        }
+
+        Self {
+            tokens,
+            char_to_idx,
+        }
+    }
+
+    // --- 业务策略方法 ---
+
+    /// 将字符转为索引，不存在返回 -1 (保持与原 Python 库行为一致)
+    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
+        }
+    }
+
+    /// 将索引转为字符引用，零拷贝。若越界返回 None
+    pub fn index_to_char_ref(&self, index: usize) -> Option<&str> {
+        self.tokens.get(index).map(|cow| cow.as_ref())
+    }
+
+    pub fn is_valid_char(&self, char_str: &str) -> bool {
+        self.char_to_idx.get(char_str).is_some()
+    }
+    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(),)
+    }
+}
+
+
+
 // =====================================================================
 // 1. 辅助定义的枚举与结构体
 // =====================================================================
@@ -74,28 +135,6 @@ pub struct ModelMetadata {
 
 impl ModelMetadata {
     // --- 优雅的工厂模式构造器 ---
-    /// 从预设的旧版字符集创建
-    pub fn from_builtin_old() -> Self {
-        Self::from_static_slice(
-            CHARSET_OLD,
-            false,
-            Resize::DynamicWidth(64),
-            1,
-            Normalization::ZeroToOne,
-        )
-    }
-
-    /// 从预设的 Beta 版字符集创建
-    pub fn from_builtin_beta() -> Self {
-        Self::from_static_slice(
-            CHARSET_BETA,
-            false,
-            Resize::DynamicWidth(64),
-            1,
-            Normalization::MinusOneToOne,
-        )
-    }
-
     /// 通用的静态切片转换构造器
     pub fn from_static_slice(
         slice: &[&'static str],
@@ -113,19 +152,8 @@ impl ModelMetadata {
             normalization,
         }
     }
-
+    pub fn from_json_str(json_str: &str) -> Result<Self> {
-    /// 从外部外部 JSON 文件动态加载字符集
+        let dto: ModelMetadataDto = serde_json::from_str(json_str)
-    pub fn from_json_file<P: AsRef<Path>>(path: P) -> Result<Self> {
-        let path = path.as_ref();
-        if !path.exists() {
-            return Err(anyhow!("模型元数据配置文件不存在: {:?}", path));
-        }
-
-        let mut file = File::open(path)?;
-        let mut content = String::new();
-        file.read_to_string(&mut content)?;
-
-        let dto: ModelMetadataDto = serde_json::from_str(&content)
             .map_err(|e| anyhow!("JSON 反序列化失败，请检查字段是否完整: {}", e))?;
 
         // 1. 将 DTO 的字符串数组转化为强类型的 Charset
@@ -163,4 +191,10 @@ impl ModelMetadata {
             normalization: dto.normalization,
         })
     }
+    /// 机制 2：从内存字节流加载（极大地方便 include_bytes! 或网络下载）
+    pub fn from_json_bytes(bytes: &[u8]) -> Result<Self> {
+        let json_str = std::str::from_utf8(bytes)
+            .map_err(|e| anyhow!("JSON 字节流不是合法的 UTF-8 编码: {}", e))?;
+        Self::from_json_str(json_str)
+    }
 }

ddddocr-core/src/models/ocr/mod.rs

@@ -0,0 +1,9 @@
+mod builder;
+mod executor;
+pub mod metadata;
+pub mod color_filter;
+mod token_filter;
+
+pub use builder::OcrBuilder;
+pub use executor::{Ocr, OcrResult};
+// pub use ddddocr_tract::session::OcrSession;

ddddocr-core/src/models/ocr/token_filter.rs

@@ -0,0 +1,146 @@
+use std::borrow::Cow;
+
+/// 字符集范围限制枚举
+pub struct ValidationCtx<'a> {
+    pub text: &'a str,   // 当前 Token 的文本内容
+    pub token_id: usize, // 当前 Token 的 ID 索引
+}
+
+/// 统一的约束接口
+pub trait TokenFilter {
+    fn matches(&self, ctx: &ValidationCtx) -> bool;
+    /// 预估容量提示，帮助精准开辟 Vec 内存
+    fn estimated_capacity(&self) -> usize {
+        128
+    }
+    /// 【新引入的架构级核心方法】
+    /// 统一接管全量字符集的密集遍历、CTC Blank放行、去重、排序及空交集退化兜底
+    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 {
+            println!("警告：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)
+        }
+    }
+}
+
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum CharRestrict {
+    Digit,
+    Lowercase,
+    Uppercase,
+    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,
+        }
+    }
+}
+
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum IdRestrict {
+    TopN(usize),
+    IdRange(std::ops::Range<usize>),
+    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> {
+    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()
+    }
+}
+// =====================================================================
+// 声明式宏：替代 `+` 运算符，解决组合扩展痛苦
+// =====================================================================
+#[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 $crate::TokenFilter
+    };
+
+    // 场景 B：如果用户传入了多个规则，自动织成一张静态组合网
+    ($($filter:expr),+ $(,)?) => {
+        &$crate::MultiOrRestrict {
+            filters: vec![ $( &$filter as &dyn $crate::TokenFilter ),+ ]
+        }
+    };
+}

ddddocr-core/src/utils/image_io.rs

@@ -3,7 +3,7 @@ use base64::{Engine as _, engine::general_purpose};
 use image::{DynamicImage, GenericImageView, ImageBuffer, ImageFormat, Luma, Rgb, RgbImage, Rgba};
 use std::fs;
 use std::path::{Path, PathBuf};
-use tract_onnx::prelude::tract_ndarray::{Array3, ArrayD, ArrayViewD};
+use ndarray::{Array3, ArrayD, ArrayViewD};
 #[derive(Debug)]
 pub enum ColorMode {
     RGB,

ddddocr-core/src/utils/image_proc.rs

@@ -1,7 +1,8 @@
 use image::{ImageBuffer, Luma};
+use ndarray::{Array2, Array3, ArrayView2, ArrayView3, azip};
 use std::cmp::{max, min};
-use tract_onnx::prelude::tract_ndarray::{Array2, Array3, ArrayView2, ArrayView3, azip};
 
+// 模拟openCV
 /// 1. 计算两个数组的绝对差值 (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
@@ -72,6 +73,9 @@ pub fn find_contours_and_max(labelled: &ImageBuffer<Luma<u32>, Vec<u32>>) -> Opt
         Some(max_label)
     }
 }
+/// 根据目标连通域标签，计算其在图像中的外接矩形边界框（对应 `cv2.boundingRect`）
+///
+/// 返回格式: `(min_x, min_y, width, height)`
 pub fn bounding_rect(
     labelled: &ImageBuffer<Luma<u32>, Vec<u32>>,
     max_label: u32,
@@ -95,13 +99,22 @@ pub fn bounding_rect(
     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)
 }
+
+/// 高性能转换：将 `ndarray` 2D 灰度视图规整为 `image::ImageBuffer` 格式
+///
+/// 放弃低效的逐像素显式嵌套循环，采用原生内存池直接构造，减少寻址开销
 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 {
@@ -126,7 +139,7 @@ pub fn rgb_to_opencv_hsv(r: u8, g: u8, b: u8) -> (u8, u8, u8) {
     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 mut h = if delta == 0.0 {
+    let h = if delta == 0.0 {
         0.0
     } else if max == r_f {
         let mut diff = (g_f - b_f) / delta;

ddddocr-core/src/utils/image_processor.rs

@@ -1,5 +1,7 @@
-use image::{DynamicImage, GrayImage, imageops::FilterType};
+use image::{DynamicImage, GrayImage, imageops::FilterType, Rgb, ImageBuffer};
-use anyhow::Result;
+use anyhow::{anyhow, Result};
+use crate::models::ocr::color_filter::HsvRange;
+use crate::utils::image_proc::rgb_to_opencv_hsv;
 
 /// 对应 Python 的 convert_to_grayscale
 /// 将图像转换为灰度图 (L模式)
@@ -35,3 +37,4 @@ pub fn resize_image(
 //         FilterType::Lanczos3
 //     )
 // }
+

ddddocr-core/src/utils/mod.rs

@@ -0,0 +1,7 @@
+pub mod image_io;
+pub mod image_processor;
+pub mod image_proc;
+mod tensor_transform;
+// 对外统一暴露干净的 API 语义层
+pub use image_proc::*;
+pub use tensor_transform::normalize_ocr_logits;

ddddocr-core/src/utils/tensor_transform.rs

@@ -0,0 +1,38 @@
+use ndarray::s;
+use crate::error::{DdddError,Result};
+use crate::OcrOutput;
+/// 🌟 核心层复用资产：将异构的动态维度矩阵转化为标准 OCR 2D Logits 矩阵
+pub fn normalize_ocr_logits(array: ndarray::ArrayD<f32>, shape: &[usize]) -> Result<OcrOutput> {
+    let (steps, classes, data_dyn_view) = match shape.len() {
+        3 => {
+            if shape[1] == 1 {
+                (shape[0], shape[2], array)
+            } else if shape[0] == 1 {
+                (shape[1], shape[2], array)
+            } else {
+                // 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())
+            }
+        }
+        2 => (shape[0], shape[1], array),
+        1 => (1, shape[0], array),
+        _ => {
+            return Err(DdddError::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(|_| DdddError::DimensionMismatch {
+            expected: format!("无法将形状调整为 [{}, {}]", steps, classes),
+            actual: shape.to_vec(),
+        })?
+        .to_owned();
+
+    Ok(OcrOutput::Logits(matrix_cow))
+}

ddddocr-tract/Cargo.toml

@@ -0,0 +1,24 @@
+[package]
+name = "ddddocr-tract"
+version = { workspace = true }
+edition = { workspace = true }
+license = { workspace = true }
+
+[dependencies]
+
+ddddocr-core = { path = "../ddddocr-core" } # 引入兄弟库
+tract-onnx = "0.21.10"
+anyhow = "1.0.102"
+image = { workspace = true }
+base64 = "0.22.1"
+imageproc = { version = "0.26.2", default-features = true }
+serde = { workspace = true }
+serde_json = "1.0.150"
+ndarray = { workspace = true } # 继承自工作空间
+thiserror = { workspace = true } # 刚好可以开始接入你需要的标准库错误处理
+
+
+
+[features]
+default = []
+embed-models = [] # 这是一个留给有特殊需求、且自己下载了模型放入 models/ 目录的人的后门

ddddocr-tract/src/det/mod.rs

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

ddddocr-tract/src/det/session.rs

@@ -0,0 +1,80 @@
+use crate::loader::{ModelLoader, ModelSession, ModelType};
+use anyhow::Context;
+use ddddocr_core::error::{DdddError, Result};
+use ddddocr_core::{DetEngine, DetOutput, InferenceEngine};
+use ndarray::Ix3;
+use std::path::Path;
+use tract_onnx::prelude::{Graph, IntoTensor, RunnableModel, Tensor, TypedFact, TypedOp, tvec};
+#[derive(Debug)]
+pub struct DetSession {
+    pub session: RunnableModel<TypedFact, Box<dyn TypedOp>, Graph<TypedFact, Box<dyn TypedOp>>>,
+}
+
+impl ModelSession for DetSession {
+    fn get_model_type(&self) -> ModelType {
+        todo!()
+    }
+    fn desc(&self) -> String {
+        "Detection Model 加载成功".to_string()
+    }
+}
+
+impl DetSession {
+    pub fn new<P>(model_path: P) -> Result<Self>
+    where
+        P: AsRef<Path>,
+    {
+        let session = ModelLoader::model_for_path(&model_path)?.session;
+        Ok(Self { session })
+    }
+
+    pub fn model_from_bytes(model_bytes: &[u8]) -> Result<Self> {
+        let session = ModelLoader::model_from_bytes(model_bytes)?.session;
+        Ok(Self { session })
+    }
+    // pub fn inference(&self, tensor: Tensor) -> anyhow::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.ocr.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.swap_remove(0).into_tensor())
+    // }
+}
+
+impl InferenceEngine for DetSession {
+    type Output = DetOutput; // 明确绑定 OCR 小枚举
+    fn inference(&self, input_array: ndarray::Array4<f32>) -> Result<Self::Output> {
+        // 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.ocr.run(tvec!(tensor.into()))?;
+        let tensor = Tensor::from(input_array);
+
+        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.swap_remove(0).into_tensor())
+        let raw_tensor = result.swap_remove(0).into_tensor();
+        let array_d = raw_tensor
+            .into_array::<f32>()
+            .context("Tract 实体张量无法转换为 ndarray::ArrayD")?;
+        // 提前利用克隆(Clone)备份好当前未转维度前的真实 shape (Vec<usize>)
+        let actual_shape = array_d.shape().to_vec();
+
+        let array3 =
+            array_d
+                .into_dimensionality::<Ix3>()
+                .map_err(|_| DdddError::DimensionMismatch {
+                    expected: "3D 检测矩阵 [Batch, Box_Count, Box_Attributes]".to_string(),
+                    actual: actual_shape, // 优雅降维失败时动态捕获
+                })?;
+        Ok(DetOutput::Detection(array3))
+
+        // 在引擎内部消化掉 DatumType 强耦合
+    }
+}
+
+impl DetEngine for DetSession {}

ddddocr-tract/src/lib.rs

@@ -0,0 +1,6 @@
+mod det;
+pub mod loader;
+mod ocr;
+
+pub use det::session::DetSession;
+pub use ocr::session::OcrSession;

ddddocr-tract/src/loader.rs

@@ -0,0 +1,52 @@
+use anyhow::Context;
+use ddddocr_core::error::Result;
+use std::io::Cursor;
+use tract_onnx::onnx;
+use tract_onnx::prelude::*; // 引入核心层的统一错误类型
+/// OCR 模型：包含路径和字符集
+
+pub enum ModelType {
+    Ocr,
+    Det,
+    Custom,
+}
+// 定义统一的 trait
+pub trait ModelSession {
+    fn get_model_type(&self) -> ModelType;
+    fn desc(&self) -> String;
+}
+
+pub struct ModelLoader {
+    pub session: RunnableModel<TypedFact, Box<dyn TypedOp>, Graph<TypedFact, Box<dyn TypedOp>>>,
+}
+
+impl ModelLoader {
+    pub fn model_for_path<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()
+            .with_context(|| "优化 Tract 模型图失败")?
+            .into_runnable()
+            .with_context(|| "构建可运行 Tract 实例失败")?;
+        Ok(Self { session })
+    }
+    /// 策略 B：从内存字节流加载模型（配合 include_bytes! 使用）
+    pub fn model_from_bytes(model_bytes: &[u8]) -> Result<Self> {
+        // 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)
+            .with_context(|| "从内存字节流解析 ONNX 模型失败")?
+            .into_optimized()
+            .with_context(|| "优化 Tract 模型图失败")?
+            .into_runnable()
+            .with_context(|| "构建可运行 Tract 实例失败")?;
+
+        Ok(Self { session })
+    }
+}

ddddocr-tract/src/ocr/mod.rs

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

ddddocr-tract/src/ocr/session.rs

@@ -0,0 +1,125 @@
+use crate::loader::ModelLoader;
+use anyhow::Context;
+use ddddocr_core::error::{DdddError, Result};
+use ddddocr_core::{InferenceEngine, ModelMetadata, OcrEngine, OcrOutput};
+use ndarray::s;
+use std::path::Path;
+use tract_onnx::prelude::DatumType;
+use tract_onnx::prelude::{Graph, IntoTensor, RunnableModel, Tensor, TypedFact, TypedOp, tvec};
+
+pub struct OcrSession {
+    pub session: RunnableModel<TypedFact, Box<dyn TypedOp>, Graph<TypedFact, Box<dyn TypedOp>>>,
+    pub model_metadata: ModelMetadata,
+}
+impl OcrSession {
+    pub fn new<P>(model_path: P, model_metadata: ModelMetadata) -> Result<Self>
+    where
+        P: AsRef<Path>,
+    {
+        let session = ModelLoader::model_for_path(model_path)?.session;
+        Ok(Self {
+            session,
+            model_metadata,
+        })
+    }
+
+    pub fn model_from_bytes(model_bytes: &[u8], model_metadata: ModelMetadata) -> Result<Self> {
+        let session = ModelLoader::model_from_bytes(model_bytes)?.session;
+        Ok(Self {
+            session,
+            model_metadata,
+        })
+    }
+}
+impl OcrEngine for OcrSession {
+    fn metadata(&self) -> &ModelMetadata {
+        &self.model_metadata
+    }
+}
+impl InferenceEngine for OcrSession {
+    type Output = OcrOutput;
+    /// 对应 Python 的 _inference
+    fn inference(&self, input_array: ndarray::Array4<f32>) -> Result<Self::Output> {
+        // 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.ocr.run(tvec!(tensor.into()))?;
+        let tensor = Tensor::from(input_array);
+
+        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.swap_remove(0).into_tensor())
+        let raw_tensor = result.swap_remove(0).into_tensor();
+        // 在引擎内部消化掉 DatumType 强耦合
+        match raw_tensor.datum_type() {
+            DatumType::I64 => {
+                let array_d = raw_tensor
+                    .into_array::<i64>()
+                    .context("Tract 无法获取 i64 内存视图")?;
+                // 🌟 提前提取真实维度
+                let actual_shape = array_d.shape().to_vec();
+                // 转成标准的 Array1 传给 core
+                let array1 = array_d
+                    .to_owned()
+                    .into_dimensionality::<ndarray::Ix1>()
+                    .map_err(|_| DdddError::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();
+                println!("模型输出shape数据: {:?}", shape);
+                let view = raw_tensor
+                    .to_array_view::<f32>()
+                    .context("Tract 无法获取 f32 内存视图")?;
+
+                // 1. 极其纯粹的、无拷贝的多维 Shape 压扁清洗
+                let (steps, classes, data_dyn_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())
+                        }
+                    }
+                    // 形状: [Steps, Classes] -> 已经剥离了 Batch 维度
+                    2 => (shape[0], shape[1], view.into_dyn()),
+                    // 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], view.into_dyn()),
+                    _ => {
+                        return Err(DdddError::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(|_| DdddError::DimensionMismatch {
+                        expected: format!("无法将形状调整为 [{}, {}]", steps, classes),
+                        actual: shape.to_vec(),
+                    })?
+                    .to_owned(); // 转换为 Owned，断开与 tract 内存生命周期的绑定，方便传递给 core
+
+                Ok(OcrOutput::Logits(matrix_cow))
+            }
+            _ => Err(
+                // anyhow::anyhow!("不支持的模型输出数据类型: {:?}",raw_tensor.datum_type())
+                DdddError::UnknownOutputFormat,
+            ),
+        }
+    }
+}

ddddocr-tract/tests/char_slice.rs

@@ -1,3 +1,10 @@
+use std::borrow::Cow;
+use std::fs::File;
+use std::path::Path;
+use anyhow::anyhow;
+use ddddocr_core::models::ocr::metadata::Charset;
+use ddddocr_core::models::ocr::metadata::{Normalization, Resize};
+
 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",
@@ -517,212 +524,77 @@ pub const CHARSET_BETA: &[&str] = &[
 
 pub const CHARSET_OLD: &[&str] = &["", "笤", "谴", "膀", "荔"];
 
-use std::borrow::Cow;
-use std::collections::{HashMap, HashSet};
 
-/// 字符集范围限制枚举
-pub struct ValidationCtx<'a> {
-    pub text: &'a str,   // 当前 Token 的文本内容
-    pub token_id: usize, // 当前 Token 的 ID 索引
-}
 
-/// 统一的约束接口
-pub trait TokenFilter {
-    fn matches(&self, ctx: &ValidationCtx) -> bool;
-    /// 预估容量提示，帮助精准开辟 Vec 内存
-    fn estimated_capacity(&self) -> usize {
-        128
-    }
-    /// 【新引入的架构级核心方法】
-    /// 统一接管全量字符集的密集遍历、CTC Blank放行、去重、排序及空交集退化兜底
-    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
+// pub fn from_builtin_old() -> Self {
-        let mut temp_indices = Vec::with_capacity(estimated_capacity.max(16));
+//     Self::from_static_slice(
+//         CHARSET_OLD,
+//         false,
+//         Resize::DynamicWidth(64),
+//         1,
+//         Normalization::ZeroToOne,
+//     )
+// }
+// 
+// /// 从预设的 Beta 版字符集创建
+// pub fn from_builtin_beta() -> Self {
+//     Self::from_static_slice(
+//         CHARSET_BETA,
+//         false,
+//         Resize::DynamicWidth(64),
+//         1,
+//         Normalization::MinusOneToOne,
+//     )
+// }
 
-        // 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
+// /// 从外部外部 JSON 文件动态加载字符集(在后续优化中移除)
-            if token_str.is_empty() || idx == 0 {
+// pub fn from_json_file<P: AsRef<Path>>(path: P) -> anyhow::Result<Self> {
-                temp_indices.push(idx);
+//     let path = path.as_ref();
-                continue; // 关键：直接跳过，防止后续 matches 匹配成功导致重复 push 产生 Bug
+//     if !path.exists() {
-            }
+//         return Err(anyhow!("模型元数据配置文件不存在: {:?}", path));
-
+//     }
-            // 规则 B: 组装无拷贝上下文
+// 
-            let ctx = ValidationCtx {
+//     let mut file = File::open(path)?;
-                text: token_str,
+//     let mut content = String::new();
-                token_id: idx,
+//     file.read_to_string(&mut content)?;
-            };
+// 
-
+//     let dto: ModelMetadataDto = serde_json::from_str(&content)
-            // 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）
+//         .map_err(|e| anyhow!("JSON 反序列化失败，请检查字段是否完整: {}", e))?;
-            if self.matches(&ctx) {
+// 
-                temp_indices.push(idx);
+//     // 1. 将 DTO 的字符串数组转化为强类型的 Charset
-                has_any_match = true;
+//     let tokens: Vec<Cow<'static, str>> =
-            }
+//         dto.charset.into_iter().map(|s| Cow::Owned(s)).collect();
-        }
+//     let charset = Charset::new(tokens);
-
+// 
-        // 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）
+//     // 2. 解析 resize 策略（重现 Python 的复杂条件判断）
-        if !has_any_match {
+//     if dto.resize.len() != 2 {
-            println!("警告：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。");
+//         return Err(anyhow!(
-            None
+//                 "'resize (or image)' 字段必须是包含两个元素的数组，例如 [-1, 64]"
-        } 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();
+//     let r0 = dto.resize[0];
-            temp_indices.dedup();
+//     let r1 = dto.resize[1];
-            Some(temp_indices)
+// 
-        }
+//     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)
-#[derive(Debug, Clone, PartialEq, Eq)]
+//         } else {
-pub enum CharRestrict {
+//             // 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
-    Digit,
+//             Resize::DynamicWidth(r1 as u32)
-    Lowercase,
+//         }
-    Uppercase,
+//     } else {
-    CustomList(Vec<String>),
+//         // 正常的固定宽高
-}
+//         Resize::Fixed(r0 as u32, r1 as u32)
-
+//     };
-impl TokenFilter for CharRestrict {
+// 
-    fn matches(&self, ctx: &ValidationCtx) -> bool {
+//     Ok(Self {
-        match self {
+//         charset,
-            Self::Digit => ctx.text.len() == 1 && ctx.text.as_bytes()[0].is_ascii_digit(),
+//         word: dto.word,
-            Self::Lowercase => ctx.text.len() == 1 && ctx.text.as_bytes()[0].is_ascii_lowercase(),
+//         resize,
-            Self::Uppercase => ctx.text.len() == 1 && ctx.text.as_bytes()[0].is_ascii_uppercase(),
+//         channel: dto.channel,
-            Self::CustomList(vec) => vec.iter().any(|t| t == ctx.text),
+//         normalization: dto.normalization,
-        }
+//     })
-    }
+// }
-    fn estimated_capacity(&self) -> usize {
-        match self {
-            Self::Digit => 16,
-            Self::Lowercase | Self::Uppercase => 32,
-            Self::CustomList(vec) => vec.len() + 1,
-        }
-    }
-}
-
-#[derive(Debug, Clone, PartialEq, Eq)]
-pub enum IdRestrict {
-    TopN(usize),
-    IdRange(std::ops::Range<usize>),
-    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> {
-    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()
-    }
-}
-// =====================================================================
-// 声明式宏：替代 `+` 运算符，解决组合扩展痛苦
-// =====================================================================
-#[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 $crate::TokenFilter
-    };
-
-    // 场景 B：如果用户传入了多个规则，自动织成一张静态组合网
-    ($($filter:expr),+ $(,)?) => {
-        &$crate::MultiOrRestrict {
-            filters: vec![ $( &$filter as &dyn $crate::TokenFilter ),+ ]
-        }
-    };
-}
-
-// ==========================================
-// 3. 字符集核心结构体 (重命名为 Charset)
-// ==========================================
-#[derive(Debug, Clone)]
-pub struct Charset {
-    // 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>>,
-    // 反向查找表，保证字符转索引为 O(1)
-    pub char_to_idx: HashMap<Cow<'static, str>, usize>,
-    // 当前处于激活状态的有效索引缓存 (用于 CTC 解码前的过滤加速)
-    // pub valid_indices: HashSet<usize>,
-}
-
-impl Charset {
-    // 内部底层统一收拢构造
-    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)
-            // char_to_idx.entry(token.to_string()).or_insert(idx);
-        }
-
-        Self {
-            tokens,
-            char_to_idx,
-        }
-    }
-
-    // --- 业务策略方法 ---
-
-    /// 将字符转为索引，不存在返回 -1 (保持与原 Python 库行为一致)
-    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
-        }
-    }
-
-    /// 将索引转为字符引用，零拷贝。若越界返回 None
-    pub fn index_to_char_ref(&self, index: usize) -> Option<&str> {
-        self.tokens.get(index).map(|cow| cow.as_ref())
-    }
-
-    pub fn is_valid_char(&self, char_str: &str) -> bool {
-        self.char_to_idx.get(char_str).is_some()
-    }
-    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(),)
-    }
-}

ddddocr-tract/tests/ocr_test.rs

@@ -1,9 +1,12 @@
-use ddddocr_rs::models::slide::Slide;
+use ddddocr_core::models::det::DetectionResult;
-use ddddocr_rs::{DdddOcr, DdddOcrBuilder}; // 假设你的包名是这个
+use ddddocr_core::{DetBuilder,  Detector, ModelMetadata, Ocr,  Slider}; // 假设你的包名是这个
+use ddddocr_tract::{DetSession,OcrSession};
 use image::{DynamicImage, Rgb};
 use std::fs;
 use std::path::Path;
-use ddddocr_rs::models::det::DetectionResult;
+mod char_slice;
+use char_slice::CHARSET_BETA;
+use ddddocr_core::models::ocr::metadata::{Normalization, Resize};
 
 fn load_image<P: AsRef<Path>>(path: P) -> anyhow::Result<image::DynamicImage> {
     // 1. 先将泛型转为具体的 &Path 引用
@@ -17,8 +20,8 @@ fn load_image<P: AsRef<Path>>(path: P) -> anyhow::Result<image::DynamicImage> {
 }
 /// 将检测结果绘制在图像上并保存
 fn save_debug_image(
-    dynamic_img: &DynamicImage,          // 【优化点 1】直接传入解码好的引用，拒绝重复解码
+    dynamic_img: &DynamicImage, // 【优化点 1】直接传入解码好的引用，拒绝重复解码
-    bboxes: &[DetectionResult],          // 【修改点 1】类型改为自定义结构体切片
+    bboxes: &[DetectionResult], // 【修改点 1】类型改为自定义结构体切片
     output_path: &str,
 ) -> anyhow::Result<()> {
     // 删除了原本的 let dynamic_img = image::load_from_memory(image_bytes)?;
@@ -60,24 +63,38 @@ fn save_debug_image(
     img.save(output_path)?;
     Ok(())
 }
+
 #[test]
 fn test_full_classification() {
     // 1. 初始化模型
-    let ocr = DdddOcrBuilder::new().build().expect("模型加载失败");
+    let ocr = OcrSession::new(
+        "D:\\CNWei\\CNW\\Rust\\ddddocr-rs\\models\\common_sml2h3_f32.onnx",
+        ModelMetadata::from_static_slice(
+            CHARSET_BETA,
+            false,
+            Resize::DynamicWidth(64),
+            1,
+            Normalization::MinusOneToOne,
+        ),
+    )
+    .expect("模型加载失败");
 
     // 2. 加载测试图片
-    let img = image::open("samples/code2.png").expect("测试图片不存在");
+    let img = image::open("D:/CNWei/CNW/Rust/ddddocr-rs/samples/code2.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");
+    let result = Ocr::new(&ocr)
+        .predict(&img)
+        .expect("识别过程出错")
+        .into_text();
 
     println!("识别结果: {}", result);
     assert!(!result.is_empty());
 }
 #[test]
 fn test_det_load() -> anyhow::Result<()> {
-    let det = DdddOcrBuilder::new().det().build()?;
+    let det = DetSession::new("D:\\CNWei\\CNW\\Rust\\ddddocr-rs\\models\\common_det.onnx")?;
-    let image_path = "samples/det1.png";
+    let image_path = "D:/CNWei/CNW/Rust/ddddocr-rs/samples/det1.png";
     let image_bytes =
         fs::read(image_path).map_err(|e| anyhow::anyhow!("无法读取图片 {}: {}", image_path, e))?;
 
@@ -88,22 +105,19 @@ fn test_det_load() -> anyhow::Result<()> {
         .map_err(|e| anyhow::anyhow!("图片解码失败: {}", e))?;
 
     // 【修改点 2】传入统一的 &DynamicImage 引用
-    let bboxes = det.detection(&img)?;
+    let bboxes = Detector::new(&det).predict(&img)?;
-    println!(":?{}", det);
+    // println!("{:?}", det);
     println!("检测到的目标数量: {}", bboxes.len());
 
     if bboxes.is_empty() {
         println!("未检测到任何目标。");
     } else {
         // 如果 save_debug_image 报错，记得去把它的入参类型和内部访问也改为 DetectionResult
-        save_debug_image(&img, &bboxes, "samples/result.jpg")?;
+        save_debug_image(&img, &bboxes, "D:/CNWei/CNW/Rust/ddddocr-rs/samples/result.jpg")?;
 
         for (i, bbox) in bboxes.iter().enumerate() {
             // 【修改点 3】将原来的 bbox[0].. 索引访问改为结构体字段访问
-            println!(
+            println!("目标 [{}]: {}", i, bbox);
-                "目标 [{}]: x1={}, y1={}, x2={}, y2={}, 分数={:.4}, 类别ID={}",
-                i, bbox.x1, bbox.y1, bbox.x2, bbox.y2, bbox.score, bbox.class_id
-            );
         }
     }
     Ok(())
@@ -111,12 +125,12 @@ fn test_det_load() -> anyhow::Result<()> {
 
 #[test]
 fn test_real_slide_match() {
-    let engine = Slide::new();
+    let engine = Slider::new().unwrap();
 
     // 1. 加载你准备好的测试图
     // 假设图片放在项目根目录下的 assets 文件夹
-    let target_img = load_image("samples/hua.png").expect("请确保 samples/hua.png 存在");
+    let target_img = load_image("D:/CNWei/CNW/Rust/ddddocr-rs/samples/hua.png").expect("请确保 samples/hua.png 存在");
-    let bg_img = load_image("samples/huatu.png").expect("请确保 samples/huatu.png 存在");
+    let bg_img = load_image("D:/CNWei/CNW/Rust/ddddocr-rs/samples/huatu.png").expect("请确保 samples/huatu.png 存在");
 
     // 2. 执行匹配
     // 如果是那种带有明显阴影边缘的复杂滑块，建议 simple_target 传 false
@@ -128,9 +142,7 @@ fn test_real_slide_match() {
 
     // 3. 打印结果
     println!("-------------------------------------------");
-    println!("滑块匹配测试结果:");
+    println!("{}", result);
-    println!("检测坐标: [x: {}, y: {}]", result.target_x, result.target_y);
-    println!("置信度: {:.4}", result.confidence);
     println!("耗时: {:?}", duration);
     println!("-------------------------------------------");
 
@@ -142,12 +154,12 @@ fn test_real_slide_match() {
 
 #[test]
 fn test_real_slide_comparison() {
-    let engine = Slide::new();
+    let engine = Slider::new().unwrap();
 
     // 1. 加载你准备好的测试图
     // 假设图片放在项目根目录下的 assets 文件夹
-    let target_img = load_image("samples/ken.jpg").expect("请确保 samples/ken.jpg 存在");
+    let target_img = load_image("D:/CNWei/CNW/Rust/ddddocr-rs/samples/ken.jpg").expect("请确保 samples/ken.jpg 存在");
-    let bg_img = load_image("samples/kenyuan.jpg").expect("请确保 samples/kenyuan.jpg 存在");
+    let bg_img = load_image("D:/CNWei/CNW/Rust/ddddocr-rs/samples/kenyuan.jpg").expect("请确保 samples/kenyuan.jpg 存在");
 
     // 2. 执行匹配
     // 如果是那种带有明显阴影边缘的复杂滑块，建议 simple_target 传 false

examples/simple_usage.rs

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

src/lib.rs

@@ -1,184 +0,0 @@
-mod charset;
-
-mod model_metadata;
-pub mod models;
-pub mod utils;
-
-use anyhow::{Result, anyhow};
-use image::DynamicImage;
-use std::fmt::{Display, Formatter};
-
-// 关键点：直接使用 tract 重导出的 ndarray
-use crate::charset::CharRestrict;
-use crate::model_metadata::ModelMetadata;
-use crate::models::det::DetectionResult;
-use crate::utils::color_filter::{ColorPreset, HsvRange};
-use models::det::Det;
-use models::loader::ModelSession;
-use models::ocr::Ocr;
-
-pub enum ModelSpec {
-    /// 默认 OCR (使用内置路径)
-    OcrModel,
-    DetModel,
-    /// 自定义 OCR (路径由用户提供)
-    CustomOcrModel {
-        path: String,
-        model_metadata: ModelMetadata,
-    },
-}
-impl ModelSpec {
-    // 将默认路径定义为内部关联常量
-    const DEFAULT_OCR_PATH: &'static str = "models/common_sml2h3_f32.onnx";
-    const DEFAULT_DET_PATH: &'static str = "models/common_det.onnx";
-}
-pub enum Runtime {
-    Ocr(Ocr),
-    Det(Det),
-}
-impl Runtime {
-    // 统一获取描述的方法
-    pub fn desc(&self) -> String {
-        match self {
-            Runtime::Ocr(s) => s.desc(), // 调用 Ocr 结构体的方法
-            Runtime::Det(s) => s.desc(), // 调用 Det 结构体的方法
-        }
-    }
-}
-pub struct DdddOcrBuilder {
-    mode: ModelSpec,
-}
-
-impl DdddOcrBuilder {
-    pub fn new() -> Self {
-        Self {
-            mode: ModelSpec::OcrModel,
-        }
-    }
-
-    /// 切换为检测模式
-    pub fn det(mut self) -> Self {
-        self.mode = ModelSpec::DetModel;
-        self
-    }
-
-    /// 设置自定义 OCR 路径
-    pub fn custom_ocr(mut self, path: String, model_metadata: ModelMetadata) -> Self {
-        // 直接重写枚举，替换掉之前的 Ocr 或 Det
-        self.mode = ModelSpec::CustomOcrModel {
-            path,
-            model_metadata,
-        };
-        self
-    }
-
-    /// 核心初始化逻辑
-    pub fn build(self) -> Result<DdddOcr> {
-        let runtime = match self.mode {
-            ModelSpec::OcrModel => Runtime::Ocr(Ocr::new(
-                ModelSpec::DEFAULT_OCR_PATH.into(),
-                ModelMetadata::from_builtin_beta(),
-            )?),
-            ModelSpec::DetModel => Runtime::Det(Det::new(ModelSpec::DEFAULT_DET_PATH.into())?),
-            ModelSpec::CustomOcrModel {
-                path,
-                model_metadata,
-            } => Runtime::Ocr(Ocr::new(path, model_metadata)?),
-        };
-
-        Ok(DdddOcr { runtime })
-    }
-}
-
-pub struct DdddOcr {
-    runtime: Runtime,
-}
-
-impl Display for DdddOcr {
-    fn fmt(&self, f: &mut Formatter<'_>) -> std::fmt::Result {
-        write!(f, "DdddOcr(session: {})", self.runtime.desc())
-    }
-}
-
-impl DdddOcr {
-    pub fn classification(&self, img: &DynamicImage) -> Result<String> {
-        match &self.runtime {
-            // Runtime::Ocr(s) => s.predict(img).run(),
-            // Runtime::Ocr(s) => s.predictor().probability(false).predict(img),
-            // Runtime::Ocr(s) => {
-            //     let predictor = s.predictor();
-            //     let restricted = predictor.charset_restrict(&CharRestrict::Lowercase);
-            //     let a = restricted.valid_tokens();
-            //     println!("{:?}", a);
-            //     Ok("".to_string())
-            // }
-            Runtime::Ocr(s) => {
-                let res = s.predictor().probability(true).predict(img)?;
-                println!("{}", res);
-                Ok(res.to_string())
-            }
-            // Runtime::Ocr(s) => s.predictor().charset_restrict(&CharRestrict::Digit).predict(img),
-            // Runtime::Ocr(s) => s.predictor().color_filter(&ColorPreset::Custom(vec![
-            //     // 错误：下界 (82, 221, 14) 没问题
-            //     //       但上界的 H 通道写成了 240，超过了 180 的法定上限！
-            //     HsvRange::new((82, 221, 14), (240, 203, 82)),
-            // ])).predict(img),
-            Runtime::Det(_) => Err(anyhow::anyhow!("当前模型是检测模型，无法执行 OCR")),
-        }
-    }
-    pub fn detection(&self, img: &DynamicImage) -> Result<Vec<DetectionResult>> {
-        match &self.runtime {
-            Runtime::Det(s) => s.predict(img),
-            Runtime::Ocr(_) => Err(anyhow::anyhow!("当前模型是 OCR 模型，无法执行检测")),
-        }
-    }
-}
-
-// struct Classification {}
-// #[derive(Debug)]
-// struct ClassificationBuilder {
-//     img: DynamicImage,
-//     png_fix: bool,
-//     color_filter_colors: Option<Vec<ColorRange>>,
-//     color_filter_custom_ranges: Option<Vec<ColorRange>>,
-// }
-// impl ClassificationBuilder {
-//     pub fn new(img: DynamicImage) -> Self {
-//         ClassificationBuilder {
-//             img,
-//             png_fix: false,
-//             color_filter_colors: None,
-//             color_filter_custom_ranges: None,
-//         }
-//     }
-//     pub fn png_fix(mut self, value: bool) -> Self {
-//         self.png_fix = value;
-//         self
-//     }
-//     pub fn color_filter_colors(mut self, value: Vec<ColorRange>) -> Self {
-//         self.color_filter_colors = Some(value);
-//         self
-//     }
-//     pub fn color_filter_custom_ranges(mut self, value: Vec<ColorRange>) -> Self {
-//         self.color_filter_custom_ranges = Some(value);
-//         self
-//     }
-//     pub fn build(self) -> Classification {
-//         Classification {}
-//     }
-// }
-
-#[cfg(test)]
-mod tests {
-    #[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");
-    }
-}

src/models/base.rs

@@ -1,40 +0,0 @@
-pub trait ModelArgs {
-    // 获取模型路径
-    fn model_path(&self) -> &str;
-
-    // 获取字符集（由于 Det 没有，所以返回 Option）
-    fn charset(&self) -> Option<&str>;
-}
-
-pub struct HasCharset {
-    pub charset: String,
-} // 给 Ocr 和 Custom 用
-pub struct NoCharset; // 给 Det 用
-
-pub struct Model<T> {
-    pub path: String,
-    pub metadata: T,
-}
-// 针对有字符集的模型 (Ocr / Custom)
-impl ModelArgs for Model<HasCharset> {
-    fn model_path(&self) -> &str {
-        &self.path
-    }
-    fn charset(&self) -> Option<&str> {
-        Some(&self.metadata.charset)
-    }
-}
-
-// 针对没有字符集的模型 (Det)
-impl ModelArgs for Model<NoCharset> {
-    fn model_path(&self) -> &str {
-        &self.path
-    }
-    fn charset(&self) -> Option<&str> {
-        None
-    }
-}
-
-pub type OcrModel = Model<HasCharset>;
-pub type DetModel = Model<NoCharset>;
-pub type CustomModel = Model<HasCharset>; // Ocr 和 Custom 逻辑一致，可以复用

src/models/loader.rs

@@ -1,40 +0,0 @@
-use anyhow::Context;
-use image::DynamicImage;
-use tract_onnx::onnx;
-use tract_onnx::prelude::*;
-// 关键点：直接使用 tract 重导出的 ndarray
-use crate::utils::image_io::png_rgba_white_preprocess;
-use crate::utils::image_processor::{convert_to_grayscale, resize_image};
-use std::collections::HashMap;
-use tract_onnx::prelude::tract_ndarray::s;
-
-/// OCR 模型：包含路径和字符集
-
-pub enum ModelType {
-    Ocr,
-    Det,
-    Custom,
-}
-// 定义统一的 trait
-pub trait ModelSession {
-    fn get_model_type(&self) -> ModelType;
-    fn desc(&self) -> String;
-}
-
-pub struct ModelLoader {
-    pub session: RunnableModel<TypedFact, Box<dyn TypedOp>, Graph<TypedFact, Box<dyn TypedOp>>>,
-}
-
-impl ModelLoader {
-    pub fn load_model<P>(model_path: P) -> anyhow::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 })
-    }
-}

src/models/mod.rs

@@ -1,5 +0,0 @@
-pub mod base;
-pub mod loader;
-pub mod ocr;
-pub mod det;
-pub mod slide;

src/utils/mod.rs

@@ -1,4 +0,0 @@
-pub mod image_io;
-pub mod image_processor;
-pub mod cv_ops;
-pub mod color_filter;