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]
-name = "ddddocr-rs"
+[workspace]
+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"
@@ -13,8 +20,4 @@ imageproc = { version = "0.26.2", default-features = true }
 serde = { version = "1.0.228", features = ["derive"] }
 serde_json = "1.0.150"
 ndarray="0.16.1"
-
-
-[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
+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/slide.rs

@@ -1,5 +1,5 @@
-use crate::utils::cv_ops;
-use crate::utils::cv_ops::{abs_diff, min_max_loc, ndarray_to_luma8, rgb_to_gray};
+use crate::utils::image_proc;
+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::{Result, anyhow};
 use image::DynamicImage;
@@ -11,7 +11,7 @@ use imageproc::morphology::{close, open};
 use imageproc::region_labelling::{Connectivity, connected_components};
 use imageproc::template_matching::{MatchTemplateMethod, match_template};
 use std::fmt;
-use tract_onnx::prelude::tract_ndarray::{ArrayView2, ArrayView3};
+use ndarray::{ArrayView2, ArrayView3};
 #[derive(Debug)]
 pub struct SlideResult {
     pub target: [i32; 2],
@@ -111,11 +111,11 @@ impl Slider {
 
         // // 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],
@@ -206,7 +206,7 @@ impl Slider {
         // 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);
@@ -251,7 +251,7 @@ impl Slider {
         // 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

@@ -15,4 +15,34 @@ pub(crate) const MODEL_DOWNLOAD_HELP: &str = "\
       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

@@ -1,5 +1,6 @@
 use crate::models::det::executor::Detector;
-use crate::models::det::session::DetSession;
+// use ddddocr_tract::det::session::DetSession;
+use crate::DetEngine;
 
 pub struct DetBuilder {
     use_gpu: bool,
@@ -15,7 +16,7 @@ impl DetBuilder {
         self.device_id = device_id;
         self
     }
-    fn build(self, session: &DetSession) -> Detector<'_> {
+    fn build(self, session: &dyn DetEngine) -> Detector<'_> {
         Detector {
             session,
             use_gpu: self.use_gpu,

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

@@ -1,12 +1,12 @@
 use anyhow::{Context, Result};
 use image::{imageops::FilterType, DynamicImage, GenericImageView};
 use std::fmt;
-use tract_onnx::prelude::tract_ndarray::{prelude::*, s, Array2, Array3, Array4, Axis};
-use tract_onnx::prelude::{Tensor};
+use ndarray::{prelude::*, s, Array2, Array3, Array4, Axis};
+// use tract_onnx::prelude::{Tensor};
 
 
-use crate::models::det::session::DetSession;
-
+// use ddddocr_tract::det::session::DetSession;
+use crate::{DetEngine, DetOutput};
 #[derive(Debug, Clone, Copy)]
 pub struct DetectionResult {
     pub x1: i32,
@@ -28,9 +28,9 @@ impl fmt::Display for DetectionResult {
     }
 }
 
-#[derive(Debug)]
+
 pub struct Detector<'a> {
-    pub(crate) session: &'a DetSession,
+    pub(crate) session: &'a dyn DetEngine,
     #[allow(dead_code)]
     pub(crate) use_gpu: bool,
     #[allow(dead_code)]
@@ -38,7 +38,7 @@ pub struct Detector<'a> {
 }
 
 impl<'a> Detector<'a> {
-    pub fn new(session: &'a DetSession) -> Self {
+    pub fn new(session: &'a dyn DetEngine) -> Self {
         Detector {
             session,
             use_gpu: false,
@@ -51,7 +51,7 @@ impl<'a> Detector<'a> {
         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();
 
@@ -89,7 +89,7 @@ impl<'a> Detector<'a> {
             }
         }
 
-        Ok((array.into(), r))
+        Ok((array, r))
     }
 
     /// 3. demo_postprocess (逻辑与 Python 一致)
@@ -258,10 +258,8 @@ impl<'a> Detector<'a> {
         // let outputs = self.session.session.run(tvec!(input_tensor.into()))?;
         let outputs = self.session.inference(input_tensor)?;
         // let output_array = outputs[0]
-        let output_array = outputs
-            .to_array_view::<f32>()?
-            .to_owned()
-            .into_dimensionality::<Ix3>()?;
+        // 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, .., ..]);

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

@@ -1,7 +1,6 @@
 mod builder;
 mod executor;
-mod session;
 
 pub use builder::DetBuilder;
 pub use executor::{DetectionResult, Detector};
-pub use session::DetSession;
+// pub use ddddocr_tract::det::session::DetSession;

ddddocr-core/src/models/mod.rs

@@ -1,3 +1,2 @@
-pub mod loader;
 pub mod ocr;
 pub mod det;

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

@@ -1,7 +1,8 @@
 use crate::models::ocr::executor::Ocr;
-use crate::models::ocr::session::OcrSession;
+// 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
@@ -48,14 +49,14 @@ impl OcrBuilder {
         self.charset_restrict = Some(Box::new(restrict));
         self
     }
-    pub fn build(self, session: &OcrSession) -> Ocr<'_> {
+    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.model_metadata.charset.tokens;
+        let tokens = &session.metadata().charset.tokens;
         let final_charset_indices = match &self.charset_restrict {
             Some(restrict) => restrict.apply_to_charset(tokens),
             None => None,

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

@@ -1,4 +1,4 @@
-use crate::utils::cv_ops::rgb_to_opencv_hsv;
+use crate::utils::image_proc::rgb_to_opencv_hsv;
 use anyhow::anyhow;
 use image::{DynamicImage, ImageBuffer, Rgb};
 use std::str::FromStr;

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

@@ -1,7 +1,7 @@
 use crate::models::ocr::metadata::Resize;
 
-use crate::models::ocr::session::OcrSession;
 use crate::models::ocr::color_filter::{HsvRange, apply_to_image};
+// use ddddocr_tract::session::{ModelOutput, OcrSession};
 use crate::utils::image_io::png_rgba_white_preprocess;
 use crate::utils::image_processor::{convert_to_grayscale, resize_image};
 use anyhow::Result;
@@ -9,8 +9,15 @@ use image::DynamicImage;
 use serde::Serialize;
 use std::borrow::Cow;
 use std::fmt;
-use tract_onnx::prelude::tract_ndarray::{ArrayView2, Ix2, s};
-use tract_onnx::prelude::{DatumType, Tensor, tract_ndarray};
+// use tract_onnx::prelude::tract_ndarray::{ Ix2, s};
+// use tract_onnx::prelude::{DatumType, Tensor, tract_ndarray};
+// ！！！【核心纠正】：彻底弃用 tract_ndarray，全线转用标准 ndarray
+use ndarray::ArrayView2;
+// pub enum ModelOutput {
+//     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
+// }
+use crate::{OcrEngine, OcrOutput};
 #[derive(Debug, Clone, Serialize)]
 pub enum OcrResult {
     /// 纯文本分支（对应 probability = false）
@@ -96,7 +103,7 @@ impl fmt::Display for OcrResult {
 }
 
 pub struct Ocr<'a> {
-    pub(crate) session: &'a OcrSession,
+    pub(crate) session: &'a dyn OcrEngine,
     pub(crate) png_fix: bool,
     pub(crate) probability: bool,
     /// 颜色过滤：保留的颜色列表
@@ -109,7 +116,7 @@ pub struct Ocr<'a> {
 impl<'a> Ocr<'a> {
     // 初始化任务，设置默认参数
 
-    pub fn new(session: &'a OcrSession) -> Self {
+    pub fn new(session: &'a dyn OcrEngine) -> Self {
         Ocr {
             session,
             png_fix: false, // 默认值
@@ -150,25 +157,25 @@ impl<'a> Ocr<'a> {
         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() {
-            DatumType::I64 => self.process_i64_tensor(raw_tensor)?,
-            DatumType::F32 => self.process_f32_tensor(raw_tensor)?,
-            _ => OcrResult::Unsupported {
-                message: format!("不支持的模型输出数据类型: {:?}", raw_tensor.datum_type()),
-            },
-        };
+        // let ocr_output = match raw_tensor.datum_type() {
+        //     DatumType::I64 => self.process_i64_tensor(raw_tensor)?,
+        //     DatumType::F32 => self.process_f32_tensor(raw_tensor)?,
+        //     _ => OcrResult::Unsupported {
+        //         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);
-
-        Ok(ocr_output)
+        let ocr_output = self.process_model_output(raw_tensor);
+        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.session.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)
@@ -198,12 +205,12 @@ impl<'a> Ocr<'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;
@@ -212,14 +219,14 @@ impl<'a> Ocr<'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;
@@ -228,13 +235,14 @@ impl<'a> Ocr<'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(tensor)
+        Ok(array4)
+        // Ok(tensor)
 
         // let h = 64u32;
         // let w = (current_img.width() as f32 * (h as f32 / current_img.height() as f32)) as u32;
@@ -255,10 +263,61 @@ impl<'a> Ocr<'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> Ocr<'a> {
     fn is_valid_indices(&self, idx: usize) -> bool {
-        if idx >= self.session.model_metadata.charset.size() {
+        if idx >= self.session.metadata().charset.size() {
             return false;
         }
 
@@ -270,7 +329,7 @@ impl<'a> Ocr<'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.session.model_metadata.charset;
+        let charset = &self.session.metadata().charset;
         let tokens = &charset.tokens;
         match &self.final_charset_indices {
             Some(indices) => indices
@@ -284,7 +343,7 @@ impl<'a> Ocr<'a> {
     pub fn valid_size(&self) -> usize {
         match &self.final_charset_indices {
             Some(indices) => indices.len(),
-            None => self.session.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
@@ -296,7 +355,7 @@ impl<'a> Ocr<'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;
 
@@ -341,98 +400,98 @@ impl<'a> Ocr<'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> {
-        // 1. 拿到底层的动态维度只读视图
-        let view = raw_tensor.to_array_view::<i64>()?;
-
-        // 2. 索要底层连续的只读切片引用
-        let slice = view
-            .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"))?;
-
-        // 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);
-
-        // 4. 组装返回
-        if self.probability {
-            Ok(OcrResult::Probability {
-                text: final_text,
-                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
-            })
-        } else {
-            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
-    fn process_f32_tensor(&self, raw_tensor: Tensor) -> anyhow::Result<OcrResult> {
-        let shape = raw_tensor.shape();
-        println!("模型输出shape数据: {:?}", shape);
-        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
-        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(anyhow::anyhow!("不支持的输出维度: {:?}", shape)),
-        };
-        let matrix_cow = data_dyn_view
-            .to_shape(Ix2(steps, classes))
-            .map_err(|e| anyhow::anyhow!("转换为2D静态矩阵失败: {:?}", e))?;
-
-        let matrix_view: ArrayView2<f32> = matrix_cow.view();
-
-        // 2. 根据业务参数明确分流
-        if self.probability {
-            // 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) =
-                self.compute_f32_full_probability(matrix_view);
-            // 5. 执行 CTC 解码
-            let final_text = self.ctc_decode_to_string(&predicted_indices);
-
-            Ok(OcrResult::Probability {
-                text: final_text,
-                probabilities: probabilities_list,
-                confidence: confidence as f64,
-            })
-        } 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）
-            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))
-        }
-    }
+    // 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
+    //     let view = raw_tensor.to_array_view::<i64>()?;
+    //
+    //     // 2. 索要底层连续的只读切片引用
+    //     let slice = view
+    //         .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"))?;
+    //
+    //     // 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);
+    //
+    //     // 4. 组装返回
+    //     if self.probability {
+    //         Ok(OcrResult::Probability {
+    //             text: final_text,
+    //             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
+    //         })
+    //     } else {
+    //         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
+    // fn process_f32_tensor(&self, raw_tensor: Tensor) -> anyhow::Result<OcrResult> {
+    //     let shape = raw_tensor.shape();
+    //     println!("模型输出shape数据: {:?}", shape);
+    //     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
+    //     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(anyhow::anyhow!("不支持的输出维度: {:?}", shape)),
+    //     };
+    //     let matrix_cow = data_dyn_view
+    //         .to_shape(Ix2(steps, classes))
+    //         .map_err(|e| anyhow::anyhow!("转换为2D静态矩阵失败: {:?}", e))?;
+    //
+    //     let matrix_view: ArrayView2<f32> = matrix_cow.view();
+    //
+    //     // 2. 根据业务参数明确分流
+    //     if self.probability {
+    //         // 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) =
+    //             self.compute_f32_full_probability(matrix_view);
+    //         // 5. 执行 CTC 解码
+    //         let final_text = self.ctc_decode_to_string(&predicted_indices);
+    //
+    //         Ok(OcrResult::Probability {
+    //             text: final_text,
+    //             probabilities: probabilities_list,
+    //             confidence: confidence as f64,
+    //         })
+    //     } 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）
+    //         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))
+    //     }
+    // }
     /// 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.session.model_metadata.charset;
+        let charset = &self.session.metadata().charset;
         let tokens = &charset.tokens;
         // let valid_indices = &charset.valid_indices;
 

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

@@ -1,10 +1,9 @@
 mod builder;
 mod executor;
-mod session;
 pub mod metadata;
 pub mod color_filter;
 mod token_filter;
 
 pub use builder::OcrBuilder;
 pub use executor::{Ocr, OcrResult};
-pub use session::OcrSession;
+// pub use ddddocr_tract::session::OcrSession;

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

@@ -1,7 +1,7 @@
 use image::{DynamicImage, GrayImage, imageops::FilterType, Rgb, ImageBuffer};
 use anyhow::{anyhow, Result};
 use crate::models::ocr::color_filter::HsvRange;
-use crate::utils::cv_ops::rgb_to_opencv_hsv;
+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)

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

@@ -1,8 +1,8 @@
 use anyhow::Context;
+use ddddocr_core::error::Result;
 use std::io::Cursor;
 use tract_onnx::onnx;
-use tract_onnx::prelude::*;
-
+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 {
@@ -21,27 +21,31 @@ pub struct ModelLoader {
 }
 
 impl ModelLoader {
-    pub fn model_for_path<P>(model_path: P) -> anyhow::Result<Self>
+    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()?
-            .into_runnable()?;
+            .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]) -> anyhow::Result<Self> {
+    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()?
-            .into_runnable()?;
+            .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

@@ -2,8 +2,8 @@ use std::borrow::Cow;
 use std::fs::File;
 use std::path::Path;
 use anyhow::anyhow;
-use ddddocr_rs::models::ocr::metadata::Charset;
-use ddddocr_rs::models::ocr::metadata::{Normalization, Resize};
+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",

ddddocr-tract/tests/ocr_test.rs

@@ -1,11 +1,12 @@
-use ddddocr_rs::models::det::DetectionResult;
-use ddddocr_rs::{DetBuilder, DetSession, Detector, ModelMetadata, Ocr, OcrSession, Slider}; // 假设你的包名是这个
+use ddddocr_core::models::det::DetectionResult;
+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;
 mod char_slice;
 use char_slice::CHARSET_BETA;
-use ddddocr_rs::models::ocr::metadata::{Normalization, Resize};
+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
@@ -79,7 +80,7 @@ fn test_full_classification() {
     .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::new(&ocr)
@@ -93,7 +94,7 @@ fn test_full_classification() {
 #[test]
 fn test_det_load() -> anyhow::Result<()> {
     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))?;
 
@@ -112,7 +113,7 @@ fn test_det_load() -> anyhow::Result<()> {
         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
@@ -128,8 +129,8 @@ fn test_real_slide_match() {
 
     // 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 bg_img = load_image("samples/huatu.png").expect("请确保 samples/huatu.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("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
@@ -157,8 +158,8 @@ fn test_real_slide_comparison() {
 
     // 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 bg_img = load_image("samples/kenyuan.jpg").expect("请确保 samples/kenyuan.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("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

src/lib.rs

@@ -1,9 +0,0 @@
-mod algo;
-mod error;
-pub mod models;
-pub mod utils;
-
-pub use crate::algo::{SlideResult, Slider};
-pub use crate::models::det::{DetBuilder, DetSession, DetectionResult, Detector};
-pub use crate::models::ocr::{Ocr, OcrBuilder, OcrResult, OcrSession};
-pub use models::ocr::metadata::ModelMetadata;

src/models/det/session.rs

@@ -1,43 +0,0 @@
-use crate::models::loader::{ModelLoader, ModelSession, ModelType};
-use anyhow::{Context, Result};
-use std::path::Path;
-use tract_onnx::prelude::{tvec, Graph, IntoTensor, RunnableModel, Tensor, TypedFact, TypedOp};
-
-#[derive(Debug)]
-pub struct DetSession {
-    pub(crate) 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, anyhow::Error>
-    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, anyhow::Error> {
-        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())
-    }
-}

src/models/ocr/session.rs

@@ -1,53 +0,0 @@
-use crate::models::ocr::metadata::ModelMetadata;
-use crate::models::loader::{ModelLoader, ModelSession, ModelType};
-use anyhow::Context;
-use anyhow::Result;
-use std::path::Path;
-use tract_onnx::prelude::{tvec, Graph, IntoTensor, RunnableModel, Tensor, TypedFact, TypedOp};
-
-pub struct OcrSession {
-    pub session: RunnableModel<TypedFact, Box<dyn TypedOp>, Graph<TypedFact, Box<dyn TypedOp>>>,
-    pub model_metadata: ModelMetadata,
-}
-impl ModelSession for OcrSession {
-    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 OcrSession {
-    pub fn new<P>(model_path: P, model_metadata: ModelMetadata) -> Result<Self, anyhow::Error>
-    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, anyhow::Error> {
-        let session = ModelLoader::model_from_bytes(model_bytes)?.session;
-        Ok(Self {
-            session,
-            model_metadata,
-        })
-    }
-    /// 对应 Python 的 _inference
-    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())
-    }
-}

src/utils/mod.rs

@@ -1,3 +0,0 @@
-pub mod image_io;
-pub mod image_processor;
-pub mod cv_ops;