refactor(slide,det): 重构目标检测引擎并统一图像输入类型为 DynamicImage以及滑块匹配与比较引擎为 Rust 实现

- 统一 `predict` 和 `get_bbox` 接口参数为 `&DynamicImage`，消除多步处理时的重复图像解码开销。
- 引入轻量级 `DetectionResult` 结构体和固定大小数组 `[f32; 6]` 替代旧的嵌套 `Vec`，彻底消除后处理中的内存碎片。
- 优化 `preproc` 预处理逻辑，使用连续内存切片批量操作替代原有的逐像素迭代遍历。
- 移除多余的 `multiclass_nms_class_agnostic` 转发层，合并并精简 NMS 聚合函数。
- 优化 `calculate_center` 几何中心点计算函数，提高泛型语义并复用于两种匹配模式
- 在执行核心算法前增加尺寸与通道边界守卫（Guard Clauses），提升库的防防御性编程能力与崩溃安全性
- 移除多余的错误二次包装（map_err），改由 Rust 原生 Result 错误传播机制直接向上层抛出

src/charset.rs

@@ -517,81 +517,152 @@ pub const CHARSET_BETA: &[&str] = &[
 
 pub const CHARSET_OLD: &[&str] = &["", "笤", "谴", "膀", "荔"];
 
-pub fn get_default_charset() -> Vec<String> {
-    CHARSET_BETA.iter().map(|&s| s.to_string()).collect()
-}
-
 use std::borrow::Cow;
 use std::collections::{HashMap, HashSet};
 
-use std::ops::{Add, Deref};
+/// 字符集范围限制枚举
-// 字符集范围类型
+pub struct ValidationCtx<'a> {
-/// 字符集范围限制组合子枚举
+    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 CharsetRestrict {
+pub enum CharRestrict {
-    /// 纯整数 0-9
     Digit,
-
-    /// 纯小写字母 a-z
     Lowercase,
-
-    /// 纯大写字母 A-Z
     Uppercase,
-
-    // /// 过滤模式：删除所有 ASCII 字母和数字（通常用于仅保留汉字、特殊标点）
-    // // ExcludeAlphanumeric,
-    // /// 自定义单字字符集，例如 "0123456789+-x/="
-    // Single(String),
-    /// 直接设置完整的 Token 白名单（支持多字 Token），例如 vec!["html".to_string()]
     CustomList(Vec<String>),
+}
 
-    /// 核心组合子：满足左边或右边任意一个条件即可（即 A + B 的并集逻辑）
+impl TokenFilter for CharRestrict {
-    /// 使用 Box 打破 Rust 编译期对递归枚举的无限大小限制
+    fn matches(&self, ctx: &ValidationCtx) -> bool {
-    Or(Box<CharsetRestrict>, Box<CharsetRestrict>),
-}
-impl From<i32> for CharsetRestrict {
-    fn from(value: i32) -> Self {
-        match value {
-            0 => Self::Digit,
-            1 => Self::Lowercase,
-            2 => Self::Uppercase,
-            // 3 => Self::LowercaseUppercase,
-            // 4 => Self::LowercaseDigit,
-            // 5 => Self::UppercaseDigit,
-            // 6 => Self::LowercaseUppercaseDigit,
-            // 7 => Self::DefaultCharsetLowercaseUppercaseDigit,
-            _ => panic!("invalid charset range: {}", value),
-        }
-    }
-}
-impl CharsetRestrict {
-    /// 💡 辅助构造函数：直接在源头把用户的长字符串切碎，伪装成基础积木
-    pub fn from_chars(custom_str: &str) -> Self {
-        let tokens = custom_str.chars().map(|c| c.to_string()).collect();
-        CharsetRestrict::CustomList(tokens)
-    }
-    // 内部递归收集器：利用硬编码切片快速无损展开
-    pub(crate) fn matches(&self, s: &str) -> bool {
         match self {
-            CharsetRestrict::Digit => s.len() == 1 && s.as_bytes()[0].is_ascii_digit(),
+            Self::Digit => ctx.text.len() == 1 && ctx.text.as_bytes()[0].is_ascii_digit(),
-            CharsetRestrict::Lowercase => s.len() == 1 && s.as_bytes()[0].is_ascii_lowercase(),
+            Self::Lowercase => ctx.text.len() == 1 && ctx.text.as_bytes()[0].is_ascii_lowercase(),
-            CharsetRestrict::Uppercase => s.len() == 1 && s.as_bytes()[0].is_ascii_uppercase(),
+            Self::Uppercase => ctx.text.len() == 1 && ctx.text.as_bytes()[0].is_ascii_uppercase(),
-            CharsetRestrict::CustomList(vec) => vec.iter().any(|t| t == s),
+            Self::CustomList(vec) => vec.iter().any(|t| t == ctx.text),
-            CharsetRestrict::Or(left, right) => left.matches(s) || right.matches(s),
+        }
+    }
+    fn estimated_capacity(&self) -> usize {
+        match self {
+            Self::Digit => 16,
+            Self::Lowercase | Self::Uppercase => 32,
+            Self::CustomList(vec) => vec.len() + 1,
         }
     }
 }
-// =====================================================================
-// 5. 优雅的魔法：重载 + 运算符 (实现 std::ops::Add)
-// =====================================================================
 
-/// 支持 `CharsetRestrict::Digit + CharsetRestrict::Lowercase`
+#[derive(Debug, Clone, PartialEq, Eq)]
-impl Add for CharsetRestrict {
+pub enum IdRestrict {
-    type Output = Self;
+    TopN(usize),
+    IdRange(std::ops::Range<usize>),
+    IdList(Vec<usize>),
+}
 
-    fn add(self, rhs: Self) -> Self::Output {
+impl TokenFilter for IdRestrict {
-        CharsetRestrict::Or(Box::new(self), Box::new(rhs))
+    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 ),+ ]
+        }
+    };
 }
 
 // ==========================================
@@ -639,14 +710,12 @@ impl Charset {
         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()
     }
-    
 }
 
 // ==========================================
@@ -654,10 +723,6 @@ impl Charset {
 // ==========================================
 impl std::fmt::Display for Charset {
     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        write!(
+        write!(f, "Charset [Total Size: {}", self.size(),)
-            f,
-            "Charset [Total Size: {}",
-            self.size(),
-        )
     }
 }

src/lib.rs

@@ -1,20 +1,21 @@
 mod charset;
 
+mod model_metadata;
 pub mod models;
 pub mod utils;
-mod model_metadata;
 
-use anyhow::Result;
+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::get_default_charset;
+use crate::charset::CharRestrict;
-use crate::models::ocr::ColorRange;
+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;
-use crate::model_metadata::ModelMetadata;
 
 pub enum ModelSpec {
     /// 默认 OCR (使用内置路径)
@@ -64,7 +65,10 @@ impl DdddOcrBuilder {
     /// 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.mode = ModelSpec::CustomOcrModel {
+            path,
+            model_metadata,
+        };
         self
     }
 
@@ -76,7 +80,10 @@ impl DdddOcrBuilder {
                 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)?),
+            ModelSpec::CustomOcrModel {
+                path,
+                model_metadata,
+            } => Runtime::Ocr(Ocr::new(path, model_metadata)?),
         };
 
         Ok(DdddOcr { runtime })
@@ -96,11 +103,30 @@ impl Display for DdddOcr {
 impl DdddOcr {
     pub fn classification(&self, img: &DynamicImage) -> Result<String> {
         match &self.runtime {
-            Runtime::Ocr(s) => s.predict(img).run(),
+            // 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: &[u8]) -> Result<Vec<Vec<i32>>> {
+    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")),
@@ -108,39 +134,39 @@ impl DdddOcr {
     }
 }
 
-struct Classification {}
+// struct Classification {}
-#[derive(Debug)]
+// #[derive(Debug)]
-struct ClassificationBuilder {
+// struct ClassificationBuilder {
-    img: DynamicImage,
+//     img: DynamicImage,
-    png_fix: bool,
+//     png_fix: bool,
-    color_filter_colors: Option<Vec<ColorRange>>,
+//     color_filter_colors: Option<Vec<ColorRange>>,
-    color_filter_custom_ranges: Option<Vec<ColorRange>>,
+//     color_filter_custom_ranges: Option<Vec<ColorRange>>,
-}
+// }
-impl ClassificationBuilder {
+// impl ClassificationBuilder {
-    pub fn new(img: DynamicImage) -> Self {
+//     pub fn new(img: DynamicImage) -> Self {
-        ClassificationBuilder {
+//         ClassificationBuilder {
-            img,
+//             img,
-            png_fix: false,
+//             png_fix: false,
-            color_filter_colors: None,
+//             color_filter_colors: None,
-            color_filter_custom_ranges: None,
+//             color_filter_custom_ranges: None,
-        }
+//         }
-    }
+//     }
-    pub fn png_fix(mut self, value: bool) -> Self {
+//     pub fn png_fix(mut self, value: bool) -> Self {
-        self.png_fix = value;
+//         self.png_fix = value;
-        self
+//         self
-    }
+//     }
-    pub fn color_filter_colors(mut self, value: Vec<ColorRange>) -> Self {
+//     pub fn color_filter_colors(mut self, value: Vec<ColorRange>) -> Self {
-        self.color_filter_colors = Some(value);
+//         self.color_filter_colors = Some(value);
-        self
+//         self
-    }
+//     }
-    pub fn color_filter_custom_ranges(mut self, value: Vec<ColorRange>) -> Self {
+//     pub fn color_filter_custom_ranges(mut self, value: Vec<ColorRange>) -> Self {
-        self.color_filter_custom_ranges = Some(value);
+//         self.color_filter_custom_ranges = Some(value);
-        self
+//         self
-    }
+//     }
-    pub fn build(self) -> Classification {
+//     pub fn build(self) -> Classification {
-        Classification {}
+//         Classification {}
-    }
+//     }
-}
+// }
 
 #[cfg(test)]
 mod tests {

src/model_metadata.rs

@@ -1,4 +1,4 @@
-use crate::charset::{CHARSET_BETA, CHARSET_OLD, Charset, CharsetRestrict};
+use crate::charset::{CHARSET_BETA, CHARSET_OLD, Charset};
 use anyhow::{Result, anyhow};
 use serde::Deserialize;
 use std::borrow::Cow;
@@ -10,6 +10,26 @@ use std::path::Path;
 // 1. 辅助定义的枚举与结构体
 // =====================================================================
 
+#[derive(Debug, Clone, Copy, Deserialize)]
+#[serde(rename_all = "snake_case")] // 支持 json 中写 "zero_to_one" 或 "minus_one_to_one"
+pub enum Normalization {
+    /// 映射到 [0.0, 1.0] -> pixel / 255.0
+    ZeroToOne,
+    /// 映射到 [-1.0, 1.0] -> (pixel / 255.0 - 0.5) / 0.5
+    MinusOneToOne,
+}
+
+impl Normalization {
+    /// 统一归一化计算逻辑
+    #[inline(always)]
+    pub fn normalize(&self, pixel: f32) -> f32 {
+        match self {
+            Normalization::ZeroToOne => pixel / 255.0,
+            Normalization::MinusOneToOne => (pixel / 255.0 - 0.5) / 0.5,
+        }
+    }
+}
+
 /// 图像缩放策略枚举
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
 pub enum Resize {
@@ -29,6 +49,13 @@ struct ModelMetadataDto {
     #[serde(alias = "image")]
     resize: Vec<i32>,
     channel: u8,
+    /// 新增：允许在配置文件中指定归一化策略。
+    /// 使用 serde(default) 可以在不配置时提供一个默认值（比如默认 ZeroToOne）
+    #[serde(default = "default_normalization")]
+    normalization: Normalization,
+}
+fn default_normalization() -> Normalization {
+    Normalization::ZeroToOne
 }
 
 #[derive(Debug, Clone)]
@@ -41,18 +68,32 @@ pub struct ModelMetadata {
     pub resize: Resize,
     /// 图像通道数 (1 或 3)
     pub channel: u8,
+    /// 新增：传递给核心业务使用的归一化配置
+    pub normalization: Normalization,
 }
 
 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::Fixed(64, 64), 1)
+        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::Fixed(64, 64), 1)
+        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
@@ -61,6 +102,7 @@ impl ModelMetadata {
         word: bool,
         resize: Resize,
         channel: u8,
+        normalization: Normalization,
     ) -> Self {
         let tokens: Vec<Cow<'static, str>> = slice.iter().map(|&s| Cow::Borrowed(s)).collect();
         Self {
@@ -68,6 +110,7 @@ impl ModelMetadata {
             word,
             resize,
             channel,
+            normalization,
         }
     }
 
@@ -117,6 +160,7 @@ impl ModelMetadata {
             word: dto.word,
             resize,
             channel: dto.channel,
+            normalization: dto.normalization,
         })
     }
 }

src/models/det.rs

@@ -4,6 +4,20 @@ use image::{DynamicImage, GenericImageView, imageops::FilterType};
 use tract_onnx::prelude::tract_ndarray::{Array2, Array3, Array4, Axis, prelude::*, s};
 use tract_onnx::prelude::{Graph, RunnableModel, Tensor, TypedFact, TypedOp, tvec};
 
+
+
+#[derive(Debug, Clone, Copy)]
+pub struct DetectionResult {
+    pub x1: i32,
+    pub y1: i32,
+    pub x2: i32,
+    pub y2: i32,
+    pub score: f32,
+    pub class_id: u32,
+}
+
+
+
 pub struct Det {
     session: RunnableModel<TypedFact, Box<dyn TypedOp>, Graph<TypedFact, Box<dyn TypedOp>>>,
 }
@@ -20,14 +34,14 @@ impl Det {
         let session = ModelLoader::load_model(&model_path)?.session;
         Ok(Self { session })
     }
-    pub fn predict(&self, image_bytes: &[u8]) -> Result<Vec<Vec<i32>>> {
+    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_bytes)
+        self.get_bbox(image)
     }
     /// 2. preproc: 纯 Rust 实现 (替代 OpenCV)
-    fn preproc(&self, img: &DynamicImage, input_size: (u32, u32)) -> Result<(Tensor, f32)> {
+    fn preproc(&self, image: &DynamicImage, input_size: (u32, u32)) -> Result<(Tensor, f32)> {
         let (target_h, target_w) = input_size;
-        let (img_w, img_h) = img.dimensions();
+        let (img_w, img_h) = image.dimensions();
 
         // 计算缩放比例 (Letterbox)
         let r = (target_h as f32 / img_h as f32).min(target_w as f32 / img_w as f32);
@@ -35,7 +49,7 @@ impl Det {
         let new_w = (img_w as f32 * r) as u32;
 
         // Resize 图像
-        let resized = img.resize_exact(new_w, new_h, FilterType::Triangle);
+        let resized = image.resize_exact(new_w, new_h, FilterType::Triangle);
         // 2. 关键：将 DynamicImage 显式转换为 RgbImage (Rgb<u8>)
         let resized_rgb = resized.to_rgb8();
         // 创建 114 灰度填充的背景
@@ -45,21 +59,24 @@ impl Det {
         // 将 resize 后的图像覆盖到左上角 (类似于原始代码中的 padded_img[:h, :w])
         image::imageops::overlay(&mut base_img, &resized_rgb, 0, 0);
 
+        // 优化：直接获取底层的扁平 raw buffer，比 enumerate_pixels() 快得多
+        let raw_samples = base_img.as_flat_samples();
+        let slice = raw_samples.as_slice();
+
         // 构造 NCHW Tensor
         let mut array = Array4::<f32>::zeros((1, 3, target_h as usize, target_w as usize));
-        for (x, y, pixel) in base_img.enumerate_pixels() {
+
-            let x = x as usize;
+        // 用连续的 stride 步长进行写入，提高 CPU 缓存利用率
-            let y = y as usize;
+        for y in 0..target_h as usize {
-            // 核心对标 Python 的 BGR 逻辑：
+            for x in 0..target_w as usize {
-            // pixel[0] 是 R, pixel[1] 是 G, pixel[2] 是 B
+                let idx = (y * target_w as usize + x) * 3;
-            // 如果模型需要 BGR：
+                // BGR 赋值
-            // array[[0, 0, y as usize, x as usize]] = pixel[0] as f32;
+                array[[0, 0, y, x]] = slice[idx + 2] as f32; // B
-            // array[[0, 1, y as usize, x as usize]] = pixel[1] as f32;
+                array[[0, 1, y, x]] = slice[idx + 1] as f32; // G
-            // array[[0, 2, y as usize, x as usize]] = pixel[2] as f32;
+                array[[0, 2, y, x]] = slice[idx] as f32;     // R
-            array[[0, 0, y, x]] = pixel[2] as f32; // B
-            array[[0, 1, y, x]] = pixel[1] as f32; // G
-            array[[0, 2, y, x]] = pixel[0] as f32; // R
             }
+        }
+
 
         Ok((array.into(), r))
     }
@@ -161,13 +178,14 @@ impl Det {
     }
 
     /// 5. multiclass_nms
+    //multiclass_nms_class_agnostic
     pub fn multiclass_nms(
         &self,
         boxes: &Array2<f32>,  // [25200, 4] -> xyxy 格式
         scores: &Array2<f32>, // [25200, 80] -> 已经乘以 objectness 的得分
         nms_thr: f32,
         score_thr: f32,
-    ) -> Vec<Vec<f32>> {
+    ) -> Vec<[f32; 6]> {
         let mut candidates = Vec::new();
 
         // 1. 筛选高分框 (单次遍历完成 Argmax 和 Threshold 过滤)
@@ -213,17 +231,17 @@ impl Det {
             .map(|k_idx| {
                 let (orig_idx, score, cls_id) = candidates[k_idx];
                 let b = boxes.row(orig_idx);
-                vec![b[0], b[1], b[2], b[3], score, cls_id as f32]
+                [b[0], b[1], b[2], b[3], score, cls_id as f32]
             })
             .collect()
     }
     /// 6. get_bbox (完全解耦 OpenCV)
-    pub fn get_bbox(&self, image_bytes: &[u8]) -> Result<Vec<Vec<i32>>> {
+    pub fn get_bbox(&self, dynamic_img: &DynamicImage) -> Result<Vec<DetectionResult>> {
         // 使用 utils crate 解码
-        let dynamic_img = image::load_from_memory(image_bytes).context("Failed to decode utils")?;
+        // let dynamic_img = image::load_from_memory(image_bytes).context("Failed to decode utils")?;
         let (orig_w, orig_h) = dynamic_img.dimensions();
 
-        let (input_tensor, ratio) = self.preproc(&dynamic_img, (416, 416))?;
+        let (input_tensor, ratio) = self.preproc(dynamic_img, (416, 416))?;
 
         // tract 推理
         let outputs = self.session.run(tvec!(input_tensor.into()))?;
@@ -236,7 +254,13 @@ impl Det {
         let pred = predictions.slice(s![0, .., ..]);
 
         let boxes = pred.slice(s![.., 0..4]);
-        let scores = &pred.slice(s![.., 4..5]) * &pred.slice(s![.., 5..]);
+        let obj_conf = pred.slice(s![.., 4..5]);
+        let cls_conf = pred.slice(s![.., 5..]);
+        let obj_broadcast = obj_conf
+            .broadcast(cls_conf.dim())
+            .context("ndarray broadcasting failed for scores calculation")?;
+        let scores = &obj_broadcast * &cls_conf;
+        // let scores = &pred.slice(s![.., 4..5]) * &pred.slice(s![.., 5..]);
 
         let mut boxes_xyxy = Array2::<f32>::zeros(boxes.raw_dim());
         for i in 0..boxes.nrows() {
@@ -247,17 +271,19 @@ impl Det {
         }
 
         let detections = self.multiclass_nms(&boxes_xyxy, &scores, 0.45, 0.1);
-
+        let final_results = detections
-        Ok(detections
             .into_iter()
             .map(|d| {
-                vec![
+                DetectionResult{
-                    (d[0] as i32).max(0).min(orig_w as i32),
+                    x1: (d[0] as i32).max(0).min(orig_w as i32),
-                    (d[1] as i32).max(0).min(orig_h as i32),
+                    y1: (d[1] as i32).max(0).min(orig_h as i32),
-                    (d[2] as i32).max(0).min(orig_w as i32),
+                    x2: (d[2] as i32).max(0).min(orig_w as i32),
-                    (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],
+                    class_id: d[5] as u32,
+                }
             })
-            .collect())
+            .collect();
+        Ok(final_results )
     }
 }

src/models/ocr.rs

@@ -1,19 +1,108 @@
-use crate::charset::CharsetRestrict;
+use crate::charset::{TokenFilter, ValidationCtx};
-use crate::model_metadata::ModelMetadata;
+use crate::model_metadata::{ModelMetadata, Resize};
 use crate::models::base::ModelArgs;
 use crate::models::loader::{ModelLoader, ModelSession, ModelType};
+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 image::DynamicImage;
+use anyhow::{Result, anyhow};
+use image::{DynamicImage, ImageBuffer, Rgb};
+use serde::Serialize;
+use std::borrow::Cow;
 use std::collections::HashSet;
-use tract_onnx::prelude::tract_ndarray::s;
+use std::fmt;
+use tract_onnx::prelude::tract_ndarray::{ArrayView2, Ix2, s};
 use tract_onnx::prelude::{
     DatumType, Graph, IntoTensor, RunnableModel, Tensor, TypedFact, TypedOp, tract_ndarray, tvec,
 };
+// 引入 cv_ops 模块中的 OpenCV HSV 转换算子
+use crate::utils::cv_ops::rgb_to_opencv_hsv;
 
-// 颜色过滤的自定义范围：(低值RGB, 高值RGB)
+/// 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）
-pub type ColorRange = ((u8, u8, u8), (u8, u8, u8));
+#[derive(Debug, Clone, Serialize)]
+pub enum OcrResult {
+    /// 纯文本分支（对应 probability = false）
+    Text(String),
+    /// 包含全量概率的分支（对应 probability = true）
+    Probability {
+        text: String,
+        /// 满额概率矩阵 [Steps, Classes]
+        probabilities: Vec<Vec<f32>>,
+        /// 全局平均置信度
+        confidence: f64,
+    },
+    /// 不支持的模型或未知输出
+    Unsupported { message: String },
+}
+impl OcrResult {
+    /// 消费自身，直接提取最终文本
+    pub fn into_text(self) -> String {
+        match self {
+            OcrResult::Text(text) => text,
+            OcrResult::Probability { text, .. } => text,
+            OcrResult::Unsupported { message } => {
+                // 作为库，这里可以返回空，或者直接携带错误信息，取决于你的设计
+                format!("Error: {}", message)
+            }
+        }
+    }
+}
+impl fmt::Display for OcrResult {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        match self {
+            OcrResult::Text(text) => {
+                // 纯文本分支，直接输出文本内容
+                write!(f, "{}", text)
+            }
+            OcrResult::Probability {
+                text,
+                probabilities,
+                confidence,
+            } => {
+                // 概率分支，友好地展示文本以及百分比形式的置信度
+                // 1. 基本信息
+                write!(f, "{} (置信度: {:.2}%)", text, confidence * 100.0)?;
+
+                // 2. 概率矩阵流式安全打印
+                write!(f, " [概率矩阵预览: ")?;
+
+                let max_steps_to_show = 10;
+                let take_steps = probabilities.iter().take(max_steps_to_show);
+
+                for (i, step_probs) in take_steps.enumerate() {
+                    if i > 0 {
+                        write!(f, ", ")?;
+                    }
+
+                    // 为了防止单行内部数据过长，单行也做一下截断保护（比如每行最多显示前 3 个概率）
+                    let max_classes_to_show = 3;
+                    write!(f, "[")?;
+                    for (j, prob) in step_probs.iter().take(max_classes_to_show).enumerate() {
+                        if j > 0 {
+                            write!(f, ", ")?;
+                        }
+                        write!(f, "{:.4}", prob)?;
+                    }
+                    if step_probs.len() > max_classes_to_show {
+                        write!(f, ", ..")?;
+                    }
+                    write!(f, "]")?;
+                }
+
+                // 如果总 Step 数量超过 10，末尾追加 .. 表示截断
+                if probabilities.len() > max_steps_to_show {
+                    write!(f, ", ..")?;
+                }
+                write!(f, "]")
+            }
+            OcrResult::Unsupported { message } => {
+                // 错误分支，直观输出异常原因
+                write!(f, "未识别成功: {}", message)
+            }
+        }
+    }
+}
 
 pub struct Ocr {
     pub session: RunnableModel<TypedFact, Box<dyn TypedOp>, Graph<TypedFact, Box<dyn TypedOp>>>,
@@ -21,7 +110,7 @@ pub struct Ocr {
 }
 impl ModelSession for Ocr {
     fn get_model_type(&self) -> ModelType {
-        todo!()
+        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()
@@ -29,43 +118,51 @@ impl ModelSession for Ocr {
 }
 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,
         })
     }
-    pub fn predict<'a>(&'a self, image: &'a DynamicImage) -> OcrBuilder<'a> {
+    /// 对应 Python 的 _inference
-        OcrBuilder::new(self, image)
+    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 OcrBuilder<'a> {
+pub struct OcrPredictor<'a> {
     ocr: &'a Ocr,
-    image: &'a DynamicImage,
     /// 是否修复PNG格式问题
     png_fix: bool,
     /// 是否返回概率信息
-    #[allow(dead_code)]
     probability: bool,
     /// 颜色过滤：保留的颜色列表
-    color_filter_colors: Option<Vec<ColorRange>>,
+    color_filter: Result<Option<Vec<HsvRange>>, String>,
-    /// 颜色过滤：自定义RGB范围
+
-    color_filter_custom_ranges: Option<Vec<ColorRange>>,
     /// 字符集范围
-    charset_restrict: Option<CharsetRestrict>,
+    charset_restrict: Option<Vec<usize>>,
 }
 
-impl<'a> OcrBuilder<'a> {
+impl<'a> OcrPredictor<'a> {
     // 初始化任务，设置默认参数
-    pub fn new(ocr: &'a Ocr, image: &'a DynamicImage) -> Self {
+    pub fn new(ocr: &'a Ocr) -> Self {
         Self {
             ocr,
-            image,
             png_fix: false, // 默认值
             probability: false,
-            color_filter_colors: None,
+            color_filter: Ok(None),
-            color_filter_custom_ranges: None,
             charset_restrict: None,
         }
     }
@@ -73,92 +170,280 @@ impl<'a> OcrBuilder<'a> {
         self.png_fix = value;
         self
     }
-    pub fn color_filter_colors(mut self, value: Vec<ColorRange>) -> Self {
+    pub fn probability(mut self, value: bool) -> Self {
-        self.color_filter_colors = Some(value);
+        self.probability = value;
-        self
-    }
-    pub fn color_filter_custom_ranges(mut self, value: Vec<ColorRange>) -> Self {
-        self.color_filter_custom_ranges = Some(value);
-        self
-    }
-    pub fn charset_restrict(mut self, restrict: CharsetRestrict) -> Self {
-        self.charset_restrict = Some(restrict);
         self
     }
 
-    pub fn run(&self) -> anyhow::Result<String> {
+    pub fn color_filter(mut self, filter: &dyn ColorFilter) -> Self {
-        let tensor = self.preprocess_image(self.image, self.png_fix)?;
+        // 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() {
-        let raw_tensor = self.inference(tensor)?;
+            Ok(new_ranges) => self.color_filter = Ok(new_ranges),
-        let raw_indices = self.extract_indices_from_tensor(&raw_tensor)?;
+            Err(err_msg) => self.color_filter = Err(err_msg), // 校验失败，Builder 正式中毒
-
-        // 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);
-
-        println!("最终识别出的验证码是: {}", final_text);
-        Ok(final_text)
         }
 
-    /// 对应 Python 的 _preprocess_image
+        self
-    /// 负责：透明背景修复 -> 灰度化 -> 按比例 Resize -> 归一化 -> 4维张量转换
+    }
-    fn preprocess_image(&self, img: &DynamicImage, png_fix: bool) -> anyhow::Result<Tensor> {
+
-        // A. 修复 PNG 透明背景 (内部逻辑你之前已实现)
+    pub fn charset_restrict(mut self, restrict: &dyn TokenFilter) -> Self {
-        let _ = if png_fix && img.color().has_alpha() {
+        let charset = &self.ocr.model_metadata.charset;
-            png_rgba_white_preprocess(img)
+        let tokens = &charset.tokens;
-        } else {
+        self.charset_restrict = restrict.apply_to_charset(tokens);
-            img.clone()
+        self
+    }
+}
+impl<'a> OcrPredictor<'a> {
+    pub fn predict(self, image: &DynamicImage) -> anyhow::Result<OcrResult> {
+        println!("当前颜色过滤器状态: {:?}", self.color_filter);
+        // =====================================================================
+        // 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 {
+            Err(err_msg) => {
+                return Err(anyhow::anyhow!(
+                    "颜色过滤器初始化失败，全链路短路: {}",
+                    err_msg
+                ));
+            }
+            Ok(None) => {
+                // 核心优化点：直接借用原图，不发生任何克隆
+                Cow::Borrowed(image)
+            }
+            Ok(Some(ranges)) => {
+                // 只有真正需要过滤时，才在内部提取像素并生成清洗后的 Owned 新图
+                let filtered_img = filter_image(image, ranges)?;
+                Cow::Owned(filtered_img)
+            }
+        };
+        let tensor = self.preprocess_image(&img_cow)?;
+
+        let raw_tensor = self.ocr.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 h = 64u32;
+        // let raw_indices = self.ocr.extract_indices_from_tensor(&raw_tensor)?;
-        let w = (img.width() as f32 * (h as f32 / img.height() as f32)) as u32;
+        // // 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 gray_img = convert_to_grayscale(img);
+        // let final_text = self.ctc_decode_to_string(&raw_indices);
-        let resized = resize_image(&gray_img, w, h);
-        // resized.save("debug_preprocessed.png").unwrap();
-        // 1. 预处理：转灰度 -> Resize -> 归一化
-        // let resized = img.resize_exact(w, h, FilterType::Lanczos3).to_luma8();
 
-        // 使用 tract_ndarray 构造，避免版本冲突
+        Ok(ocr_output)
-        let array =
+    }
-            tract_ndarray::Array4::from_shape_fn((1, 1, h as usize, w as usize), |(_, _, y, x)| {
+    /// 对应 Python 的 _preprocess_image
-                let pixel = resized.get_pixel(x as u32, y as u32)[0] as f32;
+    /// 负责：透明背景修复 -> 灰度化 -> 按比例 Resize -> 归一化 -> 4维张量转换
-                (pixel / 255.0 - 0.5) / 0.5
+    fn preprocess_image(&self, img: &DynamicImage) -> anyhow::Result<Tensor> {
-            });
+        // 1. 获取模型元数据配置
+        let meta = &self.ocr.model_metadata;
+        let norm = &meta.normalization; // 获取归一化器
 
-        let tensor = Tensor::from(array);
+        // A. 修复 PNG 透明背景 (内部逻辑你之前已实现)
+        let current_img = if self.png_fix && img.color().has_alpha() {
+            // 只有满足条件才去触发分配，生成新图
+            Cow::Owned(png_rgba_white_preprocess(img))
+        } else {
+            // 正常情况下，仅仅是再次安全借用，无开销
+            Cow::Borrowed(img)
+        };
+
+        // 3. 管道节点 2: 根据 Resize 策略计算目标宽高并进行缩放
+        let (target_w, target_h) = match meta.resize {
+            Resize::Fixed(w, h) => (w, h),
+            Resize::DynamicWidth(h) => {
+                // 高度固定，宽度根据原始比例动态计算：W_target = W_orig * (H_target / H_orig)
+                let w =
+                    (current_img.width() as f32 * (h as f32 / current_img.height() as f32)) as u32;
+                (w, h)
+            }
+            Resize::Square(size) => {
+                // 单字识别模型，直接缩放为正方形
+                (size, size)
+            }
+        };
+        // 执行缩放
+        let resized_img = resize_image(&current_img, target_w, target_h);
+
+        // 4. 管道节点 3: 颜色通道转换（单通道灰度 vs 三通道 RGB）与 4D 张量填充
+        let tensor = 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(
+                    (1, 1, target_h as usize, target_w as usize),
+                    |(_, _, y, x)| {
+                        let pixel = gray_img.get_pixel(x as u32, y as u32)[0] as f32;
+                        // pixel / 255.0 // 严格对齐 Python 归一化 [0.0, 1.0]
+                        // (pixel / 255.0 - 0.5) / 0.5
+                        norm.normalize(pixel)
+                    },
+                );
+                Tensor::from(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(
+                    (1, 3, target_h as usize, target_w as usize),
+                    |(_, c, y, x)| {
+                        let pixel = rgb_img.get_pixel(x as u32, y as u32)[c] as f32;
+                        // pixel / 255.0 // 严格对齐 Python 归一化 [0.0, 1.0]
+                        // (pixel / 255.0 - 0.5) / 0.5
+                        norm.normalize(pixel)
+                    },
+                );
+                Tensor::from(array)
+            }
+
+            _ => return Err(anyhow::anyhow!("不支持的通道数配置: {}", meta.channel)),
+        };
 
         Ok(tensor)
+
+        // let h = 64u32;
+        // let w = (current_img.width() as f32 * (h as f32 / current_img.height() as f32)) as u32;
+        // let gray_img = convert_to_grayscale(&current_img);
+        // let resized = resize_image(&gray_img, w, h);
+        // // resized.save("debug_preprocessed.png").unwrap();
+        // // 1. 预处理：转灰度 -> Resize -> 归一化
+        // // let resized = img.resize_exact(w, h, FilterType::Lanczos3).to_luma8();
+        //
+        // // 使用 tract_ndarray 构造，避免版本冲突
+        // let array =
+        //     tract_ndarray::Array4::from_shape_fn((1, 1, h as usize, w as usize), |(_, _, y, x)| {
+        //         let pixel = resized.get_pixel(x as u32, y as u32)[0] as f32;
+        //         (pixel / 255.0 - 0.5) / 0.5
+        //     });
+        //
+        // let tensor = Tensor::from(array);
+        //
+        // Ok(tensor)
     }
-    /// 对应 Python 的 _inference
+}
-    fn inference(&self, tensor: Tensor) -> anyhow::Result<Tensor> {
+impl<'a> OcrPredictor<'a> {
-        // tract 的 run 会返回一个 Vec<TValue>，我们通常只需要第一个输出
+    fn is_valid_indices(&self, idx: usize) -> bool {
-        // let result = self.session.run(tvec!(tensor.into()))?;
+        if idx >= self.ocr.model_metadata.charset.size() {
-        let mut result = self
+            return false;
-            .ocr
-            .session
-            .run(tvec!(tensor.into()))
-            .context("执行模型推理失败")?;
-        println!("模型输出原始数据: {:?}", result);
-        Ok(result.remove(0).into_tensor())
         }
 
-    /// 核心解析逻辑：将模型输出的各种维度/类型的 Tensor 转为字符索引序列
+        match &self.charset_restrict {
-    fn extract_indices_from_tensor(&self, raw_tensor: &Tensor) -> anyhow::Result<Vec<i64>> {
+            Some(v) => v.binary_search(&idx).is_ok(),
+            None => true,
+        }
+    }
+    /// 【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 tokens = &charset.tokens;
+        match &self.charset_restrict {
+            Some(indices) => indices
+                .iter()
+                .filter_map(|&idx| tokens.get(idx).map(|cow| cow.as_ref()))
+                .collect(),
+            // 如果是 None，现场映射出全量 Token 视图给外部
+            None => tokens.iter().map(|cow| cow.as_ref()).collect(),
+        }
+    }
+    pub fn valid_size(&self) -> usize {
+        match &self.charset_restrict {
+            Some(indices) => indices.len(),
+            None => self.ocr.model_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
+    fn compute_f32_full_probability(
+        &self,
+        matrix_view: ArrayView2<f32>,
+    ) -> (Vec<Vec<f32>>, f32, Vec<i64>) {
+        let steps = matrix_view.nrows();
+        let classes = matrix_view.ncols();
+
+        // 1. 预分配满额概率矩阵内存
+        let mut prob_matrix = tract_ndarray::Array2::<f32>::zeros((steps, classes));
+        let mut predicted_indices = Vec::with_capacity(steps);
+        let mut confidence_sum = 0.0f32;
+
+        // 2. 融合单次遍历
+        for (step_idx, row) in matrix_view.outer_iter().enumerate() {
+            // 寻找当前 Step 的最大值和最大值索引 (Argmax)
+            let (row_max_idx, max_logit) = row
+                .iter()
+                .enumerate()
+                .max_by(|(_, a), (_, b)| a.total_cmp(b))
+                .map(|(idx, &val)| (idx, val))
+                .unwrap_or((0, 0.0));
+
+            predicted_indices.push(row_max_idx as i64);
+
+            // 计算单行 exp 溢出防范和
+            let mut exp_sum = 0.0f32;
+            for &val in row.iter() {
+                exp_sum += (val - max_logit).exp();
+            }
+
+            // 归一化 Softmax 顺序写入
+            for (class_idx, &val) in row.iter().enumerate() {
+                prob_matrix[[step_idx, class_idx]] = (val - max_logit).exp() / exp_sum;
+            }
+
+            // 当前 Step 最大概率在线累加
+            confidence_sum += 1.0f32 / exp_sum;
+        }
+
+        // 3. 统计全局平均置信度
+        let confidence = if steps > 0 {
+            confidence_sum / steps as f32
+        } else {
+            1.0
+        };
+
+        // 4. 将矩阵转化为标准安全序列化格式 [Steps, Classes]
+        let probabilities_list: Vec<Vec<f32>> =
+            prob_matrix.outer_iter().map(|row| row.to_vec()).collect();
+
+        (probabilities_list, confidence, predicted_indices)
+    }
+    /// 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 datum_type = raw_tensor.datum_type();
-        println!("模型输出datum_type数据: {:?}", datum_type);
-
-        match raw_tensor.datum_type() {
-            // 情况 1: huashi666 式模型，直接输出 i64 索引 (通常是模型内部做好了 Argmax)
-            DatumType::I64 => {
-                let view = raw_tensor.to_array_view::<i64>()?;
-                Ok(view.iter().cloned().collect())
-            }
-
-            // 情况 2: sml2h3 原版模型，输出 F32 概率矩阵
-            DatumType::F32 => {
         let view = raw_tensor.to_array_view::<f32>()?;
-                let (steps, classes, data_view) = match shape.len() {
+
+        // 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
@@ -173,102 +458,56 @@ impl<'a> OcrBuilder<'a> {
                     (shape[1], shape[2], sliced.into_dyn())
                 }
             }
-                    2 => {
             // 形状: [Steps, Classes] -> 已经剥离了 Batch 维度
-                        (shape[0], shape[1], view.into_dyn())
+            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 array_2d = data_view.to_shape((steps, classes))?;
+        let matrix_cow = data_dyn_view
-                //
+            .to_shape(Ix2(steps, classes))
-                // 对每一行执行 Argmax (寻找概率最大的字符索引)
+            .map_err(|e| anyhow::anyhow!("转换为2D静态矩阵失败: {:?}", e))?;
-                let indices = array_2d
+
+        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)| {
+                        .max_by(|(_, a), (_, b)| a.total_cmp(b))
-                                a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal)
-                            })
                         .map(|(idx, _)| idx as i64)
                         .unwrap_or(0)
                 })
                 .collect();
-                Ok(indices)
+
-            }
+            let final_text = self.ctc_decode_to_string(&predicted_indices);
-            _ => Err(anyhow::anyhow!(
+            Ok(OcrResult::Text(final_text))
-                "不支持的模型输出数据类型: {:?}",
-                raw_tensor.datum_type()
-            )),
         }
     }
     /// 获取有效字符索引列表 (用于外部验证或过滤)
-    pub fn get_valid_indices(&self) -> HashSet<usize> {
-        let (_, valid_indices) = self.valid_indices();
-        valid_indices
-    }
-
-    fn valid_indices(&self) -> (bool, HashSet<usize>) {
-        let charset = &self.ocr.model_metadata.charset;
-        let tokens = &charset.tokens;
-        /// 根据传入的 CharsetRestrict 枚举策略，动态更新有效索引
-        // 1. 🧠 零克隆防御战：在局部判断并动态构建专属的白名单判定表
-        let mut valid_indices = HashSet::new();
-        let mut has_any_match = false;
-        if let Some(ref policy) = self.charset_restrict {
-            // 🧠 性能精算：根据限制策略的类型，智能分配合适的初始容量，1个字节都不浪费！
-            let estimated_capacity = match policy {
-                CharsetRestrict::Digit => 16,
-                CharsetRestrict::Lowercase | CharsetRestrict::Uppercase => 32,
-                CharsetRestrict::CustomList(vec) => vec.len() + 1, // 动态匹配列表大小
-                _ => 128, // 组合子（Or）等复杂情况，给个 128 黄金保底值
-            };
-            // 🚀 精准开辟内存，完美避开 8120 个槽位的巨大空置浪费
-            valid_indices = HashSet::with_capacity(estimated_capacity);
-
-            for (idx, token) in tokens.iter().enumerate() {
-                let token_str = token.as_ref();
-                // CTC Blank 空字符串无条件放行，其余交给超高性能的 matches
-                if token_str.is_empty() {
-                    valid_indices.insert(idx);
-                } else if policy.matches(token_str) {
-                    valid_indices.insert(idx);
-                    has_any_match = true;
-                }
-            }
-
-            // 终极防御：如果除了 Blank 之外，没有任何一个字符被匹配到
-            if !has_any_match {
-                valid_indices = (0..tokens.len()).collect();
-                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。");
-            }
-        }
-        (has_any_match, valid_indices)
-    }
-    /// 🌟 【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 get_valid_tokens(&self) -> Vec<&str> {
-        let charset = &self.ocr.model_metadata.charset;
-        let tokens = &charset.tokens;
-        self.get_valid_indices()
-            .iter()
-            .map(|&idx| tokens[idx].as_ref())
-            .collect()
-    }
-    pub fn valid_size(&self) -> usize {
-        self.get_valid_indices().len()
-    }
     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 tokens = &charset.tokens;
         // let valid_indices = &charset.valid_indices;
 
-        let (has_any_match, valid_indices) = self.valid_indices();
-
         // 对应 _ctc_decode_indices 的逻辑：去重、去 blank (0)
         let mut res = String::new();
         let mut prev_idx: i64 = -1;
@@ -291,11 +530,11 @@ impl<'a> OcrBuilder<'a> {
                 Err(_) => continue,
             };
 
-            // 4. 终极性能：既然你的有效索引库必然有全量或部分数据，这里直接进行 O(1) 包含校验
+            // 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！
-            // 注意：去掉原本错误的 `!`
+            // 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 has_any_match {
+            if let Some(ref indices) = self.charset_restrict {
-                if !valid_indices.contains(&u_idx) {
+                if indices.binary_search(&u_idx).is_err() {
-                    continue; // 不在有效字符集内，安全跳过
+                    continue;
                 }
             }
 

src/models/slide.rs

@@ -27,7 +27,7 @@ impl Slide {
         Self
     }
 
-    /// 对应 Python: slide_match
+    /// 对应 Python: slide_match 滑块匹配接口
     pub fn slide_match(
         &self,
         target_image: &DynamicImage,
@@ -38,9 +38,8 @@ impl Slide {
         let background_array = image_to_ndarray(background_image);
 
         self.perform_slide_match(target_array.view(), background_array.view(), simple_target)
-            .map_err(|e| anyhow!("滑块匹配失败: {}", e))
     }
-    /// 对应 Python: slide_comparison
+    /// 对应 Python: slide_comparison 差异比较接口
     /// 用于比较带坑位的图片与原始背景图，定位差异点
     pub fn slide_comparison(
         &self,
@@ -53,7 +52,6 @@ impl Slide {
 
         // 2. 执行比较逻辑 (对应 _perform_slide_comparison)
         self.perform_slide_comparison(target_array.view(), background_array.view())
-            .map_err(|e| anyhow!("滑块比较执行失败: {}", e))
     }
     /// 对应 Python: _perform_slide_comparison
     pub fn perform_slide_comparison(
@@ -61,7 +59,7 @@ impl Slide {
         target: ArrayView3<u8>,
         background: ArrayView3<u8>,
     ) -> Result<SlideResult> {
-        let (h, w, _) = target.dim();
+        // 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
@@ -77,6 +75,26 @@ impl Slide {
         //     }
         // }
         // 1. 计算差异数组 (复用 cv2::absdiff)
+
+        let (th, tw, tc) = target.dim();
+        let (bh, bw, bc) = background.dim();
+
+        // 1. 比较模式下的严格尺寸校验
+        if th != bh || tw != bw || tc != bc {
+            return Err(anyhow!(
+                "比较模式要求两张图分辨率与通道数完全一致！Target: [{}x{}x{}], Background: [{}x{}x{}]",
+                tw,
+                th,
+                tc,
+                bw,
+                bh,
+                bc
+            ));
+        }
+        if th == 0 || tw == 0 {
+            return Err(anyhow!("输入图像尺寸不能为0"));
+        }
+
         let diff_array = abs_diff(&target, &background);
 
         // 2. 转换为灰度数组 (复用你的 cv2.cvtColor)
@@ -130,6 +148,30 @@ impl Slide {
         background: ArrayView3<u8>,
         simple_target: bool, // 增加这个参数
     ) -> Result<SlideResult> {
+        let (th, tw, tc) = target.dim();
+        let (bh, bw, bc) = background.dim();
+
+        // 1. 严格的鲁棒性校验（防止底层的 imageproc 算子崩溃）
+        if th == 0 || tw == 0 || bh == 0 || bw == 0 {
+            return Err(anyhow!("输入图像的宽度或高度不能为0"));
+        }
+        if th > bh || tw > bw {
+            return Err(anyhow!(
+                "尺寸不匹配：滑块模板(target)尺寸 [{}x{}] 不能大于背景图(background) [{}x{}]",
+                tw,
+                th,
+                bw,
+                bh
+            ));
+        }
+        if tc != bc {
+            return Err(anyhow!(
+                "目标图与背景图的通道数不一致 (target: {}, bg: {})",
+                tc,
+                bc
+            ));
+        }
+
         // 1. 统一灰度化
         let target_gray = rgb_to_gray(target);
         let background_gray = rgb_to_gray(background);

src/utils/color_filter.rs

@@ -0,0 +1,249 @@
+use std::str::FromStr;
+use anyhow::anyhow;
+use image::{DynamicImage, ImageBuffer, Rgb};
+use crate::utils::cv_ops::rgb_to_opencv_hsv;
+
+#[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 &&
+            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> {
+    // 1. 统一转换为连续内存的 RGB8 缓冲区 (对应 Python 的 Image 到 RGB/BGR 数组转换)
+    let rgb_img = image.to_rgb8();
+    let (width, height) = rgb_img.dimensions();
+    let mut raw_pixels = rgb_img.into_raw();
+
+    // 2. 密集计算核心：原地流式迭代修改
+    // 每次取出 3 个 u8 字节，分别代表 [R, G, B]，无多余掩膜矩阵内存分配
+    for chunk in raw_pixels.chunks_exact_mut(3) {
+        let r = chunk[0];
+        let g = chunk[1];
+        let b = chunk[2];
+
+        // 像素级转换为 OpenCV 标准的 HSV
+        let (h, s, v) = rgb_to_opencv_hsv(r, g, b);
+
+        // 模拟 Python 的多范围 mask bitwise_or 并在 mask == 0 处刷白
+        // 如果该像素没有命中任何一个配置的颜色区间，立刻原地刷白 [255, 255, 255]
+        if !is_pixel_matched(hsv_ranges, h, s, v) {
+            chunk[0] = 255;
+            chunk[1] = 255;
+            chunk[2] = 255;
+        }
+    }
+
+    // 3. 将扁平字节数组重新打包回 DynamicImage 容器
+    let filtered_buffer = ImageBuffer::<Rgb<u8>, Vec<u8>>::from_raw(width, height, raw_pixels)
+        .ok_or_else(|| anyhow!("图像缓冲重新组装失败，维度与数据大小不匹配"))?;
+
+    Ok(DynamicImage::ImageRgb8(filtered_buffer))
+}
+impl HsvRange {
+    pub const fn new(lower: (u8, u8, u8), upper: (u8, u8, u8)) -> Self {
+        Self { lower, upper }
+    }
+}
+impl HsvRange {
+    /// 验证当前 HSV 范围是否合法
+    /// 对应 Python 逻辑：H 在 0-180，S/V 在 0-255，且下界 <= 上界
+    pub fn validate(&self) -> Result<(), String> {
+        // 1. 校验 H 通道边界 (OpenCV 中 H 范围是 0-180)
+        if self.lower.0 > 180 || self.upper.0 > 180 {
+            return Err("H通道值必须在 0-180 范围内".to_string());
+        }
+
+        // 2. 校验下界不能大于上界
+        if self.lower.0 > self.upper.0 || self.lower.1 > self.upper.1 || self.lower.2 > self.upper.2 {
+            return Err("HSV范围下界不能大于上界".to_string());
+        }
+
+        Ok(())
+    }
+}
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum ColorPreset {
+    Red,
+    Blue,
+    Green,
+    Yellow,
+    Orange,
+    Purple,
+    Cyan,
+    Black,
+    White,
+    Gray,
+    Custom(Vec<HsvRange>),
+}
+
+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 { lower: (170, 50, 50), upper: (180, 255, 255) },
+            ],
+            ColorPreset::Blue   => &[HsvRange { lower: (100, 50, 50), upper: (130, 255, 255) }],
+            ColorPreset::Green  => &[HsvRange { lower: (40, 50, 50), upper: (80, 255, 255) }],
+            ColorPreset::Yellow => &[HsvRange { lower: (20, 50, 50), upper: (40, 255, 255) }],
+            ColorPreset::Orange => &[HsvRange { lower: (10, 50, 50), upper: (20, 255, 255) }],
+            ColorPreset::Purple => &[HsvRange { lower: (130, 50, 50), upper: (170, 255, 255) }],
+            ColorPreset::Cyan   => &[HsvRange { lower: (80, 50, 50), upper: (100, 255, 255) }],
+            ColorPreset::Black  => &[HsvRange { lower: (0, 0, 0), upper: (180, 255, 50) }],
+            ColorPreset::White  => &[HsvRange { lower: (0, 0, 200), upper: (180, 30, 255) }],
+            ColorPreset::Gray   => &[HsvRange { lower: (0, 0, 50), upper: (180, 30, 200) }],
+            ColorPreset::Custom(ranges) => ranges,
+        }
+    }
+    /// 校验逻辑：在这里实现完美的“责任分离”
+    pub fn validate(&self) -> Result<(), String> {
+        match self {
+            // 1. 快捷变体：完全绕过，根本不校验，0 运行时开销放行！
+            ColorPreset::Custom(ranges) => {
+                // 2. 只有 Custom 变体需要接受严格的参数政审
+                for r in ranges {
+                    r.validate()?;
+                }
+                Ok(())
+            }
+            _ => Ok(()),
+        }
+    }
+}
+
+impl FromStr for ColorPreset {
+    type Err = String;
+    fn from_str(s: &str) -> Result<Self, Self::Err> {
+        match s.to_lowercase().as_str() {
+            "red" => Ok(ColorPreset::Red),
+            "blue" => Ok(ColorPreset::Blue),
+            "green" => Ok(ColorPreset::Green),
+            "yellow" => Ok(ColorPreset::Yellow),
+            "orange" => Ok(ColorPreset::Orange),
+            "purple" => Ok(ColorPreset::Purple),
+            "cyan" => Ok(ColorPreset::Cyan),
+            "black" => Ok(ColorPreset::Black),
+            "white" => Ok(ColorPreset::White),
+            "gray" => Ok(ColorPreset::Gray),
+            _ => Err(format!("不支持的颜色预设: {}", s)),
+        }
+    }
+}
+
+// =====================================================================
+// 3. 颜色约束特征（Trait）与组合子设计模式
+// =====================================================================
+
+pub struct PixelCtx {
+    pub hsv: (u8, u8, u8),
+}
+
+/// 统一的颜色约束接口
+pub trait ColorFilter {
+    /// 将自身的有效约束平铺追加到统一的目标容器中
+    fn append_ranges(&self, target: &mut Vec<HsvRange>);
+    /// 预估范围数量，借助原生内置的 len() 实现 O(1) 完美控容
+    fn estimated_count(&self) -> usize;
+    /// 将自身的有效约束平铺追加到统一目标容器中
+    /// 验证当前过滤器是否合法，默认直接放行（Ok(())）
+    fn validate_self(&self) -> Result<(), String> {
+        Ok(())
+    }
+    /// 【新扩展的架构方法】将自身安全的合并到已有的普通容器中，并完成去重和排序
+    /// 完美的责任分离：Builder 不再需要关心怎么分配内存、怎么排序去重
+    fn collect_to_vec(&self) -> Result<Option<Vec<HsvRange>>, String> {
+        // 1. 触发自检
+        self.validate_self()?;
+
+        let total_capacity = self.estimated_count();
+        if total_capacity == 0 {
+            return Ok(None);
+        }
+
+        // 2. 永远一击必中分配精准内存，不需要再考虑追加和扩容！
+        let mut v = Vec::with_capacity(total_capacity.max(16));
+
+        // 2. 倒入数据
+        self.append_ranges(&mut v);
+
+        // 3. 原地完成排序与去重
+        v.sort_unstable();
+        v.dedup();
+
+        Ok(Some(v))
+    }
+}
+
+impl ColorFilter for ColorPreset {
+    fn append_ranges(&self, target: &mut Vec<HsvRange>) {
+        // 直接利用我们第一步写好的 matches() 拿到切片，整块高速拷贝倒入目标容器
+        target.extend_from_slice(self.matches());
+    }
+    fn estimated_count(&self) -> usize {
+        // 直接获取切片长度
+        self.matches().len()
+    }
+    fn validate_self(&self) -> Result<(), String> {
+        // 直接调用我们在第一步中为 ColorPreset 实现的精细化分流校验
+        // 快捷变体在这里会直接返回 Ok(()), 只有 Custom 才会去真正校验
+        self.validate()
+    }
+}
+
+
+/// 多路颜色“或”逻辑组合子（并集网络）
+pub struct MultiOrColorRestrict<'a> {
+    pub filters: Vec<&'a dyn ColorFilter>,
+}
+
+impl<'a> ColorFilter for MultiOrColorRestrict<'a> {
+    fn append_ranges(&self, target: &mut Vec<HsvRange>) {
+        // 管道递延：依次指挥内部每一个子过滤器把数据倒进目标容器
+        for f in &self.filters {
+            f.append_ranges(target);
+        }
+    }
+    fn estimated_count(&self) -> usize {
+        // 数量累加：$O(1)$ 地把所有子过滤器的预估容量加起来
+        self.filters.iter().map(|f| f.estimated_count()).sum()
+    }
+
+    fn validate_self(&self) -> Result<(), String> {
+        // 递归政审：只要其中一个子过滤器校验失败（比如某个 Custom 变体非法），立刻熔断
+        for f in &self.filters {
+            f.validate_self()?;
+        }
+        Ok(())
+    }
+}
+
+// =====================================================================
+// 4. 声明式宏：一语定乾坤
+// =====================================================================
+
+#[macro_export]
+macro_rules! color_any_of {
+    ($only:expr) => {
+        &$only as &dyn $crate::ColorFilter
+    };
+    ($($filter:expr),+ $(,)?) => {
+        &$crate::MultiOrColorRestrict {
+            filters: vec![ $( &$filter as &dyn $crate::ColorFilter ),+ ]
+        }
+    };
+}
+

src/utils/cv_ops.rs

@@ -1,6 +1,6 @@
-use std::cmp::{max, min};
 use image::{ImageBuffer, Luma};
-use tract_onnx::prelude::tract_ndarray::{azip, Array2, Array3, ArrayView2, ArrayView3};
+use std::cmp::{max, min};
+use tract_onnx::prelude::tract_ndarray::{Array2, Array3, ArrayView2, ArrayView3, azip};
 
 /// 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> {
@@ -13,7 +13,6 @@ pub fn abs_diff(a: &ArrayView3<u8>, b: &ArrayView3<u8>) -> Array3<u8> {
     diff
 }
 
-
 /// RGB 到灰度转换
 pub fn rgb_to_gray(rgb: ArrayView3<u8>) -> Array2<u8> {
     let (h, w, _) = rgb.dim();
@@ -67,9 +66,16 @@ pub fn find_contours_and_max(labelled: &ImageBuffer<Luma<u32>, Vec<u32>>) -> Opt
             max_label = label;
         }
     }
-    if max_label == 0 { None } else { Some(max_label) }
+    if max_label == 0 {
+        None
+    } else {
+        Some(max_label)
+    }
 }
-pub fn bounding_rect(labelled: &ImageBuffer<Luma<u32>, Vec<u32>>,max_label: u32) -> (u32, u32, u32, u32) {
+pub fn bounding_rect(
+    labelled: &ImageBuffer<Luma<u32>, Vec<u32>>,
+    max_label: u32,
+) -> (u32, u32, u32, u32) {
     // 5. 计算最大区域的边界框 (对应 cv2.boundingRect)
     let mut min_x = labelled.width();
     let mut max_x = 0;
@@ -85,14 +91,13 @@ pub fn bounding_rect(labelled: &ImageBuffer<Luma<u32>, Vec<u32>>,max_label: u32)
         }
     }
 
-
     let w = max_x - min_x;
     let h = max_y - min_y;
     (min_x, min_y, w, h)
 }
-pub fn calculate_center(max_loc: (u32, u32), tw: usize, th: usize) -> (i32, i32) {
+pub fn calculate_center(top_left: (u32, u32), width: usize, height: usize) -> (i32, i32) {
-    let center_x = max_loc.0 as i32 + (tw as i32 / 2);
+    let center_x = top_left.0 as i32 + (width as i32 / 2);
-    let center_y = max_loc.1 as i32 + (th as i32 / 2);
+    let center_y = top_left.1 as i32 + (height as i32 / 2);
     (center_x, center_y)
 }
 pub fn ndarray_to_luma8(array: ArrayView2<u8>) -> ImageBuffer<Luma<u8>, Vec<u8>> {
@@ -105,3 +110,52 @@ pub fn ndarray_to_luma8(array: ArrayView2<u8>) -> ImageBuffer<Luma<u8>, Vec<u8>>
     }
     buffer
 }
+// =====================================================================
+// 5. 核心高性能图像转换算法 (纯 Rust 编写)
+// =====================================================================
+
+#[inline(always)]
+pub fn rgb_to_opencv_hsv(r: u8, g: u8, b: u8) -> (u8, u8, u8) {
+    // 1. 规避高昂的除法，直接转为 f32 进行比对
+    let r_f = r as f32;
+    let g_f = g as f32;
+    let b_f = b as f32;
+
+    let max = r_f.max(g_f).max(b_f);
+    let min = r_f.min(g_f).min(b_f);
+    let delta = max - min;
+
+    // 2. 计算 H (色调) - 移除负数取余陷阱，改用平铺分支
+    let mut h = if delta == 0.0 {
+        0.0
+    } else if max == r_f {
+        let mut diff = (g_f - b_f) / delta;
+        if diff < 0.0 {
+            diff += 6.0; // 规避 Rust f32 % 负数的行为
+        }
+        60.0 * diff
+    } else if max == g_f {
+        60.0 * (((b_f - r_f) / delta) + 2.0)
+    } else {
+        60.0 * (((r_f - g_f) / delta) + 4.0)
+    };
+
+    // OpenCV 的 H 量化：H / 2
+    // 注意：OpenCV 底层使用截断还是四舍五入与特定版本有关，
+    // 标准的 cvtColor 内部实现通常是： h * (180.0 / 360.0) -> h * 0.5
+    // 这里使用强转(截断)，若单测对齐发现差1，可改为 (h * 0.5 + 0.5) 或 round()
+    let h_opencv = (h * 0.5) as u8;
+
+    // 3. 计算 S (饱和度)
+    // OpenCV 公式: S = max == 0 ? 0 : 255 * delta / max
+    let s_opencv = if max == 0.0 {
+        0
+    } else {
+        ((255.0 * delta) / max) as u8
+    };
+
+    // 4. 计算 V (明度)
+    let v_opencv = max as u8;
+
+    (h_opencv, s_opencv, v_opencv)
+}

src/utils/image_processor.rs

@@ -11,17 +11,27 @@ pub fn convert_to_grayscale(image: &DynamicImage) -> GrayImage {
 /// 对应 Python 的 resize_image
 /// 调整图像尺寸。当前版本仅实现 keep_aspect_ratio=false
 pub fn resize_image(
-    image: &GrayImage,
+    image: &DynamicImage,
     target_width: u32,
     target_height: u32,
     // resample 参数我们直接使用 FilterType，Lanczos3 是最接近 Python LANCZOS 的
-) -> GrayImage {
+) -> DynamicImage {
-    // 使用 resize 算法进行精确缩放
-    image::imageops::resize(
-        image,
-        target_width,
-        target_height,
-        FilterType::Lanczos3
-    )
-}
 
+    // image::imageops::resize 的最高层封装
+    // FilterType::Lanczos3 与 Python Pillow 的 Image.LANCZOS 算法完全对齐，缩放质量最高
+    image.resize_exact(target_width, target_height, FilterType::Lanczos3)
+}
+// pub fn resize_image(
+//     image: &GrayImage,
+//     target_width: u32,
+//     target_height: u32,
+//     // resample 参数我们直接使用 FilterType，Lanczos3 是最接近 Python LANCZOS 的
+// ) -> GrayImage {
+//     // 使用 resize 算法进行精确缩放
+//     image::imageops::resize(
+//         image,
+//         target_width,
+//         target_height,
+//         FilterType::Lanczos3
+//     )
+// }

src/utils/mod.rs

@@ -1,3 +1,4 @@
 pub mod image_io;
 pub mod image_processor;
 pub mod cv_ops;
+pub mod color_filter;

tests/ocr_test.rs

@@ -1,8 +1,10 @@
 use ddddocr_rs::models::slide::Slide;
 use ddddocr_rs::{DdddOcr, DdddOcrBuilder}; // 假设你的包名是这个
-use image::Rgb;
+use image::{DynamicImage, Rgb};
 use std::fs;
 use std::path::Path;
+use ddddocr_rs::models::det::DetectionResult;
+
 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
     let path_ref = path.as_ref();
@@ -15,27 +17,26 @@ 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(
-    image_bytes: &[u8],
+    dynamic_img: &DynamicImage,          // 【优化点 1】直接传入解码好的引用，拒绝重复解码
-    bboxes: &Vec<Vec<i32>>,
+    bboxes: &[DetectionResult],          // 【修改点 1】类型改为自定义结构体切片
     output_path: &str,
 ) -> anyhow::Result<()> {
-    let dynamic_img = image::load_from_memory(image_bytes)?;
+    // 删除了原本的 let dynamic_img = image::load_from_memory(image_bytes)?;
     let mut img = dynamic_img.to_rgb8();
     let (width, height) = img.dimensions();
     let red = Rgb([255u8, 0, 0]);
 
     for bbox in bboxes {
-        // 基础边界检查
+        // 【修改点 2】将原来的索引 bbox[0].. 改为结构体字段访问 .x1, .y1 ..
-        let x1 = bbox[0].max(0).min(width as i32 - 1) as u32;
+        let x1 = bbox.x1.max(0).min(width as i32 - 1) as u32;
-        let y1 = bbox[1].max(0).min(height as i32 - 1) as u32;
+        let y1 = bbox.y1.max(0).min(height as i32 - 1) as u32;
-        let x2 = bbox[2].max(0).min(width as i32 - 1) as u32;
+        let x2 = bbox.x2.max(0).min(width as i32 - 1) as u32;
-        let y2 = bbox[3].max(0).min(height as i32 - 1) as u32;
+        let y2 = bbox.y2.max(0).min(height as i32 - 1) as u32;
 
         // 绘制横向线条
         for x in x1..=x2 {
             img.put_pixel(x, y1, red);
             img.put_pixel(x, y2, red);
-            // 如果要加粗，多画一行
             if y1 + 1 < height {
                 img.put_pixel(x, y1 + 1, red);
             }
@@ -47,7 +48,6 @@ fn save_debug_image(
         for y in y1..=y2 {
             img.put_pixel(x1, y, red);
             img.put_pixel(x2, y, red);
-            // 如果要加粗，多画一列
             if x1 + 1 < width {
                 img.put_pixel(x1 + 1, y, red);
             }
@@ -66,7 +66,7 @@ fn test_full_classification() {
     let ocr = DdddOcrBuilder::new().build().expect("模型加载失败");
 
     // 2. 加载测试图片
-    let img = image::open("samples/code3.png").expect("测试图片不存在");
+    let img = image::open("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");
@@ -82,17 +82,27 @@ fn test_det_load() -> anyhow::Result<()> {
         fs::read(image_path).map_err(|e| anyhow::anyhow!("无法读取图片 {}: {}", image_path, e))?;
 
     println!("图片读取成功，字节大小: {}", image_bytes.len());
-    let bboxes = det.detection(&image_bytes)?;
+
+    // 【修改点 1】将字节流解码为统一的 DynamicImage
+    let img = image::load_from_memory(&image_bytes)
+        .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)?;
     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 {
-        save_debug_image(&image_bytes, &bboxes, "samples/result.jpg")?;
+        // 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")?;
+
         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!(
-                "目标 [{}]: x1={}, y1={}, x2={}, y2={}",
+                "目标 [{}]: x1={}, y1={}, x2={}, y2={}, 分数={:.4}, 类别ID={}",
-                i, bbox[0], bbox[1], bbox[2], bbox[3]
+                i, bbox.x1, bbox.y1, bbox.x2, bbox.y2, bbox.score, bbox.class_id
             );
         }
     }