feat: 重构 CTC 解码逻辑

- 重构 ctc_decode 为关联函数并优化内存分配。
- 增加 单元测试和集成测试
This commit is contained in:
2026-05-01 21:54:33 +08:00
parent 642fed5d9f
commit 1c366b7165
4 changed files with 106 additions and 152 deletions

5
examples/simple_usage.rs Normal file
View File

@@ -0,0 +1,5 @@
fn main() {
let ocr = ddddocr_rs::DdddOcr::new("model/common.onnx").unwrap();
let img = image::open("samples/code3.png").unwrap();
println!("Result: {}", ocr.classification(&img).unwrap());
}

View File

@@ -1,18 +1,16 @@
mod model;
mod utils;
mod charset; mod charset;
mod image_io; mod image_io;
mod image_processor; mod image_processor;
mod model;
mod utils;
use crate::image_io::png_rgba_white_preprocess;
use crate::image_processor::{convert_to_grayscale, resize_image};
use anyhow::{Context, Result}; use anyhow::{Context, Result};
use image::{DynamicImage, imageops::FilterType}; use image::{DynamicImage, imageops::FilterType};
use tract_onnx::prelude::*; use tract_onnx::prelude::*;
// 关键点:直接使用 tract 重导出的 ndarray // 关键点:直接使用 tract 重导出的 ndarray
use crate::image_io::png_rgba_white_preprocess; use tract_onnx::prelude::tract_ndarray::s;
use crate::image_processor::{convert_to_grayscale, resize_image};
use tract_onnx::prelude::tract_itertools::Itertools;
pub struct DdddOcr { pub struct DdddOcr {
session: RunnableModel<TypedFact, Box<dyn TypedOp>, Graph<TypedFact, Box<dyn TypedOp>>>, session: RunnableModel<TypedFact, Box<dyn TypedOp>, Graph<TypedFact, Box<dyn TypedOp>>>,
} }
@@ -37,14 +35,14 @@ impl DdddOcr {
// 3. 解析结果 // 3. 解析结果
// let output = result[0].to_array_view::<i64>()?; // let output = result[0].to_array_view::<i64>()?;
let output = self.inference(tensor)?; let output = self.inference(tensor)?;
let output2 = self.extract_indices(&output)?; let output2 = self.process_text_output(&output)?;
Ok(self.decode_ctc(&output2)) Ok(Self::ctc_decode_indices(&output2))
} }
/// 对应 Python 的 _preprocess_image /// 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 /// 负责:透明背景修复 -> 灰度化 -> 按比例 Resize -> 归一化 -> 4维张量转换
fn preprocess_image(&self, img: &DynamicImage, png_fix: bool) -> Result<Tensor> { fn preprocess_image(&self, img: &DynamicImage, png_fix: bool) -> Result<Tensor> {
// A. 修复 PNG 透明背景 (内部逻辑你之前已实现) // A. 修复 PNG 透明背景 (内部逻辑你之前已实现)
let processed_img = if png_fix && img.color().has_alpha() { let _ = if png_fix && img.color().has_alpha() {
png_rgba_white_preprocess(img) png_rgba_white_preprocess(img)
} else { } else {
img.clone() img.clone()
@@ -54,6 +52,7 @@ impl DdddOcr {
let w = (img.width() as f32 * (h as f32 / img.height() as f32)) as u32; let w = (img.width() as f32 * (h as f32 / img.height() as f32)) as u32;
let gray_img = convert_to_grayscale(img); let gray_img = convert_to_grayscale(img);
let resized = resize_image(&gray_img, w, h); 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 // 1. 预处理:转灰度 -> Resize -> 归一化
// let resized = img.resize_exact(w, h, FilterType::Lanczos3).to_luma8(); // let resized = img.resize_exact(w, h, FilterType::Lanczos3).to_luma8();
@@ -76,12 +75,15 @@ impl DdddOcr {
.session .session
.run(tvec!(tensor.into())) .run(tvec!(tensor.into()))
.context("执行模型推理失败")?; .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()) 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 /// 核心解析逻辑:将模型输出的各种维度/类型的 Tensor 转为字符索引序列
fn extract_indices(&self, raw_tensor: &Tensor) -> Result<Vec<i64>> { fn process_text_output(&self, raw_tensor: &Tensor) -> Result<Vec<i64>> {
let shape = raw_tensor.shape(); 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() { 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) // 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)
@@ -93,32 +95,43 @@ impl DdddOcr {
// 情况 2: sml2h3 原版模型,输出 F32 概率矩阵 // 情况 2: sml2h3 原版模型,输出 F32 概率矩阵
DatumType::F32 => { DatumType::F32 => {
let view = raw_tensor.to_array_view::<f32>()?; let view = raw_tensor.to_array_view::<f32>()?;
let (steps, classes, data_view) = match shape.len() {
// 处理典型的 CTC 输出形状 [TimeSteps, Batch:1, Classes] 3 => {
if shape.len() == 3 { if shape[1] == 1 {
let steps = shape[0]; // 形状: [Steps, 1, Classes] -> 你的原有逻辑
let classes = shape[2]; (shape[0], shape[2], view.into_dyn())
} else if shape[0] == 1 {
// 将一维视图重新整理为二维 [steps, classes] // 形状: [1, Steps, Classes] -> 另一种常见导出格式
let array_2d = view.to_shape((steps, classes))?; (shape[1], shape[2], view.into_dyn())
} else {
// 对每一行执行 Argmax (寻找概率最大的字符索引) // 默认取第一个 batch: [Batch, Steps, Classes]
let indices = array_2d // 使用 slice 对应 Python 的 output[0, :, :]
.outer_iter() let sliced = view.slice(s![0, .., ..]);
.map(|row| { (shape[1], shape[2], sliced.into_dyn())
row.iter() }
.enumerate() }
.max_by(|(_, a), (_, b)| { 2 => {
a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal) // 形状: [Steps, Classes] -> 已经剥离了 Batch 维度
}) (shape[0], shape[1], view.into_dyn())
.map(|(idx, _)| idx as i64) }
.unwrap_or(0) _ => return Err(anyhow::anyhow!("不支持的输出维度: {:?}", shape)),
}) };
.collect(); let array_2d = data_view.to_shape((steps, classes))?;
Ok(indices) //
} else { // 对每一行执行 Argmax (寻找概率最大的字符索引)
Err(anyhow::anyhow!("不支持的 F32 输出形状: {:?}", shape)) let indices = array_2d
} .outer_iter()
.map(|row| {
row.iter()
.enumerate()
.max_by(|(_, a), (_, b)| {
a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal)
})
.map(|(idx, _)| idx as i64)
.unwrap_or(0)
})
.collect();
Ok(indices)
} }
_ => Err(anyhow::anyhow!( _ => Err(anyhow::anyhow!(
"不支持的模型输出数据类型: {:?}", "不支持的模型输出数据类型: {:?}",
@@ -126,20 +139,44 @@ impl DdddOcr {
)), )),
} }
} }
fn decode_ctc(&self, indices: &[i64]) -> String { fn ctc_decode_indices(predicted_indices: &[i64]) -> String {
use crate::charset::CHARSET_BETA; println!("indices模型输出原始数据: {:?}", predicted_indices);
let mut res = String::new();
let mut last_idx: i64 = -1;
for &idx in indices { use crate::charset::CHARSET_BETA;
// ddddocr 的 blank 通常是 0 // 对应 _ctc_decode_indices 的逻辑:去重、去 blank (0)
if idx != 0 && idx != last_idx { let mut res = String::new();
if let Some(&char_str) = CHARSET_BETA.get(idx as usize) { let mut prev_idx: i64 = -1;
res.push_str(char_str);
for &idx in predicted_indices {
// 1. 跳过连续重复的索引
// 2. 跳过 blank 字符 (假设索引 0 是 blank)
if idx != prev_idx && idx != 0 {
if let Ok(u_idx) = usize::try_from(idx) {
if let Some(&char_str) = CHARSET_BETA.get(u_idx) {
res.push_str(char_str);
}
} }
} }
last_idx = idx; prev_idx = idx;
} }
println!("最终识别出的验证码是: {}", res);
res res
} }
} }
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn test_ctc_decode_indices() {
// 模拟一个 DdddOcr 实例(如果 decode 不依赖 session可以设为相关函数
// 这里假设你的 decode_ctc 是公开或内部可访问的
let input = vec![1, 1, 0, 1, 2, 2, 0, 2];
// 逻辑:[1, 1] -> 1, [0] -> 跳过, [1] -> 1, [2, 2] -> 2, [0] -> 跳过, [2] -> 2
// 预期结果索引应该是 [1, 1, 2, 2] 对应的字符
// 具体的断言取决于你的 CHARSET_BETA
// let result = dddd.ctc_decode_indices(&input);
// assert_eq!(result, "AABB");
}
}

View File

@@ -1,104 +0,0 @@
mod charset;
use anyhow::{anyhow, Result};
use charset::CHARSET_BETA;
use image::{imageops::FilterType, open};
use tract_onnx::prelude::*;
// 编译时读取字典文件
fn main() -> Result<()> {
// 1. 加载并优化模型 (假设模型文件在根目录)
let model = onnx()
.model_for_path("model/common_huashi666_i64.onnx")? // 这里替换成你提取的 ddddocr 模型名
.into_optimized()?
.into_runnable()?;
// 2. 加载并处理图片 (需要缩放到模型要求的尺寸,例如 64x30)
let img = open("samples/code3.png")?;
let h = 64u32;
let w = (img.width() as f32 * (h as f32 / img.height() as f32)) as u32;
// 1. 缩放并转灰度
let resized = img.resize_exact(w, h, FilterType::Lanczos3).to_luma8();
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);
// 4. 运行推理
let result = model.run(tvec!(tensor.into()))?;
// 注意:这里需要根据 ddddocr 的要求将图片转为 Tensor
// 简化逻辑:
// let tensor: Tensor = tract_ndarray::Array4::<f32>::zeros((1, 1, 30, 64)).into();
let raw_tensor = &result[0];
// 3. 运行推理
// let result = model.run(tvec!(tensor.into()))?;
println!("模型输出原始数据: {:?}", result);
let shape = result[0].shape();
println!("模型输出shape数据: {:?}", shape);
let datum_type = result[0].datum_type();
println!("模型输出datum_type数据: {:?}", datum_type);
let predicted_indices: Vec<i64> = match raw_tensor.datum_type() {
// 情况 1: huashi666 式模型,直接输出 i64 索引
DatumType::I64 => {
raw_tensor.to_array_view::<i64>()?.iter().cloned().collect()
}
// 情况 2: sml2h3 原版模型,输出 F32 概率
DatumType::F32 => {
let view = raw_tensor.to_array_view::<f32>()?;
// 模仿 Python 的维度判断逻辑
if shape.len() == 3 {
// 假设形状是 [21, 1, 8210]
let steps = shape[0];
let classes = shape[2];
let array_2d = view.to_shape((
(steps, classes),
tract_onnx::prelude::tract_ndarray::Order::RowMajor
))?;
array_2d.outer_iter()
.map(|row| {
row.iter().enumerate()
.max_by(|(_, a), (_, b)| a.partial_cmp(b).unwrap())
.map(|(idx, _)| idx as i64).unwrap()
})
.collect()
} else {
// 其他形状处理...
vec![]
}
}
_ => return Err(anyhow!("不支持的输出类型")),
};
// let output = result[0].to_array_view::<i64>()?;
// println!("模型输出原始数据2: {:?}", output);
// let indices: Vec<i64> = output.iter().cloned().collect();
// 2. 将视图转为切片并调用函数
let code = decode_ctc(&predicted_indices);
println!("indices模型输出原始数据: {:?}", predicted_indices);
println!("最终识别出的验证码是: {}", code);
Ok(())
}
// common_huashi666_i64
fn decode_ctc(indices: &[i64]) -> String {
let mut res = String::new();
let mut last_idx: i64 = -1;
for &idx in indices {
// idx == 0 通常是 CTC 的 blank 占位符
if idx != 0 && idx != last_idx {
if let Some(&char_str) = CHARSET_BETA.get(idx as usize) {
res.push_str(char_str);
}
}
last_idx = idx;
}
res
}

16
tests/ocr_test.rs Normal file
View File

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