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

View File

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