Merge branch 'release/2.5' of https://github.com/PaddlePaddle/PaddleOCR into release/2.5

applications/高精度中文识别模型.md

+# 高精度中文场景文本识别模型SVTR
+
+## 1. 简介
+
+PP-OCRv3是百度开源的超轻量级场景文本检测识别模型库，其中超轻量的场景中文识别模型SVTR_LCNet使用了SVTR算法结构。为了保证速度，SVTR_LCNet将SVTR模型的Local Blocks替换为LCNet，使用两层Global Blocks。在中文场景中，PP-OCRv3识别主要使用如下优化策略（[详细技术报告](../doc/doc_ch/PP-OCRv3_introduction.md)）：
+- GTC：Attention指导CTC训练策略；
+- TextConAug：挖掘文字上下文信息的数据增广策略；
+- TextRotNet：自监督的预训练模型；
+- UDML：联合互学习策略；
+- UIM：无标注数据挖掘方案。
+
+其中 *UIM：无标注数据挖掘方案* 使用了高精度的SVTR中文模型进行无标注文件的刷库，该模型在PP-OCRv3识别的数据集上训练，精度对比如下表。
+
+|中文识别算法|模型|UIM|精度|
+| --- | --- | --- |--- |
+|PP-OCRv3|SVTR_LCNet| w/o |78.4%|
+|PP-OCRv3|SVTR_LCNet| w |79.4%|
+|SVTR|SVTR-Tiny|-|82.5%|
+
+aistudio项目链接: [高精度中文场景文本识别模型SVTR](https://aistudio.baidu.com/aistudio/projectdetail/4263032)
+
+## 2. SVTR中文模型使用
+
+### 环境准备
+
+
+本任务基于Aistudio完成, 具体环境如下：
+
+- 操作系统: Linux
+- PaddlePaddle: 2.3
+- PaddleOCR: dygraph
+
+下载 PaddleOCR代码
+
+```bash
+git clone -b dygraph https://github.com/PaddlePaddle/PaddleOCR
+```
+
+安装依赖库
+
+```bash
+pip install -r PaddleOCR/requirements.txt -i https://mirror.baidu.com/pypi/simple
+```
+
+### 快速使用
+
+获取SVTR中文模型文件，请扫码填写问卷，加入PaddleOCR官方交流群获取全部OCR垂类模型下载链接、《动手学OCR》电子书等全套OCR学习资料🎁
+<div align="center">
+<img src="https://ai-studio-static-online.cdn.bcebos.com/dd721099bd50478f9d5fb13d8dd00fad69c22d6848244fd3a1d3980d7fefc63e"  width = "150" height = "150" />
+</div>
+
+```bash
+# 解压模型文件
+tar xf svtr_ch_high_accuracy.tar
+```
+
+预测中文文本，以下图为例：
+![](../doc/imgs_words/ch/word_1.jpg)
+
+预测命令：
+
+```bash
+# CPU预测
+python tools/infer_rec.py -c configs/rec/rec_svtrnet_ch.yml -o Global.pretrained_model=./svtr_ch_high_accuracy/best_accuracy Global.infer_img=./doc/imgs_words/ch/word_1.jpg Global.use_gpu=False
+
+# GPU预测
+#python tools/infer_rec.py -c configs/rec/rec_svtrnet_ch.yml -o Global.pretrained_model=./svtr_ch_high_accuracy/best_accuracy Global.infer_img=./doc/imgs_words/ch/word_1.jpg Global.use_gpu=True
+```
+
+可以看到最后打印结果为
+- result: 韩国小馆    0.9853458404541016
+
+0.9853458404541016为预测置信度。
+
+### 推理模型导出与预测
+
+inference 模型（paddle.jit.save保存的模型） 一般是模型训练，把模型结构和模型参数保存在文件中的固化模型，多用于预测部署场景。 训练过程中保存的模型是checkpoints模型，保存的只有模型的参数，多用于恢复训练等。 与checkpoints模型相比，inference 模型会额外保存模型的结构信息，在预测部署、加速推理上性能优越，灵活方便，适合于实际系统集成。
+
+运行识别模型转inference模型命令，如下：
+
+```bash
+python tools/export_model.py -c configs/rec/rec_svtrnet_ch.yml -o Global.pretrained_model=./svtr_ch_high_accuracy/best_accuracy Global.save_inference_dir=./inference/svtr_ch
+```
+
+转换成功后，在目录下有三个文件：
+```shell
+inference/svtr_ch/
+    ├── inference.pdiparams         # 识别inference模型的参数文件
+    ├── inference.pdiparams.info    # 识别inference模型的参数信息，可忽略
+    └── inference.pdmodel           # 识别inference模型的program文件
+```
+
+inference模型预测，命令如下：
+
+```bash
+# CPU预测
+python3 tools/infer/predict_rec.py --image_dir="./doc/imgs_words/ch/word_1.jpg" --rec_algorithm='SVTR' --rec_model_dir=./inference/svtr_ch/ --rec_image_shape='3, 32, 320'  --rec_char_dict_path=ppocr/utils/ppocr_keys_v1.txt --use_gpu=False
+
+# GPU预测
+#python3 tools/infer/predict_rec.py --image_dir="./doc/imgs_words/ch/word_1.jpg" --rec_algorithm='SVTR' --rec_model_dir=./inference/svtr_ch/ --rec_image_shape='3, 32, 320'  --rec_char_dict_path=ppocr/utils/ppocr_keys_v1.txt --use_gpu=True
+```
+
+**注意**
+
+- 使用SVTR算法时，需要指定--rec_algorithm='SVTR'
+- 如果使用自定义字典训练的模型，需要将--rec_char_dict_path=ppocr/utils/ppocr_keys_v1.txt修改为自定义的字典
+- --rec_image_shape='3, 32, 320' 该参数不能去掉

configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_dml.yml

@@ -65,7 +65,7 @@ Loss:
       - ["Student", "Teacher"]
       maps_name: "thrink_maps"
       weight: 1.0
-      act: "softmax"
+      # act: None
       model_name_pairs: ["Student", "Teacher"]
       key: maps
   - DistillationDBLoss:

configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_dml.yml

@@ -60,7 +60,7 @@ Loss:
       - ["Student", "Student2"]
       maps_name: "thrink_maps"
       weight: 1.0
-      act: "softmax"
+      # act: None
       model_name_pairs: ["Student", "Student2"]
       key: maps
   - DistillationDBLoss:

configs/rec/rec_svtrnet.yml

@@ -83,8 +83,7 @@ Train:
           img_mode: BGR
           channel_first: False
       - CTCLabelEncode: # Class handling label
-      - RecResizeImg:
-          character_dict_path:
+      - SVTRRecResizeImg:
           image_shape: [3, 64, 256]
           padding: False
       - KeepKeys:
@@ -98,14 +97,13 @@ Train:
 Eval:
   dataset:
     name: LMDBDataSet
-    data_dir: ./train_data/data_lmdb_release/validation/
+    data_dir: ./train_data/data_lmdb_release/evaluation/
     transforms:
       - DecodeImage: # load image
           img_mode: BGR
           channel_first: False
       - CTCLabelEncode: # Class handling label
-      - RecResizeImg:
-          character_dict_path:
+      - SVTRRecResizeImg:
           image_shape: [3, 64, 256]
           padding: False
       - KeepKeys:

configs/rec/rec_svtrnet_ch.yml

+Global:
+  use_gpu: true
+  epoch_num: 100
+  log_smooth_window: 20
+  print_batch_step: 10
+  save_model_dir: ./output/rec/svtr_ch_all/
+  save_epoch_step: 10
+  eval_batch_step:
+  - 0
+  - 2000
+  cal_metric_during_train: true
+  pretrained_model: null
+  checkpoints: null
+  save_inference_dir: null
+  use_visualdl: false
+  infer_img: doc/imgs_words/ch/word_1.jpg
+  character_dict_path: ppocr/utils/ppocr_keys_v1.txt
+  max_text_length: 25
+  infer_mode: false
+  use_space_char: true
+  save_res_path: ./output/rec/predicts_svtr_tiny_ch_all.txt
+Optimizer:
+  name: AdamW
+  beta1: 0.9
+  beta2: 0.99
+  epsilon: 8.0e-08
+  weight_decay: 0.05
+  no_weight_decay_name: norm pos_embed
+  one_dim_param_no_weight_decay: true
+  lr:
+    name: Cosine
+    learning_rate: 0.0005
+    warmup_epoch: 2
+Architecture:
+  model_type: rec
+  algorithm: SVTR
+  Transform: null
+  Backbone:
+    name: SVTRNet
+    img_size:
+    - 32
+    - 320
+    out_char_num: 40
+    out_channels: 96
+    patch_merging: Conv
+    embed_dim:
+    - 64
+    - 128
+    - 256
+    depth:
+    - 3
+    - 6
+    - 3
+    num_heads:
+    - 2
+    - 4
+    - 8
+    mixer:
+    - Local
+    - Local
+    - Local
+    - Local
+    - Local
+    - Local
+    - Global
+    - Global
+    - Global
+    - Global
+    - Global
+    - Global
+    local_mixer:
+    - - 7
+      - 11
+    - - 7
+      - 11
+    - - 7
+      - 11
+    last_stage: true
+    prenorm: false
+  Neck:
+    name: SequenceEncoder
+    encoder_type: reshape
+  Head:
+    name: CTCHead
+Loss:
+  name: CTCLoss
+PostProcess:
+  name: CTCLabelDecode
+Metric:
+  name: RecMetric
+  main_indicator: acc
+Train:
+  dataset:
+    name: SimpleDataSet
+    data_dir: ./train_data
+    label_file_list:
+    - ./train_data/train_list.txt
+    ext_op_transform_idx: 1
+    transforms:
+    - DecodeImage:
+        img_mode: BGR
+        channel_first: false
+    - RecConAug:
+        prob: 0.5
+        ext_data_num: 2
+        image_shape:
+        - 32
+        - 320
+        - 3
+    - RecAug: null
+    - CTCLabelEncode: null
+    - SVTRRecResizeImg:
+        image_shape:
+        - 3
+        - 32
+        - 320
+        padding: true
+    - KeepKeys:
+        keep_keys:
+        - image
+        - label
+        - length
+  loader:
+    shuffle: true
+    batch_size_per_card: 256
+    drop_last: true
+    num_workers: 8
+Eval:
+  dataset:
+    name: SimpleDataSet
+    data_dir: ./train_data
+    label_file_list:
+    - ./train_data/val_list.txt
+    transforms:
+    - DecodeImage:
+        img_mode: BGR
+        channel_first: false
+    - CTCLabelEncode: null
+    - SVTRRecResizeImg:
+        image_shape:
+        - 3
+        - 32
+        - 320
+        padding: true
+    - KeepKeys:
+        keep_keys:
+        - image
+        - label
+        - length
+  loader:
+    shuffle: false
+    drop_last: false
+    batch_size_per_card: 256
+    num_workers: 2
+profiler_options: null

deploy/android_demo/app/src/main/cpp/native.cpp

@@ -47,7 +47,7 @@ str_to_cpu_mode(const std::string &cpu_mode) {
   std::string upper_key;
   std::transform(cpu_mode.cbegin(), cpu_mode.cend(), upper_key.begin(),
                  ::toupper);
-  auto index = cpu_mode_map.find(upper_key);
+  auto index = cpu_mode_map.find(upper_key.c_str());
   if (index == cpu_mode_map.end()) {
     LOGE("cpu_mode not found %s", upper_key.c_str());
     return paddle::lite_api::LITE_POWER_HIGH;
@@ -116,4 +116,4 @@ Java_com_baidu_paddle_lite_demo_ocr_OCRPredictorNative_release(
   ppredictor::OCR_PPredictor *ppredictor =
       (ppredictor::OCR_PPredictor *)java_pointer;
   delete ppredictor;
-}
+}
\ No newline at end of file

deploy/android_demo/app/src/main/java/com/baidu/paddle/lite/demo/ocr/OCRPredictorNative.java

@@ -54,7 +54,7 @@ public class OCRPredictorNative {
     }
 
     public void destory() {
-        if (nativePointer > 0) {
+        if (nativePointer != 0) {
             release(nativePointer);
             nativePointer = 0;
         }

deploy/cpp_infer/docs/windows_vs2019_build.md

@@ -109,8 +109,10 @@ CUDA_LIB、CUDNN_LIB、TENSORRT_DIR、WITH_GPU、WITH_TENSORRT
 
 运行之前，将下面文件拷贝到`build/Release/`文件夹下
 1. `paddle_inference/paddle/lib/paddle_inference.dll`
-2. `opencv/build/x64/vc15/bin/opencv_world455.dll`
-3. 如果使用openblas版本的预测库还需要拷贝 `paddle_inference/third_party/install/openblas/lib/openblas.dll`
+2. `paddle_inference/third_party/install/onnxruntime/lib/onnxruntime.dll`
+3. `paddle_inference/third_party/install/paddle2onnx/lib/paddle2onnx.dll`
+4. `opencv/build/x64/vc15/bin/opencv_world455.dll`
+5. 如果使用openblas版本的预测库还需要拷贝 `paddle_inference/third_party/install/openblas/lib/openblas.dll`
 
 ### Step4: 预测
 

deploy/lite/config.txt

@@ -4,4 +4,5 @@ det_db_box_thresh  0.5
 det_db_unclip_ratio  1.6
 det_db_use_dilate 0
 det_use_polygon_score 1
-use_direction_classify  1
+use_direction_classify  1
\ No newline at end of file
+rec_image_height  32
\ No newline at end of file

deploy/lite/crnn_process.cc

@@ -19,24 +19,28 @@
 
 const std::vector<int> rec_image_shape{3, 32, 320};
 
-cv::Mat CrnnResizeImg(cv::Mat img, float wh_ratio) {
+cv::Mat CrnnResizeImg(cv::Mat img, float wh_ratio, int rec_image_height) {
   int imgC, imgH, imgW;
   imgC = rec_image_shape[0];
+  imgH = rec_image_height;
   imgW = rec_image_shape[2];
-  imgH = rec_image_shape[1];
 
-  imgW = int(32 * wh_ratio);
+  imgW = int(imgH * wh_ratio);
 
-  float ratio = static_cast<float>(img.cols) / static_cast<float>(img.rows);
+  float ratio = float(img.cols) / float(img.rows);
   int resize_w, resize_h;
+
   if (ceilf(imgH * ratio) > imgW)
     resize_w = imgW;
   else
-    resize_w = static_cast<int>(ceilf(imgH * ratio));
+    resize_w = int(ceilf(imgH * ratio));
+
   cv::Mat resize_img;
   cv::resize(img, resize_img, cv::Size(resize_w, imgH), 0.f, 0.f,
              cv::INTER_LINEAR);
-
+  cv::copyMakeBorder(resize_img, resize_img, 0, 0, 0,
+                     int(imgW - resize_img.cols), cv::BORDER_CONSTANT,
+                     {127, 127, 127});
   return resize_img;
 }
 

deploy/lite/crnn_process.h

@@ -26,7 +26,7 @@
 #include "opencv2/imgcodecs.hpp"
 #include "opencv2/imgproc.hpp"
 
-cv::Mat CrnnResizeImg(cv::Mat img, float wh_ratio);
+cv::Mat CrnnResizeImg(cv::Mat img, float wh_ratio, int rec_image_height);
 
 std::vector<std::string> ReadDict(std::string path);
 

deploy/lite/ocr_db_crnn.cc

@@ -162,7 +162,8 @@ void RunRecModel(std::vector<std::vector<std::vector<int>>> boxes, cv::Mat img,
                  std::vector<std::string> charactor_dict,
                  std::shared_ptr<PaddlePredictor> predictor_cls,
                  int use_direction_classify,
-                 std::vector<double> *times) {
+                 std::vector<double> *times,
+                 int rec_image_height) {
   std::vector<float> mean = {0.5f, 0.5f, 0.5f};
   std::vector<float> scale = {1 / 0.5f, 1 / 0.5f, 1 / 0.5f};
 
@@ -183,7 +184,7 @@ void RunRecModel(std::vector<std::vector<std::vector<int>>> boxes, cv::Mat img,
     float wh_ratio =
         static_cast<float>(crop_img.cols) / static_cast<float>(crop_img.rows);
 
-    resize_img = CrnnResizeImg(crop_img, wh_ratio);
+    resize_img = CrnnResizeImg(crop_img, wh_ratio, rec_image_height);
     resize_img.convertTo(resize_img, CV_32FC3, 1 / 255.f);
 
     const float *dimg = reinterpret_cast<const float *>(resize_img.data);
@@ -444,7 +445,7 @@ void system(char **argv){
   //// load config from txt file
   auto Config = LoadConfigTxt(det_config_path);
   int use_direction_classify = int(Config["use_direction_classify"]);
-
+  int rec_image_height = int(Config["rec_image_height"]);
   auto charactor_dict = ReadDict(dict_path);
   charactor_dict.insert(charactor_dict.begin(), "#"); // blank char for ctc
   charactor_dict.push_back(" ");
@@ -473,7 +474,7 @@ void system(char **argv){
   
     std::vector<double> rec_times;
     RunRecModel(boxes, srcimg, rec_predictor, rec_text, rec_text_score,
-                charactor_dict, cls_predictor, use_direction_classify, &rec_times);
+                charactor_dict, cls_predictor, use_direction_classify, &rec_times, rec_image_height);
   
     //// visualization
     auto img_vis = Visualization(srcimg, boxes);
@@ -590,12 +591,16 @@ void rec(int argc, char **argv) {
   std::string batchsize = argv[6];
   std::string img_dir = argv[7];
   std::string dict_path = argv[8];
+  std::string config_path = argv[9];
 
   if (strcmp(argv[4], "FP32") != 0 && strcmp(argv[4], "INT8") != 0) {
       std::cerr << "Only support FP32 or INT8." << std::endl;
       exit(1);
   }
 
+  auto Config = LoadConfigTxt(config_path);
+  int rec_image_height = int(Config["rec_image_height"]);
+
   std::vector<cv::String> cv_all_img_names;
   cv::glob(img_dir, cv_all_img_names);
 
@@ -630,7 +635,7 @@ void rec(int argc, char **argv) {
     std::vector<float> rec_text_score;
     std::vector<double> times;
     RunRecModel(boxes, srcimg, rec_predictor, rec_text, rec_text_score,
-                charactor_dict, cls_predictor, 0, &times);
+                charactor_dict, cls_predictor, 0, &times, rec_image_height);
   
     //// print recognized text
     for (int i = 0; i < rec_text.size(); i++) {

deploy/lite/readme.md

@@ -34,7 +34,7 @@ For the compilation process of different development environments, please refer
 ### 1.2 Prepare Paddle-Lite library
 
 There are two ways to obtain the Paddle-Lite library：
-- 1. Download directly, the download link of the Paddle-Lite library is as follows：
+- 1. [Recommended] Download directly, the download link of the Paddle-Lite library is as follows：
 
       | Platform | Paddle-Lite library download link |
       |---|---|
@@ -43,7 +43,9 @@ There are two ways to obtain the Paddle-Lite library：
 
       Note: 1. The above Paddle-Lite library is compiled from the Paddle-Lite 2.10 branch. For more information about Paddle-Lite 2.10, please refer to [link](https://github.com/PaddlePaddle/Paddle-Lite/releases/tag/v2.10).
 
-- 2. [Recommended] Compile Paddle-Lite to get the prediction library. The compilation method of Paddle-Lite is as follows：
+    **Note: It is recommended to use paddlelite>=2.10 version of the prediction library, other prediction library versions [download link](https://github.com/PaddlePaddle/Paddle-Lite/tags)**
+
+- 2. Compile Paddle-Lite to get the prediction library. The compilation method of Paddle-Lite is as follows：
 ```
 git clone https://github.com/PaddlePaddle/Paddle-Lite.git
 cd Paddle-Lite
@@ -104,21 +106,17 @@ If you directly use the model in the above table for deployment, you can skip th
 
 If the model to be deployed is not in the above table, you need to follow the steps below to obtain the optimized model.
 
-The `opt` tool can be obtained by compiling Paddle Lite.
+- Step 1: Refer to [document](https://www.paddlepaddle.org.cn/lite/v2.10/user_guides/opt/opt_python.html) to install paddlelite, which is used to convert paddle inference model to paddlelite required for running nb model
 ```
-git clone https://github.com/PaddlePaddle/Paddle-Lite.git
-cd Paddle-Lite
-git checkout release/v2.10
-./lite/tools/build.sh build_optimize_tool
+pip install paddlelite==2.10 # The paddlelite version should be the same as the prediction library version
 ```
-
-After the compilation is complete, the opt file is located under build.opt/lite/api/, You can view the operating options and usage of opt in the following ways:
-
+After installation, the following commands can view the help information
 ```
-cd build.opt/lite/api/
-./opt
+paddle_lite_opt
 ```
 
+Introduction to paddle_lite_opt parameters:
+
 |Options|Description|
 |---|---|
 |--model_dir|The path of the PaddlePaddle model to be optimized (non-combined form)|
@@ -131,6 +129,8 @@ cd build.opt/lite/api/
 
 `--model_dir` is suitable for the non-combined mode of the model to be optimized, and the inference model of PaddleOCR is the combined mode, that is, the model structure and model parameters are stored in a single file.
 
+- Step 2: Use paddle_lite_opt to convert the inference model to the mobile model format.
+
 The following takes the ultra-lightweight Chinese model of PaddleOCR as an example to introduce the use of the compiled opt file to complete the conversion of the inference model to the Paddle-Lite optimized model
 
 ```
@@ -240,6 +240,7 @@ det_db_thresh  0.3        # Used to filter the binarized image of DB prediction,
 det_db_box_thresh  0.5    # DDB post-processing filter box threshold, if there is a missing box detected, it can be reduced as appropriate
 det_db_unclip_ratio  1.6  # Indicates the compactness of the text box, the smaller the value, the closer the text box to the text
 use_direction_classify  0  # Whether to use the direction classifier, 0 means not to use, 1 means to use
+rec_image_height  32      # The height of the input image of the recognition model, the PP-OCRv3 model needs to be set to 48, and the PP-OCRv2 model needs to be set to 32
 ```
 
  5. Run Model on phone
@@ -258,8 +259,15 @@ After the above steps are completed, you can use adb to push the file to the pho
  cd /data/local/tmp/debug
  export LD_LIBRARY_PATH=${PWD}:$LD_LIBRARY_PATH
  # The use of ocr_db_crnn is:
- # ./ocr_db_crnn Detection model file Orientation classifier model file Recognition model file Test image path Dictionary file path
- ./ocr_db_crnn ch_PP-OCRv2_det_slim_opt.nb  ch_PP-OCRv2_rec_slim_opt.nb  ch_ppocr_mobile_v2.0_cls_opt.nb  ./11.jpg  ppocr_keys_v1.txt
+ # ./ocr_db_crnn Mode Detection model file Orientation classifier model file Recognition model file  Hardware  Precision  Threads Batchsize  Test image path Dictionary file path
+ ./ocr_db_crnn system ch_PP-OCRv2_det_slim_opt.nb  ch_PP-OCRv2_rec_slim_opt.nb  ch_ppocr_mobile_v2.0_cls_slim_opt.nb  arm8 INT8 10 1  ./11.jpg  config.txt  ppocr_keys_v1.txt  True
+# precision can be INT8 for quantitative model or FP32 for normal model.
+
+# Only using detection model
+./ocr_db_crnn  det ch_PP-OCRv2_det_slim_opt.nb arm8 INT8 10 1 ./11.jpg  config.txt
+
+# Only using recognition model
+./ocr_db_crnn  rec ch_PP-OCRv2_rec_slim_opt.nb arm8 INT8 10 1 word_1.jpg ppocr_keys_v1.txt config.txt
  ```
 
 If you modify the code, you need to recompile and push to the phone.
@@ -283,3 +291,7 @@ A2: Replace the .jpg test image under ./debug with the image you want to test, a
 Q3: How to package it into the mobile APP?
 
 A3: This demo aims to provide the core algorithm part that can run OCR on mobile phones. Further, PaddleOCR/deploy/android_demo is an example of encapsulating this demo into a mobile app for reference.
+
+Q4: When running the demo, an error is reported `Error: This model is not supported, because kernel for 'io_copy' is not supported by Paddle-Lite.`
+
+A4: The problem is that the installed paddlelite version does not match the downloaded prediction library version. Make sure that the paddleliteopt tool matches your prediction library version, and try to switch to the nb model again.

deploy/lite/readme_ch.md

@@ -8,7 +8,7 @@
     - [2.1 模型优化](#21-模型优化)
     - [2.2 与手机联调](#22-与手机联调)
 - [FAQ](#faq)
-  
+
 
 本教程将介绍基于[Paddle Lite](https://github.com/PaddlePaddle/Paddle-Lite) 在移动端部署PaddleOCR超轻量中文检测、识别模型的详细步骤。
 
@@ -32,7 +32,7 @@ Paddle Lite是飞桨轻量化推理引擎，为手机、IOT端提供高效推理
 ### 1.2 准备预测库
 
 预测库有两种获取方式：
-- 1. 直接下载，预测库下载链接如下：
+- 1. [推荐]直接下载，预测库下载链接如下：
 
       | 平台 | 预测库下载链接 |
       |---|---|
@@ -41,7 +41,9 @@ Paddle Lite是飞桨轻量化推理引擎，为手机、IOT端提供高效推理
 
       注：1. 上述预测库为PaddleLite 2.10分支编译得到，有关PaddleLite 2.10 详细信息可参考 [链接](https://github.com/PaddlePaddle/Paddle-Lite/releases/tag/v2.10) 。
 
-- 2. [推荐]编译Paddle-Lite得到预测库，Paddle-Lite的编译方式如下：
+**注：建议使用paddlelite>=2.10版本的预测库，其他预测库版本[下载链接](https://github.com/PaddlePaddle/Paddle-Lite/tags)**
+
+- 2. 编译Paddle-Lite得到预测库，Paddle-Lite的编译方式如下：
 ```
 git clone https://github.com/PaddlePaddle/Paddle-Lite.git
 cd Paddle-Lite
@@ -102,22 +104,16 @@ Paddle-Lite 提供了多种策略来自动优化原始的模型，其中包括
 
 如果要部署的模型不在上述表格中，则需要按照如下步骤获得优化后的模型。
 
-模型优化需要Paddle-Lite的opt可执行文件，可以通过编译Paddle-Lite源码获得，编译步骤如下：
+- 步骤1：参考[文档](https://www.paddlepaddle.org.cn/lite/v2.10/user_guides/opt/opt_python.html)安装paddlelite，用于转换paddle inference model为paddlelite运行所需的nb模型
 ```
-# 如果准备环境时已经clone了Paddle-Lite，则不用重新clone Paddle-Lite
-git clone https://github.com/PaddlePaddle/Paddle-Lite.git
-cd Paddle-Lite
-git checkout release/v2.10
-# 启动编译
-./lite/tools/build.sh build_optimize_tool
+pip install paddlelite==2.10  # paddlelite版本要与预测库版本一致
 ```
-
-编译完成后，opt文件位于`build.opt/lite/api/`下，可通过如下方式查看opt的运行选项和使用方式；
+安装完后，如下指令可以查看帮助信息
 ```
-cd build.opt/lite/api/
-./opt
+paddle_lite_opt
 ```
 
+paddle_lite_opt 参数介绍：
 |选项|说明|
 |---|---|
 |--model_dir|待优化的PaddlePaddle模型（非combined形式）的路径|
@@ -130,6 +126,8 @@ cd build.opt/lite/api/
 
 `--model_dir`适用于待优化的模型是非combined方式，PaddleOCR的inference模型是combined方式，即模型结构和模型参数使用单独一个文件存储。
 
+- 步骤2：使用paddle_lite_opt将inference模型转换成移动端模型格式。
+
 下面以PaddleOCR的超轻量中文模型为例，介绍使用编译好的opt文件完成inference模型到Paddle-Lite优化模型的转换。
 
 ```
@@ -148,7 +146,7 @@ wget  https://paddleocr.bj.bcebos.com/dygraph_v2.0/slim/ch_ppocr_mobile_v2.0_cls
 
 转换成功后，inference模型目录下会多出`.nb`结尾的文件，即是转换成功的模型文件。
 
-注意：使用paddle-lite部署时，需要使用opt工具优化后的模型。 opt 工具的输入模型是paddle保存的inference模型
+注意：使用paddle-lite部署时，需要使用opt工具优化后的模型。 opt工具的输入模型是paddle保存的inference模型
 
 <a name="2.2与手机联调"></a>
 ### 2.2 与手机联调
@@ -234,13 +232,14 @@ ppocr_keys_v1.txt   # 中文字典
 ...
 ```
 
-2.  `config.txt` 包含了检测器、分类器的超参数，如下：
+2.  `config.txt` 包含了检测器、分类器、识别器的超参数，如下：
 ```
 max_side_len  960         # 输入图像长宽大于960时，等比例缩放图像，使得图像最长边为960
 det_db_thresh  0.3        # 用于过滤DB预测的二值化图像，设置为0.-0.3对结果影响不明显
-det_db_box_thresh  0.5    # DB后处理过滤box的阈值，如果检测存在漏框情况，可酌情减小
+det_db_box_thresh  0.5    # 检测器后处理过滤box的阈值，如果检测存在漏框情况，可酌情减小
 det_db_unclip_ratio  1.6  # 表示文本框的紧致程度，越小则文本框更靠近文本
 use_direction_classify  0  # 是否使用方向分类器，0表示不使用，1表示使用
+rec_image_height  32      # 识别模型输入图像的高度，PP-OCRv3模型设置为48，PP-OCRv2模型需要设置为32
 ```
 
  5. 启动调试
@@ -259,8 +258,14 @@ use_direction_classify  0  # 是否使用方向分类器，0表示不使用，1
  cd /data/local/tmp/debug
  export LD_LIBRARY_PATH=${PWD}:$LD_LIBRARY_PATH
  # 开始使用，ocr_db_crnn可执行文件的使用方式为:
- # ./ocr_db_crnn  检测模型文件 方向分类器模型文件  识别模型文件  测试图像路径  字典文件路径
- ./ocr_db_crnn ch_PP-OCRv2_det_slim_opt.nb  ch_PP-OCRv2_rec_slim_opt.nb  ch_ppocr_mobile_v2.0_cls_slim_opt.nb  ./11.jpg  ppocr_keys_v1.txt
+ # ./ocr_db_crnn 预测模式  检测模型文件 方向分类器模型文件  识别模型文件 运行硬件 运行精度 线程数  batchsize  测试图像路径  参数配置路径  字典文件路径 是否使用benchmark参数
+ ./ocr_db_crnn system  ch_PP-OCRv2_det_slim_opt.nb  ch_PP-OCRv2_rec_slim_opt.nb  ch_ppocr_mobile_v2.0_cls_slim_opt.nb  arm8 INT8 10 1  ./11.jpg  config.txt  ppocr_keys_v1.txt  True
+
+# 仅使用文本检测模型，使用方式如下：
+./ocr_db_crnn  det ch_PP-OCRv2_det_slim_opt.nb arm8 INT8 10 1 ./11.jpg  config.txt
+
+# 仅使用文本识别模型，使用方式如下：
+./ocr_db_crnn  rec ch_PP-OCRv2_rec_slim_opt.nb arm8 INT8 10 1 word_1.jpg ppocr_keys_v1.txt config.txt
  ```
 
  如果对代码做了修改，则需要重新编译并push到手机上。
@@ -284,3 +289,7 @@ A2：替换debug下的.jpg测试图像为你想要测试的图像，adb push 到
 Q3：如何封装到手机APP中？
 
 A3：此demo旨在提供能在手机上运行OCR的核心算法部分，PaddleOCR/deploy/android_demo是将这个demo封装到手机app的示例，供参考
+
+Q4：运行demo时遇到报错`Error: This model is not supported, because kernel for 'io_copy' is not supported by Paddle-Lite.`
+
+A4：问题是安装的paddlelite版本和下载的预测库版本不匹配，确保paddleliteopt工具和你的预测库版本匹配，重新转nb模型试试。

deploy/slim/quantization/README_en.md

@@ -73,4 +73,4 @@ python deploy/slim/quantization/export_model.py -c configs/det/ch_ppocr_v2.0/ch_
 The numerical range of the quantized model parameters derived from the above steps is still FP32, but the numerical range of the parameters is int8.
 The derived model can be converted through the `opt tool` of PaddleLite.
 
-For quantitative model deployment, please refer to [Mobile terminal model deployment](../../lite/readme_en.md)
+For quantitative model deployment, please refer to [Mobile terminal model deployment](../../lite/readme.md)

doc/doc_ch/PP-OCRv3_introduction.md

@@ -185,7 +185,7 @@ UDML（Unified-Deep Mutual Learning）联合互学习是PP-OCRv2中就采用的
 
 **（6）UIM：无标注数据挖掘方案**
 
-UIM（Unlabeled Images Mining）是一种非常简单的无标注数据挖掘方案。核心思想是利用高精度的文本识别大模型对无标注数据进行预测，获取伪标签，并且选择预测置信度高的样本作为训练数据，用于训练小模型。使用该策略，识别模型的准确率进一步提升到79.4%（+1%）。
+UIM（Unlabeled Images Mining）是一种非常简单的无标注数据挖掘方案。核心思想是利用高精度的文本识别大模型对无标注数据进行预测，获取伪标签，并且选择预测置信度高的样本作为训练数据，用于训练小模型。使用该策略，识别模型的准确率进一步提升到79.4%（+1%）。实际操作中，我们使用全量数据集训练高精度SVTR-Tiny模型（acc=82.5%）进行数据挖掘，点击获取[模型下载地址和使用教程](../../applications/高精度中文识别模型.md)。
 
 <div align="center">
     <img src="../ppocr_v3/UIM.png" width="500">