Combine rec v2 and v3

python/README.md

@@ -17,24 +17,37 @@
 
 ### 简介和说明
 - **各个版本的ONNX模型下载地址：**[百度网盘](https://pan.baidu.com/s/1PTcgXG2zEgQU6A_A3kGJ3Q?pwd=jhai) | [Google Drive](https://drive.google.com/drive/folders/1x_a9KpCo_1blxH1xFOfgKVkw1HYRVywY?usp=sharing)
-- 所有常用的参数配置都在`config.yaml`下，一目了然，更加便捷
+- 所有常用的参数配置都在[`config.yaml`](./config.yaml)下，一目了然，更加便捷
 - 每个独立的模块下均有独立的`config.yaml`配置文件，可以单独使用
 - `det`部分：
-  - `det`中`mobile`和`server`版，共用一个推理代码，直接更改配置文件中模型路径即可
-  - `det`中`v2`和`v3`两个版本代码，共用一个推理代码，直接更改配置文件中模型路径即可
+  - `det`中`mobile`和`server`版，推理代码一致，直接更改配置文件中模型路径即可
+  - `det`中`v2`和`v3`两个版本，推理代码一致，直接更改配置文件中模型路径即可
     ```yaml
     Det:
         module_name: ch_ppocr_v2_det
         class_name: TextDetector
         model_path: resources/models/ch_PP-OCRv3_det_infer.onnx
     ```
-- `rec`中`mobile`和`server`版本，共用一个推理代码，直接更改配置文件中模型路径即可
-- onnxruntime和OpenVINO调用方式如下:
+- `rec`部分：
+  - `rec`中`mobile`和`server`版本，推理代码一致，直接更改配置文件中模型路径即可
+  - `rec`中`v2`和`v3`两个版本，共用同一个推理代码。
+    - 两版本差别仅在输入shape和模型。经过测试，采用`v3 rec模型`+`[3, 48, 320]`效果最好。
+    - 目前配置文件`config.yaml`中（如下所示）已经改为该最优组合。
+    ```yaml
+    module_name: ch_ppocr_v2_rec
+    class_name: TextRecognizer
+    model_path: resources/models/ch_PP-OCRv3_rec_infer.onnx
+
+    rec_img_shape: [3, 48, 320]
+    rec_batch_num: 6
+    keys_path: resources/rec_dict/ppocr_keys_v1.txt
+    ```
+- onnxruntime和openvino调用方式如下:
     ```python
     # 基于onnxruntime引擎推理
     from rapidocr_onnxruntime import TextSystem
 
-    # 基于openvin引擎推理
+    # 基于openvino引擎推理
     from rapidocr_openvino import TextSystem
     ```
 - 值得说明的是，基于openvino推理部分中`ch_ppocr_v2_cls`部分仍然是基于onnxruntime的，原因是openvino有bug，详情见[openvino/issue](https://github.com/openvinotoolkit/openvino/issues/11501)
@@ -92,7 +105,7 @@
 
         pip install -r requirements.txt -i https://pypi.douban.com/simple/
         ```
-   - 基于OpenVINO推理所需环境安装：
+   - 基于openvino推理所需环境安装：
         ```bash
         # Windows端
         pip install openvino==2022.1.0
@@ -114,7 +127,7 @@
         # 基于onnxruntime引擎推理
         from rapidocr_onnxruntime import TextSystem
 
-        # 基于OpenVINO引擎推理
+        # 基于openvino引擎推理
         # from rapidocr_openvino import TextSystem
 
         config_path = 'config.yaml'
@@ -129,16 +142,6 @@
         ```bash
         python test_demp.py
         ```
-    - `rec`部分，如果想要使用`PPOCR-v3`的话，需要将`config.yaml`中的`Rec`部分改为如下:
-        ```yaml
-        module_name: ch_ppocr_v3_rec
-        class_name: TextRecognizer
-        model_path: resources/models/PPOCRv3/ch_PP-OCRv3_rec_infer.onnx
-
-        rec_img_shape: [3, 48, 320]
-        rec_batch_num: 6
-        keys_path: resources/rec_dict/ppocr_keys_v1.txt
-        ```
 
 ### [`config.yaml`](./config.yaml)中常用参数介绍
 - `Global`部分

python/config.yaml

@@ -57,7 +57,7 @@ Cls:
 Rec:
     module_name: ch_ppocr_v2_rec
     class_name: TextRecognizer
-    model_path: resources/models/ch_ppocr_mobile_v2.0_rec_infer.onnx
+    model_path: resources/models/ch_PP-OCRv3_rec_infer.onnx
 
     use_cuda: false
     CUDAExecutionProvider:
@@ -66,6 +66,6 @@ Rec:
         cudnn_conv_algo_search: EXHAUSTIVE
         do_copy_in_default_stream: true
 
-    rec_img_shape: [3, 32, 320]
+    rec_img_shape: [3, 48, 320]
     rec_batch_num: 6
     keys_path: resources/rec_dict/ppocr_keys_v1.txt

python/rapidocr_onnxruntime/ch_ppocr_v2_rec/__init__.py

@@ -1 +1,4 @@
+# -*- encoding: utf-8 -*-
+# @Author: SWHL
+# @Contact: liekkaskono@163.com
 from .text_recognize import TextRecognizer

python/rapidocr_onnxruntime/ch_ppocr_v2_rec/config.yaml

@@ -1,4 +1,4 @@
-model_path: resources/models/ch_ppocr_mobile_v2.0_rec_infer.onnx
+model_path: resources/models/ch_PP-OCRv3_rec_infer.onnx
 
 use_cuda: false
 # Details of the params: https://onnxruntime.ai/docs/execution-providers/CUDA-ExecutionProvider.html
@@ -9,6 +9,6 @@ CUDAExecutionProvider:
     cudnn_conv_algo_search: EXHAUSTIVE
     do_copy_in_default_stream: true
 
-rec_img_shape: [3, 32, 320]
+rec_img_shape: [3, 48, 320]
 rec_batch_num: 6
 keys_path: resources/rec_dict/ppocr_keys_v1.txt

python/rapidocr_onnxruntime/ch_ppocr_v2_rec/text_recognize.py

@@ -36,31 +36,6 @@ class TextRecognizer(object):
         self.session = session_instance.session
         self.input_name = session_instance.get_input_name()
 
-    def resize_norm_img(self, img, max_wh_ratio):
-        img_channel, img_height, img_width = self.rec_image_shape
-        assert img_channel == img.shape[2]
-
-        img_width = int((32 * max_wh_ratio))
-        max_wh_ratio = 1
-
-        h, w = img.shape[:2]
-        ratio = w / float(h)
-        if math.ceil(img_height * ratio) > img_width:
-            resized_w = img_width
-        else:
-            resized_w = int(math.ceil(img_height * ratio))
-
-        resized_image = cv2.resize(img, (resized_w, img_height))
-        resized_image = resized_image.astype('float32')
-        resized_image = resized_image.transpose((2, 0, 1)) / 255
-        resized_image -= 0.5
-        resized_image /= 0.5
-
-        padding_im = np.zeros((img_channel, img_height, img_width),
-                              dtype=np.float32)
-        padding_im[:, :, 0:resized_w] = resized_image
-        return padding_im
-
     def __call__(self, img_list: List[np.ndarray]):
         if isinstance(img_list, np.ndarray):
             img_list = [img_list]
@@ -101,6 +76,31 @@ class TextRecognizer(object):
             elapse += time.time() - starttime
         return rec_res, elapse
 
+    def resize_norm_img(self, img, max_wh_ratio):
+        img_channel, img_height, img_width = self.rec_image_shape
+        assert img_channel == img.shape[2]
+
+        img_width = int((img_height * max_wh_ratio))
+        max_wh_ratio = 1
+
+        h, w = img.shape[:2]
+        ratio = w / float(h)
+        if math.ceil(img_height * ratio) > img_width:
+            resized_w = img_width
+        else:
+            resized_w = int(math.ceil(img_height * ratio))
+
+        resized_image = cv2.resize(img, (resized_w, img_height))
+        resized_image = resized_image.astype('float32')
+        resized_image = resized_image.transpose((2, 0, 1)) / 255
+        resized_image -= 0.5
+        resized_image /= 0.5
+
+        padding_im = np.zeros((img_channel, img_height, img_width),
+                              dtype=np.float32)
+        padding_im[:, :, 0:resized_w] = resized_image
+        return padding_im
+
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()

python/rapidocr_onnxruntime/ch_ppocr_v3_rec/README.md

@@ -1,9 +0,0 @@
-
-#### 简单对比
-
-|输入尺寸|前处理函数|模型|测试集|Score|Exact_Match|Char_Match|Rec_Speed(s/条)|
-|---|---|---|---|---|---|---|---|
-|48 |resize_norm_img_svtr |ch_PP-OCRv3_rec_infer.onnx|中英印刷体|0.5011|0.3571|0.6450|0.0203|
-|48 |resize_norm_img |ch_PP-OCRv3_rec_infer.onnx|中英印刷体|0.7248|0.5417|0.9079|0.0209|
-|32 | resize_norm_img_svtr |ch_PP-OCRv3_rec_infer.onnx|中英印刷体|0.3019|0.1607|0.4431|0.0161|
-|32| resize_norm_img |ch_PP-OCRv3_rec_infer.onnx|中英印刷体|0.5537|0.3274|0.7800|0.0152|

python/rapidocr_onnxruntime/ch_ppocr_v3_rec/__init__.py

@@ -1,4 +0,0 @@
-# -*- encoding: utf-8 -*-
-# @Author: SWHL
-# @Contact: liekkaskono@163.com
-from .text_recognize import TextRecognizer

python/rapidocr_onnxruntime/ch_ppocr_v3_rec/config.yaml

@@ -1,14 +0,0 @@
-model_path: resources/models/ch_PP-OCRv3_rec_infer.onnx
-
-use_cuda: false
-# Details of the params: https://onnxruntime.ai/docs/execution-providers/CUDA-ExecutionProvider.html
-CUDAExecutionProvider:
-    device_id: 0
-    arena_extend_strategy: kNextPowerOfTwo
-    gpu_mem_limit: 2 * 1024 * 1024 * 1024
-    cudnn_conv_algo_search: EXHAUSTIVE
-    do_copy_in_default_stream: true
-
-rec_img_shape: [3, 48, 320]
-rec_batch_num: 6
-keys_path: resources/rec_dict/ppocr_keys_v1.txt

python/rapidocr_onnxruntime/ch_ppocr_v3_rec/text_recognize.py

@@ -1,130 +0,0 @@
-# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-import argparse
-import math
-import time
-from pathlib import Path
-from typing import List
-
-import cv2
-import numpy as np
-import onnxruntime
-
-try:
-    from .utils import CTCLabelDecode, read_yaml, OrtInferSession
-except:
-    from utils import CTCLabelDecode, read_yaml, OrtInferSession
-
-root_path = Path(__file__).parent
-
-
-class TextRecognizer(object):
-    def __init__(self, config):
-        self.rec_image_shape = config['rec_img_shape']
-        self.rec_batch_num = config['rec_batch_num']
-        self.character_dict_path = config['keys_path']
-        self.postprocess_op = CTCLabelDecode(self.character_dict_path)
-
-        session_instance = OrtInferSession(config)
-        self.session = session_instance.session
-        self.input_name = session_instance.get_input_name()
-
-    def resize_norm_img(self, img, max_wh_ratio):
-        img_channel, img_height, img_width = self.rec_image_shape
-        assert img_channel == img.shape[2]
-
-        img_width = int((32 * max_wh_ratio))
-        max_wh_ratio = 1
-
-        h, w = img.shape[:2]
-        ratio = w / float(h)
-        if math.ceil(img_height * ratio) > img_width:
-            resized_w = img_width
-        else:
-            resized_w = int(math.ceil(img_height * ratio))
-
-        resized_image = cv2.resize(img, (resized_w, img_height))
-        resized_image = resized_image.astype('float32')
-        resized_image = resized_image.transpose((2, 0, 1)) / 255
-        resized_image -= 0.5
-        resized_image /= 0.5
-
-        padding_im = np.zeros((img_channel, img_height, img_width),
-                              dtype=np.float32)
-        padding_im[:, :, 0:resized_w] = resized_image
-        return padding_im
-
-    def resize_norm_img_svtr(self, img, max_wh_ratio=None):
-        img_c, img_h, img_w = self.rec_image_shape
-        resized_image = cv2.resize(img, (img_w, img_h),
-                                   interpolation=cv2.INTER_LINEAR)
-
-        resized_image = resized_image.astype('float32')
-        resized_image = resized_image.transpose((2, 0, 1)) / 255
-        resized_image -= 0.5
-        resized_image /= 0.5
-        return resized_image
-
-    def __call__(self, img_list: List[np.ndarray]):
-        if isinstance(img_list, np.ndarray):
-            img_list = [img_list]
-
-        # Calculate the aspect ratio of all text bars
-        width_list = [img.shape[1] / float(img.shape[0]) for img in img_list]
-
-        # Sorting can speed up the recognition process
-        indices = np.argsort(np.array(width_list))
-
-        img_num = len(img_list)
-        rec_res = [['', 0.0]] * img_num
-
-        batch_num = self.rec_batch_num
-        elapse = 0
-        for beg_img_no in range(0, img_num, batch_num):
-            end_img_no = min(img_num, beg_img_no + batch_num)
-            max_wh_ratio = 0
-            for ino in range(beg_img_no, end_img_no):
-                h, w = img_list[indices[ino]].shape[0:2]
-                wh_ratio = w * 1.0 / h
-                max_wh_ratio = max(max_wh_ratio, wh_ratio)
-
-            norm_img_batch = []
-            for ino in range(beg_img_no, end_img_no):
-                norm_img = self.resize_norm_img_svtr(img_list[indices[ino]])
-                norm_img_batch.append(norm_img[np.newaxis, :])
-            norm_img_batch = np.concatenate(norm_img_batch).astype(np.float32)
-
-            starttime = time.time()
-            onnx_inputs = {self.input_name: norm_img_batch}
-            preds = self.session.run(None, onnx_inputs)[0]
-            rec_result = self.postprocess_op(preds)
-
-            for rno in range(len(rec_result)):
-                rec_res[indices[beg_img_no + rno]] = rec_result[rno]
-            elapse += time.time() - starttime
-        return rec_res, elapse
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument('--image_path', type=str, help='image_dir|image_path')
-    parser.add_argument('--config_path', type=str, default='config.yaml')
-    args = parser.parse_args()
-
-    config = read_yaml(args.config_path)
-    text_recognizer = TextRecognizer(config)
-
-    img = cv2.imread(args.image_path)
-    rec_res, predict_time = text_recognizer(img)
-    print(f'识别结果: {rec_res}\t cost: {predict_time}s')

python/rapidocr_onnxruntime/ch_ppocr_v3_rec/utils.py

@@ -1,115 +0,0 @@
-# -*- encoding: utf-8 -*-
-# @Author: SWHL
-# @Contact: liekkaskono@163.com
-import warnings
-
-import numpy as np
-import yaml
-from onnxruntime import (InferenceSession, SessionOptions,
-                         get_available_providers, get_device)
-
-
-class OrtInferSession(object):
-    def __init__(self, config):
-        sess_opt = SessionOptions()
-        sess_opt.log_severity_level = 4
-        sess_opt.enable_cpu_mem_arena = False
-
-        cuda_ep = 'CUDAExecutionProvider'
-        cpu_ep = 'CPUExecutionProvider'
-
-        EP_list = []
-        if config['use_cuda'] and get_device() == 'GPU' \
-                and cuda_ep in get_available_providers():
-            EP_list = [(cuda_ep, config[cuda_ep])]
-        EP_list.append(cpu_ep)
-
-        self.session = InferenceSession(config['model_path'],
-                                        sess_options=sess_opt,
-                                        providers=EP_list)
-
-        if config['use_cuda'] and cuda_ep not in self.session.get_providers():
-            warnings.warn(f'{cuda_ep} is not avaiable for current env, the inference part is automatically shifted to be executed under {cpu_ep}.\n'
-                          'Please ensure the installed onnxruntime-gpu version matches your cuda and cudnn version, '
-                          'you can check their relations from the offical web site: '
-                          'https://onnxruntime.ai/docs/execution-providers/CUDA-ExecutionProvider.html',
-                          RuntimeWarning)
-
-    def get_input_name(self, input_idx=0):
-        return self.session.get_inputs()[input_idx].name
-
-    def get_output_name(self, output_idx=0):
-        return self.session.get_outputs()[output_idx].name
-
-
-def read_yaml(yaml_path):
-    with open(yaml_path, 'rb') as f:
-        data = yaml.load(f, Loader=yaml.Loader)
-    return data
-
-
-class CTCLabelDecode(object):
-    """ Convert between text-label and text-index """
-
-    def __init__(self, character_dict_path):
-        super(CTCLabelDecode, self).__init__()
-
-        self.character_str = []
-        assert character_dict_path is not None, "character_dict_path should not be None"
-        with open(character_dict_path, "rb") as fin:
-            lines = fin.readlines()
-            for line in lines:
-                line = line.decode('utf-8').strip("\n").strip("\r\n")
-                self.character_str.append(line)
-        self.character_str.append(' ')
-
-        dict_character = self.add_special_char(self.character_str)
-        self.character = dict_character
-
-        self.dict = {}
-        for i, char in enumerate(dict_character):
-            self.dict[char] = i
-
-    def __call__(self, preds, label=None):
-        preds_idx = preds.argmax(axis=2)
-        preds_prob = preds.max(axis=2)
-        text = self.decode(preds_idx, preds_prob,
-                           is_remove_duplicate=True)
-        if label is None:
-            return text
-        label = self.decode(label)
-        return text, label
-
-    def add_special_char(self, dict_character):
-        dict_character = ['blank'] + dict_character
-        return dict_character
-
-    def get_ignored_tokens(self):
-        return [0]  # for ctc blank
-
-    def decode(self, text_index, text_prob=None, is_remove_duplicate=False):
-        """ convert text-index into text-label. """
-
-        result_list = []
-        ignored_tokens = self.get_ignored_tokens()
-        batch_size = len(text_index)
-        for batch_idx in range(batch_size):
-            char_list = []
-            conf_list = []
-            for idx in range(len(text_index[batch_idx])):
-                if text_index[batch_idx][idx] in ignored_tokens:
-                    continue
-                if is_remove_duplicate:
-                    # only for predict
-                    if idx > 0 and text_index[batch_idx][idx - 1] == text_index[
-                            batch_idx][idx]:
-                        continue
-                char_list.append(self.character[int(text_index[batch_idx][
-                    idx])])
-                if text_prob is not None:
-                    conf_list.append(text_prob[batch_idx][idx])
-                else:
-                    conf_list.append(1)
-            text = ''.join(char_list)
-            result_list.append((text, np.mean(conf_list)))
-        return result_list

python/rapidocr_onnxruntime/rapid_ocr_api.py

@@ -1,5 +1,6 @@
-# !/usr/bin/env python
 # -*- encoding: utf-8 -*-
+# @Author: SWHL
+# @Contact: liekkaskono@163.com
 import copy
 import importlib
 import sys
@@ -14,8 +15,11 @@ sys.path.append(str(root_dir))
 
 
 class TextSystem(object):
-    def __init__(self, config_path=str(root_dir / 'config.yaml')):
+    def __init__(self, config_path):
         super(TextSystem).__init__()
+        if not Path(config_path).exists():
+            raise FileExistsError(f'{config_path} does not exist!')
+
         config = self.read_yaml(config_path)
 
         global_config = config['Global']

python/tests/test_rec_v3.py

@@ -1,24 +0,0 @@
-# -*- encoding: utf-8 -*-
-# @Author: SWHL
-# @Contact: liekkaskono@163.com
-import base
-import cv2
-
-
-module_name = 'ch_ppocr_v3_rec'
-class_name = 'TextRecognizer'
-TextRecognizer = base.init_module(module_name, class_name)
-
-yaml_path = base.root_dir / base.package_name / module_name / 'config.yaml'
-config = base.read_yaml(str(yaml_path))
-config['model_path'] = str(base.root_dir / config['model_path'])
-config['keys_path'] = str(base.root_dir / config['keys_path'])
-
-text_rec = TextRecognizer(config)
-
-
-def test_rec():
-    img_path = 'test_files/text_rec.jpg'
-    img = cv2.imread(img_path)
-    rec_res, elapse = text_rec([img])
-    assert rec_res[0][0] == '韩国小馆'