rapidocr

models
  |-- ch_PP-OCRv2_det_infer.onnx
  |-- ch_ppocr_mobile_v2.0_cls_infer.onnx
  |-- ch_ppocr_mobile_v2.0_det_infer.onnx
  |-- ch_ppocr_server_v2.0_det_infer.onnx
  |-- ch_ppocr_server_v2.0_rec_infer.onnx
  |-- en_number_mobile_v2.0_rec_infer.onnx
  |-- korean_mobile_v2.0_rec_infer.onnx
  `-- japan_rec_crnn.onnx

fonts
  |-- msyh.ttc
  `-- korean.ttf

运行

接口调用方式运行

如果想要使用其他语言的识别模型，可以只对rec_model_path和keys_path作对应更改即可

!!!各个不同语言的推理代码区别只在于模型和字典文件!!!

from rapid_ocr_api import TextSystem, visualize

det_model_path = 'models/ch_ppocr_mobile_v2.0_det_infer.onnx'
cls_model_path = 'models/ch_ppocr_mobile_v2.0_cls_infer.onnx'

# 中英文识别
rec_model_path = 'models/ch_ppocr_mobile_v2.0_rec_infer.onnx'
keys_path = 'rec_dict/ppocr_keys_v1.txt'

text_sys = TextSystem(det_model_path,
                      rec_model_path,
                      use_angle_cls=True,
                      cls_model_path=cls_model_path,
                      keys_path=keys_path)

image_path = r'test_images/det_images/ch_en_num.jpg'
dt_boxes, rec_res = text_sys(image_path)
visualize(image_path, dt_boxes, rec_res)


# 只有中英文和数字识别
rec_model_path = 'models/en_number_mobile_v2.0_rec_infer.onnx'
keys_path = 'rec_dict/en_dict.txt'

text_sys = TextSystem(det_model_path,
                      rec_model_path,
                      use_angle_cls=True,
                      cls_model_path=cls_model_path,
                      keys_path=keys_path)

image_path = r'test_images/det_images/en_num.png'
dt_boxes, rec_res = text_sys(image_path)
visualize(image_path, dt_boxes, rec_res)


# 日语识别
rec_model_path = 'models/japan_rec_crnn.onnx'
keys_path = 'rec_dict/japan_dict.txt'

text_sys = TextSystem(det_model_path,
                      rec_model_path,
                      use_angle_cls=True,
                      cls_model_path=cls_model_path,
                      keys_path=keys_path)

image_path = r'test_images/det_images/japan.png'
dt_boxes, rec_res = text_sys(image_path)
visualize(image_path, dt_boxes, rec_res)


# 韩语识别
rec_model_path = 'models/korean_mobile_v2.0_rec_infer.onnx'
keys_path = 'rec_dict/korean_dict.txt'
font_path = 'fonts/korean.ttf'

text_sys = TextSystem(det_model_path,
                      rec_model_path,
                      use_angle_cls=True,
                      cls_model_path=cls_model_path,
                      keys_path=keys_path)

image_path = r'test_images/det_images/korean_1.jpg'
dt_boxes, rec_res = text_sys(image_path)
visualize(image_path, dt_boxes, rec_res, font_path)

命令行运行
```
$ bash test_demo.sh
```

参数名称	作用	建议取值范围	默认值	代码位置
box_thresh	文本检测所得框是否保留的阈值，值越大，召回率越低	[0, 1]	0.5	text_detect.py
unclip_ratio	控制文本检测框的大小，值越大，检测框整体越大	[1.6, 2.0]	1.6	text_detect.py
text_score	文本识别结果置信度，值越大，把握越大	[0, 1]	0.5	rapidOCR.py

onnxruntime-gpu版推理配置

ONNXRuntime-gpu需要严格按照与cuda、cudnn版本对应来安装，具体参考文档, 这一步关乎后面是否可以成功调用GPU

推理代码中，加载onnx模型部分，用以下对应语言代码替换即可,详细参见：官方教程

python版本

import onnxruntime as ort

model_path = '<path to model>'

providers = [
   ('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,
   }),
   'CPUExecutionProvider',
]

session = ort.InferenceSession(model_path, providers=providers)

C/C++版本

OrtSessionOptions* session_options = /* ... */;

OrtCUDAProviderOptions options;
options.device_id = 0;
options.arena_extend_strategy = 0;
options.gpu_mem_limit = 2 * 1024 * 1024 * 1024;
options.cudnn_conv_algo_search = OrtCudnnConvAlgoSearch::EXHAUSTIVE;
options.do_copy_in_default_stream = 1;

SessionOptionsAppendExecutionProvider_CUDA(session_options, &options);

推理时间粗略对比(跑完RapidOCR整个项目)

推理方式推理图像数目耗费时间

CPU 46张图像 191 s

GPU 46张图像 52.38 s

推理方式	推理图像数目	耗费时间
CPU	46张图像	191 s
GPU	46张图像	52.38 s

README.md

使用步骤

相关调节参数

onnxruntime-gpu版推理配置