Purpose

        There are many such problems , Most of the problems are trained model Preprocessing during prediction 、 Post processing parameters and inference model Preprocessing during prediction 、 Caused by inconsistent post-processing parameters . official DBnet The detection model provides yml When the file is self training, it needs to be modified accordingly to train the model .

solve

One 、 Adjust according to official instructions

The official part also explains :PaddleOCR/detection.md at release/2.5 · PaddlePaddle/PaddleOCR (github.com)

According to the code of the official website deployment prediction model ,yml The configuration is as follows

      - DetResizeForTest:
          limit_side_len: 960
          limit_type: max

Two 、 Adjust after self-test

        Training bank card 、 license plate number 、 ID card and other text detection models , According to the above configuration, the actual training effect is not good , By fixing the size of the picture, the effect is very good , Modified as follows ：

1、 modify yml file

      - DetResizeForTest:
          resize_long: 960

2、tools/infer/utility.py Add the following parameters to the file

 3、tools/infer/predict_det.py The document is modified as follows

4、 Other documents ： Modify the deployment forecast file , With hubserving Deployment as an example ：

deploy/hubserving/ocr_system/params.py

summary         

          The above modification effect , You can check my previous articles Bank card 、 license plate number 、 Id card ​​​​​​​ Text detection training results , It is much better than the general text detection model in the vertical effect .