OCR knowledge summary

OCR

Character recognition is also CV One of the main research directions , The direction of character recognition is mainly ：
1. Separate character recognition
2. Combined with text detection
3. Text end-to-end recognition
Character recognition technology ： Locate the text in the image area through text detection , Feature on the extracted region , On this basis, do special character recognition , There are many end-to-end ENd2End OCR

1. Text detection — Locate the text area in the picture （ The accuracy of positioning directly affects the subsequent Recongnition）
   Text detection in Concept ：
   1. ground truth(GT): In supervised learning Data is a marker （X,t） x It's input data , Correct t The label is ground truth
      In image recognition ： Input the alpha chart , The original drawing uses Alpha The big brother label is GT (Aplha Channels indicate the transparency and opacity of an image )
   2. detecting box: Window moved box
   3. IOU: Standard performance metrics for image segmentation , The similarity between the predicted area and the live area
   4. Text detection algorithm ：
   5. EAST/CTPN/SegLink/PixelLink/TextBoxes/TextBoxes++/TextSnake/MSR/…
2. Character recognition ：
   For text recognition without bending
   * CNN + RNN + CTC
   * CNN + seq2deq+Attention
   * CNN + LSTM + CTC Verification code recognition
   For curved text recognition ：
   In the traditional way There are a lot of invalid areas ,STN Network learning transformation parameters
   Use Deformable Convolution Deformable convolution Different shape features of the text area can be extracted

Reference article ： https://zhuanlan.zhihu.com/p/657075435