1

I train OCR model to recognize cyrillic handwritten text. I know, for example, that it confuses very often 'Б' with '6'. How can I use this information to fine tune the model ?

Just in case, my architecture is resnet50 + transformer

UPD: train data is crops with one or two words (sequence of characters). Overall my model is able to predict 82 characters.

Constantin Werner
  • 151
  • 4
  • Are you testing just on a single character? – Good Luck Mar 23 '21 at 17:46
  • @GoodLuck thanks, I've specified information in the question. – Constantin Werner Mar 23 '21 at 17:48
  • Can you provide architecture details and the idea how the model works/ learns. – Good Luck Mar 23 '21 at 18:01
  • @GoodLuck I use resnet50 to get features and then I split it into vectors and pass it to encoder/decoder. [diagram](https://sun9-33.userapi.com/impg/l7bDgXLob9f6FVo_hrM30oFmCarKaYVa-gT4bQ/_tkAY9aKVe8.jpg?size=872x350&quality=96&sign=3cf8402303818912e5818f14f2ec7a06&type=album) – Constantin Werner Mar 23 '21 at 18:01
  • [whole pipeline](https://sun9-69.userapi.com/impg/HzW_f0_xCZTIxCnPzvJsVbGQ_1v3f7DFxXfQxg/S3aAkW_R3g8.jpg?size=859x591&quality=96&sign=e14352a6af90059cdf2e99db95a7d14b&type=album) – Constantin Werner Mar 23 '21 at 18:04

0 Answers0