| demo images | TPS-ResNet-BiLSTM-Attn korean |
|---|---|
 |
서울대역 |
 |
상설 |
 |
영어전문학원 |
 |
아카데미 |
 |
대형출력인쇄 |
 |
available |
We have added Korean to the project and modified it for reading. trainings studied 500000 sets of Korean and English datasets made using synth text, and the examples of those datasets are as follows.
| Trainning images | GT text |
|---|---|
 |
항자원 |
 |
신청을 |
 |
Polizei! |
 |
때문에 |
 |
소비가 |
After securing the Korean data set, input is required to the Korean string.
We modified train.py and saved Korean consonants in the file KoreanCodec.txt so that we could call it up.
model downlaod : TPS-ResNet-BiLSTM-Attn-korean.pth Please contact me if there is a problem with the model.
If the OCR image is corrected for tilting and further learning, 99.922% of the accuracy of the Korean OCR was shown, but the model cannot be disclosed due to the research project contract.
| Predict | GT | Result |
|---|---|---|
| 배산공원 | 배산공원 | True |
| 87 | 87 | True |
| 300m | 300m | True |
| 분당소방서 | 분당소방서 | True |
| 시의회 | 시의회 | True |
[100000/100000] Train Loss: 0.00663 elapsed_time: 2609.21140
[100000/100000] valid loss: 0.00391 accuracy: 99.922, norm_ED: 0.83