Best Traineddata Feedback - Gujarati - ન - ત Confusion

[Shreeshrii]

When using tesseract 4.0 with --oem 1 (LSTM) with Gujarati traineddata, ન is being recognized as ત in the attached image.

Same image when recognized with --oem 0 is recognizing ન correctly, but has other accuracy problems.

So, it looks like that LSTM model for Gujarati has not been trained with this font.

Image and ground truth file are attached.

It would be helpful to have the ability to finetune using real images in addition to synthetic data.

guj.ag.exp0-GT.txt

[Shreeshrii]

Tested with both best/guj and best/Gujarati traineddata - psm 6 just now.

While the ન - ત Confusion is still there, Gujarati traineddata seems better than guj - it is dropping fewer words in OCR output.

[amitdo]

While the ન - ત Confusion is still there, Gujarati traineddata seems better than guj - it is dropping fewer words in OCR output.

https://github.com/tesseract-ocr/tesseract/pull/1264