lstm eval gives error 0% but tesseract fails to predict correctly

[Artur Maricato Curinga]

I'm trying to generate a model with OCR-B font to read documents MRZ, tried an overfitting test using 44 cropped characters and trained using the ocrd-training repo (https://github.com/OCR-D/ocrd-train) using train = eval img list or without the eval on training.

The training stops since lstmeval gives a 0% error rate, but testing the model with tesseract fails to predict the correct text on 15% of the images.
What is the difference between the lstmeval and the tesseract prediction? This behaviour is expected?

[Steven ZHOU]

Hi, I met the same problem. Have you found the reason or solutions? Thanks

在 2019年7月16日星期二 UTC+8上午3:00:37，Artur Maricato Curinga写道：