Robust OCR on alternating-distance handwritten grades

asked 2020-04-06 07:09:14 -0500

I have trained a keras OCR model on MNIST and have to train apply it in order to extract the grade out of images like this one (one digit at a time):

image description

Now, the problem is that the size and the distance differs from image to image, so my sliding window approach fails.

Can you give me any hint/clue on how I can extract those 3 digits that make up the grade? Thanks!

PS: Some other examples of grades:

image description

edit retag flag offensive close merge delete