I'm developing a character segmentation algorithm for license plate OCR. My algorithm includes two steps: segmentation and recognition. There is almost no problem for recognition thanks to CNN. My problem is about segmentation. I know some state-of-the-art algorithms such as binarization + connected component labeling (or contour detection), MSER, vertical and horizontal projections. Aforementioned approaches are not strong. For example, in case of character breakage they fail. I, my self use adaptive local thresholding method and analyzing connected component for segmentation. I need a deep learning approach for segmentation or any segmentation free OCR such as STN-OCR (the problem about STN-OCR is that I can not find any tutorial code for it, specially for for training such networks). Is it convention using autoencoders or GAN for theresholding?"
A also provide some license plates as below: