I'm working on a multifont OCR system. So far, I've tried convolutional neural networks, decision trees, KNN and SVMs for classification. I've also tried shape contexts. I haven't done much work in feature extraction, though. Most of my approaches have come from reading papers, which tend to be heavy on the math and light on the implementation detail.
Are there any books or online resources on this topic? I'd be especially interested in clear information on how to implement skeleton extraction, endpoint/junction finding and other feature extractors suited for OCR. I would also love detailed information on how to detect and prevent overfitting in neural networks.
(If you're interested, here's the project on GitHub - but note that it's under construction and doesn't work that well yet.)
[link] [6 comments]