This looks great. How does the approach compare to Tesseract? Would it be possible to beat the accuracy of Tesseract with this? Are there any numbers on how long it would take to process an image once trained?
However, I do believe that converting the network input to character codes (output layer size 4 bits for digits, 8 bits for letters) instead of using a softmax layer (would need 10 bits for digits, 26 bits for letters) is a novel approach that really improves performance.