Thanks for the suggestion to measure the image processing time, that could be interesting. To be honest I haven't yet tried Tesseract so I can't make any comparisons.

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.