Good neural network topology and training method for image recognition

Question

I have already done a simple project on pattern recognition.
I used Backpropagation and a multi-layer Neural network to recognize 10 alphabetical letters.
It was pretty hard and I had lots of trouble fighting local minimums and slow training.
But still, this programme worked. Of course, it could have been just my bad code :)

Right now I'd like to improve my results and use all 26 letters. Moreover the training data may be more complex (larger images at least). So I am a bit worried that I might use a wrong approach.

What do you think might help? Maybe I should use some sort of combination between genetic algorithms and Neural Networks?
Or use a more complex network topology? (Ward network and training algorithm come to mind)

I'd like to mention the fact that I am not using any OCR libraries, it's just an educational project.

It would be great if anyone could give me an advice. Thank you.

Answer 1

Try resilient propagation training.

I have had great success with vanilla Rprop training. As for ANN topology normal feed-forward works for image processing. You can try experimenting with Self-organizing map if all the letters are of same font and small size (4x4 maybe), the output could represent letters.