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.