i tested some methods for OCR and LBP operator was the best. You can gain promising result by customising it. My paper:

http://sdiwc.net/digital-library/persianarabic-handwritten-digit-recognition-using-local-binary-pattern.html

many type of technique its depend on you

Try extracting features that is font and rotation invariant. You can try Subspace algorithms that can be font invariant but variant to rotation. 

@ Mohsen, as per my knowledge the LBP operator was the best technique for face recognition and  it is useful for  document images also, cause i also working document image analysis and i got stuck in the concept of that LBP. can u help me