Handwritten Arabic Text Recognition using Deep Belief Networks

Offline Arabic handwritten text recognition task exhibits high variations in observed variables such as size, loops, slant and continuity. Learning algorithm tries to capture the statistical dependence between these variables but often fails to learn the complete distribution because of their large degree-of-freedom. However, it is possible to output a good hypothesis if either data samples for training are sufficient or features representing the data are rich enough to learn the highly non linear target function. Number of training samples are generally limited in case of handwritten scripts ruling out the first option. Therefore, in this work we propose a method to represent data in a more informative manner that enables learning algorithm to approximate the actual target function despite limited training data samples. We use Deep Belief Networks which incrementally learns complex structure of the data by representing it in a more compact and abstract manner. We use publically available AMA PAW dataset to show the efficacy of our method and significant improvement over state-of-the-art methods is reported.