Off-line Persian Handwritten Recognition Using Hidden Markov Models

Abstract

We present a system for recognition of Persian/Arabic handwritten scripts using hidden {M}arkov models (HMMs). the text is segmented to words at first and from words to characters by an appropriate algorithm using the strokes and contour of word image. Then the feature vectors are extracted from sequential vertical frames of characters. Next, a self-organizing map ({SOM}) is employed for clustering the features and reducing the size of inputs as well as smoothing the parameters of HMMs in classification phase. Finally, by using the {HMM} the characters are classified, and by concatenating the character HMMs, the word {HMM} is composed. the system is evaluated with five sorts of Persian handwritten data containing a number of 1, 025 words, and the mean correct classification rate is 97% in word level.