HANDWRITTEN ARABIC CHARACTERS RECOGNITION SYSTEM BASED ON FUZZY CLUSTERING

In this paper, hybridization approach of fuzzy sets and rough sets clustering, support vector machines (SVMs), and heuristic techniques to develop an OCR for handwritten Arabic characters is presented. The learning phase in building the classifier is done via two stages; firstly grouping the handwritten Arabic letters into groups using fuzzy-rough clustering and heuristic techniques. Secondly building a multi-stage classifier using SVMs for each group. The paper presents the extraction of robust features set of different types, data sets used, experiments design for similar and dissimilar clustering, and building of single and multi stage classifiers. The comparison between the different developed classifiers, results analysis, and the comparison with related works are also introduced Experiments were established on a database of 140000 isolated handwritten Arabic letters (500 writers each writes the 28 alphabet Arabic characters 10 times (500x10x28)), 91000 characters are used in training (65%) and 49000 characters in testing (35%)). The accuracy of the multi stage classifier system reached to 98.8%, and reduced the misclassification rate from 11.3 % (in single stage classification system) to 1.2% (in multi stage classification system).