Hybrid Approach for Telugu Handwritten Character Recognition Using k-NN and SVM Classifiers

Abstract

Research in optical character recognition (OCR) had started more than five decades ago. The recognition accuracy for printed characters is above 90%, whereas for handwritten characters is very low and less than 60% as reported in the literature. Handwritten character recognition of Indian languages is still at nascent stage of research and hence captivated our attention for further analysis. This paper describes the handwritten character recognition of Telugu language using two-stage classifiers. k-Nearest Neighbor (k-NN) and Support Vector Machines (SVM) were used for classification in this work. The use of these two classifiers one after the other in two subsequent stages increases the recognition accuracy of the system. Various features extracted from the images are block pixel count, block based directions, histograms and boundaries for both the training and test images. In the first stage k-NN classifier was used and subsequently the wrongly recognized characters were tested with SVM classifier. Again the classifiers were interchanged in the first and second stages to check the improvement of accuracy. It was found that the recognition accuracy increased to a great extent by cascading the two different classifiers. Using these two classifiers the best recognition accuracy obtained was 90.2%.