Design and Simmulation of Handwritten Multiscript Character Recognition

Abstract: The work presented in this paper focuses on recognition of isolated handwritten characters in Devanagari and Gurumukhi script. The proposed work uses four feature extraction methods like Zoning density, Projection histograms, Distance profiles and Background Directional Distribution(BDD). On the basis of these four types of features we have formed 10 feature vectors using different combinations of four basic features. This work uses two classifiers like Support Vector machines (SVM), K-Nearest Neighbor (KNN). A total of 7000 samples of characters are taken for Gurumukhi and 7200 samples for Devanagari are used and we have attain a maximum recognition accuracy of 95.79% in case of Gurumukhi recognition and 92.88% for Devanagri. In addition to it we have compared the performance of three similarity based classifiers like Euclidean distance, Manhattan distance and Normalized Histogram Intersection for Gurumukhi and Devanagari characters. Among these three Normalized Histogram Intersection gives the highest accuracy of 89.28% for Gurumukhi and Manhattan distance gave the highest recognition accuracy of 86.14% for Devanagri characters.

Keywords: Character recognition, Feature extraction, Support Vector Machine, Classification