TAR based shape features in unconstrained handwritten digit recognition

In this research, the recognition accuracy of triangle-area representation (TAR) based shape feature is measured in recognizing the totally unconstrained handwritten digits. The TAR features for different triangles of variable side lengths that are formed by taking the combinations of different contour points were computed. The set of contour points that yielded the best features was experimentally discovered. For classification a curve matching technique is used. Several experiments were conducted on real-life sample data that was collected from postal zip codes written by mail writers. The highest recognition result of 98.5 % was achieved on the training data set and 98.3% on the test data set. Key-Words: Triangle area representation, Shape descriptors, Digit recognition, Contour points, and zip codes