A Hybrid Decision Tree for Printed Sinhala Character Recognition Using SVMs

Printed character recognition is a well researched area due to its necessity in many real world applications. Printed character recognition methods have reached very high performance for many languages. However, research on recognizing printed Sinhala characters is in need of improvement, especially in terms of accuracy of recognition as well as completeness. Printed Sinhala character recognition is a very challenging task, due to large number of complex structured characters and similarity between characters. This paper proposes a multi-class classification approach to recognize Sinhala printed characters using a hybrid decision tree. The decision tree is composed of a directed acyclic graph (DAG) followed by an unbalanced decision tree (UDT). The DAG was implemented using OVO-based Support Vector Machines (SVMs) and UDT was implemented using OVA-based SVMs with RBF kernel. Basic, density, histogram of gradients (HoG) and Transition features were utilised to construct the decision tree. Several trials were conducted to evaluate the performance of the proposed model with 203 unique character classes, where each class contains 14 font styles. The overall recognition rate was observed to be 78.41%. This study has considered the full Sinhala alphabet of more than 600 characters and achieved high recognition accuracy. Further, this study has analyzed the performance of different type of features and identified the best performing features. In addition this study has resulted in building a dataset of 17052 Sinhala character images.