A Multistage Classification and Kernel-Based Recognition Scheme for Handwritten Marathi Compound Characters

Compound character recognition is a challenging task in handwritten Marathi character recognition. The frequency of occurrence of compound characters in Marathi language is more compared to other languages derived from Devanagari. This paper presents a novel approach for recognition of handwritten Marathi compound characters based using two stage classification. In the first stage, the classification is carried out by a decision tree based on various structural parameters derived from the global and local features of each character. The second stage of recognition is carried out by template matching to find similarity between the test character and the templates obtained after first stage classification. Two different types of character templates are generated. In the first case, the 16×16 resized binary character images are stored as templates. In the second case, the 16×16 resized images are used as kernels. The templates for each character are generated by convolving the kernels with themselves. The recognition accuracy for the binary templates is 90.57% and 92.59% for 16×16 and 32×32 respectively while for the convolved templates, the recognition accuracy increases to 92.89% and 95.39% for 16×16 and 32×32 respectively.