Comparison of the classifiers in Bangla handwritten numeral recognition

Handwritten Bangla numeral recognition has great prospects in Writer Identification, Postal Automation, Bangla OCR (Optical Character Recognizer) etc. In this paper we have presented the detailed comparison of classifiers for Bangla handwritten numeral recognition. For this work we have used our own database (WBSUCS character database) which consists of total 517 documents and ISI Bangla Numeral database which consists of more than 12000 numerals. For our database each writer was asked to write predefined filled in forms five times. After collecting and extracting characters from filled in forms, 400 dimensional feature vectors is computed based on gradient of the images. The feature and classifier selection is one of the most challenging tasks in the field of Pattern Recognition. As the performance of 400 dimensional feature is already established in numeral recognition field, for the present work we have focused on performance evaluation of classifiers in handling complex real time Pattern Recognition problems like Numeral Recognition. Here we have selected Support Vector Machine (SVM), Library for Large Linear (LIBLINEAR), Multilayer Perceptron (MLP), Fuzzy Un-ordered Rule Induction Algorithm (FURIA), Modified Quadratic Discriminant Function (MQDF) as the classifiers for recognition of the numerals and comparison of the results. Though all these classifier are suitable for this work but LIBLINEAR is found to be the fastest in terms of convergence criteria while MQDF outperform others in terms of recognition result for our WBSUCS character database.