A Real-Time Appearance-Based Bengali Alphabet And Numeral Signs Recognition System

Abstract

This paper presents a real-time appearance-based Bengali alphabet and numeral signs recognition system. Region of Interest (ROI) is initialized by detecting ‘Opened Hand’ followed by ‘Closed Hand’ posture from captured images using Haar cascade classifiers. Probable hand area is segmented based on Hue and Saturation values of human skin color from ROI. Segmented image has been converted to binary image and resized it to predefine resolution (150 150). Row vectors are generated from the binary images to train and/or test the system. Hand postures are classified using K-Nearest Neighbors (KNN) classifier. The system is trained using 3600 (36 10 10) images of 36 Bengali alphabets and 1000 (10 10 10) images of 10 numeral signs, from 10 performers. The system is tested using 7200 images of alphabets and 2000 images of numeral signs. Among which, 3600 images of alphabets and 1000 images of numerals were collected from who didn’t participate in training process. Experimental results show satisfactory classification accuracies in real-time.