Sign Language Recognition Resolution In Variable Background For Hearing Impaired People

Abstract

There are several methods used to detect the sign language. Researchers have used two technologies to capture hand gestures first one is to use data gloves and second one is to use computer vision based technologies. The glove based recognition requires the user to wear a cumbersome data glove to capture hand and finger movement. Vision based hand sign language recognition can be divided into two forms as static and dynamic. Vision based approach have achieved many progress but it still faces some challenges and remains an open problem due to its less adaptability in terms of varying hand size, lighting, background and camera characteristics. This thesis has evaluated the recognition percentage of the hand gestures in the form of alphabets A to Z using different feature extractor and classifiers. Feature extractors used for recognition are Wavelet Transform, Curvelet Transform and Contourlet Transform. Two classifiers namely Neural Network and K Nearest Neighbor are used for classification. The experiment was done on three data sets in which each data set has 1014 image and for the training the data set splitted in the ratio of 80 percentage and 20 percentage. The results are declared with the help of experiments performed on three data sets an overall accuracy of 95.38 percent is achieved with the combination of Wavelet Transform and k NN classifier. newline We presented a robust position invariant Sign Language Recognition framework in which a depth sensor device Kinect has been used to obtain the signers information. Similarly, non manual signs play an important role in Sign Language Recognition systems because they carry grammatical and prosodic information. We have proposed multimodal framework for Sign Language Recognition system by incorporating expression with sign gesture using two different sensors and based on the proposed work our analysis shows promising recognition result of 96.05 percent. newline newline