Implementation of Indian Sign Language Biometrics Using Scale Invariant Feature Transform

Abstract

India, having less awareness towards the hearing impaired people s leads to increase the communication gap between deaf and hard hearing community. Sign language is commonly developed for such hearing impaired people s to convey their message by generating the different sign pattern. The different hand gestures of Indian Sign Language (ISL) (A to Z) and Devnagri sign language have been used to perform reliable matching using scale invariant feature transform (SIFT). The scale invariant features transform provides the stable features irrespective of translation, rotation and scaling, the features are highly stable at clutter background and noise. In this work, the various phases of scale invariant feature transform have been implemented to extract the distinctive features of Indian Sign Language and Devnagri sign language gestures. Initially, the time constraints have been reported to each phase of SIFT algorithm to extract the number of features. Secondly, the intensity based feature extraction and matching has been performed using SIFT features. In the third phase, the scale of the original image has been varied and measures the computational time for extracting the features and reliable matching has been found for various scales. The primary part of experimental result shows the time constraint on each phase and the number of features extracted from 26 Indian Sign Language gestures. The secondary part of experimental result shows the total number of key points matching when the intensity of the image varies from low on high. The third part of the experimental result shows the reliable matching of scale varied image with respect to the original image. The experimental result of above three phases shows that SIFT algorithm can perform reliable matching for a given image irrespective of translation, rotation, and scale. Among the 520 images of Indian Sign Language gesture, 494 images were applied to test, out of which 483 images were correctly recognized by the system that yields 97.77% of the recognition accuracy