Hill-Climbing Optimization for An Efficient Indian Sign Language Recognition System using a k-means Clustering Approach for Keyframe selection

Automatic recognition system for Sign Langauge (SL) function as an advanced means of interaction, facilitating the interaction process between members of the deaf community and those who can hear. Despite significant progress in the past decade, practical implementation of Sign Language Recognition System (SLRS) remains challenging. Many existing techniques for Sign SL recognition are hindered by inefficiency, mainly because of the necessity to capture characteristics from extended sequences of frames. Furthermore, the effectiveness of the Deep Learning-based model relies on factors such as training and model parameters, and determining the ideal parameters can be a challenging endeavor. Therefore, an SLRS named Optimized Clustering based Efficient SLRS (OCESLRS) is proposed recognition of individual ISL words characterized by dynamic gestures. The suggested approach integrates a keyframe selection technique based on k-means clustering to eliminate redundant frames. Feature selection relies on the VGG-19 model, followed by the utilization of a BiLSTM network to capture sequential patterns and classification of sign words. Furthermore, the proposed OCESLRS uses Hill Climbing (HC) based optimization technique to find the optimal parameters to train the model which helped to achieved optimal results. The proposed model attained a mean accuracy of 96.84\%, along with precision, recall, and Fl-score values of 0.9705, 0.9685, and 0.9685, respectively. This performance is notably superior to the existing SLRS.