A Hybrid Deep Learning Framework for Isolated Sign Language Recognition

Abstract

This project develops a robust isolated Sign Language Recognition (SLR) system for real-time video-based interpretation of sign gestures, enabling effective communication for the deaf and hard-of-hearing community. The system implements a hybrid deep learning approach combining MobileNetV2-based CNN for lightweight spatial feature extraction with attention-enhanced LSTM for precise temporal modelling, further strengthened by insights from R3(2+1)D convolutional blocks and multimodal pose-based feature fusion. Evaluated on the benchmark WLASL (100 classes) dataset, the proposed frame-work demonstrates strong signer-independent performance. The lightweight architectures supports efficient real-time inference while maintaining high robustness against background clutter, illumination changes, and signer variations, making it suitable for deployment in assistive technologies, education, and human- computer interaction applications.