International Journal of Advanced Research in Computer and Communication Engineering

Abstract: Communication barriers between hearing-impaired individuals and the general population pose significant challenges in education, healthcare, and daily interactions. Sign language serves as an essential medium for such communities, yet the lack of widespread proficiency creates a persistent accessibility gap. Recent progress in computer vision and deep learning provides a promising pathway to automate the interpretation of sign gestures in real time. This work presents a Convolutional Neural Network (CNN)-based sign language recognition framework that accurately classifies static hand gestures from the American Sign Language (ASL) alphabet. The system integrates image preprocessing, region-of-interest (ROI) extraction, and optimized feature learning to enhance recognition efficiency under varying lighting, backgrounds, and hand orientations.

To build a robust model, multiple CNN architectures—including MobileNetV2, a custom deep CNN, and a classical LeNet-5 variant—were trained and evaluated on the ASL Alphabet Dataset. An ensemble fusion mechanism was designed to combine the predictive strengths of all three networks, producing a stable and highly accurate classification output. Post-processing with an N-gram–based decoder further improves consistency by reducing misclassification of visually similar signs. Experimental evaluation demonstrates that the proposed approach delivers strong performance across key metrics such as accuracy, precision, recall, and inference time, enabling reliable real-time deployment. The resulting system supports text-based and text-to-speech outputs, offering a practical tool for inclusive communication. Overall, the research provides a scalable and efficient solution for sign language recognition, contributing toward accessible human–computer interaction technologies.

Downloads: | DOI: 10.17148/IJARCCE.2026.15140