Lightweight Convolutional Neural Network with Self-Attention Mechanism for Bangla Sign Language Recognition

Bangla Sign Language (BdSL) recognition is a crucial step toward enabling seamless communication for the hearing-impaired community in Bangladesh. However, existing methods predominantly rely on computationally intensive models that demand high-resource environments, limiting their practical applicability and accessibility. In this research, we propose a lightweight Convolutional Neural Network (CNN) model integrated with a self-attention mechanism. The model is specifically designed to maintain a simple architecture and operate with low computational power, making it an efficient and scalable solution for real-world use cases. Despite its simplicity, the proposed model achieves an impressive accuracy of 93.47% on 38 BdSL dataset, which is comparable to state-of-the-art (SOTA) models from the prior works that rely on more complex and resource-heavy frameworks. The integration of the self-attention mechanism enhances feature representation while keeping computational costs minimal. This study not only demonstrates the feasibility of using lightweight models for BdSL recognition but also establishes a new benchmark for developing cost-effective, high-performance solutions. By achieving competitive results with reduced resource requirements, this work provides a promising path for creating accessible technologies that can empower the hearing-impaired community and facilitate their inclusion in society.