Bangla Hand Sign Recognition Using Hybrid CNN with Attention Mechanism

Bangla Sign Language (BdSL) is essential for communication among the hearing-impaired in Bangladesh, yet existing recognition systems face challenges with generalization, real-time performance, and limited feature extraction. While recent models employ deep learning and segmentation, they often lack efficient attention mechanisms and struggle with imbalanced datasets. A hybrid CNN model integrating channel Attention, hybrid pooling, and GAN-based augmentation is proposed to address these issues. The model was evaluated on KU-BdSL and BdSL 49 datasets and achieved 99.58% and 97.65% accuracy, respectively. This approach enhances robustness, contributing to inclusive, real-time communication for the hearing-impaired community.