Lightweight Attention-Based Model for Bangla Sign Language Recognition
August 7, 2026
A 298,470-parameter convolutional network achieves 96.37% accuracy on the new RSBdSL38 dataset of 10,874 expert-validated Bangla Sign Language images. The architecture uses grouped bottleneck residual blocks and multi-scale depthwise hand-feature blocks to enable on-device deployment.
HOW THIS AFFECTS YOU
●
builderYou can deploy this model on low-power personal devices for real-time sign language recognition.
●
researcherThe model achieves near-ImageNet-pretrained performance while being trained from scratch on a niche dataset.