Lightweight Attention Model Achieves 96.37% Accuracy in Bangla Sign Language Recognition
So, there's this new lightweight convolutional network designed to recognize Bangla Sign Language (BdSL), and it’s pretty remarkable! It hit a 96.37% accuracy rate on a dataset called RSBdSL38, which has 10,874 images covering all 38 BdSL signs that represent the 51 letters of the Bangla alphabet. These images were taken from real signers in three special-needs schools across Bangladesh. The model is quite efficient, with just 298,470 parameters, and uses advanced techniques like attention mechanisms and dual pooling. It’s been trained from scratch and does better or is on par with nine other models based on ImageNet, missing the top score by only 1.08%. This breakthrough could make it easier for the deaf and hard-of-hearing community in Bangladesh to access education and services. The study is available on arXiv with the ID 2608.06252.
Key facts
- New dataset RSBdSL38 contains 10,874 expert-validated images of all 38 BdSL hand signs.
- Dataset represents the 51 letters of the Bangla alphabet.
- Images recorded from real signers at three special-needs schools in Bangladesh.
- Proposed lightweight model has 298,470 parameters.
- Model achieves 96.37% accuracy (95.72% ± 0.54% over five seeds).
- Model is within 1.08 percentage points of the best ImageNet-pretrained backbone.
- Model uses grouped bottleneck residual blocks, channel and spatial attention, multi-scale depthwise hand-feature block, dual pooling, and Swish activations.
- Research announced on arXiv with identifier 2608.06252.
Entities
Institutions
- arXiv
Locations
- Bangladesh