Lightweight Bangla Sign Language Recognition

Summary & Methodology Analysis

The researchers addressed the performance overhead of heavyweight pretrained backbones by designing a custom lightweight attention-based CNN. This architecture utilizes 298,470 parameters and incorporates grouped bottleneck residual blocks, channel and spatial attention modules (which dynamically focus the model on the most relevant features in an image), and a multi-scale depthwise hand-feature block. To maintain high performance on mobile devices, the model employs dual pooling and Swish activations, training from scratch using SGD with Nesterov momentum and extensive data augmentation. The design prioritizes inference speed and footprint, avoiding the bloat of standard ImageNet-pretrained architectures. Evaluation involved rigorous benchmarking against nine ImageNet-pretrained models and six public datasets, demonstrating superior hardware efficiency on consumer-grade mobile hardware. The model was validated using Grad-CAM and Grad-CAM++ (techniques used to visualize which pixels contribute to a model's prediction) to ensure the system is learning semantically meaningful hand features. Despite the strong performance, the current iteration is restricted to static hand signs. It does not provide temporal modeling capabilities required for complex, dynamic sign language sequences. The paper does not specify the power consumption metrics of the model during operation, focusing instead on latency and footprint.