Toward equity: TSLYOLOv8s - transcribe for Tamil Sign Language recognition

Sign language relies on hand gestures, facial expressions, and body postures. Vision-based systems are crucial for recognizing these gestures to convert sign language into text or speech. Tamil Sign Language Recognition (TSLR) remains underdeveloped, with limited studies on Tamil vowels and consonants. Many Tamil words can be represented with one hand, but no benchmark dataset exists for Tamil Sign Language. To address this, we constructed a dataset with 56 classes, including vowels, consonants, and Tamil Single-Hand Alphabets (TSHA), incorporating variations in scale, background, pose, resolution, and lighting. We propose Tamil Sign Language - You Only Look Once-small (TSL-YOLOv8s), an enhanced YOLOv8s model integrating Space to Depth - Convolution, Batch Normalization, and Sigmoid Activation (SPD-CBS) with Dynamic Head (DyHead) to improve recognition accuracy without increasing computational cost. The proposed model achieves 94% accuracy, outperforming state-of-the-art methods and demonstrating its effectiveness in real-world applications.