Toward equity: TSLYOLOv8s - transcribe for Tamil Sign Language recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Linguistics

Background:

• Sign language recognition is vital for converting sign language into text or speech.
• Tamil Sign Language Recognition (TSLR) is an underdeveloped field with limited datasets and research.
• Existing research lacks a benchmark dataset for Tamil Sign Language, hindering progress.

Purpose of the Study:

• To develop a comprehensive dataset for Tamil Sign Language Recognition (TSLR).
• To propose an enhanced deep learning model for accurate TSLR.
• To address the limitations in current Tamil sign language recognition research.

Main Methods:

• Construction of a 56-class Tamil Sign Language dataset, including vowels, consonants, and Tamil Single-Hand Alphabets (TSHA).
• Development of the Tamil Sign Language - You Only Look Once-small (TSL-YOLOv8s) model, integrating Space to Depth - Convolution, Batch Normalization, and Sigmoid Activation (SPD-CBS) with Dynamic Head (DyHead).
• Evaluation of the model's performance on the constructed dataset with variations in scale, background, pose, resolution, and lighting.

Main Results:

• The proposed TSL-YOLOv8s model achieved an accuracy of 94%.
• The model demonstrated superior performance compared to state-of-the-art methods in TSLR.
• The integrated SPD-CBS and DyHead components improved recognition accuracy without increasing computational cost.

Conclusions:

• The developed TSLR dataset and TSL-YOLOv8s model provide a significant advancement for Tamil Sign Language recognition.
• The proposed model shows high effectiveness for real-world TSLR applications.
• This research lays the foundation for future developments in Tamil Sign Language technology.