A Novel Comparative Compression with Attention Based Yolo Model for Indian Sign Language Recognition

Resumen

 People with hearing disabilities face significant communication challenges, often utilizing sign language. The impact of technology in identifying sign language has been extensively researched in various languages around the globe. However, tools and techniques for conveying sign language in local dialects remain underdeveloped. This proposed work investigates the effectiveness of the YOLO model in recognizing Native Indian Sign Language. Therefore, the YOLO 8, YOLO 9, YOLO 10, and YOLO 11 models are employed to evaluate performance. Additionally, a high-accuracy YOLO model with various attention mechanisms, including channel attention, spatial attention, and split-attention, is selected and examined to analyze performance. Finally, the model's complexity is analyzed and minimized using different compression methods. Techniques such as pruning (weight pruning, structured pruning, and filter pruning), quantization (uniform quantization, non-uniform quantization, min-max quantization, and logarithmic quantization), and knowledge distillation are employed to reduce complexity while ensuring high performance. This proposed work facilitates the selection of high-performance YOLO models with suitable compression methods to decrease complexity. The performance is validated using Python-based metrics such as accuracy (93.65%), precision (95.10%), balanced accuracy (94.28%), geometric mean (94.01%), and false values, including false positives (3.12%), false negatives (3.85%), false acceptance rate (2.95%), and false rejection rate (3.42%). Additionally, higher-level evaluation metrics such as Matthew’s Correlation Coefficient (0.89) and Pearson’s Correlation Coefficient (0.91) further confirm the model’s reliability. The average computation time per frame was recorded at 17 milliseconds, demonstrating the model’s suitability for real-time deployment.