An Efficient Thai Sign Language Recognition System Using CNN-Based Transfer Learning Models

This paper presents an advanced Thai Sign Language (TSL) recognition system using CNN-based transfer learning models to overcome communication challenges for the deaf community in Thailand. Leveraging deep learning techniques, we fine-tuned five high-performing CNN models—InceptionV3, VGG-19, MobileNetV2, DenseNet201, and EfficientNetB7—to accurately recognize TSL gestures. Transfer learning was applied to improve model efficiency, effectively reducing computational requirements while achieving high levels of accuracy. Among the models evaluated, EfficientNetB7 demonstrated superior performance, achieving an accuracy of 95.68% and an F1-score of 95.68%. This high performance establishes EfficientNetB7 as the optimal model for TSL gesture recognition, capable of accurately discerning subtle variations in hand gestures. These findings underscore the potential of CNN-based transfer learning in developing reliable and scalable TSL recognition systems, which can significantly improve accessibility, reduce communication barriers, and enhance social participation for individuals with hearing impairments. Future work will explore expanding gesture recognition capabilities and enhancing the model’s adaptability to diverse real-world scenarios.