Efficient American Sign Language Recognition Using Resnet-50

This study presents a robust and efficient American Sign Language (ASL) recognition system utilizing the ResNet50 architecture with transfer learning. The system is designed to accurately classify 26 static ASL alphabet gestures while tackling challenges such as hand shape variations, varying lighting conditions, and a limited dataset size. To enhance performance and reduce overfitting, data augmentation methods including rotation and horizontal flipping were applied. By leveraging transfer learning from a pre-trained ResNet-50 model, the proposed approach benefits from shorter training times and reduced computational demands without compromising accuracy. Experimental results show the system achieves an overall accuracy of 98.34 %, along with balanced precision, recall, and F1-scores across all gesture classes. Additionally, early stopping mechanisms were implemented to prevent overfitting and ensure generalization to unseen data. This work contributes to the advancement of deep learning-based models for ASL recognition and supports the development of inclusive communication technologies that foster seamless interaction between hearing and non-hearing communities.