Real-Time American Sign Language Gesture Recognition Using Transfer Learning

The study presents a novel approach to improving American Sign Language (ASL) recognition accuracy, which is vital for the hearing-impaired. Utilizing transfer learning with the MobileNetV2 architecture and optimized training, the model achieves high accuracy. With a dataset of 274,612 RGB images and rigorous pre-processing, training was expedited on Kaggle’s Dual GPU T4 infrastructure. The proposed approach utilizes the categorical cross-entropy loss function and the Stochastic Gradient Descent optimizer. It leverages the pre-trained features of MobileNetV2, fine-tuning them specifically for ASL recognition. Remarkable results include 99.82% training accuracy, 99.14% validation accuracy, and fast real-time processing (99 milliseconds per step). This robust model promises broader application in education, assistive devices, and communication aids, fostering inclusivity for the hearing-impaired. Future research may explore dialect and gesture variations for enhanced accessibility.