Hybrid\_Asl: Cross-Domain Transfer Learning for High-Accuracy American Sign Language Recognition

Communication is one of the issues that people with hearing impairments constantly have to struggle with; this is partly because advanced assistive devices are not widely available and used. These obstacles often hinder access to essential services and the ability to engage in meaningful communication. To address this gap, the current paper presents HybridASL, a novel architecture based on deep learning specifically designed to enhance the recognition of American Sign Language (ASL) hand gestures. HybridASL utilises the concept of cross-domain transfer learning, enabling the model to transfer visual knowledge gained from large-scale datasets to the problem-specific task of ASL classification. The model structure combines convolutional and transformer-based modules, thereby capturing both local spatial patterns and long-range contextual relationships. The training pipeline involves pre-training the model using a large number of images of ASL (n = 87,000), which are annotated with labels. The model is then trained through a fine-tuning process, allowing for optimisation of classification accuracy while preserving computational efficiency. The presented dual-phase training method can be applied to models, yielding a significant improvement in model operation, albeit with prohibitively large computational requirements. The results of the empirical analysis demonstrate that HybridASL achieves excellent performance, with an accuracy of 99.98% and comparable levels of precision, recall, and F1 Scores. Not only do these results outperform those of standard convolutional neural networks (e.g. ResNet, EfficientNet), but they also show that the integration of transfer learning with adaptive model refinement is effective. Ultimately, this study demonstrates the potential adaptability of HybridASL in the development of intelligent and scalable assistive technology. The model will enhance the accessibility of communication and the overall quality of life for the deaf and hard-of-hearing community, as it provides nearly flawless ASL recognition capabilities.