Optimized Deep Learning Models for Enhanced Sign Language Alphabet Recognition

Abstract

This study enhances communication between the hearing-impaired and hearing communities through the optimization of deep learning models for alphabet recognition in sign language. The research applies hyper-parameter optimization to convolutional neural networks (CNN) and convolutional long short-term memory networks (CLSTM) to enhance models’ accuracy and reduce recognition time. A dataset of 39,000 images spanning 26 alphabet classes was used for training and evaluation, while confusion matrix was used for performance evaluation. Results indicate that the optimized CLSTM model achieved a better performance with 96.50% accuracy at 41.04 seconds recognition time, compared to the traditional CNN model with 93.57% accuracy at 62.06 seconds recognition time. The findings revealed the potential of CLSTM’s sequential modeling in capturing spatial and temporal patterns in dynamic gestures. This advancement promotes inclusivity by bridging communication gaps, providing a robust foundation for deploying efficient sign language recognition (SLR) systems. The study emphasizes the significance of leveraging deep learning to foster accessibility, with implications for broader applications in healthcare, education, and beyond. Future research should explore multimodal input and extended optimization strategies for enhanced system adaptability and scalability.