tsta

Home / سی و دومین کنفرانس ملی و دهمین کنفرانس بین المللی مهندسی زیست پزشکی ایران

Comparative Evaluation of Deep Learning Architectures for Static American Sign Language Recognition

Authors :

Shamim Najafi¹ Sedigheh Dehghani²

1- دانشگاه آزاد اسلامی واحد علوم و تحقیقات تهران 2- دانشگاه شهید بهشتی

Keywords :

Sign Language Recognition،Convolutional Neural Networks (CNN)،Long Short-Term Memory (LSTM)،Modified LeNet-5،Autoencoder Classifier Network (AEC)،Ensemble Strategies

Abstract :

Sign language is a vital form of non-verbal communication for individuals with hearing and speech impairments, yet its interpretation remains challenging. This study optimizes four deep learning models—Convolutional Neural Networks (CNN), Autoencoder Classifier Networks (AEC), modified LeNet-5, and Long Short-Term Memory (LSTM)—for static American Sign Language (ASL) recognition. Performance and generalization were enhanced using Batch Normalization, Dropout, Global Average Pooling, Learning Rate Scheduling, and alternative optimizers such as SGD and AdamW. Experimental results show that CNN-9 achieved the highest performance (99.05%), followed by AEC-7 (94.88%), LeNet-9 (91.03%), and LSTM (88.32%). These findings demonstrate the effectiveness of deep learning architectures for static ASL recognition and highlight the impact of architectural and training adjustments. Ensemble strategies further improved results: Soft Voting combining CNN-9 and AEC-7 achieved the best outcome, Weighted Soft Voting ensembles excluding LSTM-1 performed best, and Hard Voting with CNN-9, AEC-7, and LSTM-1 outperformed other configurations. Overall, ensemble methods enhance accuracy and robustness beyond individual models.