Twin Residual Network for Sign Language Recognition from Video

This study investigates transfer learning features in video-based sign language recognition, assessing various pretrained models and introducing a proposed Twin Residual Net. Notably, the proposed model outperforms ResNet-18, ResNet34, ResNet-50, ResNet-101, ResNet-151, and ResNet-201, showcasing superior capabilities in capturing object-oriented features and temporal dynamics crucial for precise sign language interpretation, attaining an average BLEU score of 0.46. The analysis emphasizes the pivotal role of transfer learning in enhancing models' proficiency by leveraging representations from extensive datasets, crucial for understanding the temporal intricacies in sign language expressions. Effective motion vector features, especially concerning object type and movement, are highlighted. The study guides practitioners and researchers in selecting models, positioning the proposed model as promising for sign language recognition tasks. Future research directions may explore fine-tuning or hybrid approaches, advancing synergy between models and sign language recognition challenges, contributing to broader advancements in computer vision and pattern recognition beyond sign language applications.