Sign language recognition in the deep learning era: A comprehensive study of model performance, robustness and deployment considerations

This paper presents a dual-domain evaluation of classical and modern architectures for Sign Language Recognition (SLR) and Traffic Sign Classification (TSC), addressing critical challenges in accessibility and autonomous systems. We conduct a comprehensive assessment of 20 SLR models, spanning CNNs, hybrid CNN-LSTM pipelines, and transformer-based frameworks, evaluated on the Sign Language MNIST and ASL Fingerspelling datasets. Performance is measured across accuracy, computational efficiency, and robustness metrics.

For TSC, we benchmark 10 models—including lightweight CNNs, vision transformers, and object detectors—on the GTSRB, BelgiumTS, and TT100K datasets. The study examines classification and detection performance under varying noise conditions to assess real-world applicability. We analyze trade-offs between model complexity, inference speed, and deployment feasibility, providing guidelines for edge-optimized implementations.