Automatic recognition of dynamic signs of Mexican sign language using deep learning

Introduction: Over four million individuals in Mexico face communication barriers due to hearing impairments. Sign language serves as an essential communication tool within the deaf community; however, automatic translation between sign and oral languages remains a significant challenge. This study proposes an approach for recognizing dynamic gestures from Mexican Sign Language (LSM) to support the development of assistive communication technologies. Methods: In collaboration with expert interpreters, an LSM corpus comprising 121 signs was developed, including a specialized lexicon focused on medical emergencies and accident scenarios. A standardized video acquisition protocol was implemented with both expert and non-expert participants. The proposed methodology consists of skeletal keypoint extraction using MediaPipe, data augmentation through frame sampling, and dataset normalization. Multiple deep learning architectures were evaluated, including ResNet, Simple RNN, LSTM, Bidirectional LSTM (BiLSTM), Gated Recurrent Units (GRU), a Transformer encoder, and a hybrid ResNet-Transformer model. Results: Among the evaluated models, the ResNet architecture achieved the best performance, obtaining an F1-score of 0.948 under subject-independent evaluation, with an average inference time of 0.468 seconds. Hyperparameter optimization analysis indicated that performance improvements were primarily driven by training dynamics and regularization strategies rather than increases in architectural depth. Discussion: The results demonstrate the effectiveness of deep learning-based approaches for dynamic LSM gesture recognition and highlight the importance of optimization strategies for robust generalization. This work contributes toward LSM-to-Spanish translation systems and provides a foundation for advancing data-driven sign language recognition technologies.