DLSC-Net: Double Level Spatio-temporal Context-aware Network for Continuous Sign Language Recognition

Continuous sign language recognition based on the deep learning is a hot topic in computer vision. Existing methods struggle to accurately capture fine-grained local features and effectively model the long-term contextual information. These limitations often lead to suboptimal performance in continuous sign language recognition. To address these issues, we propose a novel network named DLSC-Net: Double Level Spatio-temporal Context-aware Network for Sign Language recognition. The network employs a double stream structure, consisting of the Time-sequence Learning Module and the Temporal Context-aware Module. The Time-sequence Learning Module leverages Convolutional Neural Networks (CNNs) to capture short-term local feature variations, ensuring precise recognition of detailed hand gestures. In parallel, the Temporal Context-aware Module employs dilated convolutions to handle long-term dependencies, expanding the receptive field without increasing computational load, thus effectively modeling the temporal dynamics in sign language sequences. We validated the effectiveness of the DLSC-Net on the large-scale continuous sign language sentence dataset. Our results show that DLSC-Net enhances recognition accuracy, making it a promising approach for continuous sign language recognition.