Chinese sign language recognition and translation with virtual digital human dataset

Abstract

Sign language recognition and translation are crucial for communication among individuals who are deaf or mute. Deep learning methods have advanced sign language tasks, surpassing traditional techniques in accuracy through autonomous data learning. However, the scarcity of annotated sign language datasets limits the potential of these methods in practical applications. To address this, we propose using digital twin technology to build a virtual human system at the word level, which can automatically generate sign language sentences, eliminating human input, and creating numerous sign language data pairs for efficient virtual-to-real transfer. To enhance the generalization of virtual sign language data and mitigate the bias between virtual and real data, we designed novel embedding representations and augmentation methods based on skeletal information. We also established a multi-task learning framework and a pose attention module for sign language recognition and translation. Our experiments confirm the efficacy of our approach, yielding state-of-the-art results in recognition and translation.