Abstract
Sign language is an important communication tool for deaf and hearing-impaired people. The study of sign language recognition can not only promote the communication between deaf-mutes and normal people, but also push the development of intelligent human-computer interaction. Sign language recognition based on deep learning has advantages in processing large scale dataset. Most of them use 3D convolution, which is not conducive to optimization. In this paper, an improved (2+1)D-ResNet model is proposed for isolated word recognition. The model convolves the video frame sequence in space and time dimensions and optimizes the parameters respectively. Based on CELU activation function, the accuracy of sign language recognition is improved effectively. The validity of proposed algorithm is verified on CSL dataset..
https://ror.org/05vr1c885 
