Dynamic Sign Language Recognition Based on Improved R(2+1)D Algorithm

Previous sign language recognition networks faced problems with model computation and training time high requirements. To address these problems, a sign language recognition method based on 3D convolution and (2+1)D convolution is proposed. First, image enhancement is applied to the dataset based on wavelet transform to speed up the convergence of the model. Then, a dynamic sign language recognition network based on Spatio-temporal convolutional network R(3+2+1)D is designed. The residual module of the R(2+1)D network model is improved using the idea of pre-activation structure, and the original down-sampling residual unit is reconstructed using a pooling operation with step size 2 and 3D convolution with size 1 to reduce the model size and improve the recognition rate. Finally, comparison experiments are conducted on the Chinese Sign Language dataset (CSL). The experimental results show that compared with the original network, the recognition speed of this method can be improved by 68.1% and the recognition accuracy by 1.5%, while the memory occupation is reduced by 52.4%.