Lightweight sign language intelligent recognition model based on improved R-C3D

Abstract

The study proposes a continuous dynamic sign language recognition model based on an improved regional 3D convolutional network. A 3D convolutional network is taken as a special extraction sub-network, and the depth separable convolution is introduced into the 3D convolutional network to reduce computational costs. The inverted residual results are taken to avoid information loss issues. In addition, the pre-selection box size of the optimized region 3D convolutional network is shortened, and the action judgment threshold is increased to improve the action accuracy. The average accuracy of the improved 3D convolutional network was 44.2 %, which was higher than that of other types of feature extraction sub-networks. After reducing the pre-selection box, the average accuracy of the time suggestion sub-network increased from 41.6 % to 44.5 %. The loss value also decreased from 0.5 to 0.46. After increasing the action judgment threshold from 0.5 to 0.7, the loss value decreased from 0.58 to 0.17. The loss value of the 3D convolutional network in the entire improved area was only 0.15, the sign language recognition speed was 183 ms, and the average accuracy was 44.6 %, which was better than those of other sign language recognition schemes. The above results indicate that the improved regional 3D convolutional network can accurately and quickly recognize continuous sign language actions.