Abstract: Convolution neural network itself has a rich ability of expressing features and learning, but in essence, the module geometric transformation ability is fixed. Therefore, the VGG16 network structure is improved by introducing a deformable convolution kernel, and a convolution neural network structure named DCVGG is built to study the gesture recognition. In different data sets, the gesture recognition method based on deformable convolution neural network can input RGB image data directly into the network. The results show that the average recognition rate of gestures is over 97%, which can improve the performance of the network, enhance the tolerance and diversity of the convolution neural network to the sample object, and enrich the expression ability of the convolution neural network. Compared with the traditional LeNet5, VGG16 structure and traditional feature extraction by hand, DCVGG is deeper than the traditional structure, the robustness is better, the recognition rate is stronger, which can provide reference for the effective recognition of gestures in complex background, and has some extension ability.

Key words: gesture recognition, deformable convolution, convolution neural network (CNN), convolution kernel, bilinear interpolation