Mini neural network based on knowledge distillation for dynamic gesture recognition in real scenes

Abstract

At present, gesture interaction has become one of the important ways of human-computer interaction, and the human hand, as the most flexible organ in the body, will make non-contact dynamic gesture interaction more convenient and more versatile. However, because of the diversity and uncertainty of gestures in time and space, and the complex deformable body of the human hands, it finally leads to the low recognition rate of dynamic gesture recognition in real scenes. Aiming at the difficulty of hand feature recognition and positioning, and the interference problem of unconstrained environment, this paper proposes a lightweight neural network based on distillation collaborative training for dynamic gesture recognition. This method uses knowledge distillation to train the deep network and the lightweight network together to improve the recognition performance of the lightweight neural network. At the same time, LSTM is used to improve the network's recognition rate of dynamic gestures of different durations in real scenes. In this paper, the lightweight dynamic gesture recognition network was tested on 20BN-Something-Something V2, and the accuracy rate was as high as 64.3%.