Application of Generative Adversarial Networks in Gesture Recognition

Abstract

Aiming at the problem of insufficient accuracy of gesture recognition under the condition of small samples, a gesture generation method based on Generative Adversarial Networks is proposed to expand the dataset. For the expansion of gesture dataset, the idea of adversarial training in Generative Adversarial Networks is adopted, the discriminator model of deep convolution and the generator model of deep transpose convolution are designed respectively, the training process is optimized by using the way of adaptive learning rate, and the gesture is generated according to the gesture image in the gesture dataset created by the user. Then the accuracy is verified by using the real gesture image and the generated gesture image. With that, based on the complex characteristics of the algorithm for generating gesture image, it is proposed to directly generate the Fourier Descriptors of the image, so that the gesture has translation, scaling and rotation invariance, and their accuracy and training time are tested respectively. The experimental results show that comparing with generating gesture images, the training time of directly generating Fourier Descriptors is shorter and the recognition accuracy is higher.