Data augmentation computing model based on generative adversarial network

Abstract

The edge intelligent computing technology can reduce the delay and energy consumption of deep learning model reasoning through the collaborative terminal acquisition equipment and edge server. We apply the neural network to the edge computing and build a data augmentation computing model based on the sparse data volume. Then, we get an intelligent generative image after the network training to achieve the effect of enhancement computing. In this paper, we choose a relatively small number of national elements and the generative adversarial network (GAN) as the experimental data calculation set and network model. First, we normalize the preprocessing of the collected data to form the initial sample data set. Second, the model extracts the feature vector by input image to the convolution neural network (CNN) layer. After that, we use a random noise vector z which follows a Gaussian distribution as the initial input of the conditional generative adversarial network (CGAN). The feature vector extracted from the image is taken as a label and a condition constraint of the CGAN to train the parameters of the CGAN. Finally, the trained CGAN model is used to complete the data augmentation computing. A total of 350 samples were collected, and 97 sample images were actually applied for data augmentation. The enhanced data set of this model is as many as 1,700 samples, and it is found that the generated image is of good quality by using this data set for the peak signal-to-noise ratio (PSNR) detection, which is of innovative value in the standard of real samples.