Application of an Improved DCGAN for Image Generation

Abstract

With the rapid development of deep learning, image generation technology has become one of the current hot research areas. A deep convolutional generative adversarial network (DCGAN) can better adapt to complex image distributions than other methods. In this paper, based on a traditional generative adversarial networks (GANs) image generation model, first, the fully connected layer of the DCGAN is further improved. To solve the problem of gradient disappearance in GANs, the activation functions of all layers of the discriminator are LeakyReLU functions, the output layer of the generator uses the Tanh activation function, and the other layers use ReLU. Second, the improved DCGAN model is verified on the MNIST dataset, and simple initial fraction (ISs) and complex initial fraction (ISc) indexes are established from the two aspects of image quality and image generation diversity, respectively. Finally, through a comparison of the two groups of experiments, it is found that the quality of images generated by the DCGAN model constructed in this paper is 2.02 times higher than that of the GANs model, and the diversity of the images generated by the DCGAN is 1.55 times higher than that of GANs. The results show that the improved DCGAN model can solve the problem of low-quality images being generated by the GANs and achieve good results.