Image Scene Conversion Algorithm Based on Generative Adversarial Networks

Abstract

This paper presents an image scene conversion algorithm based on generative adversarial networks (GANs). First, the generator uses the generator network with cross-layer connection structure to realize the sharing of image structure information, so that the structure and edge of the generated image are consistent with the input image as far as possible. Secondly, the multi-scale global convolution network discriminator is used to determine different scales of image. Then, the combinational loss functions including GAN, L1, VGG and feature matching (FM) are designed. The network structure of the generator, the number of multi-scale discriminator and the weighted combination of multiple loss functions are evaluated and analyzed, and the optimized algorithm structure is given. Finally, through image fogging and day-to-night conversion experiment, the results show that the details of the converted image are more complete and the generated image is more realistic.