GAN-Based Day and Night Image Cross-Domain Conversion Research and Application

With the development and application of deep learning in computer vision, the performance of many basic visual tasks such as object detection and semantic segmentation has been greatly improved. However, most of networks are based on standard illumination, which results in poor performance in low illumination scenarios, and it is difficult to collect datasets with different illumination levels in restricted scenes. In this paper, GAN and related derived networks are systematically studied and summarized, and based on the idea of generation-antagonism of GAN, the design of day-night cross-domain converter is completed on the basis of the structure of CycleGAN. Based on this, Inception layer is added to optimize the structure of the converter, and the performance of the day-night cross-domain converters before and after optimization are compared through experiments. The results show that the optimized day-night converter can make the converted image more realistic. It is of great significance for enhancing the quality of datasets in restricted scenes, improving the performance of object detection and segmentation models in low illumination scenes.