Contrastive Learning Deraining Network Based on Multi-Scale CycleGAN Architecture

In outdoor vision systems like autonomous driving, traffic monitoring, and drone imaging, rain can significantly degrade image quality and affect task accuracy. To address this, we propose a multi-scale CycleGAN-based contrastive learning deraining network. Using unpaired rainy and clear images, the network enhances deraining through generative adversarial networks and contrastive learning. Generator G converts rainy images into clear ones, while generator F simulates the degradation from clear to rainy images. Discriminators Dclear and Drain evaluate the authenticity of the generated images. Experiments demonstrate that the proposed method effectively removes rain streaks and improves image quality in complex scenes, offering a solution for single-image deraining