Multiple adverse weather removal using adversarial and contrastive learning

Adverse weather removal is significant and valuable for real-world computer vision systems. Most existing algorithms focus on removing one specific type of weather and rely on paired synthesized weather datasets for training. In this paper, to train a multiple weather removal model with unpaired datasets, we introduce multi-dimension contrastive learning mechanism into generative adversarial networks. Specifically, our method consists of two blocks: multi-channel cycle translation (MCCT) block and multi-dimension contrastive learning (MDCL) block. MCCT is used to preserve the color and structural information and gain informative supervisory signals. MDCL is proposed to discriminate feature distribution among different image domains. Results of extensive experiments show that our method is superior over existing unsupervised methods and maintains a decent performance degradation when adapted to multiple weather removal tasks.