Unsupervised Haze Removal for Aerial Imagery Based on Asymmetric Contrastive CycleGAN

Aerial image dehazing is an important preprocessing step, since haze extremely degrades the imaging quality and affects subsequent applications of aerial imagery. Most current removal methods achieve encouraging performance by relying on paired synthetic data, while are limited to their generality scalability in practical tasks. To this end, paper aims learn effective unsupervised model from unpaired set clear hazy images. Motivated great advantages contrastive learning representation field, we first attempt formulate a Asymmetric Contrastive CycleGAN framework (namely ACC-GAN) maximize mutual information between domain haze-free domain. In latent space, introduced constraint ensures that restored pulled closer pushed away image, so as indirectly regularize process. Importantly, different standard CycleGAN, develop additional feature transfer network into forward path form asymmetric structure ACC-GAN, which can enhance encoded features During training, multi-dimension loss terms jointly built committee for generating dehazed results with higher naturalness better fidelity. Experimental synthesis real-world datasets indicate our method superior existing approaches, also very competitive other related supervised models.