Remote Sensing Images Dehazing Algorithm based on Cascade Generative Adversarial Networks

Abstract

The existing remote sensing image dehazing methods based on deep learning networks usually use pairs of clear images and corresponding haze images to train the model. However, pairs of clear images and their haze counterparts are extremely lacking, and synthetically haze images could not accurately simulate the real haze generation process in real-world scenarios. To address this problem, a cascade method combining two GANs (generative adversarial networks) is proposed. It contains a learning-to-haze GAN (UGAN) and learning-to-dehaze GAN (PAGAN). UGAN learns how to haze remote sensing images with unpaired clear and haze images sets, and then guides the PAGAN to learn how to correctly dehaze such images. To reduce the discrepancy between real haze and synthetic haze images, we added self-attention mechanism to PAGAN. The details can be generated using cues from all feature locations. Moreover, the discriminator could check that highly detailed features in distant portions of the images that are consistent with each other. Compared with other dehazing methods, this algorithm does not require numerous pairs of images to train the network repeatedly. And the results show that the cascaded generative adversarial networks has visual and quantitative effectiveness for the removal of haze, thin clouds.