Image haze removal by adaptive cycleGAN

Abstract

We introduce our machine-learning method to remove the fog and haze in image. Our model is based on CycleGAN, an ingenious image-to-image translation model, which can be applied to de-hazing task. The datasets that we used for training and testing are creatd according to the atmospheric scattering model. With the change of the adversarial loss from cross-entropy loss to hinge loss, and the change of the reconstruction loss from MAE loss to perceptual loss, we improve the performance measure of SSIM value from 0.828 to 0.841 on the NYU dataset. With the Middlebury stereo datasets, we achieve an SSIM value of 0.811, which is significantly better than the baseline CycleGAN model. © 2019 IEEE.