DehazeGAN: Underwater Haze Image Restoration using Unpaired Image-to-image Translation

Abstract

In this paper, we propose a Generative Adversarial Networks (GAN)-based image restoration method. Our method adopts an unpaired image-to-image translation network to learn the characteristics of underwater haze images. To enhance restoration, we propose multiple cyclic consistency losses that capture the detail of images and suppress distortion image translation. To prepare unpaired images of clean and degraded scenes, we collected images from Flickr and filter out false images using image characteristics. The proposed network is tested on public underwater images and shows promising results under severe image distortion.