Non-homogeneous Image Dehazing with Edge Attention Based on Relative Haze Density

Ruting Deng, Zhan Li, Yifan Deng, Hang Long, Zhanglu Chen, Zhiqing Kang, Zhichao Qiu

Abstract: Image dehazing is a widely used technology for recovering clear images from hazy inputs. However, most dehazing methods are designed to target a specific haze concentration, without considering the varying degrees of image degradation. Removing non-homogeneous haze from real-world images is challenging. To address this issue, this study proposes a dual-cycle framework based on relative haze density, in which inputs are regarded as both hazy images to be recovered by a restoration network (RNet) and clear images to be deteriorated by a degradation network (DNet). Edge attention blocks and multi-order derivative loss are proposed for RNet to enhance the details and colors. Furthermore, two multi-class discriminators are designed to distinguish between relative levels of haze density. Extensive experiments on both real-world and synthetic datasets demonstrate that the proposed method is superior to state-of-the-art approaches for non-homogeneous image dehazing using either supervised or unsupervised learning. This code is available at https://github.com/lizhangray/EARHD.