Self-Supervised Dehazing Network Using Physical Priors

Gwangjin Ju, Yeongcheol Choi, Donggun Lee, Jee Hyun Paik, Gyeongha Hwang, Seungyong Lee; Proceedings of the Asian Conference on Computer Vision (ACCV), 2022, pp. 3018-3033

In this paper, we propose a lightweight self-supervised dehazing network with the help of physical priors, called Self-Supervised Dehazing Network (SSDN). SSDN is a modified U-Net that estimates a clear image, transmission map, and atmospheric airlight out of the input hazy image based on the Atmospheric Scattering Model (ASM). It is trained in a self-supervised manner, utilizing recent self-supervised training methods and physical prior knowledge for obtaining realistic outputs. Thanks to the training objectives based on ASM, SSDN learns physically meaningful features. As a result, SSDN learns to estimate clear images that satisfy physical priors, instead of simply following data distribution, and it becomes generalized well over the data domain. With the self-supervision of SSDN, the dehazing performance can be easily finetuned with an additional dataset that can be built by simply collecting hazy images. Experimental results show that our proposed SSDN is lightweight and shows competitive dehazing performance with strong generalization capability over various data domains.