Single-image dehazing via depth-guided deep retinex decomposition

In this paper, we explore the problem of single-image haze removal based on retinex model and new deep retinex decomposition architecture. Reformulating dehazing as reverse retinex, we propose a depth-guided retinex decomposition network, which consists of the Decom-Net, with two-branches for the retinex decomposition on the reversed hazy image, and the Guide-Net, with depth information for guiding the estimation of ideal illumination. To promote the accuracy of retinex decomposition, we develop an effective boosted decoder with a fusion attention mechanism to optimize the illumination iteratively, giving rise to a refined reflectance. Additionally, due to the reversible relationship between haze and low-light images, our network could effectively realize dehazing in nighttime. Through sets of experiments on a variety of synthetic and natural images, we validate the effectiveness of the proposed model in haze removal, competitively in terms of visual appearance and metrics, considering both daytime and nighttime cases.