Variational Regularized Single Image Dehazing

Renjie He, Jiaqi Yang, Xintao Guo, Zhongke Shi

Abstract: Dehazing from a single hazy image has been a crucial task for both computer vision and computational photography applications. While various methods have been proposed in the past decades, estimation of the transmission map remains a challenging problem due to its ill-posed nature. Moreover, because of the inevitable noise generated during imaging process, visual artifacts could be extremely amplified in the recovered scene radiance in densely hazy regions. In this paper, a novel variational regularized single image dehazing (VRD) approach is proposed to accurately estimate the transmission map and suppress artifacts in the recovered haze-free image. Firstly, an initial transmission is coarsely estimated based on a modified haze-line model. After that, in order to preserve the local smoothness property and depth discontinuities in the transmission map, a novel non-local Total Generalized Variation regularization is introduced to refine the initial transmission. Finally, a transmission weighted non-local regularized optimization is proposed to recover a noise suppressed and texture preserved scene radiance. Compared with the state-of-the-art dehazing methods, both quantitative and qualitative experimental results on both real-world and synthetic hazy image datasets demonstrate that the proposed VRD method is capable of obtaining an accurate transmission map and a visually plausible dehazed image.