Robust regularization for single image dehazing

Abstract

Generally, conventional image dehazing methods attempt to improve the initial transmission map by using guidance information from the input image as a structural prior. However, these methods do not consider the structural differences between the guidance and the initial transmission map. Furthermore, these methods provide limited robustness against outliers. This paper tackles these problems by optimizing a nonconvex energy function that jointly leverages structural information from the transmission map and guidance. Specifically, a nonconvex regularizer, and correspondingly, an efficient energy function are proposed. The energy function is solved using a majorize-minimize algorithm. Consequently, an improved transmission map is obtained in which the edges are well-preserved. Ultimately, this improved transmission map provides a high-quality haze-free image that has vivid colors and fine details. Experiments were performed on various synthetic and real hazy image datasets. The performance of the proposed method is compared with the state-of-the-art methods by using the root mean square error (RMSE), structural similarity index (SSIM), fog aware density evaluator (FADE), and visibility level descriptor (VLD). Both the qualitative and quantitative results demonstrate the effectiveness of the proposed robust regularization.