Two-Phase Transmission Map Estimation for Robust Image Dehazing

Abstract

A robust two-phase transmission map estimation framework is proposed in this paper for single image dehazing. The proposed framework first estimates the coarse transmission map through the statistical assumption of dark channel prior (DCP). To refine the coarse transmission map, a novel image-gradient-guided high-order variational method is then proposed in the second phase. The resulting L1-regularized high-order nonsmooth optimization problem will be effectively solved using the primal-dual algorithm. Once the fine transmission map is accurately obtained, the final haze-free image could be restored based on the haze imaging model of Koschmieder. To further enhance dehazing performance, an improved tolerance mechanism is incorporated into the proposed method to suppress the undesirable artifacts usually produced by DCP in large sky regions. Numerous experiments on both synthetic and realistic images were performed to compare our proposed method with several state-of-the-art dehazing methods. Dehazing results have illustrated the superior performance of the proposed method.