Haze Removal Via Edge Weighted Pixel-To-Patch Fusion

Keyan Wang,Shanshan Zhang,Yunsong Li

DOI: https://doi.org/10.1007/s11036-017-0869-y

2017-01-01

Mobile Networks and Applications

Abstract:One of the key issues in image dehazing is how to accurately estimate the transmission map using strong priors or assumptions. By far the most common prior adopted by existing haze removal approaches is dark channel prior (DCP). Despite the remarkable progress, the existing DCP-based methods induce misestimation of the transmission while preventing halo artifacts, and may cause distortions of the recovered haze-free images in both chromaticity and contrast. This because they fail to observe the fact that the edges in a color image do not necessarily correspond to the changes in depth; instead, they try to preserve all types of edges indiscriminately in the estimated transmission map, which does not accord with the property of a realistic transmission map. In order to address this issue, we propose an efficient method for haze removal under the guidance of the depth edges. The main contribution of our study is that we present a depth edge prior to obtain the depth edges from the hazy image, and then employ a pixel-to-patch fusion scheme weighted by the depth edges to estimate the transmission directly, which can preserve the sharp discontinuity at depth edges but smooth the surface texture in the rest regions of the transmission map. The experimental results show that our approach can obtain more accurate estimation of the transmission, and consequently restore better quality haze-free images in terms of less color distortion and high contrast.