Single Image Dehazing via Large Sky Region Segmentation and Multiscale Opening Dark Channel Model

Abstract

Outdoor images acquired under poor weather conditions are usually contaminated by suspended particles and aerosols in the atmosphere. These captured images easily suffer from contrast reduction, low visibility, and color distortion. In this paper, we develop a novel single image dehazing method based on large sky region segmentation and multiscale opening dark channel model (MODCM). First, a simple but effective method for large sky region detection based on SVM classification is presented, which can be considered as the first step of atmospheric light estimation. Then, two different strategies are utilized for obtaining a more accurate estimate of the atmospheric light according to the mentioned detection result. Furthermore, MODCM can adaptively make use of different patch sizes to calculate the dark channel according to different edge levels, which can prevent halo artifacts near edges of depth discontinuity. In addition, the gradient domain guided filter is adopted to refine the initial transmission map due to its accuracy near edges. Finally, the haze-free image can be obtained through correcting the colors of the sky region and combining the sky and non-sky region. Experimental results on different kinds of hazy images indicate that our proposed approach can produce the visually desirable results with genuine color and high scene visibility, even superior than the other state-of-the-art dehazing methods.