基于超像素的均值–均方差暗通道单幅图像去雾方法  

Superpixel-based Mean and Mean Square Deviation Dark Channel for Single Image Fog Removal

作　　者：汪云飞[1] 冯国强[1] 刘华伟[1] 赵搏欣 WANG Yun-Fei1, FENG Guo-Qiang1, LIU Hua-Wei1, ZHAO Bo-Xin1(1. Department of UAV Application Engineering, College of Aeronautics and Astronautics Engineering, Air Force Engineer- ing University, Xi＇an 71003)

机构地区：[1]空军工程大学航空航天工程学院无人机运用工程系,西安710038

出　　处：《自动化学报》2018年第3期481-489,共9页Acta Automatica Sinica

基　　金：国家自然科学基金(61379104)资助~~

摘　　要：雾、霾天气会引起图像严重降质.本文假设局部雾浓度恒定,认为暗通道优先的有效性随景深递增呈指数衰减,其比例大小可间接反映雾浓度高低.在此基础上提出一种基于超像素的均值–均方差暗通道单幅图像去雾算法.首先通过超像素分割得到景深恒定的小区域,接着在每个区域内计算均值–均方差暗通道:用均值替代最小值抑制景深突变处的光晕效应,同时采用均方差对其进行修正纠正景深无限远处的偏色问题,由此生成的透射率在超像素内保持不变且更加精细、准确.实验结果表明该算法在雾浓度较大时能够显著提高大景深图像的可见性.Fog and haze can cause serious image degradation. In this paper the fog concentration is supposed to be locally constant, and the effectiveness of dark channel prior decays exponentially as the depth increases, with its proportion indirectly reflecting the level of fog density. Based on this, we propose an SMMD （superpixel-based mean-mean square deviation dark channel） algorithm for single image fog removald. Firstly, the small regions in which the depth is kept constant by superpixel segmentation are obtained. Then the mean-mean square deviation dark channel is computed in these regions. This dark channel can suppress the ＂halo effect＂ which occurs in depth mutation using mean instead of minimum. Moreover, the mean square deviation is employed to revise mean in order to rectify the cast problems which happen when depth is infinite. On the foundation of that the resulting transmittance keeps constant and is finer and more precise. Experimental results show that the algorithm can effectively improve the visibility images of large depth in field when fog concentration is big.

关 键 词：图像去雾 大气散射模型 超像素 均值 均方差 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]