基于高斯分布的大气光估计算法

摘要

This paper proposed an estimation algorithm of atmospheric light based on Gaussian distribution to solve the problem that the existing approaches estimate the atmospheric light with few candidate points and it may lead to significant statistical error in defogging results.The proposed algorithm firstly uses a brightness threshold to select the candidate points to increase the number of initial sample points.Then it uses some clustering algorithms to merge the obtained clusters to increase the number of sample points in the clusters.It also uses a proportional threshold to filter out unreasonable point clusters and regards each point cluster as a single light source,and then calculates their influence on surrounding pixels with a Gaussian distribution based model.Finally,atmospheric light map instead of aconstant value is used to restore the haze image.Experiments show that the defogging results produced by the proposed algorithm are more natural than other defogg algorithms in subjective vision.The image quality evaluation indicators also reveal the fact that the proposed algorithm improves the defoging effect efficiently in objective aspect compared with other algorithms.%针对现有去雾算法在估计大气光向量时,所采用的方法包含的大气光候选点数量较少,导致估计结果在统计意义上误差较大这一问题,提出了基于高斯分布的大气光估计算法.该算法首先使用阈值划分的方式选取候选点以增加初始样本点数量;然后引入聚类算法对原算法所得光源点簇进行合并以提高单个点簇所含样本点个数;同时,使用比例阈值过滤掉不合理的点簇,并将各点簇视为单独光源,单独计算其对周围像素的影响,其影响通过二维高斯分布函数进行建模;最后使用大气光图取代全局大气光复原图像.实验结果表明,相对于原算法,使用高斯分布大气光图复原的结果在主观视觉上看起来更加自然,且其客观图像质量评价指标也得到了提高.