基于先验知识的图像去雾算法

发布时间：2018-11-19 07:17

【摘要】：雾霾天气下悬浮在空气中的大气颗粒物对光线的传播产生了不利影响,使得成像设备获取的图像能见度低、对比度下降,给图像分割、目标跟踪、行为检测等后续计算机视觉任务造成了极大不便,直接影响到现有户外成像系统(如安防监控系统等)的正常工作,给人们的生活造成了巨大的安全隐患。因此,研究如何提高雾天降质图像复原结果质量、降低雾霾天气对现有户外成像系统的不利影响有着十分重要的现实意义。本文从雾霾天气的特点出发,详细分析了大气颗粒物对图像成像过程的影响及雾天降质图像的退化过程。通过分析研究现有的图像去雾算法,发现其中存在的不足及可改进之处。本文首先介绍了大气光散射模型,并通过理论推导证明了其中存在的不足并提出了有效的改进方法,同时还针对现有去雾算法在获取大气光值时存在误差过大这一问题进行了完善与改进,做出了一些有意义的实际工作。概括而言,本论文的主要工作及创新主要集中在如下几个方面:1)针对暗原色先验去雾算法结果色彩失真的问题,提出了一种对各颜色通道分别计算透射率的改进方法。算法利用了大气介质对各颜色可见光透射率不同的先验知识。首先依据比尔郎伯定律分析入射光频率对各颜色通道透射率的影响,推导出各通道透射率之间的比例关系,然后采用先对图像进行降采样预处理获取细化透射率之后再恢复原尺寸的方法提高算法运行效率,最后通过比例关系获取所有颜色通道上的透射率,并在各通道上分别使用对应的透射率恢复图像。实验结果表明,改进后的图像去雾算法与现有算法的结果相比,去雾结果图像色彩更加自然,消除了现有算法色彩饱和度偏高的缺点,且算法运算效率大幅提高。2)现有去雾算法在估算大气光时过于粗略导致估算结果误差较大,造成图像复原结果常常存在色彩失真的问题。针对这一问题,本文提出了一种基于聚类统计的大气光估算方法,主要利用了光源点处大气光样本点分布更加密集的先验知识。算法首先在原图中选取部分可能的大气光源点,通过无监督聚类算法对这部分大气光源点进行聚类,聚类出若干个备选大气光源点点簇,再对各点簇中所含样本点个数进行统计,选取包含点数最多的点簇求解大气光。使用点簇中各大气光样本点的亮度均值向量作为大气光的估算值,同时以各点簇的几何中心作为大气光所处位置。实验结果表明,统计聚类方式估计的大气光亮度向量和光源位置都更加准确,这使得图像复原结果在主观视觉上看起来更加自然,同时也较大地提升了各类图像质量客观评价指标。3)现有方法采用固定数量的大气光样本点进行聚类,并以包含候选点最多的点簇统计估算大气光值。由于样本点较少,导致估算的大气光值在统计意义上误差较大。为了解决这一问题,本文采用阈值划分的方式选取大气光样本点,以此提高大气光样本点数量,同时采用蚁群算法聚类大气光点簇,提高大气光值估算结果的准确度。为了提高算法的计算效率,本文先使用K均值算法对大气光候选点进行初步聚类,再使用蚁群算法改良聚类结果。实验结果证明,使用该算法估算的大气光值能使去雾结果看起来更加自然,且能进一步改善去雾结果的图像质量评价指标。4)现有去雾算法通常假定大气光值全局恒定,而实际场景中各区域的大气光值分布不均,利用这一先验知识,本文提出了基于高斯分布的大气光估计算法。算法使用阈值划分的方式选取候选点以增加初始样本点数量,同时引入聚类算法对原算法所得光源点点簇进行合并以提高单个点簇所含样本点个数。使用比例阈值过滤掉不合理的点簇,同时将各点簇视为单独光源单独计算其对周围像素的影响,并通过二维高斯分布函数对此进行建模,最后生成位置相关的大气光图代替全局大气光。实验结果表明,使用高斯分布大气光图复原的结果在主观视觉上相对于原算法看起来更加自然,且在图像质量评价指标上也得到改善。
[Abstract]:The air particles suspended in the air in the haze weather have adverse effects on the propagation of light, so that the image visibility acquired by the imaging device is low, the contrast is reduced, and the following computer visual tasks such as image segmentation, target tracking, behavior detection and the like are greatly inconvenient, the normal operation of an existing outdoor imaging system (such as a security monitoring system and the like) is directly affected, and a great potential safety hazard is caused to people's life. Therefore, it is of great practical significance to study how to improve the quality of the image restoration of the fog and reduce the adverse effect of the haze weather on the existing outdoor imaging system. In this paper, the effects of the atmospheric particulate matter on the image forming process and the degradation process of the fog-based image are analyzed in detail from the characteristics of the haze weather. By analyzing the existing image de-fog algorithm, it is found that the existing image de-fog algorithm can be improved. In this paper, the atmospheric light scattering model is introduced, and the shortcomings in the model are proved by the theory, and the effective improvement method is put forward. At the same time, the problem of the existing de-fog algorithm in obtaining the atmospheric light value is improved and improved. Some meaningful work has been made. In general, the main work and innovation of this thesis are mainly focused on the following aspects: 1) For the problem of color distortion of the dark primary-color prior de-fog algorithm, an improved method for calculating the transmittance for each color channel is proposed. the algorithm utilizes the prior knowledge that the atmospheric medium has different visible light transmittance for each color. firstly, the influence of the incident light frequency on the transmittance of each color channel is analyzed according to the Biberman's law, the proportion relation between the transmittance of each channel is deduced, and finally, the transmittance on all color channels is obtained through the proportional relation, and corresponding transmittance recovery images are respectively used on each channel. The experimental results show that the improved image defogging algorithm is more natural than the results of the existing algorithm, and the disadvantage of high color saturation of the existing algorithm is eliminated, and 2) the existing defogging algorithm is too rough to estimate the atmospheric light, and the error of the estimation result is large, so that the image restoration result is often a problem of color distortion. In order to solve this problem, a method for estimating the atmospheric light based on the statistics of the light source is proposed, which mainly uses the prior knowledge of the distribution of the atmospheric light sample at the light source point. The method comprises the following steps of: firstly, selecting a part of the possible atmospheric light source points in the original drawing, carrying out the clustering on the part of the atmospheric light source point by a non-supervised clustering algorithm, clustering a plurality of alternative atmospheric light source spots, and counting the number of sample points contained in each cluster, the cluster of points with the largest number of points is selected to solve the atmospheric light. the luminance average vector of each atmospheric light sample point in the cluster is used as the estimated value of the atmospheric light, and the geometric center of each cluster is used as the position of the atmospheric light. The experimental results show that the light intensity vector and the position of the light source are more accurate, which makes the image restoration appear more natural in the subjective vision. in that prior method, a fixed numb of atmospheric light sample points are adopt for clustering, and the atmospheric light value is estimated at the most point cluster containing the candidate points. As the sample point is small, the estimated atmospheric light value is in a statistically significant error. In order to solve this problem, the method of threshold division is used to select the sample point of the atmospheric light, so as to improve the number of the sample points of the atmospheric light, and the clustering of the atmospheric light spot by the ant colony algorithm is used to improve the accuracy of the estimation result of the atmospheric light value. In order to improve the calculation efficiency of the algorithm, the paper first uses the K-means algorithm to carry out the preliminary clustering of the atmospheric light candidate points, and then uses the ant colony algorithm to improve the clustering result. the experimental results show that the atmospheric light value estimated by the algorithm can make the defogging result look more natural and can further improve the image quality evaluation index of the defogging result. In the actual scene, the distribution of the atmospheric light in each area is not uniform, and this prior knowledge is used, and an atmospheric light estimation algorithm based on the Gaussian distribution is proposed in this paper. the algorithm selects the candidate points in a way of threshold division to increase the number of initial sample points, and simultaneously introduces a clustering algorithm to combine the light source spot clusters obtained by the original algorithm so as to improve the number of sample points contained in a single dot cluster. Unreasonable point clusters are filtered out by using a proportional threshold, and each cluster is considered as a single light source to separately calculate the effect on the surrounding pixels, and the two-dimensional Gaussian distribution function is used for modeling the cluster, and finally, a position-related atmospheric light diagram is generated in place of the global atmospheric light. The experimental results show that the result of using the Gaussian distribution atmospheric photo-graph is more natural in the subjective vision than the original algorithm, and is also improved in the image quality evaluation index.
【学位授予单位】：电子科技大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】