基于景深和稀疏编码的图像去雨算法  

An Image Rain Removal Algorithm Via Depth of Field and Sparse Coding

作　　者：肖进胜[1] 王文 邹文涛 童乐 雷俊锋[1] XIAO Jin-Sheng;WANG Wen;ZOU Wen-Tao;TONG Le;LEI Jun-Feng(School of Electronic Information, Wuhan University, Wuhan 430072)

机构地区：[1]武汉大学电子信息学院

出　　处：《计算机学报》2019年第9期2024-2034,共11页Chinese Journal of Computers

基　　金：国家重点研发计划项目(2017YFB1302401);国家自然科学基金(61471272)资助~~

摘　　要：降雨天气往往会导致室外监控视频质量下降,会使成像的图像产生畸变和模糊现象.为了改善雨天拍摄的图像的质量,该文提出了一种基于景深和稀疏编码的图像去雨算法.针对基于图像分解的去雨算法存在的低频成分中的雨痕残留和轮廓边缘丢失,以及高频部分的背景误判问题,该文利用联合双边滤波和短时傅里叶变换将图像进行分解,使得图像低频部分中的轮廓得到较好的保留,并引入景深改善低频成分中的雨痕残留和高频成分中与雨痕具有相同梯度的背景误判问题.该算法主要分为四个部分:图像分解、字典学习、基于主成分分析和支持向量机的原子聚类,景深修正.首先是利用图像分解提取出图像低频和高频成分,对于图像分解的方法,主要采用的是双边滤波和短时傅里叶变换相结合的方法,此方法对图像的轮廓和边缘保持度较高.接下来,对低频成分进行保留,对高频成分进行进一步处理.根据图像本身的纹理特性将高频成分进行分类,基于每一类再对高频成分进行分块处理,得到每一类图像的字典,从而进行字典学习.然后,利用主成分分析和支持向量机对字典进行分类,根据梯度信息分为含雨字典和非雨字典两类,应用正交匹配追踪获得基于新高频字典的稀疏系数,从而获得高频成分中非雨成分.最后,对于高频成分中和雨痕具有相同梯度的背景误判问题,通过景深,将图像高频按纹理和梯度方向进行二次分类,将高频成分中和雨条纹具有相同梯度的背景进行保留,有效提高分类准确性.同时,利用景深提取出含雨图像中的显著性特征来进一步去除低频成分中的残留雨痕.本文利用主观视觉效果以及客观指标对算法进行评估,实验结果证明主观效果得到明显的改善,客观指标也得到了提升,证明了该文基于景深和稀疏编码的图像去雨算法能够在去雨的同时较好地保�Rainfall weather often results in the deterioration of video quality, which can cause distortion of the image. In order to improve the quality of rainy images, we propose an image rain streaks removal algorithm based on the depth of field and sparse coding. Currently, the rain removal methods based on the image decomposition have attracted much attention for those methods require no restrictions on the types of rain. However, most of the decomposition-based rain removal methods are suffered from the problems of the rain residuals, the loss of contours and edges in low-frequency part, and the background mismatch in high frequency part. In this condition, we propose to decompose the image by using the combination of the bilateral filtering and the short-time Fourier transform, so that the contours of the low frequency part of the image is preserved well. Besides, we propose to use the depth of field saliency map of the image to remove the rain residues in the low-frequency components and solve the mismatching between the background and the rain streaks with the same gradient in the high-frequency components. The algorithm mainly includes four parts: image decomposition, dictionary learning, atomic clustering based on Principal Component Analysis and Support Vector Machine, image correction based on the depth of field. Firstly, the low-frequency and high-frequency components are extracted by image decomposition. We combine the bilateral filtering and the short-time Fourier transform. This method can better preserve the contours and textures. After that, the low-frequency part is retained, and the high frequency part is further processed. Secondly, the high frequency components are classified according to theirs texture. In each category, the high frequency components are further divided into blocks. Then the dictionaries for dictionary learning are constructed corresponding to each category. Thirdly, the dictionaries are classified into two categories: rainy dictionaries and non-rain dictionaries, by using Principal

关 键 词：图像去雨 景深 稀疏编码 主成分分析 双边滤波 短时傅里叶变换 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]