基于支持向量回归的无参考模糊和噪声图像质量评价方法  

No-reference quality assessment algorithm for blur and noise images using support vector regression

作　　者：桑庆兵[1] 梁狄林 吴小俊[1] 李朝锋[1] 

机构地区：[1]江南大学物联网工程学院计算机系,江苏无锡214122

出　　处：《光电子．激光》2014年第3期595-601,共7页Journal of Optoelectronics·Laser

基　　金：国家自然科学基金(61170120;60973094);国家自然科学基金(61103128);江苏省自然科学基金(BK2011147)资助项目

摘　　要：基于支持向量回归(SVR)和图像奇异值分解,提出了一种新的无参考(NR,no-reference)模糊和噪声图像质量评价(IQA)方法。首先通过对待评价图像进行高斯低通滤波生成再模糊图像,然后分别对它们进行奇异值分解并计算奇异值的改变量,最后使用奇异值的改变量作为SVR的输入,训练预并测得到图像的质量评分。在3个公开的模糊和噪声数据库上的实验结果表明,新方法预测得分与主观得分有较好的一致性,获得了较好的评价指标;对于模糊失真类型和噪声失真类型,在LIVE2数据库上的性能评价指标斯皮尔曼等级相关系数(SROCC)分别达到0.961 3和0.965 9。In this paper,we propose a new no-reference image quality assessment （IQA） algorithm ior blur and noise images using support vector regression （SVR） and singular value decomposition. The al- gorithm is composed of three steps. First, a re-blurred reference image is produced by using Gaussian low-pass filter for a test image. Then we do singular value decomposition to them and calculate the change of their singular values. Thirdly, we train the support vector regression by using change of singu- lar values and predict image quality score. Experimental results on three open blur and noise databases show that the proposed algorithm is more reasonable and stable than other methods. It has high correla- tion with human judgments and obtains a better evaluation index. So the proposed method is appropriate for no-reference blurred and noise image quality assessment. For the blur and noise distortion types,the performance indices of Spearman rank correlation coefficient （SROCC） on the LIVE2 datahase can reach 0. 9613 and 0. 9659 ,respeetively.

关 键 词：图像质量评价(IQA) 无参考 奇异值分解 支持向量回归(SVR) 高斯低通滤波 

分 类 号：TN919.8[电子电信—通信与信息系统]