基于图像相关性和结构信息的无参考图像质量评价

No-reference image quality assessment based on image correlation and structure information

摘要:

分析了非下采样Contourlet变换(NSCT)方向子带亲 戚系数 间和父子系数间的强相关性及其包含的结构信息,并基于图像发生失真会 影响这些系数间的强相关性和结构信息的假设,提出了一种新的无参考图像质量评价(NR-I QA)方法。首先,分别计算 NSCT方向子带亲戚系数间和父子系数间的互信息(MI),以此作为描述这些系数 间相关性的统计特征；其次, 分别计算NSCT方向子带亲戚系数间和父子系数间的结构相似度(SSIM),以此作为描述 图像结构信息的统计特征； 进而,结合这些系数间的MI和SSIM等统计特征,构造了相应的NR-I QA模型和图像 失真类型识别模型；最后,在LIVE及LIVE Multiply Distorted图像质量评价数据库上进行 了大量的实验仿 真。结果表明,本文评价模型的评价结果与人类主观评价具有非常高的相关性,LIVE图 像质量评价数据 库上的斯皮尔曼等级相关系数和皮尔逊线性相关系数均在0.931以上 。无论总体评价效果还是各失真类型 评价效果,与当今主流评价算法相比非常具有竞争性；而且,失真类型识别模型的识别精度 可以达到92.31%,明显高于这些主流算法。

Abstract:

The nonsubsampled contourlet transform (NSCT),which is fully shift-i nvariant,is a multi-scale,local, multi-direction and overcomplete image representation.In this paper,we analyze the strong correlation and structure information of cousins′ coefficients and father-son coefficients respectively,and put forward a no-reference image quality assessment (NR-IQA) method based on the hypothesis that the distorted image can affect the strong correlation and the structural information between these coefficients .Firstly,the mutual information of cousins′ coefficients and father-son coefficients is calculated respectivel y and used to describe the statistical characteristics of the coefficients correlation.Secondly,structure similarity (S SIM) of those subband coefficients is calculated and used to describe the statistics characteristic of im age structure information.Then,a no-reference image quality assessment model and an image distortion type recogn ition model are constructed by combining with the mutual information statistical characteristics and structu ral similarity statistical characteristics.Finally,a large number of simulation experiments are carried o ut in the LIVE and LIVE Multiply Distorted image quality evaluation databases.The simulation results show that th is method is suitable for many common image distortion types and correlates well with the human judgments of im age quality,and the Spearman′s rank ordered correlation coefficient (SROCC) and the pearson′s li near correlation coefficient (PLCC) in LIVE image quality evaluation database are more than 0.931.The overal l and every distortion type evaluation of this assessment model have highly competitive performance than oth er state-of-the-art NR-IQA algorithms,and the recognition accuracy of the model is up to 92.31%,significa ntly higher than that of other methods.