图像质量客观评价的复数矩阵结构相似度方法

Objective image quality assessment based on complex matrix structure similarity

客观评价数字图像质量的目的是为了得到与人眼视觉特性一致的评价结果,因此,图像结构的分析和比较要考虑图像中人眼敏感程度不同的各种因素。针对单一图像特征描述图像结构信息过于片面的问题,引入了图像结构信息表示的复数方法,构造了用于描述图像结构信息的复数矩阵。把复数作为一种信息合并方法,将图像中人眼敏感程度较高的局部方差分量和代表一般敏感信息的图像像素的灰度值分布作为复数的实部和虚部。研究了非实数域的结构相似度方法以及数学模型,将该方法用于度量两图像复数矩阵的结构相似度,采用包含779张失真图像的LIVE数据库以及相应的拟合函数验证了复数结构相似度方法。交叉失真和分类失真图像测试实验表明,所提出的复数矩阵结构相似度方法的整体性能与人眼视觉特性的一致性优于MSE等传统方法,和SSIM等流行的方法以及QSSIM等较新的方法相比也有较大优势。

The purpose of objectively assessing digital image quality is to obtain the assessment results consistent with human visual perception.Different sensitivity factors of human visual system needs to be considered in the analysis and comparison of image structure in order to obtain the image quality assessment results consistent with human perception.Aiming at the problem that single property of image structure information is too unilateral to describe the comprehensive image structure,a complex number method is used to describe the image structure information.A complex matrix is constructed to describe the image structure information.The complex number is taken as an information combination method; and the local variance component with high human eye sensitivity in the image and the grey scale distribution of the image pixels representing general sensing information are taken as the real part and imagery part,respectively.On this basis,the structure similarity method in non-real number domain and its mathematic model were deeply studied; and the method was used to measure the structure similarity of the complex matrixes of two images; then,the quantitative results were obtained.The LIVE database including 779 distorted images and corresponding fitting function were adopted to verify the complex matrix structure similarity method.The cross-distortion experiment and classification distortion experiment results show that using the proposed complex matrix structure similarity method,the consistency of overall performance with human visual perception is better than that using the state-of-the-art methods,such as MSE,SSIM methods and the new QSSIM method.