【作者】 桑庆兵；

【导师】 吴小俊；

【作者基本信息】 江南大学 ， 轻工信息技术与工程， 2013， 博士

【摘要】 近年来日益增长的图像应用方面的消费需求，促进了人们对客观图像质量评价算法的研究兴趣。客观图像质量评价研究的目标就是开发出能够像人眼一样自动地测量图像质量下降的评价算法，它在图像和视频处理的各个领域中都处于重要地位。根据使用参考图像信息的多少，客观图像质量评价方法可以分为全参考方法、半参考方法和无参考方法。本文主要聚焦半参考和无参考方法，其主要内容总结如下：(1)提出了一种基于小波变换域的半参考图像质量评价算法，该方法首先对图像进行2尺度小波分解，提取尺度2上的低频小波系数作为图像特征向量，为了度量向量的相似性，本文将图像特征向量看成n维欧氏空间中的一个点，然后分别计算参考图像和失真图像的低频小波系数特征向量在n维欧氏空间中所对应的点之间的欧氏距离D(X,Y)，作为为图像质量评价指标。(2)提出了一种基于灰度共生矩阵的无参考模糊图像质量评价方法。首先通过LogGabor小波变换生成相位一致图像，然后利用灰度共生矩阵计算相位一致图像的信息熵、能量、对比度、相关性和同质性5个特征，最后利用支持向量回归(Support VectorRegression，SVR)模型学习预测图像质量得分。实验结果表明，该方法能够有效地评价图像的模糊失真。(3)基于离散余弦变换系数提出了一种无参考模糊图像质量评价方法。该方法首先通过对图像进行离散余弦变换，得到图像的离散余弦变换系数作为图像质量变化的特征向量，然后利用广义回归神经网络模型对此特征向量进行训练学习，预测得到无参考模糊图像质量得分。(4)提出了一种基于奇异值分解的无参考模糊和噪声图像质量评价方法。该方法首先通过对待评价模糊图像和噪声图像进行高斯低通滤波生成再模糊图像，然后分别对待评价图像和再模糊图像分别进行奇异值分解，得到各自奇异值向量，最后构造奇异值的改变量来作为无参考模糊和噪声图像的质量评价指标。(5)提出了两种基于奇异值分解的通用无参考图像质量评价方法。第一种方法是基于奇异值的改变量构造了一种通用无参考图像质量评价指标，奇异值的改变量能够反映图像的蚀变情况。第二种提出了基于图像奇异值倒数曲线的通用无参考图像质量评价方法。图像的奇异值倒数曲线近似幂函数曲线，且随着图像失真程度的不同，奇异值倒数曲线的弯曲程度也不相同。根据奇异值倒数曲线的这一特征，本文从面积和曲率两个角度，构造了两种通用无参考图像质量评价指标。更多还原

【Abstract】 In recent years, the increasing number of demanding consumer image applications hasboosted interest in objective image quality assessment (IQA) algorithms. Objective imagequality assessment aims to automatically measure the quality degradation perceived by thehuman eyes. It is of fundamental importance to address a wide variety of problems in imageand video processing. Based on the availability of the information about the reference image,IQA models can be classified into full-reference (FR), reduced-reference (RR) and noreference (NR) IQA methods. This dissertation focuses on RR-IQA AND NR-IQA, the majorcontents are as follows in general：First, we propose a novel metric for RRIQA based on wavelet transform. We dothe wavelet decomposition of2scales to images, and extract the low frequency waveletcoefficients of the second scale as the image feature vector. In order to measure thesimilarity of vectors, we see the feature vector as a point in n-dimensional Euclidean space,and calculate the distance between referenc image feature vector and distortion image featurevector in n-dimensional Euclidean space. The distance is regarded as the metric of imagequality.Second, we propose a no reference blur image quality assessment method based on graylevel co-occurrence matrix extraction phase congruency image feature and supportvector regression (SVR). The method is composed of three steps. First, we use Log Gaborwavelet to generate phase congruency map of the image. Then we calculate the PhaseCongruency map’s features which are entropy, energy, contrast, correlation and homogeneityby gray level co-occurrence matrix. Finally, we predict no-reference blur image quality scoreby using SVR model training and learning.Third, we propose a no reference blur image quality assessment method based ondiscrete cosine transform. First of all, we do discrete cosine transform to images and extractdiscrete cosine transform coefficients as feature vector, and then used the generalizedregression neural network model to train feature vector to predict image quality. In the threepublic databases, the experimental results show that this method has a good correlation withthe subjective quality score.Fourth, we propose a new blind blur and noise index for still images using Gaussian blurand Singular Value Decomposition (SVD). The algorithm is composed of three steps. Firstly,a re-blurred image is produced by using Gaussian blur to the test image. Then the singularvalue decomposition is performed to the test image and re-blurred image. Finally, a blur andnoise index is constructed by using the change of singular values. Experimental resultsobtained on four simulated databases show that the proposed algorithm has high correlationwith human judgments when assessing blur distortion of images.Finally, we propose two universal blind image quality assessment methods. The first is auniversal blind image quality assessment method based on the change of singular value. Thechange of singular value can reflect the distortion of an image. The second is a universal blindimage quality assessment method using a reciprocal singular value curve. The reciprocalsingular value curves of natural images resemble inverse power functions. The bending degree of the reciprocal singular value curve is varies with distortion type and severity. Weconstructed two new general blind IQA indices utilizing the area and curvature of imagereciprocal singular value curves. These two methods do not require prior knowledge of anyimage or distortion, and hence do not require any process of training, hence are "completelyblind" IQA models.更多还原