现如今，随着互联网技术以及通信技术的快速发展，数字图像已经成为人们日常生活中信息传递的重要方式。据统计，2011年以来，世界产生的数字照片总量达到上百亿张，而且这个数量还在逐年增长。然而，图像在采集、存储、压缩和传输等过程中容易受到不同种类的失真干扰，从而造成图像质量的降低。所以，如何准确、可靠地评价图像的质量成为当前及未来研究中一个重要的研究热点。通常，大多数图像是由人来观看，所以最可靠的图像质量评价方式是主观质量评价，即组织观看人员根据他们的经验来对图像的质量进行主观评分，然而，随着图像数量的增多，主观质量评价的实施变得困难，而且也无法应用于实时的图像处理系统中。所以，研究者们提出客观质量评价方法，通过设计客观算法来对图像的质量做出评价。根据是否参考原始图像，现有的客观质量评价方法被分为三个类别，分别是全参考，部分参考和无参考质量评价方法。尽管目前针对这三类客观图像质量评价分别提出了大量的方法，但是客观质量评价的研究仍然不够成熟，主要表现在以下几个方面，第一，由于目前对人类视觉感知机制的理解不够深入，现有的基于度量信号失真的客观质量评价方法不能准确地模拟主观质量评价；第二，在无参考质量评价方法设计中，大多数方法仍然需要利用主观质量分数来训练质量评价模型；第三，在评价真实场景中的失真图像时，现有的客观算法的表现仍然不够理想。针对上述问题，本文从人类视觉感知和统计的角度出发对客观质量评价中部分参考和无参考质量评价展开深入的研究。 具体的研究内容分为以下四个部分： 第一，由于对人类视觉感知机制的理解不够深入，所以大多数客观质量评价方法主要通过度量信号本身的失真程度来评价图像的质量，这类方法由于没有考虑人类视觉感知机制而不能准确地评价图像的质量。通过对人类视觉感知机制的研究，基于自由能原理和稀疏表示，本文提出了一个部分参考的图像质量评价方法。在最近提出的大脑以及神经科学理论中，自由能原理指出视觉感知与理解是一个活跃的推理过程，这个过程是由大脑内部的一个生成模型来控制，通过这个生成模型，大脑可以对输入的图像进行预测，产生相应的预测图像。那么输入图像与大脑预测图像之间的差异可以反映图像的感知质量。而且，研究表明大脑在表示外界图像时与稀疏表示的方式类似。所以，结合自由能原理和稀疏表示，我们利用稀疏表示来近似大脑内部的生成模型，并建立了一个新的部分参考的图像质量评价模型。在该模型中，我们首先对原图和失真图分别进行稀疏表示，然后分别计算表示的残差，残差的信息熵用来反映图像感知质量的变化。图像的质量被定义为残差的熵之间的差异。实验结果表明，我们提出的方法比同类的方法预测准确性更高，同时我们的方法的时间复杂度更低。尽管我们的方法属于部分参考的图像质量评价方法，但是我们的方法仅仅需要参考原始图像的一个数字，最大程度地降低了所需的原始图像的信息量。 第二，现有的无参考质量评价方法大多数属于主观质量分数已知的评价方法，此类方法通常需要大量的训练样本图像以及对应的主观分数来训练质量预测模型，相比之下，主观质量分数未知的无参考评价方法仍然较少而且已有方法的性能仍然不能与主观质量分数已知的方法相比，因此，本文提出了一个高效的主观质量分数未知的无参考评价方法，在提出的方法中，我们通过度量图像的结构、自然性以及感知质量的变化来预测图像质量的失真程度。首先，我们对图像的相位一致性和梯度进行建模，提取模型参数来表示图像的梯度信息，其次，我们对图像的局部归一化系数以及相邻系数对的乘积进行建模，提取模型参数来描述图像的自然性的程度；最后，基于大脑的自由能原理，我们对输入图像和大脑对它的预测图像之间的残差进行建模，模型参数用于表示图像的感知质量。在特征提取完之后，我们利用一组无失真图像学习出一个多元的高斯模型并利用该模型来预测图像的质量。失真图像的质量被定义为它的多元高斯模型与学习到的无失真高斯模型之间的距离。实验结果表明，提出的方法相比于主观质量分数未知的无参考评价方法取得了更高的预测准确性，以及与主观质量分数已知的无参考评价方法相当的预测性能。 第三，图像在获取的过程中，由于相机参数的设置不合理，或者获取方式的随意性，获取的图像很可能受到不同种类的失真干扰，比如噪声、模糊、对比度失真等。在这些失真类型中，失焦模糊占据了很大比例，然而，专门针对这种失焦模糊的图像质量评价方法还较少，所以，在本文中，我们对失焦模糊图像的质量评价进行专门研究，提出了一个针对真实失焦模糊图像的无参考质量评价模型。首先我们通过主观实验建立了一个专门的失焦模糊图像数据库，然后提出了一种专门针对失焦模糊的图像质量评价方法。在该方法中，我们利用梯度和相位一致性特征来提取图像的结构信息，从而度量图像的模糊程度，其次，考虑到失焦模糊的在图像中的随意性，我们对图像进行显著性检测，利用显著图来加权图像的模糊程度图，最后对加权的模糊程度图进行池化，得到最后的质量分数。实验结果表明，我们的方法与主观质量评价的结果具有较高的一致性。 第四，现有的图像质量评价方法尽管在现有的数据库上取得了较高的预测准确性，但是这些数据库的失真大多数都是人为生成的失真，与真实获取到的图像具有的失真存在较大的差异。真实图像中的失真具有很高的复杂度，准确地预测其质量仍然具有很大的难度。针对如何准确地评价真实失真图像的质量，本文提出了一个无参考的基于神经网络的真实失真图像的质量估计模型。在提出的方法中，我们提取了与质量相关的两类统计特征，第一类特征是提取图像局部归一化系数的统计分布来描述视觉感知的低层次特性，第二类特征是提取图像与大脑预测图像的预测残差的统计分布来描述视觉感知的高层次特性。然后我们设计了一个四层的神经网络对图像的质量做出预测。实验结果表明，我们的方法相比于现有的无参考质量评价方法取得了更高的预测准确性。 关键词：图像质量评价；自然图像统计特性；自由能原理；稀疏表示；图像模糊

 Nowadays, as the internet and communication technology develop rapidly, digital images have become the important means of information transmission in daily life. According to the statistics, the total amount of digital images produced in the world attains tens of billions from the year of 2011, such number is still increasing year by year. However, the digital images are vulnerable to different kinds of distortions during acquisition, storage, compression and transmission. Therefore, the image quality is often affected at different levels. How to accurately evaluate the image quality becomes an important research hotspot in current days and in the future days. Generally, most of the images are consumed by humans. Then the most reliable way for evaluating the image quality is the subjective quality assessment, which refers to asking viewers to rate the image quality according to their perception of the image quality. However, subjective quality assessment is often hard to carry out because of the large amount of images. In addition, subjective quality assessment can’t be used in the real-time image processing applications. Toward this end, researchers attempt to develop objective quality assessment methods for image quality evaluation by means of objective algorithms. According to the accessibility of the original image, existing objective image quality assessment methods can be classified into three categories, which are full-reference (FR), reduced-reference (RR) and no-reference (NR) methods respectively. Although a lot of objective methods belonging to these three types have been proposed, the research on objective quality assessment is still immature, which can be manifested as follows: First, due to lack of understanding about the human visual perception mechanism, existing objective methods can’t simulate the subjective quality assessment accurately; Second, in the design of NR methods, most methods still resort to subjective scores for training the quality model; Third, to assess the real distorted images, existing objective methods still perform poorly. Focusing on the above issues, this dissertation carries out research on RR and NR methods in objective IQA from the perspective of the human visual perception mechanism and statistical modeling. The content of the dissertation can be divided into four sections detailed as follows: First, due to lack of understanding about the human visual perception mechanism, most existing methods mainly assess the image quality by measuring the distortion degree of the signal itself, this kind of methods can’t evaluate the image quality accurately without considering the human visual perception mechanism. Based on the exploration of the human visual perception mechanism, a RR image quality evaluation method based on the free-energy principle and sparse representation is proposed. The free-energy principle in recent studies of brain theory and neuroscience models the perception and understanding of the outside scene as an active inference process, in which the brain tries to account for the visual scene with an internal generative model. Specifically, with the internal generative model, the brain yields corresponding predictions for its encountered visual scenes. Then the discrepancy between the visual input and its brain prediction should be closely related to the quality of perceptions. On the other hand, sparse representation has been evidenced to resemble the strategy of the primary visual cortex in the brain for representing natural images. With the strong neurobiological support for sparse representation, in this dissertation, we approximate the internal generative model with sparse representation and propose an image quality metric accordingly, which is named FSI (Free-energy principle and Sparse representation-based Index for image quality assessment). In FSI, the reference and distorted images are respectively predicted by sparse representation at first. Then the difference between the entropies of the prediction discrepancies is defined to measure the image quality. Experimental results confirm the superiority of the proposed method over the same kind of methods. The time complexity is also lower in the meantime. Although our method belongs to RR methods, it only needs a single number from the reference image, which maximumly reduces the needed data amount from the original image for quality evaluation. Second, most of the existing NR methods belong to the opinion-aware methods, which require a large number of image samples with the associated subjective scores for training the quality model. By comparison, opinion-unaware methods are very limited and existing opinion-unaware methods still can’t be compared with the opinion-aware methods. Therefoer, a highly effective opinion-unaware NR method is proposed, in which we attempt to quantify the image quality degradations through measuring the structure, naturalness and the perception quality variations of the distorted image from the pristine images. In our method, the structure is characterized by modeling the image phase congruency (PC) and image gradients distributions. The naturalness degree is measured through the parameters that depict the distributions of the locally mean subtracted and contrast normalized (MSCN) coefficients and the products of pairs of the adjacent MSCN coefficients. The perception quality is characterized by the prediction discrepancy between the image and its brain prediction based on the free-energy principle. After feature extraction, we learn a pristine multivariate Gaussian (MVG) model with the extracted features from a set of pristine images, which is used for quality definition. The quality of a new image is defined as its MVG model variation from the pristine MVG model. Experimental results demonstrate the proposed method outperforms state-of-the-art opinion-unaware methods and delivers comparative performance with mainstream opinion-aware methods. Third, during image acquisition, the images are possibly distorted by different kinds of distortions, such as noise, blur, contrast change because of the improper camera settings or the casual photographing manner. Among the distortions, out-of-focus blur occupies a large proportion. However, the specific quality assessment to this kind of distorted images is still limited. Therefore, in this dissertation, we focus our attention on the quality assessment for the out-of-focus blurred images and propose a specific blind quality assessment model to evaluate the quality of the real out-of-focus blurred images. At first, we establish a specific out-of-focus blurred image database through subjective experiments. Then we propose a specific quality model to evaluate the out-of-focus blurred image quality. In our model, we employ the gradients and PC features to extract the structure information of the image for measuring the blurriness degree. Considering the arbitrary distribution of the out-of-focus blurr over the image, we then perform saliency detection on the image and utilize the saliency map to weight the local blurriness map accordingly. The quality of the image is finally estimated by pooling the weighted blurriness map. Experimental results demonstrate the proposed method delivers high consistency with subjective evaluation results. Fourth, although existing quality assessment methods earn high prediction performance on the existing databases, the distortions in those databases are artificially generated, which is quite different from the distortions in real camera images. The distortions in real images are much complex, which raises great challenge for precisely evaluating the image quality. For accurately evaluating the quality of real camera images, a blind quality assessment method based on neural network for real camera images is proposed in this dissertation. In our proposed method, we extract two types of quality-aware statistical features to characterize the image quality degradations. The first type of statistical features are extracted from the locally mean subtracted and contrast normalized (MSCN) coefficients, which describe the low-level characteristics in the early human vision. The second type of features are extracted from the distribution of the prediction discrepancy between the image and its brain prediction based on the free-energy principle, which is to characterize the high-level characteristics of the human visual perception. After feature extraction, we design a neural network of four layers to evaluate the image quality. Experimental results demonstrate the proposed method achieve higher prediction accuracy than the existing blind quality methods. Keywords: image quality assessment; natural scene statistics; free-energy principle; sparse representation; image blur