【作者】 何立火；

【导师】 高新波；

【作者基本信息】 西安电子科技大学 ， 模式识别与智能系统， 2013， 博士

【摘要】 视觉信息数字化波及到世界的每一个角落，人们将不断追求视觉信息的高清晰和高保真。但是视觉信息在采集、压缩、处理、传输和恢复的过程中都可能会引入失真，它们会给视觉信息的处理、分析和解译带来一定的障碍，也影响了人们正确认识客观世界的程度。因此，需要设计合理可靠的评价方法来精确度量视觉信息的感知质量，从而方便地指导视觉信息处理系统的优化、改进和提高，以最少的代价提供最优的视觉质量。本文针对视觉信息中自然场景图像的质量感知模型和评价方法这一基本问题，开展了系统深入的研究。以人类视觉系统基本特性为基础，探索视觉信息的感知特性，分析自然场景的统计规律，建模质量语义分层特性，构建视觉信息客观质量评价方法，度量视觉信息的保真度和可懂度，为视觉信息处理系统的设计和优化提供合理的依据。主要的研究内容概括如下：（1）考虑到人类视觉系统对不同的形态学成份具有不同的感知特性，并结合视觉恰可察觉差异模型，提出了一种基于形态学成份分析的全参考型图像质量评价方法。首先通过形态学成份分析把原始图像和失真图像都稀疏表示成纹理和卡通成份；然后分别建模和求解两种形态学成份的恰可察觉差异感知特性；最后计算原始图像和失真图像的感知特性的差异，并对两种形态学的误差进行融合得到失真图像的质量。实验结果表明此方法能够有效的评价不同失真类型的图像，并优于经典的图像质量评价算法。（2）提出了基于S-CIELAB颜色模型的全参考型图像质量评价方法。由于自然场景图像具有的色彩非常丰富，信息量非常大，图像失真引起的颜色信息丢失，对于图像的感知质量具有重要的影响。因此，基于人眼感知过程中颜色视觉特性，提出了基于符合人类视觉特性的S-CIELAB颜色模型的全参考型图像质量评价方法。首先将参考图像和待测图像变换到S-CIELAB颜色空间，然后对变换后的三个通道分别计算待测与原始图像间的结构相似性，最后融合三个通道得到最终图像质量。实验结果表明客观评价结果与主观评价结果达到较好的一致性。（3）提出了基于颜色分形结构特征的部分参考型图像质量评价方法。颜色信息在图像质量感知中起着至关重要的作用，但在很多情况下无法完全获取这些信息，因此提出了一种基于颜色分形结构模型的部分参考型图像质量评价方法。首先用分形结构模型分别提取原始图像和待测图像的颜色信息和结构信息，用以模拟自然图像中颜色的局部性和相似性；然后比较原始图像和失真图像特征的差异；最后运用支持向量回归将特征差异映射成图像质量。该方法有效地减少了对原始图像信息的依赖程度，具有更广泛的应用，且实验结果表明该算法的性能优于现有部分参考型质量评价方法。（4）提出了基于自然场景统计特征稀疏表示的无参考型图像质量评价方法。目前大多数无参考型图像质量评价方法只能针对一种或几种特定失真类型图像进行质量评价。因此，提出了一种针对自然场景图像的无参考型图像质量评价方法，实现了对不同类型失真图像的有效评测，具有通用性和普适性。首先将待测图像进行小波分解，提取其自然场景统计特征；然后通过稀疏表示对特征进行编码；最后通过稀疏编码系数对平均主观差异分数值进行加权求和得到视觉感知质量。实验结果表明，该方法评价结果与主观感知质量具有较高的一致性，较现有的无参考型图像质量评价方法和经典全参考型图像质量评价算法有更好的性能。（5）提出了基于视觉质量主题的真正无参考型图像质量评价方法。由于视觉显著性在质量评价中占有非常重要的地位，以视觉显著性作为先验，主导视觉隐主题的分布，提出了视觉显著性加权的分层狄利克雷过程混合模型。基于该模型提出了无参考型质量评价方法。首先对训练集图像进行质量感知特征提取并建立视觉字典；然后根据前面提出的模型计算与视觉质量有关的主题分布；最后计算失真图像的主题分布与原始图像的主题分布的差异得到最终的图像质量。在现有公开数据库上的实验验证了该方法在一致性和鲁棒性方面更具优势。上述的研究从全参考型到无参考型，对参考图像信息的依赖性逐渐降低，实际应用性逐渐增强，理论由浅入深，对视觉信息质量感知模型及评价方法进行了深入研究。本文的研究成果为视觉信息客观质量评价方法的研究开辟了新的思路，具有重要的理论意义及实用价值。更多还原

【Abstract】 Visual information digitization has spread to every corner of the world, and peoplewill continue to pursue high definition and fidelity of visual information. However, theprocesses of acquisition, compression, processing, transmission and restorationintroduce different distortions to the visual information. This gives a great obstacle tothe process of visual information processing, analysis and interpretation, preventinghuman beings from understanding the objective world. Therefore, it is necessary todesign reasonable and reliable methods to measure perceived quality of visualinformation. These methods have guiding significance for optimizing, improving andoptimizing the visual information processing system. Thus the system can provide thebest visual quality with minimum cost.In this dissertation, a systematic study about fundamental issues of the qualityperception model and assessment metrics for visual information is carried out. Basedupon the fundamental characteristics of human visual system, this study explores theperceptual characteristics, analyzes the statistical regularities of natural scenes, modelsthe hierarchical quality semantic and finally constructs several objective qualityassessment methods. These methods measure the fidelity and intelligibility of visualinformation and provide a reasonable basis for the design and optimization of visualinformation processing system. The author’s major contributions are outlined as follows(1) Considering that the Human Visual System (HVS) has different perceptualcharacteristics for different morphological components, a novel image quality metric isproposed by incorporating Morphological Component Analysis (MCA) and HVS,which is capable of assessing the image with different kinds of distortion. Firstly,reference and distorted images are decomposed into linearly combined texture andcartoon components by MCA respectively. Then these components are turned intoperceptual features by Just Noticeable Difference (JND) which integrates maskingfeatures, luminance adaptation, and Contrast Sensitive Function (CSF). Finally, thediscrimination between the feature of the reference image and that of the distortedimage is quantified using a pooling strategy before the final image quality is obtained.Experimental results demonstrate that the performance of the proposed prevails oversome existing methods.(2) Image quality assessment based on S-CIELAB model. Most natural sceneimages captured by digital devices are color images, which bear huge information. Theloss of the color information caused by the distortion has great impact on the perceptual image quality. However, most existing IQA metrics are only designed for gray images.Hence, the S-CIELAB color model, which has an excellent performance for mimickingthe perceptual processing of human color vision, is incorporated with the geometricaldistortion measurement to assess the image quality. First, the reference and distortedimages are transformed into the S-CIELAB color perceptual space, and the transformedimages are evaluated by an existing metric in three color perceptual channels. Thefidelity factors of three channels are weighted to obtain the image quality. Experimentalresults show that the proposed methods are in good consistency with human subjectivequality scores.(3) Color fractal structure model for reduced-reference colorful image qualityassessment. Most of existing methods fail to take color information into consideration,although the color distortion is significant for the increasing color images. To solve theaforementioned problem, this paper proposed a novel IQA method which focuses on thecolor distortion. In particular, we extract color features based on the model of colorfractal structure. Then the color and structure features are mapped into visual qualityusing the support vector regression. Experimental results demonstrate that the proposedmethod is highly consistent with the human perception especially on images with colordistortion.(4) Sparse representation for no-reference image quality assessment. Since theexisting no-reference image quality assessment (IQA) is designed for the image withone or few specific distortions, a universal no-reference image quality assessmentmetric is introduced. It is a simple yet effective algorithm based upon the sparserepresentation of natural scene statistics (NSS) feature and it can predict the quality ofimage with different distortions. This algorithm consists of three key steps: extractingNSS features in the wavelet domain, representing features via sparse coding, andweighting differential mean opinion scores by the sparse coding coefficients to obtainthe final visual quality values. Thorough experiments show that the proposed algorithmis consistent with subjective perception values, and outperforms representative blindimage quality assessment algorithms and some full-reference metrics.(5) Visual-quality-related topics based “completely” no-reference image qualityassessment. Because visually salient areas are critical in subjective quality assessment,visual saliency weighted hierarchical Dirichlet processes (visual-saliency wHDP) isproposed by introducing visual saliency as the prior for observations to dominate theconstruction of latent topics. According to this model, no-reference image quality assessment metric is proposed, and it includes three parts: constructing visualvocabulary by extracting the quality-aware features from a training set, obtainingdistributions of visual-quality-related topics by training visual saliency wHDP, andestimating the quality of a test image by computing the difference between thevisual-quality-related topics found in the test image and those found in the originalimages contained in the training set. Thorough experiments show that the proposedvisual-saliency wHDP robustly produces quality-related topics and obtains promisingperformance.Extensive study has been conducted from full-reference IQA to no-reference IQA,which differs in the dependence of the references and practicability. The overallresearch on visual information perceptual model and quality assessment follows thegradual progress that is from the elementary to the profound. The fruit presented in thisdissertation opens up a new way for visual information assessment, which hasextremely important theoretical significance and practical value.更多还原