【作者】 刘明娜；

【导师】 杨新；

【作者基本信息】 上海交通大学 ， 模式识别与智能系统， 2010， 博士

【摘要】 图像信息是视觉信息最重要的组成部分,是人类获取外界信息的主要途径。在数字图像的获取、处理、编码、存储、传输和重建的每一个步骤中,通常都会对图像的质量产生影响,如何评价图像质量成为图像处理、计算机视觉领域一个基本而又富有挑战的问题。本论文围绕图像失真特征的表示和提取,从多特征融合、多尺度几何分析以及人眼视觉系统建模等方面研究图像质量的客观评价问题,并将常规图像质量分析和度量方法应用在医学图像领域。主要工作和创新成果如下:1.对图像视觉质量失真的表示和失真特征提取进行了分析和比较,通过数学推导讨论了图像失真特征的不同表示方法以及他们之间的内在关联。在此基础上,引入典型相关分析方法,将多种失真特征有效融合在一起,给出基于多特征的图像质量评价算法,优化了特征选择方式,并使其可以根据失真类型给出更为准确、稳健的图像质量评价结果,能全面地综合评价图像质量。2.图像结构特征的表示和度量是图像质量评价的关键技术之一,本文引入多尺度几何分析方法来提取图像的结构特征、方向特征和多分辨率特征,逼近人眼视觉系统对图像信号的多通道分解过程。在此基础上,提出一种多尺度方向性差异模型,表示和度量图像质量的视觉失真程度。3.在充分考虑前端人眼视觉系统特性的基础上,本文引入视觉感知在频率和方向上的多通道分解特性、对比度敏感特性、视觉掩蔽效应、频带响应和误差融合模型,与图像的结构特征表示方法有效结合,提高了基于特征提取的图像质量评价方法的准确性,同时改进了基于单纯人眼视觉系统特性模型算法的复杂性,取得了较好的效果。4.在对图像信号的多通道分解过程中,基于视觉系统对不同频率的响应特性,提出了对频带误差的线性融合方法。在综合性图像数据库上与非线性频带误差融合方法相比较,证实了线性误差融合方法能够满足对视觉系统的逼近要求,节省了计算成本的同时提高了图像质量预测的准确度。5.针对临床CT成像应用中保证图像质量可辨度的条件下降低射线剂量的要求,本文研究了客观图像质量评价算法在计算机辅助评价CT图像质量中的应用,特别是对低剂量CT图像的质量评价问题。针对CT图像质量失真,如噪声增加、空间分辨率降低、对比度降低等,本文在经典的图像信噪比评价方法的基础上,增加视觉信息保真测度、结构相似性测度的应用。改变CT射线剂量,以医生判断作为标准,评估客观图像质量评价测度于CT图像的性能,进而分析CT图像的视觉失真特征。在对测试体膜和动物切片进行的大量实验中,复小波域的结构相似性测度给出了与医生专业评判相符的结果。更多还原

【Abstract】 Among the visual information received by human visual system images arethe matters of primary importance. In the procedures of image processing sys-tem, e.g., acquisition, processing, coding, storage, transmission and reproduction,digital image may be in degradation in visual quality, so evaluating image qual-ity becomes a fundamental job in the field of image processing and computervision. This thesis investigates objective image quality assessment problem fromthe viewpoint of decision fusion, structural feature extraction and multi-scale ge-ometric analysis. Furthermore, the application of general objective image qualityapproach in medical image evaluation is studied and performed. The main workand innovations are listed as follows:1. The research background is extensively reviewed, and some popular im-age quality measures are analyzed and compared from the aspect of distortiondescription and feature extraction. Moreover, the ideal of canonical correlationanalysis is introduced, which keeps the e?ective discriminating information ofmulti-feature as well as eliminates the redundant information inside them tocertain degree. A Supervised Multi-Feature (SMF) full-reference image qualityevaluation algorithm is proposed. Compared with the state-of-the-art image qual-ity metrics, the proposed method improves accuracy and robustness of qualityprediction.2. Propose a novel objective full-reference image quality assessment metricbased on multiscale geometric analysis. The multichannel behavior of the humanvision system is emulated by contourlet transform, a perceptual subband decom-position. The distortion feature and level is emulated by a multi-scale directionaldi?erence model. 3. The HVS model of the low-level perception used in this metric includessubband decomposition, contrast masking, entropy masking, and error pool-ing. Extensive validation experiments are performed on two professional imagedatabases.The proposed method displayed a higher prediction accuracy and ro-bustness across extensive distortion types and a broad range of distortion level,exhibiting a generally better performance.4 .In the error pooling stage, a linear fusion scheme for subband distortionis proposed to trade the frequency properties of the HVS and computation cost.The nonlinear fusion scheme of the Minkowski summation is also implemented forcomparison. Finally, the advantages and limitations of these are compared anddiscussed. Owing to the the employment of the frequency properties of HVS andthe Weber-Fechner law, the linear fusion scheme of subband distortion proved tobe a preferable alternative for the Minkwoski summation.5. Computed tomography (CT) is an essential imaging modality. To solvethe problem of increasing radiation exposure from CT scanner and image qual-ity, general objective image quality methods are applied in CT image qualityevaluation. After analyzing perceptual feature of CT image, several popular ob-jective image quality metrics, which focus on the similar perceptual features, aretested on the CT image of phantom and animals. A lot of experiments are per-formed. Compared with the subjective ratings from two professional radiationphysicians, the complex wavelet-based structural similarity metric presents thebest CT image quality prediction results.更多还原