【作者】 李浩；

【导师】 侯春萍；

【作者基本信息】 天津大学 ， 信息与通信工程， 2019， 硕士

【摘要】 随着多媒体技术和互联网的发展,图像逐渐成为一种简单高效的信息承载方式,在人们的日常生活中发挥着至关重要的作用。然而,图像在获取、处理、传输和存储的各个阶段都有可能产生质量退化效应,不仅影响用户的视觉体验,而且不利于后续的应用。因此,通过分析和建模提出有效的图像质量评价方法具有重要意义。图像信号通过人眼接收并传递给视觉皮层,人眼视觉系统决定了图像以何种形式被解释和理解,从而产生美观与糟糕等主观判断。图像统计分析从统计学角度出发,挖掘普遍存在且有效的统计规律。本文结合人眼视觉系统(HVS)特性和图像统计分析,探索并建立与人眼主观感知高度契合的无参考图像质量评价方法。本文的研究工作主要包括以下两个方面:针对色调映射图像,提出了一种基于全局和局部视觉感知的无参考质量评价方法。首先,分析色调映射图像特点,模拟HVS从全局和局部两个层次感知图像信号;其次,在全局感知方面,结合统计分析,采用颜色矩描述色彩信息,设计明暗分布特征来衡量整体曝光效应,利用信息熵来计算全局的信息量;然后,在局部感知方面,针对图像块进行处理,计算块的对比差异和局部熵,并结合多通道分解机制在离散小波变换(DWT)域上分解图像信号并计算其能量;最后,结合全局和局部两个层次上的特征,利用机器学习方法对特征进行回归处理得到评价模型。实验结果表明,本文所提出的算法在ESPL-LIVE HDR数据库上的皮尔森线性相关系数(PLCC)和斯皮尔曼秩相关系数(SRCC)比目前最优算法还要高出3%,并且与主观评分有较高的一致性。针对真实失真图像,提出了一种基于结构、纹理和色彩信息的无参考质量评价方法。考虑到真实失真图像的复杂性,采用HVS敏感的多属性进行视觉感知,每种属性的感知不是单一神经细胞响应的结果,因此每种属性采用多种特征进行统计表示。其中,主要从结构、纹理和色彩信息三方面表征视觉多属性感知特性。在结构信息方面,采用梯度直方图和曝光度来联合描述;在纹理信息方面,对图像进行DWT分解求解对数能量;在色彩信息方面,分为色度感知和饱和度感知,其中,色度信息从全局和局部两个层次上获取。最后,结合上述三种信息,利用支持向量回归训练得到一个质量敏感的预测模型。实验结果表明,本文提出的算法在CID2013数据库上的PLCC达到了0.8709,SRCC达到了0.8059,与主观感知分数有极高的一致性,优于现有评价算法。更多还原

【Abstract】 With the development of multimedia technology and the Internet,images have gradually become a simple and efficient way to carry information,playing a vital role in people’s daily lives.However,the image may have quality degradation effects at various stages of acquisition,processing,transmission,and storage,which not only affects the user’s visual experience but also is not conducive to subsequent application.Therefore,it is important to propose effective image quality assessment(IQA)methods through analysis and modeling.The image signal is received by the human eye and transmitted to the visual cortex.The human visual system(HVS)determines the form in which the image is interpreted and understood,resulting in subjective judgments such as beauty and badness.Image statistical analysis explores ubiquitous and effective statistical laws from a statistical perspective.This thesis combines characteristics of the HVS and statistical analysis of images to explore and establish blind image quality assessment(BIQA)methods that are highly compatible with the subjective perception of human eyes.The research work of this thesis mainly includes the following two aspects:A no-reference quality evaluation method based on global and local visual perception is proposed for tone-mapped images.Firstly,the characteristics of tonemapped images are analyzed,and the HVS is simulated to perceive image signals from global and local levels.Global perception is combined with statistical analysis.Specifically,color moments are used to describe color information,light and dark distribution features are designed to measure the overall exposure degree and information entropy is used to calculate the global information.In terms of local perception,images are divided into many blocks.Then,the contrast difference and local entropy of the block are calculated,and the image signal is decomposed and calculated in the discrete wavelet transformation(DWT)domain in combination with the multichannel decomposition mechanism.Combining the characteristics of the global and local levels,the machine learning method is used to regress the features to obtain the evaluation model.The experimental results on the ESPL-LIVE HDR database show that the proposed algorithm has a 3% higher Pearson Linear Correlation Coefficient(PLCC)and Spearman Rank-order Correlation Coefficient(SRCC)than the current state-of-the-art algorithm,and has a high consistency with the subjective score.A no-reference quality evaluation method based on structure,texture and color information is proposed for authentically distorted images.Considering the complexity of authentically distorted images,the HVS sensitive multi-attributes are used for visual perception.The perception of each attribute is not the result of a single nerve cell response.Therefore,each attribute is statistically represented by various features.Among them,the HVS multi-attribute perception characteristics are mainly characterized by three aspects: structure,texture,and color information.In terms of structural information,gradient histogram and exposure are utilized.In the aspect of texture information,DWT decomposition is performed and its logarithmic energy is computed.Concerning color information,it is divided into chroma perception and saturation perception,in which chroma information is obtained from global and local levels.Finally,combined with the above three kinds of information,a quality-sensitive prediction model is obtained by using support vector regression.The experimental results show that the proposed algorithm has a PLCC of 0.8709 and an SRCC of 0.8059 on the CID2013 database,which is highly consistent with the subjective perception score and is superior to the existing evaluation algorithm.更多还原