【作者】 冯丹丹；

【导师】 张瑞峰；

【作者基本信息】 天津大学 ， 电磁场与微波技术， 2018， 硕士

【摘要】 随着多媒体技术的快速发展,图像质量评价已经成为图像处理领域内一项很有意义的研究课题。图像的获取、压缩、传送和处理等过程会对图像造成多种失真,因此评价混合失真图像具有重要的现实意义。近年来,生物学家对人类视觉系统的研究有了重要进展,质量评价研究人员模拟人类视觉特性提出的质量评价模型取得了很好的评价效果。本文在总结前人研究成果的基础上,提出了两种基于人类视觉特性的评价混合失真类型图像的客观图像质量评价方法。首先,本文提出了一种基于双树复小波变换与LBP(Local Binary Pattern)算子的无参考混合失真图像质量评价算法。双树复小波变换相较于传统小波变换拥有更多的方向选择性且具有近似的平移不变性。本算法先对图像进行双树复小波变换,再利用LBP算子对小波系数的幅值信息进行统计。在统计过程中为了突出图像的显著部分,对第一尺度下六个方向的小波系数的幅值信息进行求和处理后作为LBP统计过程中的权重值。在两个混合失真数据库(MDID2013和MLIVE)上的实验结果表明,利用所提算法进行评价的结果与主观评价结果拥有很好的一致性。然后,本文提出了一种基于稀疏表示与分解残差的全参考混合失真图像质量评价方法。近年来由于稀疏表示能够很好地反映人类视觉特性而成为图像评价领域研究的热点。但目前基于稀疏表示的图像质量评价算法主要适用于评估单失真类型图像,并未考虑不同失真类型对图像不同成份所造成的不同影响,同时也忽略了失真对分解残差的影响以及残差对稀疏系数的辅助作用。针对这些问题,本文算法将图像分解为纹理与卡通两部分,计算纹理稀疏系数与卡通稀疏系数,并考虑失真对局部残差与全局残差的影响。在MDID2013和MLIVE数据库上的实验结果表明,利用该方法的评价结果与主观评分有较好的一致性。更多还原

【Abstract】 Image information has been widely applied with the dramatic development of multimedia,and image quality assessment has become a very meaningful subject in the image process field.The research of assessing multiply-distorted images has very important realistic significance since the steps of acquisition,compression,and transmission might introduce multiple distortions.In recent years biologists have made important progress in the study of human visual system.Researchers of image quality assessment have proposed many methods based on the human visual characteristics,and a good result was achieved.Based on the review to precious works,this paper proposes two image quality assessment methods for multiply-distorted image based on the human visual characteristics.Firstly,this paper proposes a no-reference quality assessment method for multiply-distorted image based on dual-tree complex wavelet transform and local binary pattern.Compared with traditional wavelet transform,dual-tree complex wavelet transform have more directional selectivity and approximate translation invariance.This algorithm processes image with dual-tree complex wavelet transform before extracting statistical features of the amplitude of wavelet transform coefficients with local binary patterns operator.The method uses the sum of the amplitude of wavelet coefficients in six directions of the first scale as the weight of the statistical processing for purpose of highlighting the significant part of the image.The experimental results on two multiply distorted image databases(MDID2013 and MLIVE)show that the proposed method has better consistent alignment with subjective assessment.Secondly,this paper proposes a full-reference quality assessment method for multiply-distorted image based on sparse representation and residual.The research of assessing multiply-distorted images has very important realistic significance since the steps of acquisition,compression,and transmission might introduce multiple distortions.Sparse representation has become the hotspot in field of image quality assessment owing to the ability that could effectively represent the human perception vision for the past few years.However currently sparse representation-based image quality assessment methods mainly work for images with single type of distortion while neglecting to consider the discrepant influence that the various distortions cause on the different components of image.At the same time,they also ignore to think about the effect of distortion on sparse residual and the supplementary role of residual.To tackle these problems,this paper proposes a full-reference image quality assessment method for multiply distorted image.We decompose the image into two parts,i.e.texture and the carton,and calculate the texture sparse coefficient and the carton sparse coefficient respectively.Meanwhile we consider the local residual and the global residual as well.The experimental results on MDID2013 and MLIVE indicate that the proposed method achieves better consistent alignment with subjective assessment.更多还原