【作者】 李楠；

【导师】 丁勇；

【作者基本信息】 浙江大学 ， 电路与系统， 2016， 硕士

【摘要】 在数字图像的获取、压缩、存储和传输过程中,由于存在获取设备的缺陷、压缩编码、存储错误和传输错误等问题,使得图片的质量下降,增加了人对图像内容识别的障碍。另一方面,在数字图像处理等领域中,图像质量的好坏直接代表了算法的性能优劣。而且图像质量作为一种重要的评价指标,可以用作优化图像处理系统参数的重要反馈。但是由于人对于数字图像的主观质量评价实验复杂,受实验环境影响大,且不可重复等各种因素,难以被大型应用场景采纳使用。客观图像质量评价作为一种预测数字图像质量,得出与人的主观评价结果相一致的机器算法,开始被研究人员广泛关注。稀疏处理是当前信号处理领域的研究热点,其优势体现在对信号进行高效表示,利用少量非零变量表征原始的大量数据,降低信号处理的复杂性。本文研究了稀疏处理的原理,并基于稀疏处理方法提出了自然图像质量评价方法,概括为:(1)调研了稀疏处理方法的基本原理,研究了核独立分量分析算法的原理和相关应用,利用核独立分量分析对数据进行非线性映射使得线性不可分的信号非线性可分的思想,设计了基于核独立分量分析的客观图像质量评价方法。方法对提取出的特征进行了基于自然图像统计学的统计分析,利用相关系数与图像质量产生映射关系。经过实验验证,核独立分量分析对图像数据进行分解得到的分量,作为有效特征可以较为精确的预测图像质量。(2)由于分量的独立性对于图像质量预测的精确性有直接影响,如何提取出更加独立有效的特征成为了基于核独立分量分析的图像质量评价方法的关键。使用频域距离作为块匹配方法的匹配要求,能够对图像采样数据进行高效的筛选。之后,经过主成分分析和核独立分量分析分解图像采样数据得到特征。经过实验证实,利用此种方法获得的特征经过质量综合,能较为明显的提升图像质量的预测精度。最后,论文总结了稀疏处理应用于图像质量评价框架中的思想和关键,说明稀疏处理对于自然图像质量评价的研究意义,总结了论文中基于稀疏处理方法中的分量分析的图像质量评价方法,对未来工作进行了展望。更多还原

【Abstract】 On one hand,image quality degrades because of imperfect acquisition,compression coding,storage and transition faults which increases the difficulty for human recognition.On the other hand,in research fields on digital image processing,processed image quality comparison directly demonstrates the performance of algorithms.Also,image quality is an essential assessment standard and can be embedded into image processing systems for feedback to optimize parameters.However,since experiments are complicated,sensitive to experimental environment,and unable to be duplicated,subject image quality assessment(IQA)by human is difficult to be applied in large application scenarios.Objective IQA,as evaluation algorithms to predict image quality in accordance with objective assessment results,starts to draw researchers’ wide attention.Sparse processing is a hot research topic in signal processing.Sparse processing methods are efficient to deal with large amount of complex data with less non-zero variables to represent original data and this makes it easy to be analyzed.In this.paper,sparse processing theory is discussed and IQA methods based on sparse processing is proposed.The summary is given below.(1)A brief survey on sparse processing theory is done.The theory and application of kernel independent component analysis is studied.A full-reference objective IQA method is proposed based on kernel independent component analysis based on the fact that kernel independent component analysis can do non-linear decomposition in reproduced Hilbert kernel space.The proposed method uses correlation between two sets of features from reference image and distorted image to map onto evaluated quality value based on natural image statistics.Experimental results demonstrate that features extracted based on proposed methods can predict image quality well.(2)During experiment,it is found out that the independency of components among one set directly affects the precision of image quality prediction.This leads to the question that how to extract more independent and effective features relates to the performance of methods based on independent component analysis.Block matching methods using frequency domain distance as matching condition select image samples efficiently.After further decomposition of principal component analysis and kernel independent component analysis,the precision is improved obviously.At the end of this paper,the thought and key to apply sparse processing to IQA framework is summarized.Sparse processing is promising in IQA research and the proposed methods in this paper based on component analysis is summarized.And future work is given.更多还原