基于空间分布与统计学习的全参考图像质量评价

【作者】 王芳；

【导师】 刘暐；

【作者基本信息】 上海师范大学 ， 应用统计（专业学位）， 2019， 硕士

【摘要】 当今社会是一个互联网科技与媒体盛行的社会,我们每天从网络上海量的文字、图片、视频资源中获取各方各面的信息,它们早已成为人们了解外界所不可或缺的渠道,而这些资源的质量直接影响着所传递信息的质量,比如文字的准确度、图片的保真度以及视频的清晰度等等。本文从图像入手,针对那些有原始未失真图像作参考的图片,研究了三种度量其失真度的方法,得到质量评价模型,使评价结果尽可能地与人眼主观评价保持一致:第一种是基于图像的空间分布,给出最小差异(MSDI)、最大差异(MDD)、最小近邻差异(MSND)、最大近邻差异(MDND)和平均近邻差异(AVND)5种像素点矩阵的变换方式,通过比较变换后的失真图像和原始图像的差异来对其进行打分,这种差异是利用KS检验来度量的;第二种基于稀疏表示,通过K-SVD算法将图像矩阵分解为字典矩阵和系数矩阵,并将系数矩阵作为图像特征,计算两图像特征差异;第三种基于深度卷积神经网络,使用神经网络模型提取图片特征,并对比两图片特征的差异,从而得到目标图像的质量分数。后两种方法都利用了第一种方法的矩阵变换思想,首先将图像矩阵做变换,再进行相应计算。我们研究的对象是LIVE图像数据库里的JP2000压缩、JPEG压缩、白噪声失真、高斯模糊和快速衰落5种失真类型的图片。经实验表明,本文提出的方法在每种类型的失真图像上所得到的图像质量评分都能与人眼主观评分保持比较高的相似性,且基于稀疏表示和经过神经网络特征提取的IQA方法效果都有明显的提高,与现有的客观质量评价法相比也表现出了良好的准确性。更多还原

【Abstract】 Internet technology and media are popular in this era.We have obtained all kinds of information from the texts,pictures and video resources through the Internet,which have become an indispensable channel to open our eyes.Consequently,the quality of these resources is closely related to the effect of the information delivered,such as the accuracy of text,image fidelity and video resolution.Image is the entry point for this paper.Three methods to measure the degree of distortion of images with reference images are observed,which make the evaluation results as consistent as possible with the subjective evaluation of human eyes.The followings are the three image quality assessment models: the first one is based on the spatial distribution of the image which focuses on five transformation methods of pixel point matrix,namely most similar differences(MSDI),most dissimilar differences(MDD),most similar neighbour differences(MSND),most dissimilar neighbor differences(MDND)and average neighbor differences(AVND),and compares the differences between the original image and transformed image by KS-test.The second model is based on sparse representation.The image matrix is decomposed into dictionary matrix and coefficient matrix by the K-SVD algorithm,and the coefficient matrix is used as the image feature to calculate the difference between the two image features.The third one is based on the deep convolutional neural network.The neural network model is used to extract image features and compares the differences between the two images,so as to obtain the quality score of the target image.The last two models make use of the idea of matrix transformation of the first model: first,transform the image matrix,and then carry out the corresponding calculations.Research objects of this paper are five types of distorted images in the LIVE image database,namely JP2 K,JPEG,WN,GBlur and Fastfading.Experiments show that the image quality scores of each type of distorted image observed by the proposed method can maintain high similarity with the subjective score of human eyes.In addition,the effects of the methods based on sparse representation and neural network have been dramatically improved,which also show good accuracy compared with the existing objective quality assessment methods.更多还原