Subjective And Objective Image Quality Assessment Based On The Human Visual System

Since the invention of the camera in the last century,we have entered the visual information age.Meanwhile,along with the development of the multimedia technology and new handheld devices such as mobile phones,how to provide endusers with high quality visual information has become a research hotspot.Digital image is one of the most efficient information media,and is used widely in our daily life.However,digitial image inevitably can be corrupted by multiple distortions during image processings such as acquisition,compression,transmission,enhance and refactoring.Image distortion will cause the degeneration of image visual quality,thus affect the visual experience of enduser.To provide enduser with the best visual experience,it is important to accurately assess the quality of digital image.The human eye directly records and perceives visual information in the images.Therefore,human subjective image quality assessment(IQA)should be the most accurate and effective measure of image quality.However,subjective IQA is time-consuming,expensive,and cumbersome.It is also affected by the observers'background knowledge,mood,fatigue,and other personal factors.The shortcomings mentioned above can cause undefined IQA results,which are unsuitable for real-time applications.It is therefore necessary to develop new objective IQA methods.Employing objective IQA method,we can automatically measure the quality of image and select usable images using computers instead of the human eye.Hence,objective IQA method is more suitable for practical application.In this thesis,we focused on full-reference IQA method and no-reference IQA method.Meanwhile,we propose one remote sensing IQA databases which can be used for verifying the performance of IQA method.The main motivations and contributions of this thesis are as follows:(1)Three blind IQA methods are proposed:1)a completely blind IQA based on gray-scale fluctuations(GFQA)is proposed;2)a completely blind IQA via image gray-scale fluctuations and fractal dimension analysis(BVGF)is proposed;3)a blind IQA via probabilistic latent semantic analysis(BVLS)is proposed.GFQA employs the frequency of particular gray-scale fluctuation values in the gray-scale fluctuation primitive map(GFM)to measure the image quality.On the basis of GFQA,BVGF employs fractal dimension for IQA.BVLS extracts four distortion-affected features to form the visual words in the dictionary,and the latent topics in the images can be discovered via the dictionary.The discrepancy between the frequency of the topics in an unfamiliar image and a large number of pristine images is applied to measure the image quality.The proposed methods are both blind IQA methods which remove the effects of human opinion scores,thus they are more practical and easier to apply.Experimental results show that the newly proposed methods correlates well with human subj ective judgments of diversely distorted images.(2)A no-reference IQA based on sparse representation(SR_IQA)is proposed.The proposed method first divides the GFM of each image into patches of fixed size and the patches rearranged into column vectors.The column vector is regarded as a structural element of the image.By using sparse coding,the structural elements can be represented by sparse representation coefficients and a trained dictionary.Then,by using the sparse representation coefficients,the probability vector for observing different elements in the trained dictionary can be obtained.Finally,the quality prediction model is trained using support vector regression(SVR).The experimental results show that the proposed method can accurately predict human perceptual image quality,and is competitive in comparison with the existing no-reference IQA methods.(3)Remote sensing IQA via image usability:1)a usability-based subjective remote sensing IQA database is proposed;2)a no-reference IQA based on image usability(UB_IQA)is proposed.The proposed remote sensing IQA database can be used as credible verification platform for objective IQA method.The proposed no-reference IQA method takes fixed sized blocks from real remote sensing images and extracts the scale invariant feature.Then,the proposed method employs K-Means clustering algorithm to cluster the features and uses the clustering centers to form a visual word dictionary.After that,we extract scale invariant features from the training images and calculate the appearance frequency of different visual words in each training image.Probability vectors are used as the quality prediction feature.Finally,we use the support vector machine(SVM)to train the quality prediction model,the two inputs are the probability vectors and the corresponding human opinion scores of the training images.The experimental results show that the quality predictions of the new no-reference IQA method correlate well with human subjective scores in the usability-based subj ective remote sensing IQA database.(4)An IQA via spatial structural analysis(SA_IQA)is proposed.Modeling of image structural similarity has been regarded as suitable for achieving perceptual quality predictions.However,most structural similarity-based IQA methods focus on spatial contrast without fully considering the spatial structural distribution.Hence,we propose an IQA method that considers both spatial contrast and structural distributions the spatial structural distribution.SA_IQA calculates the spatial structural information variation matrices(SSVMs)between the GFMs of distorted and pristine images.And,the quality prediction model is trained using SVR.The experimental results show that the proposed method can accurately predict human perceptual image quality.Meanwhile,SA_IQA is suitable for remote sensing IQA.