Reseach On Image Quality Assessment Based On Visual Perception And Local Feature Extraction

With the adventure of high-definition digital era, person tends to pursuit of highdefinition and fidelity of visual information. But the acquisition process of visualinformation is likely to produce some kinds of distortions, which would reduce thequality of visual information and have a great influence on people’s access to and use ofvisual information. The change of visual quality needs to be known in the process ofoptimization. Therefore, the thesis study on the quality evaluation of visual informationhas important theoretical significance and practical value.Building upon work in the study of traditional image quality assessment methodbased on structure feature extraction, the thesis extracts the structure characteristics ofthe image by using gradient singular value decomposition, and builds an evaluationmodel based on local singular value decomposition. In order to extract all of thestructure information, the thesis establishes the evaluation model according to thedistribution of variance and gradient of the image. Experimental results demonstratethat the algorithms described above can evaluate the image quality effectively.To overcome the limitations of traditional image quality assessment methods,which not well be consistent with subjective human evaluation, this thesis proposed analgorithm based on wavelet analysis. The arithmetic simulates nerve network of humanvision system, which decomposes an image into four sub-band images by wavelettransform. Dividing the sub-band image into blocks at size of8×8, then using fastindependent component analysis to train the image blocks. After that, extracting eachimage block sparse character matrix to calculate the sparse feature fidelity of the imageand build the sparse fidelity quality evaluation model. On this basis, adding weight tothe model of image sparse fidelity by partitioning method and get the evaluation model.The experiment shows that the proposed method can effectively simulate the weightedvisual cortex of the human visual system perception mechanisms, which compensate fordeficiencies of existing image quality assessment methods.In this thesis, a no-reference image quality assessment based on the characteristicsof natural distribution is proposed. The method analysis the local amplitude and localentropy of wavelet image, and extract the natural statistical characteristic vector of theimage through the statistical distribution under the condition of different distortions. Then a mapping relationship between the statistical property and the object quality isestablished through the use of support vector regression and classification method.Theexperimental results show that the proposed method has a good consistency with thesubjective perception quality, and the performance of the proposed algorithms is superiorto some existing methods.