Image Quality Assessment Based On Nature Scene Statistics

The image is a visual symbol which is the visualization with similarity and vividness of object perceived by human. However, the process of acquisition, compression, processing, transmission and restoration introduce different distortions to the image, and thus affecting the human perception seriously. Therefore, it is necessary to design the effective objective image quality assessment(IQA) algorithms which can be utilized as the criteria for parameter optimation of image processing algorithm, performance evaluation of image analyzing system, and quality of service monitoring of image recognition apparatus. Thus the system can provide high quality images for the human.In this thesis, a systematic study about nature scene statistics of reference image will be changed with the distortions. The author's major contributions are as follows.(1) Image quality assessment based on Shearlet transform is proposed. By combining the ability of Shearlet transform to capture the visual perception features and the properties of human visual system to describe the variation of image quality induced by distortions. Experimental results illustrate that the proposed has a high consistency with human subjective quality score.(2) Image quality assessment based on generalized Gaussian distribution is proposed. The model parameters of Shearlet coefficients of image without information loss are calculated based on generalized Gaussian distribution, and the obtained model is used as the standard distribution. Then the distorted image quality score is acquired by introducing a simple difference metric between the standard distribution and the distribution of Shearlet coefficients of distorted image. Experimental results verify the rationality and validity of the proposed.(3) Photographed image quality assessment based on visual perception is proposed. Under the principle of photographed image distortion, the color uniformity of the distorted image is described with the statistical characteristics of hue and saturation, the exposure level of the distorted image is characterized with the statistical characteristics of brightness and the change in the structure information between the distorted image pixels is measured with gradient information of grayscale image. The above factors are integrated to predict the quality of photographed images. Experimental results on photographed image database demonstrate the proposed reflect the real feel of human perception.Extensive study has been conducted from image quality prediction with partial information of reference image to no reference image quality assessment. Furthermore, the object of the designed algorithms from artificially distorted images to distorted images generated by multi-cameras photograph. The overall research on image quality assessment based on nature scene statistics follows the gradual process that is from elementary to the profound and the practicability is enhanced gradually.