Research On Perceptual And Statistical Image Quality Assessment And Its Application

With the fast popularity of social networking tools, such as twitter, wechat, facebook and QQ, images have penetrated into our daily lives. Since the year of 2011,the whole world has generated over ten billion digital images every year. The research for analysis, screening, recognition, assessment, surveillance, restoration and enhancement of huge amount of visual data has been a hot topic, and aroused broad attention of researchers in multimedia, image processing, computer vision and pattern recognition fields. Based on perceptual features of the human visual system, this thesis mainly explores the image quality assessment and its applications. These problems covering the psychovisual, statistical information and image processing fields are nowadays popular and difficult. The researches for image quality assessment and its applications are able to deeply reveal the psychovisual properties, complete natural scene statistical models, and improve the performance of image processing systems.This thesis mainly includes three respects as follows:(1) Study the mechanism of full- and reduced-reference image quality assessment, for modeling the image degradation problem caused by various kinds of distortion types and varying viewing distances. Full- or reduced-reference assessment works to predict image quality with the known(complete or partial) original image. Supposing that the human visual system is highly adapted to image structural information, this thesis proposes the full-reference local tuned global model, by analyzing the low-level visual feature, computing the gradient magnitude similarity between the original and distorted images, and combining two strategies of global and local poolings.Subjective opinion scores are remarkably different when seeing the same image at different distances. This study is more approximating to the real condition, and is more challenging. Due to the fact that increasing the viewing distance will reducethe ability to distinguish details, this thesis proposes the optimal scale selection model for image quality assessment, by removing partial high-frequency information in the wavelet subbands based on the known viewing distance, and properly scaling images according to the trigonometric function among ideal visual angle, viewing distance and image resolution as well as the aspect ratio of the image.Under the condition that only the partial original information or some extracted features are available in transmission, reduced-reference quality metrics become the optimal choice. Supposing that the human visual system is very sensitive to image structural variation, this thesis proposes the structural degradation model based reduced-reference quality metric, by extracting features from the original and distorted images by different convolution kernels due to the fact that images with various distortion types and levels will appear different spatial frequency decreases, and using nonlinear strategies to measure the distance of above-mentioned features.(2) Study natural scene statistics and the joint effect across multiple distortions,for modeling no-reference image quality assessment. No-reference assessment works to predict image quality without access to the original image. Based on several new statistical features for no-reference assessment, this thesis proposes the free energy based robust no-reference quality metric, by establishing the linear relationship between free energy features and structural degradation features, computing other two types of features based on the human visual system and natural scene statistics, and learning a module with machine learning tools on extracted features.Compared with single distortion type, image degradation caused by multiple distortion types also includes the joint effect across different categories of distortions.Combining image processing methods and free energy based brain theory, this thesis proposes the six-step blind quality metric, by computing each underlying degradation level with classical methods for noise estimation, JPEG compression evaluator and blurriness measure, calculating the joint effect based on the free energy principle, and making a linear fusion for aforesaid measures.Different from distortion measures, the quality assessment for low dynamic range images that are created from high dynamic range images via tone mapping operators is another kind of new and difficult problem. From the respects of information, natu-ralness and structure, this thesis proposes the blind tone-mapped image quality index,by measuring the information amount of the intermediate images processed by brightening and darkening the low dynamic range image, computing the naturalness of the low dynamic range image, estimating the ability of the low dynamic range image to preserve the primary structures, and learning a module with machine learning tools on extracted features.(3) Study the quality metric based image processing technologies. Most existing contrast enhancement methods generally require the human assist, such as parameter tuning and output selection. The complete histogram modification framework and automatic contrast enhancement method are of higher values. Integrating subjective and objective assessment, this thesis proposes the automatic contrast enhancement technique based on saliency preservation, by completing the existing histogram modification framework with a new effective transfer mapping, and combining entropy increment and saliency preservation to devise the reduced reference quality metric and to optimize the histogram modification framework.Remote controlling and cloud computing have intrigued a growing number of applications, and related screen content image/video technologies are thus becoming current hot topics. Existing metrics fail in assessing the visual quality of screen content images. Based on the assumption that human eyes pay more attention on the text areas in screen content images, this thesis proposes structure-induced quality metric and accordingly devises a new ratio-distortion optimization strategy for improving the efficiency of screen content video coding, by adopting structural degradation model to search for the saliency regions, using structural similarity metric to quantify distortions, and making an integration of the above two.