Study On The Influence Of Color And Human Visual System On Image Quality Evaluation

Images are the main way of human to understand the world in their work and lives. The measurement of image fidelity and intelligibility is the important standard of the process of human cognitive objective things. However, images may suffer different types and degrees of distortion in the image life cycle, leading to the mistakes in the process of image analysis and understanding. Objective image quality metrics aim to simulate human visual system to realize the measure of digital image quality by computers. Therefore, it becomes to be an important topic.This paper mainly focuses on the basic problems of objective image quality assessment. We explore the characteristics of the saliency and color space and analyze the import on saliency applied to image quality assessment, and study proper saliency adding strategies. Meanwhile, we study the influence of color information to image quality. By analyzing each color channel and making quality model, we propose a no-reference image quality assessment based on color information, which provides practical applications of image quality assessment. The main contents are summarized as follows:(1) According to imports on saliency applied to image quality assessment, a proper saliency adding strategies is proposed. Based on the characteristics of the attention mechanism of the human visual system, the influence of distortion type and degree is exploded. Experimental results demonstrate that the variation in saliency highly depends on the distortion type and degree. What is more, we propose different saliency adding strategies depended on distortion type. Meanwhile, we explore effects of saliency extraction algorithms for image quality assessment, and put forward some suggestions to choose the proper one.(2) No-reference image quality metrics based on the Lab space are proposed. Getting rid of the bondage of the situation that no-reference image quality metrics only utilize features from the luminance channel, we carry on the analysis of color channel information in the Lab space, and find features that can reflect image quality. We propose two methods combining luminance features and the natural scene statistics features or two-dimension entropy features in the b channel, respectively. The process of the performance experiment on the five public databases and database independence experiment demonstrates our methods on the precision are superior to the existing metrics and have stable performance on algorithm parameters.(3) A blind image quality framework based on proper color channel is proposed. On the basis of natural scene statistics in the luminance channel and making use of mathematical models to construct a generalized pattern of color channel information, we extract features to reflect image quality variation. Additionally, we conclude that different distortion types fit different color channels. Therefore, we design a blind image quality assessment framework based on color statistic, which chooses proper color channels to assess image quality. Experimental results demonstrate that the proposed framework improves algorithm accuracy by using different color channel information. Moreover, the implementation of our framework is not limited by the actual image quality assessment features or the classifier. More effective features can further improve the performance.