Image Quality Assessment Based On Information Theory

As the result of the imperfection of the imaging system and the unadvanced processing technology, the introduction of distortion is inevitable during the process of image acquisition, compression, processing, transmission and restoration, which brings difficulties to the following image and video processing, analysis and understanding. Thus, image quality assessment(IQA) is vital importance in different kinds of image processing systems.The essence of image distortion is the loss of the image information and human is the final observers of images. A systematic research about IQA is carried out based on the information theory and the human visual system, and the major contributions are as follows.(1) An IQA model based on mutual information in pixel domain is proposed. Firstly, the segmentation based on mutual information is applied to decompose the image into nonoverlapped patches that have minimum intra-regional diversity and maximum inter-regional diversity, which explores the relationship between segmented patches. Secondly, their relationship is described using the framework of information theory. Based on relative entropy, mutual information and conditional entropy, the image is decomposed to the saliency information which is good at simulating human visual perception, the specific information which provides a good approximation to image detail and the entanglement information which is sensitive to image structuration. Finally the degradations of image information in the distorted image are mapped together to obtain final image quality. Experimental results on standard databases demonstrate that these information is good representatives of image quality, and correlates well with human subjective perception.(2) An IQA model based on natural scene statistics(NSS) of image information is proposed. This method has a two-step framework which includes the training step and testing step. In the training step, firstly the image is decomposed into the saliency information, specific information and entanglement information using the framework of information theory. Then, the NSS of these information is measured to describe image local contrast, the image structure and the multi-scale and multi-direction image properties. Finally, the mapping relationship between the NSS features and the human subjective perception constructs the image quality prediction model. In the testing step, the image quality score of the test image is measured by inputting the test image into the quality prediction model. Experimental results illustrate that the proposed has a high consistency with human subjective.(3) An IQA model of contrast-distorted images based on image information is proposed. Firstly, the image is abstracted as decomposed into the saliency information, specific information and entanglement information using the framework of information theory. Then, the probability distribution of saliency information is measured to describe human visual perception; the probability distribution of specific information is measured to describe the image details; the probability distribution of entanglement information is measured to describe the image structure. Finally the probability distributions of image information are mapped together to obtain the final image quality. The experimental results on the contrast distortion database demonstrate that the proposed model has an excellent performance and a high consistency with human subjective perception.