Research On Objective Image Quality Assessment Based On Human Vision Characteristics

With the rapid development of multimedia technology and modern communication technology, people have been in a new era of pursuit of high resolution and high fidelity of visual information. Since the information carried by image is more intuitive, rich, and efficient, it has become indispensable visual information for people to perceive the objective world. However, the images will inevitably introduce various types of distortion in the process of image acquiring, processing, encoding, conveying and storage, causing the change of image quality, which is not conducive for people to correctly understanding the objective world. Therefore, the study of reliable and efficient objective image quality assessment method has very important practical significance and application prospects.Aiming at the problem that the traditional image quality evaluation metric which is based on human visual system is not fully considered the higher level aspects of visual characteristics, this paper proposes a method of image quality evaluation based on eye movement saliency map. In the analysis of the importance of visual attention to the image quality perception, this method design eye movement experiment to collect eye-tracking data,and then use the eye-tracking data to calculate image visual saliency map, which has been as a factor weighting with image distortion index map to obtain the image quality measure.Experiments on LIVE image database show that the consistency between the proposed algorithm and the subjective perception result better than traditional methods.For the existing quality evaluation metric based on image statistical features of natural scene can't effectively simulate the characteristics of human visual perception, a method of image quality assessment based on extracting image joint gradient and phase congruency features is proposed. First, the image gradient magnitude and the Gaussian response of Laplace are calculated, followed by analyzing the relationship between statistical properties and the image distortion level, getting the joint statistical characteristics of gradient through joint adaptive normalization. Then, from the point of visual perception, the phase congruency features are extracted. Finally, extreme learning machine is adopted to build the relationship between image features and subjective quality scores. Experiments on three image databases indicate that the proposed metric correlates well with visual perception result and feature extraction is effective.