Visual Distortion And Saliency Computational Models For Stereoscopic Image And The Applications On Image Quality Assessment

With the development of stereoscopic image technology, stereoscopic image qualityassessmenthasbecome a hot research, andhavewidespread deepenedinto sectors, suchas medical,education, military and so on. But in the course of capture, all kinds of factors bring kinds ofdistortions, influence the quality of videos and images, thus, how to evaluate the distortions canpromote the development of steore video system. In this paper, the model of distortion map andsaliency map based on human visual system, as well as stereoscopic image quality assessmentalgorithmareestablished.The mainresearchcontents areas follows.(1) According to the characteristics of binocular vision, establish the model of distortion mapand visual attention for stereoscopic image, which are applied in image quality assessment. Firstly,the images from left and right eyes are combined into a cyclopean image, establish the model ofthe distortion map between the original and the distorted cyclopean images by using mean squarederror (MSE), then, based on the distortion map, combined with characteristics of the human visualsystem (HVS), such as luminance, contrast masking and contrast sensitivity function, obtaindetected distortion map; then using mathematical and statistical properties of local image blocks,such as standard deviation, kurtosis and skewness, obtain appearance distortion map. Secondly,based on the2D saliency model, combined with center bias, depth cue (foreground cue andbackground), to form three-dimensional visual attention model. Finally, based on2D metric, anobjective quality assessment method of stereoscopic images is proposed based on detecteddistortion map and appearance distortion map; then acquire3D saliency map by visual attentionmodel, and use the3D saliency map pooling2D image quality assessment method, an objectivequality assessmentmethodofstereoscopic imagesis proposedbased on3Dsaliencymap.(2) Two objective quality assessment methods of stereoscopic images are proposed based ondistortion map and saliency map. Firstly, combine with2D image quality assessment method, anobjective quality assessment method based on spectral and distortion analysis is proposed.Secondly, by analyzing the effects of appearance distortion map and3D saliency map onperceptual quality, separate stereoscopic image into four parts, and evaluate them independently,thenallevaluationresults areintegratedinto an overallscore.(3) Proposed two objective quality metrics based on gradient. Firstly, by using the threshold toobtain sensitive areas, and compute the horizontal, vertical and inter-view gradient information foreach pixel in sensitive areas, then construct the3D structure tensor matrix, the corresponding max eigenvalue and max eigenvectors are extracted from3D structure tensor matrix for each pixel. Bycomparing the similarity of max eigenvalue and max eigenvectors, the values of objectiveassessment are obtained for the original and distorted stereoscopic images. Secondly, Recreated3Dgradient model based on gradient information in all directions, use disparity space image to reflectthe depth information, and acquire gradient magnitude of the original and distorted stereoscopicimages. Finally, the values of objective assessment are predicted by comparing the similarity ofgradientmagnitudefortheoriginaland distortedstereoscopicimages.