Research On Video Quality Assessment Metric Based On Human Visual System

With the progress and gradual promotion of multimedia technology, single-view/stereo videois becoming one of the most important medias to obtain information. However, the process ofvideo transmission and video compression will inevitably bring some categories of distortions, andthe distortions will influence the consumers’ subjective perception seriously. Therefore, the studyof video quality assessment is important and necessary. This thesis studies human visualcharacteristics, and establishes visual perception models to evaluate video quality. The researchcontents are as follows:(1) Analyzing the sensitivity of the human eyes to the frequency and the influence offrequency on image composition, a video quality assessment method based on frequency domain isput forward in this thesis. Firstly, DCT is used for image segmentation to extract the low-frequencysub-image and mid-frequency sub-image, to which human eyes are sensitive. And thesesub-images are evaluated to derive spatial-domain quality of the video sequences. Secondly,considering the asymmetric perceptual characteristics of time domain and the sensitivity to thedistorted frame, low-pass function and worst-case pooling strategy are used to establish thetemporal perception model, and this model is applied to the time-domain weighting of the videospatial-domain quality. CC and SROCC of the proposed method are0.7976and0.7843respectively. These experiment results show that, compared with other classical methods, theperformance of proposed method in terms of evaluation and matching subjective ratings are better.(2) Considering the sensitivity of the human eyes to the image content, a video qualityassessment metric based on region partition and a video quality assessment metric based on thevisual perception are proposed in this thesis. The former divides the video frames into three regions,and considering the sensitivity to image content, different weights are assigned for different regionsto obtain spatial-domain quality. Then, time-domain perception characteristics are considered intime-domain pooling. Based on the former, the latter extracts the distorted edge region as theregion of interest in combination with JND, and the perceptual weighting is utilized in the region ofinterest. Then, spatial-temporal pooling uses the temporal perception model which is proposed inpart (1). CC and SROCC of the metric based on region partition are0.7383and0.7204respectively. These experiment results show that the metric based on region partition earns a goodperformance. CC and SROCC of the metric based on visual perception are0.8268and0.8242respectively. These results show that, the metric based on visual perception yields betterperformance and its evaluation results are more consistent with the subjective ratings, comparedwith other classical methods. (3) Taking the characteristics of compression distortions into account, a stereo video qualityassessment method for compression distortions is proposed. It contains two models, one is thequality assessment model of left-right viewpoint, the other is the quality assessment model of depthperception. Left-right viewpoint evaluation model uses JND to extract the region of interest inpreviously extracted salient points. The perceptual weighting is utilized in the region of interest,and the linear weighting is used for left-right viewpoint pooling by considering binocular maskingeffect. Depth perception evaluation model uses3D-DWT to extract the low-frequency structureinformation of disparity space images(DSI) indicating depth perception of human eyes. CC of theproposed method is above0.92, and SROCC is above0.91. Especially, CC of the proposed methodfor JPEG2000distortion is above0.95. The experimental results show that the performance of theproposed method in evaluation and matching the subjective ratings is outstanding.