Research On Stereo Video Quality Assessment Model Based On Human Visual Perception

The rapid development of 3D image/video technology provides technical support for sensory requirements of human. The expectation of human beings’ visual perception is higher. Thus, stereo video quality assessment(SVQA) has been becoming an important topic in 3D video field for seeking effective solutions to accurately evaluate quality. Considering visual perception characteristics in human visual system(HVS), we need to establish an effective SVQA model.Based on the traditional stereo image quality assessment(SIQA) framework, a stereo image quality assessment model is proposed. Then, in order to research a more effective model which conforms to the human visual perception, binocular fusion and binocular rivalry are considered, and a new RR SIQA model based on binocular HVS is proposed. In the next place, considering temporal movement characteristics of stereo video, a temporal weighting model for stereo video is proposed. Combining temporal weighting model and SIQA model, the final SVQA model is established. The research contents are as follows:(1)Research on stereoscopic image is the basis for stereoscopic video. Through analyzing the image multi-channel characteristics and Contrast sensitivity Functions of HVS, a reduced-reference(RR) stereoscopic SIQA model based on wavelet transform and singular value decomposition(SVD) is proposed. In this model, since the multi-channel characteristics of left and right stereo images by wavelet decomposition and singular value of images contains its properties, SVD is used on sub-bands of wavelet decomposition. We get the singular value extracted image features as a eigenvalues of the RR model. Combining each sub-band information entropy and contrast sensitivity feature, we get the final quality prediction. The experimental results show that the correlation coefficients of the assessment and the subjective perception are satisfactory. In particular, the evaluation of all the type of distortion is good. It indicates that the stereoscopic objective model is consistent with human visual properties and availability.(2) Considering the processing of binocular perception in watching stereo image, a new RR SIQA model based on binocular HVS is proposed. The stereos images are divided into binocular fusion portion and binocular rivalry portion with the guide of internal generative mechanism. Then the binocular fusion portion is fused to cyclopean view. By extracting reduced feature of cyclopean view and binocular rivalry portion, the quality indicators of binocular fusion portion and binocular rivalry portion can be obtained. Experimental result shows that the proposed model outperforms some other models, and its correlation coefficients of assessment versus the subjective perception are satisfactory over five types of distortion.(3) Considering the human visual perception of temporal masking and combining the above SIQA model of binocular vision in the HVS this paper proposes a RR SVQA method based on temporal RR quality evaluation model. Firstly, analyzing temporal movement characteristics of stereo video, the RR frames have been extracted. Secondly, according to video binocular fusion and binocular rivalry in the process of binocular perception, the stereo video RR frames are divided into binocular fusion portion and binocular rivalry portion respectively, and the RR indicators are obtained from the two portion. Finally, we compare the RR indicators of original and distortion image, and pooling the quality indicators of RR frames to gain the stereo video quality. The test results of compression artifact distortion on NAMA3DS1-COSPAD1 stereo video database show that the proposed method has high correlation with the DMOS scores and consistent human stereo visual properties excellently. Correlation Coefficient of the proposed method has achieved 0.9487.