Research On Quality Of Experience For Three-Dimensional Video System

In the context of remarkable progress of Internet and big data era,the visual effects have advanced by leaps and bounds.The rising of 3D video conferencing system,3D navigation,medical treatment and other advanced techniques,have brought great convenience to our life,but they inevitably bring new challenge and problems.In the systems for digital video acquisition,compression,transmission,and display,it is inevitable to bring in various distortions more or less,leading to a decrease in visual quality,which contributes to the degradation of the overall visual experience.Therefore,the experience quality in different stages of three-dimensional video system can be effectively monitored by quality of experience(QoE)assessment,which has great practical value.Based on visual perception,we study the QoE from the perspective single-channel image,stereoscopic image and single-channel video.The research contents are as follows:(1)According to the binocular visual perception characteristics in watching stereo image,and digging into feature extraction deeply,a novel no-reference stereoscopic image quality assessment(SIQA)based on binocular fusion and rivalry is proposed.Firstly,Gabor filtering features are extracted from the binocular fusion image formed by left view and right view image.Secondly,the natural features are abstracted from absolute difference map.Finally,two parts of features are incorporated to form stereoscopic image feature information and support vector regression is used to predict the objective scores.It is experimentally demonstrated that correlation coefficients are all about 0.94,better than the other representative image quality assessment methods.It is indicated that the proposed method has better correlation with human subjective visual quality evaluation compared with some other previous methods.(2)From the perspective of the movement information of the video,a motion perception of human vision is established based on multi-channel characteristics of Human Visual System(HVS).A novel multiband video quality assessment algorithm inspired by biological visual perception was proposed,which consists of motion perception quality index and spatial quality index.First,the original and the distorted video sequence were decomposed to five frequency bands with DOG filter bank to simulate the HVS.Then,the motion energy model is applied to evaluate the temporal distortion severity of each frequency component,which produces the motion perception quality index,and the gradient similarity measure is used to evaluate the spatial distortion of video sequence to get the spatial quality index.Finally,the random forest classification algorithm is adopted to establish the mapping between the quality vectors and the subjective opinion scores,accordingly predicting objective value of every video sequence.It is shown by the experiments that the proposed method has many significant improvements than the advanced methods using now,and PLCC value has reached to 0.84 on LIVE database.The proposed model has higher consistency with subjective perception.(3)Starting from the perspective that brain may be interpret things in terms of manifold.The geometry distortion of different distorted images can be effectively expressed by low dimensional manifold and can keep the invariance of nature in the manifold.According to the basic principle of manifold and online learning,a novel color IQA method based on online manifold learning is proposed.By means of the visual saliency model,maximum visual saliency and significant difference map are defined,which are used to select visual important areas.What's more,feature basic matrix is obtained by orthogonal locality preserving projection online learning.At last,the reference and distorted visual important patches are mapped to low dimensional submanifold using feature basic matrix and the manifold feature matrices are acquired,followed by the final color image quality evaluation value.There is prediction deviation in the case of asymmetrically distorted stereoscopic images and some previous SIQA methods are based on the luminance component of the images,so a new perceptual SIQA algorithm driven by tensor factorization and manifold learning is proposed for stereoscopic images.We firstly apply Tucker decomposition to RGB images to reduce dimensions along color channel to produce the training sets,and the projection matrix is obtained through manifold learning.Then,we adopt the manifold feature similarity to views,yielding the single estimates.Finally,considering the binocular rivalry behavior of visual perception,an adaptive binocular combination model based on local energy ratio is improved to predict the overall quality.The devised metric has better performance and achieves high consistent alignment with subjective assessment,compared with the state-of-the-art IQA metrics.