| With the rapid development of mobile communication technology and the improvement of the performance of terminal display equipment,video has gradually replaced image as the primary information communication medium.The research on professional generated content video and user generated video has attracted widespread attention.However,in the process of data collection,transmission,display and storage,the video is inevitably affected by the complex external environment,which causes its quality to decline and affects the user experience.There are obvious differences in the causes and forms of quality damage for videos under different generation backgrounds.For professional generated content such as stereo video and virtual reality,the quality damage is concentrated in the transmission and display process.For user-generated content,quality damage is mainly caused by the limitations of shooting equipment and shooting technology in the data acquisition process.Therefore,accurate and objective evaluation of the quality of specific generated content has become an important subject in the field of computer vision.In addition,the establishment of different computational models based on visual perception theory plays a key role in the quality evaluation of specific generated content.In this context,this thesis studies video quality evaluation based on visual perception.Firstly,two objective evaluation methods of stereo video quality based on motion perception model are proposed.On the one hand,according to the motion masking and the spatical-temporal property,the key frame extraction strategy and dynamic texture description based on motion perception are proposed,and the efficient stereoscopic video quality evaluation is established.Based on the above research and binocular suppression theory,a binocular fusion model suitable for asymmetric distortion perception is derived.Then,the proposed dynamic texture feature is improved to realize the linkage with the temporal domain on the premise of preserving the spatial structure.Finally,the stack autoencoder is introduced to construct the perceptual depth model for data regression,which improves the accuracy of the objective quality evaluation model.Secondly,this thesis proposes an objective evaluation method of stereoscopic video quality based on multi-domain joint perception to complete the quality estimation problem of compression loss.Considering binocular competition and crosstalk,the spatical-temporal domain of stereo video is represented as interframe cross map.Then,the multi-scale amplitude and phase statistical features in the DCT domain are extracted in the spatial and spatical-temporal domains.Finally,the quality regression strategy combined with support vector regression and dictionary learning is established by combining the characteristics of spatial domain,spatical-temporal domain and temporal domain quality perception features,which improves the performance of model.Thirdly,this thesis proposes a virtual reality quality evaluation method based on graph-level representation to solve the problem of quality loss estimation caused by compression coding.Combining with the general visual perception process of virtual reality,a two-stage quality evaluation algorithm of visual perception and spatial structure perception is proposed.According to the low geometric distortion of cube projection and spatial relations of viewpoints,the first stage uses shallow network with residual structure to extract quality perception features of local viewpoints.In the second stage,a graph structure is established to describe the spatial structure loss according to the relative position of viewpoints,so as to achieve high performance endto-end virtual reality quality evaluation.Fourthly,this thesis proposes an attention-based quality evaluation for user generated content to solve the problem of impaired video quality evaluation caused by complex factors such as filming equipment and filming technology limitations.Specifically,this method explores the effect of visual memory on quality perception and uses Transformer to establish long-term dependence on the semantic features between frames.At the same time,non-local blocks are integrated into the shallow network to extract quality perception features at the global level.Then,shallow quality perception features are fused with deep semantic features,which improves the comprehensiveness and effectiveness of user generated content quality evaluation.The above four aspects of research focus on the evaluation methods of professional generated content and user generated content driven by different visual perception theories,which are carried out in detail from the visual theory and subjective perception.Abundant experiments verify the performance of the proposed methods.This study further verifies the driving effect of visual perception model on quality evaluation and provides research direction for exploring general video quality evaluation methods. |
