Video Quality Assessment Based On Visual Perception

With the rapid growth of multimedia applications,more and more efforts have been put into the field of video quality assessment(VQA).In this work,in order to extract more accurate visual quality which reflects the subjective perception of human-eyes from a video sequence,human visual system(HVS)characteristics are considered,and the work is based on the concept of visual perception.The whole procedure of VQA is divided into two phases:single-frame image quality assessment(IQA)in the spatial domain,and quality pooling in the temporal domain.During the first phase,existed IQA model is used to compute a frame-level score,and the score of each frame is integrated into a single sequence-level score by temporal pooling strategy.Based on the above procedure,the goal is to extend an existed IQA model into a VQA model.Therefore,a spatial domain and a temporal domain VQA model are proposed respectively.In the proposed spatial domain VQA model,a more appropriate weighting scheme for the quality map computed by an IQA model is supplied,which matches HVS characteristics by incorporating visual attention model and optical flow to consider visual gaze and motion information in a video sequence.Due to the limitations of the saliency map information computed by a visual attention model in the VQA task,a uniformity of distortion distribution concept is proposed to emphasize the contribution of video distortions to the saliency map result.In order to improve the mean-based pooling scheme used by most IQA models which does not consider HVS characteristics,an adaptive spatial pooling scheme is proposed which incorporates percentile pooling and the uniformity of distortion distribution index following the idea of balancing global and local quality.In the proposed temporal VQA model,as a typical way to extend an IQ A model to a VQA model is to take the mean value of the frame-level scores computed by the IQA model,which does not take HVS characteristics into consideration,visual psychological effect is modeled based on the concept of visual perception,and a memory effect model is proposed to process the frame-level scores to match HVS characteristics.Then,the concept of distortion uniformity is extended to temporal domain,and is integrated with the proposed temporal pooling strategy based on percentile pooling to integrate the frame-level scores into one objective video quality score.Existed IQA models are combined with the proposed VQA models to evaluate the performance with other IQA and VQA models in two public-available video databases.Analyses on the experiment results and statistical significance results indicate that among the models in the experiment,one of the proposed models shows best prediction performance,and it performs statistically better than or equal with two state-of-the-art VQA models.