Visual Perception Characteristics And Video Quality Assessment Based On EEG Signals

As the main carrier of visual information,video usually suffers from quality degradation in the process of acquisition,transmission and display,resulting in the wrong expression.Therefore,effective video quality assessment(VQA)is important in information communication,algorithm optimization and visual experience improvement.However,due to the complexity of human visual characteristics,it is difficult to reproduce the real quality perception process through mathematical models.Therefore,by collecting the electroencephalography(EEG)signals triggered by different quality of videos,this thesis analyzes the potential variations of EEG signals under the visual perception characteristics according to the different perceptual responses caused by the phased distortion videos with different distortion duration and videos with or without significant target.Also,a VQA model is established based on EEG samples from multiple subjects to solve the problem of training data scarcity and negative transfer.The purpose of this thesis is to explore a VQA method in accordance with human visual cognitive process by referring to the video perception characteristics of human vision.The main content is summarized as follows.1)A phased distortion video perception quality assessment method based on EEG is proposed.The distortion perception of human vision is affected by the distortion duration of the video.Therefore,this method first selects a set of parameters whose visual effects are near the visual distortion perception threshold to generate multiple videos with different distortion levels and distortion durations as experimental stimuli.Then,based on the improved Odd Ball paradigm,the phased distortion VQA subjective experiment is designed,during which the behavioral data and EEG signals are collected.Finally,EEG signals are processed,analyzed,compared and mapped into concrete quality scores by constructing objective and subjective indexes.The experimental results show that P300 can characterize perceptive video distortion,and its amplitude is positively correlated with the degree of video distortion.For the same degree of distortion,the longer the duration is,the higher the peak achieves.Moreover,there exists a distortion perception threshold under the combined effect of distortion duration and distortion degree,within which the sensitivity of human vision to video distortion increases rapidly.2)A significant target video perception quality assessment method based on EEG is proposed.The selective attention mechanism of human visual system leads to the phenomenon that human perception of video distortion can be influenced by significant target in video.Therefore,this method first designs an eye-tracking experiment and analyzes the eye movement patterns of videos to obtain a group of contrast videos with and without significant targets.Then,the significant target video VQA subjective experiment is designed.Five quality levels are set to generate video stimuli,and the EEG signals triggered by the video with or without significant target in different distortion levels are obtained.Finally,the EEG signals of the two groups of videos are compared and analyzed.The experimental results show that the perception of video distortion is affected by the significant target in the video,and the extent of influence is related to the degree of video distortion.In the case of slight video distortion,there is a significant difference in P300 amplitude between the videos with or without significant target.3)An inter-subject EEG transfer learning based on adversarial framework for video quality classification is proposed.Due to individual differences,deep learning and EEG based VQA methods face the dilemma of data scarcity,and the training strategy based on multiple subjects is easily affected by negative transfer.Therefore,this method first trains a common feature extractor using the data of multiple source subjects and part of the data of target subject to extract the components related to video quality in EEG samples.Then,a subject discriminator is constructed to conduct adversarial learning with the common feature extractor to obtain the domain invariant features of the source and target subject domains.Finally,the features including video quality information are used to predict the video quality.The experimental results show that the proposed method can reduce the distribution differences between the subjects and improve the prediction performance of video quality.