Video Quality Assessment Based On Spatio-Temporal Statistics

Nowadays digital video is becoming more and more popular in our modern lives as different video applications are developing in a fast speed,such as digital television,digital cinema,and digital video conference.However,as the result of the inappropriate progress during video production,video compression,and video transmission,inevitable distortion is introduced to video,leading to annoying quality degradation.Thus,it is necessary to keep surveillance in time to guarantee the quality of the video service.Recently objective video quality assessment by computer vision has become a hot research area.This paper incorporates human visual perception properties,and carries out a deep research on objective video quality assessment.The contributions are as following:(1)A VQA model based on visual saliency distribution is proposed.In the proposed method,video saliency is detected based on Random Walk with Restart(RWR).Salient map is obtained by finding the steady-state distribution of the random walker.The saliency map is separated to salient region and unsalient region.And for salient region the gradient modulus similarity and luminance similarity is calculated as the quality index of the attention area.In addition to salient region,gradient similarity is also used to compute quality degradation for unsalient region but with a relatively small weight as unsalient region also contributes to the visual quality perception.Quality index for salient and unsalient region is combined as the frame quality index.Lastly,with percentile pooling strategy,all frame quality indices are pooled to obtain the video quality.Experimental results on LIVE video database and EPFL video database demonstrate the proposed model has excellent performance and has a good consistency with human perception.(2)A no reference VQA model based on the statistics of the coefficients of three dimensional discrete cosine transform(3D-DCT)is proposed.In the proposed method,firstly the coefficients of video blocks are calculated.And then four features including natural video statistics shape parameters,spectrum ratio,kurtosis of the alternating current bands,and the measurement of the blocking effect are extracted based on the 3D-DCT coefficients.Frame-level features are pooled to form the sequence-level features via temporal pooling strategy.The sequence-level features are then regarded as the input of neural network with subjective quality score.After training the trained neural network can be used to automatically predict the quality of video.The proposed algorithm is tested on LIVE video database and experimental results demonstrate the proposed model has excellent performance and has a good consistency with human perception.(3)A no reference VQA model based on natural video statistics is proposed.The proposed model doesn't need the appearance of subjective quality score.Firstly the frame in video is separated into patches,then the feature of natural scene statistics of high quality video is extracted,and the most representative scene patches in the scene are selected.Multivariate Gaussian Model is used to fit the statistics of these selected patches.Video quality is obtained by the difference between the parameters of high quality video MVG and test video MVG.The proposed algorithm is tested on LIVE video database,and experimental result demonstrates the proposed algorithm has a good consistency with human subjective perception.