Research On Compressed Video Quality Enhancement Methods Based On Spatiotemporal Information

Recently,video resources are spread more and more widely on the Internet.In order to reduce the bandwidth required for transmission and maintain the video quality as high as possible at the same time,researchers have proposed many video compression methods and formulated video coding standards.However,both traditional video compression method and deep learning based video compression model use different compression modes for image frames at different times during the compression.This can lead to quality fluctuations in the video.In recent years,researchers have proposed many deep neural networks to exploit the quality fluctuations in compressed videos to improve the overall video quality.The input frames of these networks contain the target frame to be enhanced and several reference frames.In these networks,motion estimation and motion compensation methods are often used to extract information between video frames at different times.However,due to the different quality of the target frame to be enhanced and the reference frame and the possible occlusion problem,the accuracy of the motion estimation in the network will be affected.In addition,these methods do not take into account the different quality of image frames and the different amounts of information contained in different positions in the image frames.To deal with these problems,we estimate the motion information between reference frames additionally.We also introduce a blended attention mechanism to weight the attention of features at different spatiotemporal locations.On the other hand,deformable convolution can also be used to extract spatiotemporal information in input frames and perform implicit motion estimation and motion compensation,but current methods do not attempt to combine it with the peak sampling strategy of input frames.We combine the deformable convolution and peak sampling strategy,expanding the deformable convolution into the fully spatiotemporal deformable convolution to improve the overall performance of the combined network.We train our proposed network on publicly available largescale video datasets and test it on standard video test sequences.Experiments verify the effectiveness and robustness of our proposed network.