Reaserch On Frame Rate Up-Conversion Of Compressed Video Based On Convolutional Neural Network

With the advent of the digital age,large-scale images and videos have become an important medium for information dissemination.Because of mobile Internet,people have become more enthusiastic about images and videos.However,the capacity of high-definition videos is increasing.Watching high-definition video requires a lot of traffic and because of limited bandwidth,the playback quality of high-definition video cannot meet people's need.Watching HD videos not only consumes traffic,but also affects the quality of playback due to limited bandwidth.Therefore,it is necessary for video compression before transmission.In order to reduce the bandwidth,it is necessary to transmit the compressed video at low frame rate.At the receiving end,the compressed video with low frame rate is restored to the original frame rate by interpolation.This is the frame rate up-conversion technology.The technology can not only improve the frame rate of video,but also reduce the discontinuity of the video in the process of video playback.It can greatly improve the user's visual experience.Therefore,the research of frame rate up-conversion has the important practical significance.At present,the popular frame rate up-conversion technology is motio estimation and motion compensation.This method generates the interpolated frames by calculating the motion vectors of each macro block.Although frame rate up-conversion technology has gradually matured,it still produces a lot of blocking artifacts in video when the motion of the block is violent.In addition,using the current mainstream compression standards to compressed video will reduce in the loss of information,which also increases the difficulty of generating interpolated frame by using frame rate up-conversion technology.In order to solve this problem,the paper proposes two algorithms for frame rate up-convcrsion of compressed video.The first algorithm uses a deep residual network to process the interpolated frame after the traditional frame rate up-conversion algorithm,which is equivalent to a post-filter.The deep residual network can effectively eliminate blocking artifacts.The network consists of three parts:feature extraction,feature recursive analysis and image restoration.The second algorithm uses two deep convolution networks to extract the features of forward and backward motion compensated frames generated by bilateral motion estimation and combines the output of the two networks into an initial interpolation frame.Then the initial interpolation frame is transmitted to a deep residual network in order to improve the performance of the initial interpolation frame.Finally,the deep residual network outputs the final interpolation frame.The experimental results show that compared with the traditional frame rate up-conversion algorithm,the two proposed algorithms can improve both subjective and objective quality of interpolated frame and effectively solve the problem of blocking artifacts.