Research Of HEVC Multiple Description Video Coding Based On Deep Learning

With the advent of the 5G era,the development of video products has been redirected to achieve high-resolution(such as 3840×2160,7680×4320)and high frame rate(such as 60 fps,120fps),resulting in the drastic growth of the volume of the video data.In order to ensure that the video data can be transmitted and stored effectively within the limited bandwidth,the video data needs to be compressed significantly.Therefore,a new generation of the High Efficiency Video Coding(HEVC)standards is emerging as the times require.While maintaining the quality of the reconstructed video,HEVC reduces the bit rate by about 50% compared to the previous generation of video coding standards H.264/AVC.However,HEVC improving the performance of compression but attenuating the correlation between codewords,when the video is transmitted on a lossy network,which causes the quality of the reconstructed video at the decoder side has been degraded severely.Hence,it is necessary to study how to improve the fault tolerance of HEVC video coding,and in the field of fault-tolerant coding,multiple description coding(MDC)is an effective means to solve the above problems.In recent years,due to the ability of deep learning to extract deep features of images,it has been applied to the field of computer vision gradually,and has achieved good visual effects.Therefore,this paper starts from improving the fault tolerance of HEVC and the quality of video at the decoder side,combining multiple description coding with the technology of deep learning,researches were carried out from the two aspects of constructing the framework of multiple description coding and improving the quality of side reconstruction at the decoder side.Two schemes of HEVC multiple description video coding based on deep learning are proposed.The content is as follows:First of all,to address the frame loss problem caused by dividing the source video into odd/even sequences by time down-sampling at the encoder side,the multiple description video coding based on Frame Prediction-Convolutional Neural Network(FP-CNN)is proposed.FP-CNN can make full use of the time correlation of the video,and predict the missing frames in the corresponding sequence,subtract the predicted frame from the encoded video frame of the corresponding sequence to obtain residual information,and form a description with the encoded information of the current sequence.The formed multiple descriptions can make full use of the relevant information between the descriptions at the decoding side,high-quality video reconstruction can still be achieved under unreliable network transmission.The result of experimental show that compared with the existing algorithm,the PSNR of the proposed method for high-resolution video can be improved by 4.4 d B at the same bit rate.Second,with regard to the problem that the serious degradation of video quality due to only one description received by the decoder,Side Decoding-Video Super Resolution Network(SD-VSRNet)is proposed to improve the reconstruction quality of the side decoding.the source video at the encoder side is divided into two descriptions by spatial down-sampling.When only one description is received at the decoder side,SD-VSRNet uses the spatial correlation of the video to perform super-resolution reconstruction to obtain a full-resolution video.The experimental results show that SD-VSRNet on the super-resolution dataset can improve 3.151 d B compared with the existing algorithm,obtain high-quality side decoding reconstruction video on HEVC test sequence.