Research On Quality Improvement Technology Of Low Bit Rate Video Coding Based On Deep Learning

Video coding is an important means of compressing video.Currently,the latest generation of video coding standard is HEVC.HEVC not only obtains a high compression ratio,but also greatly reduces the size of the transmission stream.However,at low bit rate,because the encoder uses large quantization parameters,it will introduce a large compression distortion,in which the image details are lost,serious artifacts will appear in the video image,causing serious distortion of the image,making the view feel great impact.Aiming at the problem of excessive compression distortion caused by large quantization parameters in low-rate video coding process,the algorithm of low-rate video coding quality improvement based on deep learning is studied.Deep learning is developed rapidly in recent years.In a discipline,deep learning has achieved very good processing results in enhancing image quality and reducing video bitstream compared to previous traditional methods.In our actual application scenario,we will encounter a video with a large transmission resolution in the case of insufficient bandwidth.The existing video coding standard is not enough to provide a perfect parameter to get a good compressed video..In order to solve this problem,this paper proposes a video framework based on video image reconstruction neural network.First,we redefine the coding framework,first reduce the amount of information in the video airspace,downsample the high-resolution video to the lowresolution video,and compress the low-resolution video to the encoder to obtain the damaged low-resolution video.The decoded low resolution video image reconstruction is then restored to the original resolution.The video image reconstruction neural network is the most important component of the coding framework.The performance of the network determines the quality of the video image.Through the analysis of the experiment,it is finally decided that the video image reconstruction neural network mainly includes two aspects: image restoration and image super resolution.The main function of the image restoration network is to remove the image distortion and blockiness elimination caused by the encoder at low bit rate.The main function of the image super-resolution network is to expand the resolution of the damaged video image after decoding to the target resolution and improve the video.Like reconstruction quality,preserve the details of the image as much as possible.By constructing the training set and the integration of the loss function,the performance of the video image reconstruction processing is optimized by continuous training and testing.During the test,we guarantee that the encoder and the new coding framework use the same rate coding test when encoding the video.After testing different types of test video,the coding framework based on video image reconstruction can be obtained than the HEVC coding framework.An average PSNR gain of 0.5 d B yields an average PSNR gain of about 0.4 d B over the Bicubic method.In addition,the most important point is that the coding framework based on video image reconstruction can effectively remove the compression distortion caused by video coding at low bit rate,mainly including block effect,ringing effect,artifact,compared with HEVC.Blurring and other noise,and can maintain the edge information and texture details of the image,greatly improve the subjective quality,and can greatly improve the video quality.