Research On Convolutional Neural Network Based In-Loop Filtering Technique For Video Compression

With the rapid development of information technology,the popularity of mobile devices,multimedia applications are widespread,and a variety of multimedia content greatly enriches people's daily lives.Among them,digital video is one of the most commonly used information carriers in multimedia applications.The original video is difficult to be applied pratically due to the huge amount of data.In order to efficiently store,transmit and utilize video data,it is necessary to compress the video data.The International Standards Organization began to establish international standards for video coding in the 1980 s,and is currently conducting research on a new generation of video coding standard called VVC.Loop filtering is one of the key technologies of video coding.In order to achieve a higher compression ratio,lossy compression is generally applied in video compression,which inevitably lead to the degration of video quality that compressed artifacts such as blocking artifacts,ringing effects,and blur are introduced in compressed video.Loop filtering technique is generally applied in mainstream coding standard to process the reconstructed videos,which effectively improves the video quality.The pixels after processed by loop filtering can be used as reference pixels for the subsequent compression process,which can further improve the compression performance.In recent years,convolutional neural networks have achieved remarkable results in many computer vision tasks,providing new ideas for further improving video compression performance.According to the principles of video compression and convolutional neural networks,this paper focuses on the application of convolutional neural networks in loop filtering techniques.Based on the new generation of video coding standard VVC,this paper conducts in-depth research on the convolutional neural network based loop filtering technique from the aspects of performance and speed.The main innovations and contributions are as follows:(1)This paper proposes a convolutional neural networks based loop filtering algorithm.In order to learn the mapping between video images before and after compression,this paper designs a new convolutional neural network structure called dense residual convolutional neural network(DRN).The proposed DRN can not only make full use of the multi-level features,but also effectively perform feature reuse and feature fusion.By integrating this algorithm into the latest international video coding standard VVC,the coding performance is further improved.(2)This paper studies the optimization acceleration of the proposed convolutional neural network based loop filtering algorithm.Although the convolutional neural network based loop filtering algorithm contributes to coding performance improvement,the complexity introduced by it is difficult to meet the requirements of practical application of video compression.Considering the computational efficiency of the proposed model,this paper proposes a lightweight and efficient convolutional neural network model.