Research On Video Codec Technology Based On Deep Learning

In recent years,with the rapid development of the Internet and the popularity of mobile Internet devices,people's life is full of multimedia information everywhere.HEVC technology standard has been able to basically meet the current needs of people for the transmission of multimedia information.However,with the commercial popularization of5 G network and the rapid development of artificial intelligence,the total amount of multimedia information has shown an explosive growth.Therefore,the exploration of more efficient video codec technology is the focus of current research.In recent years,because of the rapid development of GPU performance,on the basis of deep learning technology of artificial intelligence has been rapid development,because of the depth of the learning technology has been successfully applied to computer vision,natural language processing,speech recognition and other fields,a lot of research is trying to introduce deep learning technology video codec field,and made some significant achievements,but there are still many problems worth to study.This paper explores three important modules in the video codec framework and combines the deep learning technology to improve the performance of the existing codec framework.This paper deals with three modules in the video codec framework: intra-frame prediction,inter-frame prediction and loop filter.After several attempts,the deep learning technology has been successfully applied to loop filtering and video post-processing,which has surpassed the current research results and achieved the industry leading indicators.The specific research content of this paper is divided into the following three parts:first,it introduces the basic theoretical knowledge about deep learning.Secondly,it introduces the exploration of the two modules of in-frame prediction and inter-frame prediction in video codec with deep learning technology.Third,introduced the use of end-to-end depth convolution neural network to loop filter and video post-processing of video coding,a single frame is based on the existing neural network input to carry on the design,does not take into account before and after the frame of information,but in video coding,the current frame of reference the encoding is the frame of information,so the introduction of the current frame context information,help the current frame information reconstruction,achieve better quality enhancement effect.To make full use of the front and rear frame in the video information,this paper introduced the efficient light flow network to do motion compensation,and then combines ConvLSTM network to information fusion frame before and after,and design the efficient distillation network to filter,at the same time,introduces the structure of against production networks,to further enhance the subjective indicators of the video results,achieved remarkable results.