Optimization Of H.266 Intra Prediction Algorithm Based On Long Short-term Memory Networks

With the rapid promotion of high-definition(HD)and ultra-high-definition(UHD)video applications,the amount of data required for video transmission and storage is increasing.A large number of new tools have been applied in the upcoming H.266 encoding standard,which has greatly improved the video compression rate,but there is still a lot of room for improvement in the standard.Considering the great advantages of deep learning in the expression of massive data,the paper focuses on the problem of long short-term memory(LSTM)network used to optimize the H.266 intra prediction process,and proposes two improved frames Internal prediction optimization algorithm.The main work and innovations are as follows:(1)In view of the problem of large prediction residuals in the intra prediction process of the H.266 encoder,the paper proposes an intra mode optimization algorithm based on LSTM network.For the horizontal and vertical modes in angle prediction,the paper trains an LSTM network model,which uses this model to make a second prediction of the intra prediction residuals,compensates the prediction results of the intra mode in the standard model,and reduces the final prediction Residuals.The paper discusses the depth and size of the LSTM network,and uses statistical methods to optimize the network parameters.The experimental results show that,compared with the H.266 standard reference model VTM 2.0,the proposed scheme can reduce the BD-rate by 0.34% on average.(2)In view of the problem that the standard model has poor prediction ability for pixels far away from the reference line in larger coding blocks,the paper proposes an improved algorithm for intra prediction of multiple reference lines based on LSTM network.The other adjacent reference lines of the current coding block improve the prediction method based on the single reference line to the prediction method based on the multiple reference lines,which provides a richer context for intra prediction,and through the LSTM network The residuals in the lower right area have been compensated to improve the accuracy of intra prediction.Experimental results show that,compared with the H.266 standard reference model VTM 2.0,this scheme can reduce the BD-rate by 0.18% on average.