Research On Fast VVC Inter Prediction And Quantization Based On Multi-Stage Decision

With rapid social and economic development,people’s living standards are constantly improving.People could only watch black-and-white TV some years ago.Now people can watch 4K,8K,and ultra-HD videos always and everywhere,which is the witness of the rapid development of the video coding field.However,the video data will increase explosively,which poses significant challenges to storage and transmission under limited bandwidth.So the latest versatile video coding(VVC)introduced by the joint video exploration team(JVET)brings extremely high compression efficiency.Compared with the previous high-efficiency video coding(HEVC),VVC can save nearly half of the bitrate while maintaining the same video quality.However,the complexity of the encoder has increased sharply,which is difficult to code in real-time and design the system on the chip of the VVC encoder.Due to the high complexity of affine motion estimation(AME)in inter prediction and ratedistortion traversal in quantization,we propose a fast algorithm based on the multi-stage decisions for inter prediction and quantization to reduce the complexity of the VVC encoder.For inter prediction,we accelerate the AME and inter mode decision process.For quantization,we detect allzero blocks in advance to skip the quantization process and then reduce the complexity of the encoder.The main innovations and contributions of this article are shown in detail:1.In terms of inter prediction:(1)We propose a three-stage scheme to decrease the complexity of the AME algorithm.The three-stage is the optimal inter-mode decision based on block partition,the fast algorithm in affine motion search(AMS),and the fast inter-mode decision based on a decision tree.(2)Three steps are proposed for the fast algorithm in AMS.Firstly,the AMS is skipped when the two control point motion vectors(CPMV)are parallel.Then in the iterated CPMV update process,the early termination is made when the user-defined variable T is smaller than the previous iteration.Finally,rate-distortion cost comparison is carried out from every four directions modified to every two directions in the fine granularity CPMV search process.(3)Considering that there are some coding units(CU)traversing the translational motion estimation(TME)and AME simultaneously,we analyze the reasons for the optimal mode selection and find eight key features to predict the optimal inter-mode based on the decision tree model.These features include the sum of absolute transformed distortion(SATD), standard deviation,gradient,and so on.2.In terms of quantization:(1)The whole process of all-zero block detection is divided into three stages,which include genuine all zero blocks detection based on the hard decision quantization(HDQ),pseudo all zero blocks detection based on the adaptive threshold,and fully connected neural network based pseudo all zero block detection.Most of all zero blocks can be detected in advance by a certain step.(2)In pseudo all-zero block detection based on the adaptive threshold,an adaptive threshold is introduced by analyzing the quantization parameter and transform unit(TU)sizes.When the transform coefficient position index is smaller than the adaptive threshold,it will be detected as pseudo all zero block.(3)In fully connected neural network based all zero block detection,novel context-level syntax elements are added to the network except for some statical features such as SATD,pre-quantization,and so on.These features can distinguish all zero blocks and non all zero blocks,which helps to improve the accuracy of network training.Experimental results show that the proposed fast AME algorithm based on the multi-stage decisions achieves 10.20% and 10.33% encoding time-saving on average under Low Delay B(LDB)and Random Access(RA)configuration,while the BD-Rate loss is only 0.12% and 0.14%,respectively.And the proposed all-zero block detection algorithm based on the multi-stage decisions achieves 7.505% and 7.049% encoding time-saving with only 0.470% and 0.578% BD-Rate loss under LDB and RA configuration on average,respectively.It is seen that our proposed fast algorithm based on the multi-stage decisions for inter prediction and quantization can greatly reduce the complexity of the encoder with negligible BD-Rate loss.