Research And Implementation Of Deep-reinforcement-learning-based Rate Control Algorithm For VVA

With the rapid development of multimedia communication technology,people’s demand for ultra-high-resolution video quality is highly augmenting.However,the ultrahigh-definition video exacerbates the demand of bandwidth during the encoding and transmission process,it takes a huge pressure for storage devices and bandwidth resources.Rate control can effectively solve the conflict between limited bandwidth and great video data in video coding.Versatile Video Coding(VVC)is the latest generation of video coding technology.It has introduced more and more new features,which VVC’s rate control has not fully considered.Thus,there are many coding performances for VVC rate control to improve.According to the above background,a deep reinforcement learning based rate control optimization algorithm for VVC was proposed in this thesis based on the analysis of the impact of the new features of VVC on rate control.The specific work can be concluded as follows:(1)The key parameters of VVC rate control were studied,based on analyzing the impact of the new features of VVC on rate control.And these key parameters will regard as input parameters of proposed algorithm in next.Firstly,analysis of characteristics of VVC’s partitioning,intra prediction and inter prediction impact on rate control.Secondly,the coding parameters related to rate control was selected Finally,the grey relation analysis was used to further choose several key parameters with the highest degree of correlation with quantization parameter,which provides reasonable and reliable parameters for proposed algorithm in next.(2)A frame-level quantization parameter predict model was proposed with deep reinforcement learning,which combined with frame-level rate control process,to achieve the overall rate control optimization.The prediction of quantization parameters is an optimal process,and deep enhancement learning be good at solving the optimal decisionmaking problem.Therefore,deep enhancement learning was used to predict the quantization parameter for VVC’s rate control.Firstly,a frame-level quantization parameter prediction model based on deep reinforcement learning is proposed,which combined with the new characteristics of VVC analyzed in(1).Secondly,the reward and punishment function has been designed for this thesis research according to the information of rate and distortion.Finally,combining the above research with the traditional rate control process,the overall implementation process of the proposed rate control algorithm is provided.(3)Algorithm implementation and performance test.Firstly,the describing of algorithm’s implementation including three aspects: the extraction of prediction model parameters,the realization of the prediction model in the VVC encoder,and the application of the algorithm in the encoder Secondly,experimental testing and performance analysis on multiple optimization details of this article and the overall algorithm.Experimental results show that the proposed scheme can achieve coding performance improvement(BDBR=-1.4024%)with negligible time loss(ATC=1.14%).In this thesis,a deep-reinforcement-learning-based rate control algorithm for VVC was proposed,which provide contribution to related scientific research and practical application value for VVC.