Research On VVC Inter-Frame Coding Algorithms Based On Machine Learning

Video is an important means for humans to obtain information beyond their current time and space,and its importance is self-evident.To reduce the transmission and storage costs of videos,many encoding technologies have been developed and applied.However,with the advent of the ultra-high-definition video era,previous encoding technologies are no longer able to meet the needs of ultra-high-definition videos.Therefore,the latest generation of video encoding standard,Versatile Video Coding(VVC),has been developed and released.VVC is the latest generation of video encoding standard designed for ultra-high-definition videos.Compared to the previous generation of video encoding standard,VVC has added many new encoding technologies and tools,achieving nearly twice the compression efficiency improvement.However,its computational complexity has also increased exponentially with the introduction of new encoding tools and techniques.Therefore,reducing the computational complexity of the VVC encoding process has become a major challenge that needs to be addressed.The block partition module,which is a key module that determines the video encoding quality and computational complexity,accounts for most of the VVC encoding time.In the block partition module,VVC introduces a new block partition structure,the Quad Tree Multitype Tree(QTMT)block partition structure,which has a more flexible Coding Unit(CU)partitioning method.This makes the CU after partitioning have a lower distortion rate,thereby improving the encoding quality of VVC,while also increasing the computational complexity of the encoding process.Many related studies have also been used for CU pre-prediction or pretermination,but most of these studies only focus on the characteristics of the image itself,ignoring the process parameters in the encoder.These parameters sometimes can better determine the partitioning mode of the current CU.Moreover,previous research did not focus on the QTMT structure in VVC.To address these issues,this paper proposes the following:1)This paper calculates and extracts data from the original encoder,and conducts analysis and data visualization work.2)This paper proposes a solution using Random Forest(RF)to predict the partitioning mode of the current CU,and constructs the CU classification problem in VVC as a six-class problem.3)This paper combines the encoder’s encoding process parameters and image features.4)This paper embeds the machine learning model into the VVC encoding software and conducts relevant encoding experiments.This paper conducted encoding experiments on specific video sequences on the VVC reference software platform VTM-13.0.The experimental results show that the proposed solution can effectively reduce the computational complexity of the VVC inter-frame encoding process,and for video quality losses of 0.92% to 4.68%,the average complexity is reduced by30.3% to 52.1%,which is better than other advanced algorithms.