Research On Optimization Algorithm For Geometry Prediction Mode In VVC

VVC/H.266(Versatile Video Coding)is a new generation video coding standard.Different from the traditional rectangular block division structure,Geometric Prediction merge Mode(GPM)introduced in VVC uses a non-rectangular block division structure to divide the coding blocks into wedges or triangles in the inter-prediction process,and brings more flexible than traditional division structure.In addition,GPM with Motion Vector Refinement allows motion vector refinement in Merge mode with MVD(MMVD),and can further improve the accuracy of prediction.However,this technique significantly increases the computational complexity,resulting in a significant increase in coding time.This thesis conducts research on the above issues,and the main contributions are as follows:(1)The relationship between GPM mode selection and CU characteristics is studied,and a fast decision algorithm for inter-prediction geometry prediction merge mode based on CU gradient is proposed.It is found that there are distinct sub-block pixel differences in CUs which using GPM mode.The gradient mean is introduced to measure such differences,and the GPM division process is terminated early for the CU without sub-block pixel differences.For the CU having sub-block pixel differences,since the GPM division mode is highly related to the texture direction of the CU,the gradient orientation is introduced to make an early decision on the quantization angles in the GPM mode to skip unnecessary division modes and reduce the computational complexity.The experimental results show that the proposed algorithm can reduce the coding time by 13.5% on average under the random access configuration compared with the VTM8.0 reference model,and the BD-rates of Y,U and V components only increase by 0.14%,0.13% and 0.31% respectively.(2)Furthermore,the fast decision problem of GPM with motion vector refinement is studied,and it is found that the probability of choosing a motion step in MMVD is closely related to the motion intensity of the CU,where a CU with intense motion often chooses a large step,while a CU with less intense motion often chooses a small step.A fast algorithm for GPM with motion vector refinement based on perceptual hash(p-hash)is proposed,which uses p-hash to calculate the Hamming code distance between the reference CU and the current CU to measure the motion intensity of the current CU.According to the motion intensity of CU,the motion step in the process of motion vector refinement is decided in advance,and unnecessary refinement methods are skipped.The experimental results show that the proposed algorithm can reduce the coding time by 9% on average under the random access configuration compared with the VTM8.0 reference model,and the BD-rates of Y,U and V components increase by only 0.01%,0.08% and 0.07% respectively.