CNN-based Fast Split Mode Decision Algorithm for Versatile Video Coding (VVC) Inter Prediction

Versatile Video Coding (VVC) is the latest video coding standard developed by Joint Exploration Team (JVET). In VVC, quadtree plus multi-type tree (QT+MTT) structure of unit (CU) partition adopted, and its computational complexity considerably high due to brute-force search for recursive rate-distortion (RD) optimization. this paper, we aim reduce time inter-picture prediction mode since inter accounts a large portion total encoding time. The problem can be defined as classifying split each CU. To classify effectively, novel convolutional neural network (CNN) called multi-level (MLT-CNN) architecture introduced. For boosting classification performance, utilize additional information including while training CNN. overall algorithm MLT-CNN inference process implemented on VVC Test Model (VTM) 11.0. CUs size 128×128 inputs sequences are encoded at random access (RA) configuration with five QP values {22, 27, 32, 37, 42}. experimental results show that proposed 11.53% average, 26.14% maximum an average 1.01% increase in Bjøntegaard delta bit rate (BDBR). Especially, method shows higher performance A B classes, reducing 9.81%~26.14% 0.95%~3.28% BDBR increase.