A classification‐prediction joint framework to accelerate QTMT‐based CU partition of inter‐mode VVC

Aiming at accelerating the inter coding of versatile video (VVC), existing deep-learning-based methods utilize a single convolutional neural network (CNN) to directly predict quadtree plus multi-type tree (QTMT)-based partition whole unit (CTU). However, these adopt one prediction for unevenly distributed CTUs and ignore that different have difficulties, leading performance degradation computation waste. To overcome limitations, classification-prediction joint framework is proposed accelerate VVC in this letter, which combines classification process through networks with appropriate capacities. achieve effective CTU, QTMT-based first modeled as homogeneity map (PHM), value reflecting each 8× 8 unit. Second, module classifies into classes according their difficulty, then sub-networks capacities are utilized PHM corresponding CTU class. Finally, decision (DT) adopted determine optimal split modes based on predicted PHM. Experimental results show approach achieves 44.5% time saving 1.94% BD-BR increase, outperforming state-of-the-art approaches.