Block-Dependent Partition Decision for Fast Intra Coding of VVC

Aiming at accelerating the intra coding process of versatile video coding (VVC), previous efforts are made to predict the quad-tree plus multi-type tree (QTMT) partition structure, in which the prediction is modeled as a classification procedure. The existing deep learning-based methods usually represent the block partition as a structural output and utilize a single convolutional neural network (CNN) for the prediction of all blocks. However, they take different blocks equally, i.e., one-for-all, which ignores that the blocks with different complexity of partition structures are unevenly distributed and have different prediction difficulties. To address this problem, we propose a novel block-dependent partition decision (BDPD) framework to adaptively process different blocks by networks with different capacities. Specifically, we design a partition homogeneity map (PHM) to represent the QTMT-based block partition, which combines different partition directions to effectively reflect the complexity of the partition structure. On this basis, we propose a class-based prediction to distinguish blocks with different complexity of the partition structure and adopt appropriate FCN models to predict PHM, incorporating block classification and PHM prediction. The blocks are classified into different classes according to their coarse texture and neighboring PHMs. Then different fully convolutional network (FCN) models are utilized to predict PHM for different classes. The FCN models with different capacities are trained in the corresponding class, respectively, which achieves higher performance with less computation on the extremely unbalanced natural video. Finally, an adaptive partition decision based on predicted PHMs is adopted to conduct partition decisions for a better trade-off between rate-distortion performance and encoding complexity. Experimental results show that our approach achieves 45.7%~74.5% encoding complexity reduction with 0.78%~3.38% BD-BR increase, outperforming state-of-the-art approaches.