The latest H266/VVC standard has received numerous praises for its excellent compression efficiency. However, its extremely high computational complexity has become a hindrance to the VVC adaptation industry ecosystem, while also increasing the difficulty of hardware design and application costs. To address this issue, we designed an efficient intra-coding scheme based on neural networks, which consists of three parts: Firstly, we designed a neural network-based reverse prediction algorithm that uniquely utilizes the CNN’s prediction results for lower-level blocks to determine the QTMT partitioning of upper-level blocks, cleverly solving the adaptation problem of existing models to complex VVC partitioning patterns—a decision-making logic that has not been fully explored. Secondly, we designed a pruning algorithm, which is the first to dynamically couple the real-time RDO cost of BT segmentation with the TT segmentation direction, achieving adaptive decision-making. Finally, we designed a complexity pre-screening module. On the basis of analyzing whether the CU texture is smooth, this module designs four sets of adaptive thresholds for non-square CUs introduced in VVC. These thresholds can dynamically adjust local and global thresholds based on CU size, enabling size sensitive texture evaluation to determine whether the current block needs further partitioning. The experimental results show that, compared with traditional VTM4.0, our method reduces the average encoding time by 49.21%, while the BD-BR increase is 1.61%, and the BD-PSNR decreases by 0.06 dB, fully demonstrating its superiority and performance balance.
Shi et al. (Mon,) studied this question.