Multitask Learning-Based Early MTT Partition Decision for Versatile Video Coding

Versatile Video Coding (VVC) introduces a new block partition structure called Multi-Type Tree (MTT), which includes four partitioning modes: horizontal-vertical binary tree partitioning, horizontal-vertical ternary tree partitioning. This new block partition structure significantly improves compression performance, but at the same time greatly increases the computational complexity of VVC. To reduce the computational complexity of MTT in VVC inter-frame coding, a Multitask learning-Based early MTT partition decision for Versatile Video Coding is proposed. Firstly, for each Coding Unit (CU), two types of features related to the optimal MTT partitioning are extracted, namely encoding parameter features and encoding intermediate information features. Secondly, to reduce the number of neural network parameters, the horizontal or vertical partitioning in MTT is jointly learned, and lightweight neural networks are constructed to decide whether to skip the horizontal or vertical partitioning of binary or ternary trees. Experimental results show that under the Random Access (RA) configuration, the proposed method can reduce the VVC inter-frame computational complexity by an average of 27.79%, while only increasing the Bjontegaard delta bit rate (BDBR) by 1.14%.