Multimedia Systems Computer Networks and Communications

A fast coding unit partition algorithm for VVC intra coding based on a lightweight two-stage neural network decision

Zhang, Qiuwen, Liang, Xiaodi, Zhao, Jinchao
Springer Nature · 2026-11-16 · DOI 10.1007/s00530-026-02797-7

Abstract

Versatile Video Coding (VVC) significantly improves video compression efficiency by introducing a combined partitioning structure comprising quadtrees and multi-type trees (QT + MTT), but this also results in higher encoding complexity. In particular, the Rate-Distortion Optimisation (RDO) search during the coding unit (CU) partitioning process has become the main computational bottleneck of the encoder. To address this issue, this paper proposes a two-stage neural network-based decision-making algorithm for rapid CU partitioning. First, the Pearson Correlation Coefficient (PCC) is employed to select representative CU features. Then, a lightweight neural network trained with teacher–student knowledge distillation is used to determine whether a CU should be further split. Subsequently, a dual-head convolutional neural network predicts QT and MTT partition modes, and a candidate mode pruning strategy is applied to reduce the RDO search space. Experimental results under the Joint Video Experts Team (JVET) common test conditions for the All-Intra (AI) configuration show that the proposed method achieves 56.77% encoding time reduction with only 2.58% Bjøntegaard Delta Bit Rate increase, thereby effectively reducing the encoding complexity of VVC intra-frame encoding.

Read at publisher →