Fast CU Decision Algorithm Based on CNN and Decision Trees for VVC

• Published in: Electronics (Basel) 2023-07, Vol.12 (14), p.3053
• Main Authors: Li, Hongchan, Zhang, Peng, Jin, Baohua, Zhang, Qiuwen
• Format: Article
• Language: English
• Subjects: Accuracy; Algorithms; Analysis; Classifiers; Coding; Complexity; Data compression; Decision tree; Decision trees; Deep learning; Efficiency; Image coding; Lagrange multiplier; Methods; Neural networks; Software; Video compression
• ISSN: 2079-9292; 2079-9292
• DOI: 10.3390/electronics12143053

Compared with the previous generation of High Efficiency Video Coding (HEVC), Versatile Video Coding (VVC) introduces a quadtree and multi-type tree (QTMT) partition structure with nested multi-class trees so that the coding unit (CU) partition can better match the video texture features. This partition structure makes the compression efficiency of VVC significantly improved, but the computational complexity is also significantly increased, resulting in an increase in encoding time. Therefore, we propose a fast CU partition decision algorithm based on DenseNet network and decision tree (DT) classifier to reduce the coding complexity of VVC and save more coding time. We extract spatial feature vectors based on the DenseNet network model. Spatial feature vectors are constructed by predicting the boundary probabilities of 4 × 4 blocks in 64 × 64 coding units. Then, using the spatial features as the input of the DT classifier, through the classification function of the DT classifier model, the top N division modes with higher prediction probability are selected, and other division modes are skipped to reduce the computational complexity. Finally, the optimal partition mode is selected by comparing the RD cost. Our proposed algorithm achieves 47.6% encoding time savings on VTM10.0, while BDBR only increases by 0.91%.