A Fast Coding Unit Partitioning Decision Algorithm for Versatile Video Coding Based on Gradient Feedback Hierarchical Convolutional Neural Network and Light Gradient Boosting Machine Decision Tree

Video encoding technology is a foundational component in the advancement of modern technological applications. The latest standard in universal video coding, H.266/VVC, features a quad-tree with nested multi-type tree (QTMT) partitioning structure, which represents an improvement over its predecessor, High-Efficiency Video Coding (H.265/HEVC). This configuration facilitates adaptable block segmentation, albeit at the cost of heightened encoding complexity. In view of the aforementioned considerations, this paper puts forth a deep learning-based approach to facilitate CU partitioning, with the aim of supplanting the intricate CU partitioning process observed in the Versatile Video Coding Test Model (VTM). We begin by presenting the Gradient Feedback Hierarchical CNN (GFH-CNN) model, an advanced convolutional neural network derived from the ResNet architecture, enabling the extraction of features from 64 x 64 coding unit (CU) blocks. Following this, a hierarchical network diagram (HND) is crafted to depict the delineation of partition boundaries corresponding to the various levels of the CU block's layered structure. This diagram maps the features extracted by the GFH-CNN model to the partitioning at each level and boundary. In conclusion, a LightGBM-based decision tree classification model (L-DT) is constructed to predict the corresponding partition structure based on the prediction vector output from the GFH-CNN model. Subsequently, any errors in the partitioning results are corrected in accordance with the encoding constraints specified by the VTM, which ultimately determines the final CU block partitioning. The experimental results demonstrate that, in comparison with VTM-10.0, the proposed algorithm achieves a 48.14% reduction in complexity with only a 0.83% increase in bitrate under the top-three configuration, which is negligible. In comparison, the top-two configuration resulted in a higher complexity reduction of 63.78%, although this was accompanied by a 2.08% increase in bitrate. These results demonstrate that, in comparison to existing solutions, our approach provides an optimal balance between encoding efficiency and computational complexity.