Low-Complexity CU Partitioning for 3D-HEVC Depth Maps via SE Attention and Ensemble Learning

Efficient Coding Unit (CU) partitioning is critical to reducing the computational complexity of 3D High Efficiency Video Coding (3D-HEVC), especially for depth maps. While many existing methods rely on either handcrafted features or deep learning, they often suffer from limited feature expressiveness and insufficient focus on structurally significant regions. To address these limitations, we propose a symmetry-aware, two-stage CU partitioning framework. First, a Convolutional Neural Network (CNN) equipped with a channel-wise Squeeze-and-Excitation (SE) attention mechanism is employed to extract multi-scale texture features, highlighting symmetry-relevant patterns. These deep features are then unified via Spatial Pyramid Pooling (SPP) and combined with handcrafted descriptors—such as neighborhood RDcost, directional gradients, and variance—and fed into a Bagged Tree classifier for final prediction. Additionally, a weighted voting strategy is adopted to replace conventional majority voting in the ensemble, enhancing robustness near decision boundaries. Experimental results show that the proposed method achieves an average 52.49% reduction in encoding time, with only a 0.39% increase in Bjøntegaard delta bitrate (BDBR), achieving an excellent trade-off between complexity and performance.