Confidence-guided progressive point cloud denoising via MSAA-EdgeConv

Point cloud denoising is essential for improving the quality and reliability of noisy 3D point clouds, yet existing reconstruction-, distribution-, and displacement-based methods may suffer from geometric information loss, high computational complexity, or limited adaptation to spatially varying noise. To address these limitations, we propose a confidence-guided progressive refinement point cloud denoising network that refines noisy point clouds through stage-dependent coarse-to-fine geometric learning. At each stage, the network dynamically re-extracts local features using multi-scale attention augmented edge convolution (MSAA-EdgeConv) and jointly predicts point displacements and confidence scores. The confidence scores act as learned per-point weights that adaptively scale displacement updates, while stage-specific scale factors adjust the correction magnitude throughout refinement. Extensive experiments on synthetic benchmarks (PUNet and PCNet) and real-world datasets (Kinect v1, Kinect v2, and 3DCSR) demonstrate strong denoising performance and generalization across diverse noise conditions. We obtained 18.15 for CD and 6.83 for HD on the 10 K-point PU-Net dataset, showing better results than the 3DMambaIPF and C2AENet methods, which are current SOTA methods. On PUNet, the proposed models achieve the best CD across all tested noise levels and point densities, while on Kinect v1 and Kinect v2 they obtain the lowest CD without dataset-specific fine-tuning.