Multi-Head Self-Attention Enhanced Score Network for Point Cloud Denoising

Point cloud denoising is essential for 3D vision applications. Existing score-based methods use dynamic graph convolutional networks for feature extraction but struggle to model long-range dependencies, causing geometric distortions and residual noise. We propose a denoising framework that integrates multi-head self-attention (MHSA) with graph convolutional layers to capture both local and global features. Graph convolutions extract local geometric details, while MHSA models long-range semantic correlations for precise denoising under global constraints. Experiments on the PUNet and Paris-rue-Madame datasets show our method outperforms other deep learning approaches across various noise levels and point resolutions, achieving superior denoising accuracy and geometric fidelity.