Efficient and Lightweight Multi-Scale Point Cloud Denoising

Point cloud denoising is an essential preprocessing step in 3D vision tasks such as autonomous driving, robotic perception, and 3D reconstruction. However, uneven point density, complex geometry, and uncontrolled noise make it difficult for existing methods to balance accuracy and efficiency. To address this challenge, we propose a lightweight multi-scale point cloud denoising network (LPCDN) that enhances inference efficiency while maintaining high denoising quality. LPCDN adopts complementary large- and small-scale branches: large-scale modeling improves robustness to density variations, whereas small-scale modeling preserves fine-grained geometric structures. Random sampling is applied in dense local regions to reduce computation, while representative sampling is employed for large-scale structures to retain global consistency. A multi-scale feature fusion mechanism with channel attention further strengthens adaptability and representation capability. Extensive experiments on both synthetic and real-world datasets show that LPCDN achieves competitive accuracy compared with state-of-the-art methods, while significantly accelerating inference. In addition, the integration of Automatic Mixed Precision (AMP) in training and inference reduces runtime by about 30% without accuracy loss, confirming the effectiveness and deployability of the proposed approach.