HeteroDenoise: Intensity-Guided Point Cloud Denoising Under Heterogeneous Noise

In optical metrology and 3D surface measurement—core areas in instrumentation and measurement—spatially varying noise in point clouds poses a significant challenge to accurate surface reconstruction and dimensional analysis. Most point cloud acquisition systems reconstruct object surfaces using reflected optical signals, which are often corrupted by noise arising from surface material properties. Such noise is typically heterogeneous, exhibiting large variations in both type and intensity within a single scan. However, most existing denoising algorithms assume homogeneous or single-type noise, resulting in degraded performance under realistic, complex noise conditions. To address this issue, we propose HeteroDenoise, a novel contrastive learning-based denoising framework designed specifically for heterogeneous noise with high distributional variability. The proposed method incorporates a contrastive feature extraction module that enhances sensitivity to local noise characteristics, enabling robust structural representation. Additionally, we introduce a spatially adaptive noise intensity prediction module that guides local iterative refinement of point positions based on estimated noise levels. Extensive evaluations on both synthetic and real scanned point clouds demonstrate that HeteroDenoise significantly outperforms existing methods in denoising quality and downstream surface reconstruction accuracy.