Efficient Non-Local Point Cloud Denoising Using Curvature Entropy and $\gamma$γ-Norm Minimization

Area of Science:

• Computer Vision
• Geometric Processing
• Signal Processing

Background:

• Non-local similarity (NLS) is effective for point cloud denoising.
• Existing NLS methods face challenges with high algorithmic complexity and inaccurate low-rank matrix estimation.

Purpose of the Study:

• To propose an efficient and accurate point cloud denoising framework, named PCD-gammaCE.
• To address the limitations of existing non-local denoising methods.

Main Methods:

• Developed a structure descriptor using Curvature Entropy (CE) and Angle Subdivision (AS) to capture Non-Local Similar Structure (NLSS) details and control complexity.
• Introduced gamma-norm minimization for low-rank denoising model, ensuring robust and nearly unbiased estimation of the rank function.

Main Results:

• The PCD-gammaCE framework demonstrates superior performance on synthetic and raw scanned point clouds.
• Achieved a 99.90% reduction in processing time compared to Weighted Nuclear Norm Minimization (WNNM).
• Showed significant improvements in Mean Square Error (MSE) and Chamfer Distance (CD).

Conclusions:

• The proposed PCD-gammaCE method offers an efficient and accurate solution for point cloud denoising.
• The gamma-norm minimization and Curvature Entropy-based approach effectively handles noise while preserving structural details.