3D Mesh Denoising via Geometric Feature Mapping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing denoising algorithms for 3D meshes assume statistical characteristics of noise, which may not accurately represent real-world noise sources, leading to ineffective noise removal in scanned data.
Innovation Solution
A data-driven approach that involves training a model to map noisy normals to ground-truth normals using a dataset of noisy and ground-truth facet data, employing techniques like bilateral filtering and regression trees to generate feature vectors and update vertex positions, without assuming noise distributions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If existing denoising algorithms assume statistical characteristics of noise, then the denoising process can be simplified, but the accuracy of noise removal deteriorates because real-world noise sources do not conform to these assumptions
Solution Approach 1:
The patent changes the fundamental parameter of noise representation from statistical distributions to geometric features. Instead of assuming noise follows Gaussian or other statistical distributions, the method represents noise as deviations in geometric properties (normals, curvatures, edge orientations) and uses data-driven learning to map noisy geometric features to clean ones, achieving both simplicity and accuracy
Solution Approach 2:
The patent substitutes traditional statistical filtering mechanisms with a data-driven geometric feature mapping system. Rather than using statistical assumptions about noise distributions, the system learns from training data how geometric features transform from noisy to clean states, replacing mechanical statistical filtering with learned geometric transformations
2Productivity
If traditional denoising methods are used, then the processing speed may be maintained, but the effectiveness of noise removal deteriorates due to inaccurate noise assumptions
Solution Approach 1:
The patent performs preliminary action by pre-computing geometric features (normals, curvatures, edge orientations) and their relationships during a training phase. This pre-learning of feature mappings enables fast processing during actual denoising operations, maintaining high processing speed while achieving effective noise removal through learned geometric transformations rather than iterative statistical filtering
3Power
If statistical noise assumptions are made, then the denoising algorithm can be computationally efficient, but the adaptability to diverse noise sources deteriorates
Solution Approach 1:
The patent creates a universal denoising system that works across diverse noise sources by learning geometric feature mappings from training data that encompasses multiple noise types. The learned model becomes multi-functional, adapting to different noise sources (sensor noise, reconstruction errors, material artifacts) without requiring source-specific statistical assumptions, achieving both computational efficiency and broad adaptability
Data Source
Figure 1~2
Figure 3
Figure 4~5
AI summary
In this disclosure, a solution for denoising a curve mesh is proposed. For a curve mesh including a polygonal facet, a noisy normal and a ground-truth normal of a first facet in the mesh is obtained. Then, based on the noisy normal, a first geometric feature of the first facet is determined from a plurality of neighboring facets of the first facet in the mesh. Next, based on the first geometric feature and the ground-truth normal, a mapping from the first geometric feature to the ground-truth normal of the first facet is determined for denoising the mesh.