Point Cloud Denoising Using Multiscale Distribution Scores
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing point cloud denoising methods based on deep learning fail to accurately estimate displacement due to considering only distance relationships between noise points and clean points, leading to inaccurate denoising results with potential value abnormality and feature loss.
Innovation Solution
A denoising method using a multi-scale distribution score, incorporating a feature extraction module and displacement prediction module, which enhances feature expression through multiscale noise perturbation and iteratively learns displacement based on neighborhood distribution scores.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If a nearest distance-based objective function is used for iterative training of neural network, then the training process is simple, but the displacement estimation becomes inaccurate leading to value abnormality and feature loss
Solution Approach 1:
The patent changes the objective function parameter from nearest distance to distribution score (gradient of log probability). This parameter change transforms the training objective to consider neighborhood distribution characteristics, thereby improving displacement estimation accuracy while avoiding value abnormality and feature loss.
Solution Approach 2:
The patent introduces distribution score as an intermediary concept between the noisy point cloud and the displacement prediction. This intermediary (distribution score) mediates the learning process by providing a more informative gradient signal that reflects the underlying data distribution, leading to more accurate denoising results.
2Device complexity
If only distance relationship between noise point and clean point is considered, then the computational complexity is low, but the denoising result has value abnormality and shrinkage
Solution Approach 1:
The patent adds a new dimension to the computational approach by considering the distribution characteristics of neighborhood points rather than just pairwise distances. This dimensional expansion from considering only direct distance relationships to considering spatial distribution patterns improves the reliability of denoising results without excessive complexity increase.
Solution Approach 2:
The patent changes the computational parameter from simple distance metrics to distribution score computation. This parameter transformation allows the system to capture more nuanced spatial relationships and distribution patterns, improving denoising quality while maintaining reasonable computational complexity through efficient gradient-based computation.
3Ease of manufacture
If traditional image denoising algorithm based on deep learning is applied to point cloud, then the methodology is straightforward, but the algorithm fails due to point cloud features of disorder, unstructured, uneven distribution and vast amount of data
Solution Approach 1:
The patent applies local quality by focusing on neighborhood distribution characteristics around each point rather than treating the entire point cloud uniformly. This local approach adapts to the unstructured and uneven distribution nature of point cloud data, allowing the algorithm to handle disorder and variability effectively while maintaining denoising reliability.
Solution Approach 2:
The patent changes the fundamental parameter representation from image-based pixel grids to point cloud distribution scores. This parameter transformation enables the algorithm to naturally handle the unordered, unstructured nature of point cloud data while maintaining deep learning effectiveness, overcoming the direct applicability issues of traditional image denoising methods.
4Manufacturing precision
If iterative training with displacement prediction is performed, then the denoising process is refined, but value abnormality, shrinkage and aggregation occur
Solution Approach 1:
The patent implements feedback through the distribution score computation that continuously monitors and guides the iterative training process. The distribution score provides feedback signals that reflect whether points are moving toward appropriate positions, preventing value abnormality, shrinkage, and aggregation by adjusting displacement predictions based on neighborhood distribution characteristics.
Solution Approach 2:
The patent changes the training objective parameter from direct displacement minimization to distribution score maximization. This parameter change stabilizes the iterative process by providing a more robust gradient signal that prevents pathological behaviors like value abnormality, shrinkage, and aggregation while maintaining denoising refinement effectiveness.
Data Source
AI summary
A denoising method based on a multiscale distribution score for a point cloud includes: constructing a two-layer network model based on multiscale perturbation and point cloud distribution, where the two-layer network model includes a feature extraction module for extracting a feature of the point cloud and a displacement prediction module for predicting a displacement of a noise point; constructing a point cloud noise model for improving a denoising effect and retaining a sharp feature and avoiding reducing quality of point cloud data; extracting a global feature h by inputting the point cloud data into the feature extraction module; iteratively learning the displacement of the noise point by the displacement prediction module according to a feature obtained by the feature extraction unit; and defining a loss function of network training, and completing convergence under the condition that the loss function reaches a set threshold or a maximum number of iterations.