CN-122023176-A - Three-dimensional point cloud denoising method and system based on deep learning

CN122023176ACN 122023176 ACN122023176 ACN 122023176ACN-122023176-A

Abstract

The invention provides a three-dimensional point cloud denoising method and system based on deep learning, which relate to the technical field of point cloud denoising and comprise the steps of acquiring original three-dimensional point cloud data; the method comprises the steps of carrying out local feature coding on original three-dimensional point cloud data by combining dynamic edge convolution attention enhancement to obtain multi-scale local features, carrying out global feature extraction on the original three-dimensional point cloud data by combining multi-scale parallel processing to obtain multi-scale global features, carrying out feature fusion on the multi-scale local features and the multi-scale global features to generate enhancement features, decoding the enhancement features, and carrying out displacement prediction by combining the original three-dimensional point cloud data to obtain denoised three-dimensional point cloud data. The method solves the problems that the prior denoising method is difficult to accurately preserve the real geometric structure and local detail of the point cloud while suppressing noise due to neglecting the neighborhood density difference.

Inventors

YANG LIU
SU WEI
MA ZHENG
LIU HENG

Assignees

西南交通大学

Dates

Publication Date: 20260512
Application Date: 20260129

Claims (10)

1. The three-dimensional point cloud denoising method based on deep learning is characterized by comprising the following steps of: Acquiring original three-dimensional point cloud data; carrying out local feature coding on the original three-dimensional point cloud data by combining with dynamic edge convolution attention enhancement to obtain multi-scale local features; Carrying out global feature extraction on the original three-dimensional point cloud data by combining multi-scale parallel processing to obtain multi-scale global features; Performing feature fusion on the multi-scale local features and the multi-scale global features to generate enhanced features; And after the enhancement features are decoded, carrying out displacement prediction by combining the original three-dimensional point cloud data to obtain the denoised three-dimensional point cloud data.
2. The depth learning-based three-dimensional point cloud denoising method according to claim 1, wherein the local feature encoding of the original three-dimensional point cloud data by combining dynamic edge convolution attention enhancement to obtain multi-scale local features comprises: Setting a multi-scale neighborhood parameter; performing neighbor searching on each point in the original three-dimensional point cloud data based on the multi-scale neighbor parameters to obtain neighbor graphs under different scales; Splicing the center point coordinates of each neighborhood graph with the coordinates of each neighborhood point to obtain edge characteristics under different scales; extracting the characteristics of each edge characteristic through a dynamic edge convolution module to obtain first local characteristics under different scales; carrying out feature extraction on each first local feature through a dynamic edge convolution module to obtain second local features under different scales; And performing element addition operation based on the second local features under different scales to generate multi-scale local features.
3. The depth learning-based three-dimensional point cloud denoising method according to claim 2, wherein the step of the dynamic edge convolution module for feature extraction is as follows: carrying out feature extraction on the input features of the dynamic edge convolution module through an MLP network to obtain transformation features; Carrying out neighborhood feature aggregation on the transformation features by adopting maximum pooling to obtain single-scale local features; calculating the channel weight of the single-scale local feature through the channel attention; Screening high-value channels through channel weights, and extracting first feature submatrices corresponding to the high-value channels; Calculating the L2 norm of the node based on the first feature submatrix to obtain a spatial importance matrix, and generating spatial weights through the spatial importance matrix; screening high-value nodes through space weights, and extracting second feature submatrices corresponding to the high-value nodes; Recalculating channel weights through the second feature submatrices, and learning the single-scale local features through the spatial weights and the recalculated channel weights to obtain channel-space calibration features; self-adaptive attention learning is performed through channel-space calibration features, and the output features of the dynamic edge convolution module are generated by combining attention masks of geometric priors.
4. A depth learning based three-dimensional point cloud denoising method according to claim 3, wherein the adaptive attention learning by channel-space calibration features and generating the output features of the dynamic edge convolution module in combination with the attention mask of geometric priors comprises: Calculating the local density and normal vector consistency of each point through the original three-dimensional point cloud data; generating an attention mask by local density and normal vector consistency; multiplying the attention mask by the channel-space calibration feature node by node to obtain a geometric enhancement feature; acquiring a noise scene and a scene mask based on the local density, and calculating scene self-adaptive weights according to the noise scene; calculating self-adaptive weights through the scene self-adaptive weights and scene masks; and carrying out self-adaptive weighting on the geometric enhancement features through self-adaptive weights to generate the output features of the dynamic edge convolution module.
5. The depth learning-based three-dimensional point cloud denoising method according to claim 1, wherein the performing global feature extraction on original three-dimensional point cloud data in combination with multi-scale parallel processing to obtain multi-scale global features comprises: performing neighbor searching on each point in the original three-dimensional point cloud data based on the multi-scale neighbor parameters to obtain neighbor graphs under different scales; for each neighborhood graph, calculating the relative coordinates of each point and the neighborhood point of each point, and splicing the relative coordinates with the original coordinates of each point to obtain enhanced coordinate features under different scales; each enhanced coordinate feature sequentially passes through a global feature encoder, global attention pooling and global feature projection to perform feature extraction, so as to obtain single-scale global features under different scales; and performing element addition operation on the single-scale global features under different scales to obtain the multi-scale global features.
6. The depth learning-based three-dimensional point cloud denoising method according to claim 1, wherein after the enhancement features are decoded, displacement prediction is performed in combination with original three-dimensional point cloud data to obtain denoised three-dimensional point cloud data, and the method comprises the steps of: Decoding the enhancement features through a decoder, and obtaining residual displacement vectors by combining gradual dimension reduction and feature mapping; correcting the original three-dimensional point cloud data through the residual displacement vector to obtain corrected three-dimensional point cloud data; and carrying out local smoothing on the corrected three-dimensional point cloud data through the poisson curved surface to obtain denoising three-dimensional point cloud data.
7. Three-dimensional point cloud denoising system based on deep learning, which is characterized by comprising: the acquisition unit is used for acquiring original three-dimensional point cloud data; The feature coding unit is used for carrying out local feature coding on the original three-dimensional point cloud data by combining with the dynamic edge convolution attention enhancement to obtain multi-scale local features; The feature extraction unit is used for carrying out global feature extraction on the original three-dimensional point cloud data by combining multi-scale parallel processing to obtain multi-scale global features; the feature fusion unit is used for carrying out feature fusion on the multi-scale local features and the multi-scale global features to generate enhanced features; And the feature decoding unit is used for carrying out displacement prediction by combining the original three-dimensional point cloud data after decoding the enhanced features to obtain the denoised three-dimensional point cloud data.
8. The depth learning based three-dimensional point cloud denoising system of claim 7, wherein the feature encoding unit comprises: the setting subunit is used for setting a multi-scale neighborhood parameter; the first searching subunit is used for carrying out neighbor searching on each point in the original three-dimensional point cloud data based on the multi-scale neighbor parameters to obtain neighbor graphs under different scales; the splicing subunit is used for splicing the coordinates of the central point of each neighborhood graph with the coordinates of each neighborhood point to obtain edge characteristics under different scales; The first feature extraction subunit is used for extracting features of each edge feature through the dynamic edge convolution module to obtain first local features under different scales; The second feature extraction subunit is used for extracting the features of each first local feature through the dynamic edge convolution module to obtain second local features under different scales; and the first addition subunit is used for performing element addition operation based on the second local features under different scales to generate multi-scale local features.
9. The depth learning based three-dimensional point cloud denoising system of claim 7, wherein the feature extraction unit comprises: The second searching subunit is used for carrying out neighbor searching on each point in the original three-dimensional point cloud data based on the multi-scale neighbor parameters to obtain neighbor graphs under different scales; the enhancement subunit is used for calculating the relative coordinates of each point and the neighborhood point of each neighborhood graph, and splicing the relative coordinates with the original coordinates of each point to obtain enhancement coordinate characteristics under different scales; The third feature extraction subunit is used for extracting the features of each enhanced coordinate feature sequentially through a global feature encoder, global attention pooling and global feature projection to obtain single-scale global features under different scales; And the second adding subunit is used for carrying out element adding operation on the single-scale global features under different scales to obtain the multi-scale global features.
10. The depth learning based three-dimensional point cloud denoising system of claim 7, wherein the feature decoding unit comprises: the decoding subunit is used for decoding the enhancement features through a decoder and combining the gradual dimension reduction and feature mapping to obtain residual displacement vectors; the correction subunit is used for correcting the original three-dimensional point cloud data through the residual displacement vector to obtain corrected three-dimensional point cloud data; and the smoothing subunit is used for carrying out local smoothing on the corrected three-dimensional point cloud data through the poisson curved surface to obtain denoised three-dimensional point cloud data.

Description

Three-dimensional point cloud denoising method and system based on deep learning Technical Field The invention relates to the technical field of point cloud denoising, in particular to a three-dimensional point cloud denoising method and system based on deep learning. Background The three-dimensional point cloud is used as a core data form in the three-dimensional vision field, the geometric integrity and the accuracy of the three-dimensional point cloud directly determine the reliability of downstream tasks such as automatic driving, three-dimensional reconstruction, industrial detection and the like, but the original point cloud is easily limited by sensor noise, environmental interference and equipment accuracy in the acquisition process, various noises are inevitably introduced, and the usability of the data is seriously affected. The traditional three-dimensional point cloud denoising technology often depends on a unified characteristic input form, ignores the essential difference of point cloud neighborhood distribution in dense and sparse areas, namely, all points in the high-density area are given the same weight, so that local geometric fluctuation characteristics disappear due to excessive average, abnormal points are easily generated due to large factor fluctuation in the sparse area, and training stability is damaged. With the development of the deep learning technology in the three-dimensional vision field, the neural network-based point cloud denoising method gradually replaces the traditional method, but the existing scheme still does not effectively solve the problems that in a feature processing link, a self-adaptive feature extraction mechanism cannot be designed aiming at the neighborhood density difference, high-density region features are excessively averaged and sparse region information is weakened still exist, and in the utilization of directional information, the fine modeling of the geometric directional relation is lacking, so that complex structures such as inclined surfaces, curved surfaces and the like are difficult to accurately capture. In addition, the existing method still has the problem of insufficient cooperation between noise detection and point position adjustment, so that the real geometric structure and detail information of the point cloud are difficult to keep to the maximum extent while noise is effectively restrained, and the application requirements of high-precision scenes cannot be met. Disclosure of Invention The invention aims to provide a three-dimensional point cloud denoising method and system based on deep learning so as to solve the problems. In order to achieve the above purpose, the technical scheme adopted by the invention is as follows: In a first aspect, the present application provides a three-dimensional point cloud denoising method based on deep learning, including: Acquiring original three-dimensional point cloud data; carrying out local feature coding on the original three-dimensional point cloud data by combining with dynamic edge convolution attention enhancement to obtain multi-scale local features; Carrying out global feature extraction on the original three-dimensional point cloud data by combining multi-scale parallel processing to obtain multi-scale global features; Performing feature fusion on the multi-scale local features and the multi-scale global features to generate enhanced features; And after the enhancement features are decoded, carrying out displacement prediction by combining the original three-dimensional point cloud data to obtain the denoised three-dimensional point cloud data. In a second aspect, the present application further provides a three-dimensional point cloud denoising system based on deep learning, including: the acquisition unit is used for acquiring original three-dimensional point cloud data; The feature coding unit is used for carrying out local feature coding on the original three-dimensional point cloud data by combining with the dynamic edge convolution attention enhancement to obtain multi-scale local features; The feature extraction unit is used for carrying out global feature extraction on the original three-dimensional point cloud data by combining multi-scale parallel processing to obtain multi-scale global features; the feature fusion unit is used for carrying out feature fusion on the multi-scale local features and the multi-scale global features to generate enhanced features; And the feature decoding unit is used for carrying out displacement prediction by combining the original three-dimensional point cloud data after decoding the enhanced features to obtain the denoised three-dimensional point cloud data. The beneficial effects of the invention are as follows: (1) According to the method, by designing a dynamic edge convolution attention enhancement mechanism, accurate coding of multi-scale local features is realized, and the problems of feature overaverage or information weakening caused by nei