A point cloud iterative denoising method based on hilbert joint feature coding
By employing Hilbert joint feature encoding and exponential decay iterative denoising methods, the problems of information loss and iteration non-convergence in existing point cloud denoising algorithms are solved, generating high-quality clean point clouds and reducing computational complexity.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANJING UNIV
- Filing Date
- 2024-04-22
- Publication Date
- 2026-06-09
AI Technical Summary
Existing point cloud denoising algorithms suffer from the problem of losing detailed features due to global max pooling operations when removing noise. Furthermore, iterative denoising methods fail to converge when the number of iterations is insufficient, or fail to converge to the real object surface in the later stages of iteration.
By employing Hilbert joint feature encoding, combining local details and global morphological representations, and through an exponentially decaying iterative denoising process, the adjustment magnitude of noisy point clouds in the later stages of iteration is reduced, resulting in clean point clouds.
It improves computational accuracy, reduces computational complexity, generates high-quality clean point clouds, and enhances the real-time performance and scalability of the algorithm.
Smart Images
Figure CN118351011B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of point cloud denoising technology, and in particular to an iterative point cloud denoising method based on Hilbert joint feature encoding. Background Technology
[0002] Recent advances in laser scanning and image processing have facilitated the acquisition of 3D point clouds from real-world scenes. However, due to limitations in the precision of acquisition devices (such as laser scanners) and the matching ambiguity inherent in image reconstruction, 3D point clouds obtained from scanning real-world scenes are often contaminated by noise and outliers. Despite significant advancements in 3D sensing technology in recent years, obtaining point clouds with low noise interference and complete detail remains an expensive and time-consuming task. Noisy point clouds severely hinder their application in various point cloud understanding domains, such as scene segmentation, object rendering, and surface reconstruction. Therefore, point cloud denoising has become an indispensable and crucial component of point cloud understanding. Point cloud denoising aims to remove noise and outlier interference from point clouds while preserving the shape characteristics of the objects themselves, and simultaneously maintaining the morphological properties of the point cloud data by retaining sharp edge details and surface geometric details.
[0003] With the development of deep learning, some researchers have used deep learning methods to remove point cloud noise. However, existing point cloud denoising algorithms have the following two problems: (1) The denoising encoding process usually uses local neighborhood graph convolution to extract features, and only uses global max pooling operation to obtain global features from the output local features. However, global max pooling operation only retains the most prominent channel features. While extracting the global structural information of the point cloud, it inevitably loses a lot of information, making it difficult to recover fine-grained details in the final point cloud output. (2) Some methods use iterative methods to continuously adjust the offset of the point cloud to remove noise. However, due to the influence of the number of iteration steps, if the number of iteration steps is small, the point cloud offset will not converge and cannot reach the optimal solution. Some methods use a constant coefficient iteration process, which makes it impossible for the denoised point cloud to converge to the surface of the real object in the later stages of iteration. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides an iterative point cloud denoising method based on Hilbert joint feature encoding, which improves computational accuracy and reduces computational complexity.
[0005] This invention patent uses Hilbert joint feature encoding to effectively model the local details and global morphological representation of point clouds, and employs an exponentially decaying iterative denoising process to reduce the adjustment magnitude of noisy point clouds in the later stages of iteration, thereby generating clean point clouds.
[0006] The technical solution adopted in this invention is as follows:
[0007] An iterative point cloud denoising method based on Hilbert joint feature encoding includes:
[0008] For point cloud data to be denoised The query point cloud and key-value point cloud are obtained based on Gaussian random perturbation, wherein the query point cloud is... The key value point cloud is in and Let N be the random perturbation matrix, N be the number of points in the point cloud, and D be the dimension of the point cloud.
[0009] The coordinates of the key-value point cloud are expanded using Hilbert space to obtain an ordered point cloud sequence.
[0010] Based on the ordered point cloud sequence Obtain key-value point cloud encoded features;
[0011] The positional offset of the query point cloud is calculated using a denoising decoder;
[0012] Based on the positional offset of the queried point cloud, point cloud noise is removed using an iterative denoising method.
[0013] The ordered point cloud sequence Obtain key-value point cloud encoded features, including:
[0014] The ordered point cloud sequence is encoded through a pre-feature extraction layer. Features of each point
[0015] based on Point cloud features representing local details of key-value point clouds are obtained through the Hilbert detail encoding attention module;
[0016] based on Point cloud features representing the global shape of the key-value point cloud are obtained through the Hilbert shape encoding attention module;
[0017] The point cloud features representing local details and the point cloud features representing global shape are restored to key-value point cloud encoded features using Hilbert space restoration operation.
[0018] The application denoising decoder calculates the positional offset of the query point cloud, including:
[0019] For the query point cloud, key point cloud, and key point cloud encoded features, neighborhood interpolation based on distance weights is applied to recover the query point cloud features.
[0020] The position offset of the query point cloud is obtained by applying a decoder based on the features of the query point cloud.
[0021] The further technical solution is as follows:
[0022] The coordinates of the key-value point cloud are expanded using Hilbert space to obtain an ordered point cloud sequence. include:
[0023] The point set is obtained by using point cloud spatial normalization.
[0024]
[0025] in and These are the center coordinates and sphere radius of the key value point cloud, respectively:
[0026]
[0027] in N is the number of points in the point cloud, and the subscript i is the point number. ε = 10 -6 ;
[0028] Compute point set The corresponding M-order Hilbert order in For the expansion curve of Hilbert space of order M, It is the floor function;
[0029] The ordered point cloud sequence is obtained by reconstructing the point cloud set according to the M-order Hilbert order.
[0030]
[0031] in, The subscripts i and j are the point numbers, and [·] represents an ordered sequence.
[0032] The ordered point cloud sequence is encoded through a pre-feature extraction layer. Features of each point for:
[0033]
[0034]
[0035] in: C represents the number of output features; The feature aggregation function consists of a multilayer perceptron with feature numbers ranging from 3·D to C and a max pooling function; For point exist The neighborhood index; C is the point-by-point fusion function, which aggregates the center coordinates. Local neighborhood coordinates Relative coordinate offset from local neighborhood To explicitly capture the local shape representation of point clouds.
[0036] The basis Point cloud features representing local details of the key-value point cloud are obtained through the Hilbert detail encoding attention module, including:
[0037] Local features are obtained by grouping and partitioning GP. in N / G and G are the number of groups and the number of points within a group, respectively;
[0038] An attention module is used to calculate the unnormalized similarity weight for each point.
[0039]
[0040] in, This is a multilayer perceptron shared within the group, where the superscript T stands for matrix transpose;
[0041] The similarity weights of each point within a group are balanced using the Softmax activation function δ, resulting in normalized similarity weights.
[0042]
[0043] in,
[0044] Based on the normalized similarity weights Weighting the point cloud features within a group yields the detail coding features.
[0045]
[0046] in,
[0047] Ordered point cloud features obtained after Hilbert detail encoding attention module for:
[0048]
[0049]
[0050] in, GR stands for group recovery. This is a multilayer perceptron used for feature extraction, where r is the feature scaling ratio; BN is the batch normalization operation.
[0051] The basis Point cloud features for global shape representation of key-value point clouds are obtained through a Hilbert shape encoding attention module, including:
[0052] Features of ordered point clouds After uniform sampling and partitioning, N / U ordered point cloud sub-image features and coordinates are obtained. Where UP represents the uniform sampling partitioning operation, and U is the uniform sampling coefficient;
[0053] Point cloud features after Hilbert shape encoding attention module for:
[0054]
[0055]
[0056]
[0057]
[0058] in, To share the multilayer perceptron, δ′ is the Softmax activation function, and UR is the uniform sampling restoration operation.
[0059] The step of obtaining the position offset of the query point cloud by applying a decoder based on the query point cloud features includes:
[0060] The decoder is used to obtain the position offset of the query point cloud.
[0061]
[0062] in, It is a four-layer learnable multilayer perceptron with input features D+C and output features D, and the last layer of the perceptron does not use an activation function; The query point cloud features recovered by the application based on distance-weighted neighborhood interpolation:
[0063]
[0064] In the formula, Indicates query point In key value point cloud Neighborhood index in The j-th key-value point cloud encoding feature is defined here.
[0065] The step of removing point cloud noise based on the position offset of the queried point cloud using an iterative denoising method includes:
[0066] The coordinates of the denoised point cloud are continuously fine-tuned through T-step iterations based on exponential decay of momentum, making it increasingly closer to the true point cloud. The iteration process for step t∈{1,...,T} is as follows:
[0067]
[0068]
[0069] in, Let t be the coordinates of the query point cloud after the t-th iteration. To initially query the point cloud coordinates, Let be the momentum matrix after the t-th iteration. Let α be the initial momentum matrix, α be the iteration step size, β be the iteration decay coefficient, and γ be the momentum coefficient. It is a key-value point cloud The query point cloud is the result of the (t-1)th iteration. The offset obtained after the noise reduction decoder;
[0070] The final denoised point cloud is
[0071] The beneficial effects of this invention are as follows:
[0072] This invention combines Hilbert joint encoding and other related techniques to address the complex point cloud denoising problem. It designs a complete iterative point cloud denoising scheme based on Hilbert joint feature encoding. The scheme effectively models the local details and global morphological representation of the point cloud using Hilbert joint feature encoding, and employs an exponentially decaying iterative denoising process to reduce the adjustment amplitude of noisy point clouds in the later stages of iteration, thereby generating clean point clouds. This effectively reduces the computational complexity and resource consumption of the algorithm, while significantly improving accuracy. It also outperforms other existing methods in terms of real-time performance, scalability, and load balancing.
[0073] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. Attached Figure Description
[0074] Figure 1 This is a flowchart illustrating the method of an embodiment of the present invention. Detailed Implementation
[0075] The specific embodiments of the present invention are described below with reference to the accompanying drawings.
[0076] See Figure 1 This embodiment of an iterative point cloud denoising method based on Hilbert joint feature encoding includes:
[0077] S1. For the point cloud data to be denoised The query point cloud and key-value point cloud are obtained based on Gaussian random perturbation, wherein the query point cloud is... The key-value point cloud is ,in, , and The perturbation matrix is random; in this embodiment, the perturbation is applied only to the query point cloud. N is the number of points in the point cloud, and D is the dimension of the point cloud.
[0078] S2. The coordinates of the key-value point cloud are expanded using Hilbert space to obtain an ordered point cloud sequence. Specifically, it includes:
[0079] S21. Obtain the normalized point set using point cloud spatial normalization. :
[0080]
[0081] in, and These are the center coordinates and sphere radius of the key value point cloud, respectively:
[0082]
[0083] in, N is the number of points in the point cloud, and the subscript i is the point number. ε = 10 -6 ;
[0084] S22, Calculate the point set The corresponding M-order Hilbert order in, For the expansion curve of Hilbert space of order M, It is the floor function;
[0085] S23. Reassemble the point cloud set according to the M-order Hilbert order to obtain the ordered point cloud sequence. :
[0086]
[0087] in, The subscripts i and j are the point numbers, and [·] represents an ordered sequence.
[0088] S3, Based on the ordered point cloud sequence Obtain key-value point cloud encoded features, including:
[0089] S31. Encode ordered point cloud sequences through a pre-feature extraction layer. Features of each point
[0090]
[0091] in: C represents the number of output features; The feature aggregation function consists of a multilayer perceptron with feature numbers ranging from 3·D to C and a max pooling function; For point exist The neighborhood index is denoted by c; c is the point-by-point fusion function, which aggregates the center coordinates. Local neighborhood coordinates Relative coordinate offset from local neighborhood To explicitly capture the local shape representation of point clouds.
[0092] S32, based on Point cloud features representing local details of the key-value point cloud are obtained through the Hilbert detail encoding attention module, including:
[0093] (1) Local features are obtained by grouping and partitioning (GP). in N / G and G are the number of groups and the number of points within a group, respectively;
[0094] (2) Use an attention module to calculate the unnormalized similarity weight for each point.
[0095]
[0096] in, This is a multilayer perceptron shared within the group, where the superscript T indicates matrix transpose;
[0097] (3) Use the Softmax activation function δ to balance the similarity weights of each point in the group to obtain the normalized similarity weights.
[0098]
[0099] in,
[0100] (4) Based on the normalized similarity weights Weighting the point cloud features within a group yields the detail coding features.
[0101]
[0102] in, All All are multilayer perceptrons shared within the group;
[0103] Therefore, the ordered point cloud features obtained after passing through the Hilbert detail encoding attention module for:
[0104]
[0105]
[0106] In the above formula, GR stands for group reverse. This is a multilayer perceptron used for feature extraction, where r is the feature scaling ratio; BN is the batch normalization operation.
[0107] S33, based on Point cloud features for global shape representation of key-value point clouds are obtained through a Hilbert shape encoding attention module, including:
[0108] Features of ordered point clouds After uniform sampling and partitioning, N / U ordered point cloud sub-image features and coordinates are obtained. , where UP represents uniform sampling partitioning operation, and U is the uniform sampling coefficient;
[0109] The operation of the shape attention module is similar to that of the detail attention module. The point cloud features after passing through the Hilbert shape encoding attention module... for:
[0110]
[0111]
[0112]
[0113]
[0114] in, To share the multilayer perceptron, δ′ is the Softmax activation function, and UR is the uniform sampling reverse operation.
[0115] It is understood that the Hilbert detail encoding attention module and the Hilbert shape encoding attention module in this embodiment specifically use a combination of Hilbert space unfolding curves and attention modules to achieve the corresponding functions. Furthermore, different partitioning methods are used in the Hilbert detail and shape attention modules to explicitly guide the model to learn the local and global point cloud representations of each point in detail.
[0116] S34. The point cloud features representing local details and global shape are restored to key-value point cloud encoded features using Hilbert space restoration operations.
[0117] S4. Calculate the position offset of the query point cloud using a denoising decoder, including:
[0118] S41, Regarding the query point cloud Key-value point cloud Key-value point cloud encoding features Application of distance-weighted neighborhood interpolation to recover query point cloud features
[0119]
[0120] in, Indicates query point In key value point cloud Neighborhood index in.
[0121] S42, Based on the query point cloud features The decoder is used to obtain the position offset of the query point cloud.
[0122]
[0123] in, It is a four-layer learnable multilayer perceptron with input features D+C and output features D, and the last layer of the perceptron does not use an activation function;
[0124] S5. Based on the positional offset of the queried point cloud, remove point cloud noise using an iterative denoising method, including:
[0125] The coordinates of the denoised point cloud are continuously fine-tuned through T-step iterations based on exponential decay of momentum, making it increasingly closer to the true point cloud. The iteration process for step t∈{1,…,T} is as follows:
[0126]
[0127]
[0128] in, Let t be the coordinates of the query point cloud after the t-th iteration. To initially query the point cloud coordinates, Let be the momentum matrix after the t-th iteration. Let α be the initial momentum matrix, α be the iteration step size, β be the iteration decay coefficient, and γ be the momentum coefficient. It is a key-value point cloud The query point cloud is the result of the (t-1)th iteration. The offset obtained after the noise reduction decoder;
[0129] The final denoised point cloud is
[0130] It will be understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A point cloud iterative denoising method based on Hilbert joint feature encoding, characterized in that, include: For point cloud data to be denoised The query point cloud and key-value point cloud are obtained based on Gaussian random perturbation, wherein the query point cloud is... The key-value point cloud is ,in , and , The random perturbation matrix, For the number of points in the point cloud, The dimension of the point cloud; The coordinates of the key-value point cloud are expanded using Hilbert space to obtain an ordered point cloud sequence. ; Based on the ordered point cloud sequence Obtain key-value point cloud encoded features; The positional offset of the query point cloud is calculated using a denoising decoder; Based on the positional offset of the queried point cloud, point cloud noise is removed using an iterative denoising method. The ordered point cloud sequence Obtain key-value point cloud encoded features, including: The ordered point cloud sequence is encoded through a pre-feature extraction layer. Features of each point : based on Point cloud features representing local details of key-value point clouds are obtained through the Hilbert detail encoding attention module; based on Point cloud features representing the global shape of the key-value point cloud are obtained through the Hilbert shape encoding attention module; The point cloud features representing local details and the point cloud features representing global shape are restored to key-value point cloud encoded features using Hilbert space restoration operation. The application denoising decoder calculates the positional offset of the query point cloud, including: For the query point cloud, key point cloud, and key point cloud encoded features, neighborhood interpolation based on distance weights is applied to recover the query point cloud features. The position offset of the query point cloud is obtained by applying a decoder based on the features of the query point cloud; The basis Point cloud features representing local details of the key-value point cloud are obtained through the Hilbert detail encoding attention module, including: Grouping Obtain local features ,in , and These are the number of groups and the number of points within each group, respectively. An attention module is used to calculate the unnormalized similarity weight for each point. : ， in, , For a multilayer perceptron shared within the group, superscript T This is the matrix transpose. Use the Softmax activation function By balancing the similarity weights of each point within a group, we obtain the normalized similarity weights. : in, ; Based on the normalized similarity weights Weighting the point cloud features within a group yields the detail coding features. : in, ; Ordered point cloud features obtained after Hilbert detail encoding attention module for: in, , For group recovery, Multilayer perceptrons are used for feature extraction. is the feature scaling ratio, and BN is the batch normalization operation.
2. The point cloud iterative denoising method based on Hilbert joint feature encoding according to claim 1, characterized in that, The coordinates of the key-value point cloud are expanded using Hilbert space to obtain an ordered point cloud sequence. ,include: The point set is obtained by using point cloud spatial normalization. : ,in and These are the center coordinates and sphere radius of the key value point cloud, respectively: in , , , The index represents the number of points in the point cloud. The point number, ; Compute point set corresponding Hilbert's order ,in , for Hilbert space expansion curve of order 1, It is the floor function; According to the above The ordered point cloud sequence is obtained by recombining the point cloud set according to the Hilbert order. : ， in, subscript , These are the point numbers, It represents an ordered sequence.
3. The point cloud iterative denoising method based on Hilbert joint feature encoding according to claim 2, characterized in that, The ordered point cloud sequence is encoded through a pre-feature extraction layer. Features of each point for: ， in: , Indicates the number of output features; For feature aggregation functions, which are derived from the feature number... arrive It consists of a multilayer perceptron and a max-pooling function; For point exist Neighborhood index in; This is a point-by-point fusion function, which uses the coordinates of the aggregation center. Local neighborhood coordinates Relative coordinate offset from local neighborhood This allows for the explicit capture of local shape representations of point clouds.
4. The point cloud iterative denoising method based on Hilbert joint feature encoding according to claim 1, characterized in that, The basis Point cloud features for global shape representation of key-value point clouds are obtained through a Hilbert shape encoding attention module, including: Features of ordered point clouds After uniform sampling and partitioning, the following results were obtained. Features and coordinates of an ordered point cloud sub-image UP represents the uniform sampling partitioning operation. The sampling coefficient is uniform. Point cloud features after Hilbert shape encoding attention module for: in, To share the multilayer sensor, The Softmax activation function is used. This is for uniform sampling and restoration operations.
5. The point cloud iterative denoising method based on Hilbert joint feature encoding according to claim 4, characterized in that, The step of obtaining the position offset of the query point cloud by applying a decoder based on the query point cloud features includes: The decoder is used to obtain the position offset of the query point cloud. : ,in, , The input features are Output features are A four-layer learnable multilayer perceptron, with the last layer of the perceptron not using an activation function; The query point cloud features recovered by the application based on distance-weighted neighborhood interpolation: In the formula, , Indicates query point In key value point cloud Neighborhood index in For the first Key-value point cloud encoding features.
6. The point cloud iterative denoising method based on Hilbert joint feature encoding according to claim 5, characterized in that, The step of removing point cloud noise based on the position offset of the queried point cloud using an iterative denoising method includes: pass The coordinates of the denoised point cloud are continuously fine-tuned through iterative steps based on the exponential decay of momentum, making them increasingly closer to the true point cloud. The iterative process is as follows: in, For the first The query point cloud coordinates after step-by-step iterative adjustment To initially query the point cloud coordinates, For the first The momentum matrix after one iteration. The initial momentum matrix, The iteration step size, For the iterative decay coefficient, The momentum coefficient, It is a key-value point cloud And query point cloud as the first After one iteration The offset obtained after the noise reduction decoder; The final denoised point cloud is .