An end-to-end point cloud denoising model construction method and system
By employing an end-to-end training paradigm and fusing multiple supervisory signals, combined with dynamic graph convolutional networks and multilayer perceptrons, the problem of insufficient geometric constraints in existing point cloud denoising models is solved, achieving high-precision and high-fidelity point cloud denoising results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CE CENT FOR ENG RES TEST & APPRAISAL
- Filing Date
- 2025-11-03
- Publication Date
- 2026-07-24
AI Technical Summary
Existing deep learning-based point cloud denoising models ignore geometric constraints during the denoising process, resulting in a lack of explicit modeling of the local geometric characteristics of the point cloud, an inability to maintain sharpness, and insufficient generalization ability.
An end-to-end training paradigm is adopted, and a combined loss function is constructed by fusing multiple supervision signals. By combining dynamic graph convolutional networks and multilayer perceptrons, and utilizing multi-scale dynamic neighborhood sampling and geometric constraint loss, a deep learning model is built to denoise point clouds.
It achieves globally optimal denoising performance, improves denoising accuracy and reliability, significantly reduces dependence on manually labeled data, enhances the model's generalization ability and robustness, and protects the sharp features and edge structures of point clouds.
Smart Images
Figure CN121328618B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of deep learning technology, and in particular to an end-to-end point cloud denoising model construction method and system. Background Technology
[0002] In recent years, with the development of deep learning technology in the field of 3D vision, a variety of point cloud denoising methods based on neural networks have emerged.
[0003] Existing deep learning-based point cloud denoising models typically use standard point cloud processing networks (such as PointNet++, DGCNN, etc.) as the backbone network to directly regress the coordinates of the denoised points, which still has obvious limitations in terms of architecture design and training paradigm.
[0004] The limitations of existing denoising models are mainly reflected in the fact that they typically only focus on the overall matching degree between point clouds, ignoring the geometric constraints that the denoising results should satisfy. This leads to a lack of explicit modeling of the local geometric characteristics of the point clouds and a failure to effectively embed geometric prior knowledge into the network structure. This purely data-driven approach is prone to causing the network to exhibit "over-smoothing," especially in areas with obvious geometric features such as edges and corners. It often fails to maintain the original sharpness and may result in geometric distortion or structures that violate physical laws after denoising. Summary of the Invention
[0005] Based on the above analysis, the embodiments of the present invention aim to provide an end-to-end point cloud denoising model construction method and system to solve the problems of insufficient generalization ability and low fidelity of output point cloud geometric features of existing denoising models.
[0006] On one hand, embodiments of the present invention provide an end-to-end point cloud denoising model construction method, including the following steps: After preprocessing the collected clean point cloud, point cloud samples are generated by injecting noise to construct a sample set; A deep learning model is constructed, and a combined loss function is built by fusing multiple supervision signals. The deep learning model includes, in sequence, a feature extraction module, a denoising-guided prediction module, and a point cloud attribute prediction module, which are used to output the predicted coordinates and predicted normal vectors of the denoised point cloud. The deep learning model is trained end-to-end based on the sample set and the combined loss function, and the trained deep learning model is used as the point cloud denoising model.
[0007] Based on the above method, a further improvement is made by fusing multiple supervision signals to construct a combined loss function, including: The supervised loss is constructed based on manually labeled signals, the self-supervised loss is constructed based on data-generated signals, and the geometric constraint loss is constructed based on knowledge-driven signals. The manually labeled signals are derived from the real labels of clean point clouds, the data-generated signals are derived from the pseudo labels generated by the geometric transformation of the data itself, and the knowledge-driven signals are derived from geometric prior knowledge.
[0008] Based on further improvements to the above methods, the feature extraction module in the deep learning model adopts a dynamic graph convolutional network to extract local geometric features from the input point cloud samples; the denoising and guided prediction module adopts a first multilayer perceptron to output denoising and guided vectors for each point cloud based on the local geometric features of the point cloud according to the score matching theory; the point cloud attribute prediction module adopts a second multilayer perceptron to output coordinate displacement vectors and prediction normal vectors based on the local geometric features and denoising and guided vectors of the point cloud, wherein the coordinate displacement vectors are added to the coordinates of the input point cloud samples to obtain the predicted coordinates.
[0009] Further improvements to the above method involve end-to-end training of the deep learning model based on the sample set and the combined loss function, including: Obtain point cloud samples from the current training batch from the sample set, apply random geometric transformations to the point cloud samples, and obtain geometrically transformed samples; During training, two forward propagations are performed: point cloud samples are input into the deep learning model for supervised learning to obtain the first predicted point cloud; geometric transformation samples are input into the deep learning model for self-supervised learning to obtain the second predicted point cloud. In backpropagation, the supervised loss and geometric constraint loss are calculated based on the first prediction point cloud; the self-supervised loss is calculated based on the first prediction point cloud and the second prediction point cloud; and then the network parameters of the deep learning model are optimized based on the calculated total loss.
[0010] Based on a further improvement of the above method, the self-supervised loss is obtained by applying the same geometric transformation to the first predicted point cloud, using the transformed first predicted point cloud as a pseudo-label for the second predicted point cloud, and calculating the difference between the second predicted point cloud and its pseudo-label.
[0011] Based on further improvements to the above method, the geometric constraint loss includes: normal consistency loss, projection consistency loss, and curvature smoothing loss; among which, The normal consistency loss is obtained by calculating the difference between the normal vector of the first predicted point cloud and the normal vector estimated by principal component analysis; The projection consistency loss is obtained by calculating the distance between the first predicted point cloud and its projection points on the local tangent plane; The curvature smoothing loss is obtained by calculating the curvature difference between the first predicted point cloud and its nearest neighbor points in the neighborhood point set.
[0012] Based on a further improvement of the above method, the supervised loss is obtained by calculating the difference between the first predicted point cloud and the corresponding clean point cloud.
[0013] Based on the further improvement of the above method, the combined loss function also includes manifold reconstruction loss, which is obtained by calculating the distance between the first predicted point cloud and the local surface fitted by the point cloud sample using the moving least squares method.
[0014] Based on further improvements to the above method, the collected clean point cloud is preprocessed, including: Based on the ratio of local density to global density of each point cloud, adaptive neighborhood parameters for each clean point cloud are obtained, and then dynamic neighborhood sampling is performed to obtain the neighborhood point set of each point cloud. The coordinates of each point cloud are normalized based on the neighborhood point set of each point cloud.
[0015] On the other hand, embodiments of the present invention provide an end-to-end point cloud denoising model construction system, including: The sample set construction module is used to preprocess the collected clean point cloud and then generate point cloud samples by injecting noise to construct the sample set. The model building module is used to build a deep learning model and construct a combined loss function by fusing multiple supervision signals. The deep learning model includes, in turn, a feature extraction module, a denoising-guided prediction module, and a point cloud attribute prediction module, which are used to output the predicted coordinates and predicted normal vectors of the denoised point cloud. The model training module is used to perform end-to-end training of the deep learning model based on the sample set and the combined loss function, and the trained deep learning model is used as a point cloud denoising model.
[0016] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: 1. By adopting an end-to-end training paradigm, feature extraction, denoising-guided prediction, and displacement field learning are unified under a single optimization framework, avoiding the error accumulation problem of traditional multi-stage methods and achieving globally optimal denoising performance. Moreover, through denoising-guided prediction based on score matching theory, an optimization direction with a statistical theoretical basis is provided for point cloud displacement, enabling the deep learning model to learn noise distribution characteristics more accurately and improving the accuracy and reliability of denoising.
[0017] 2. By integrating multiple supervisory signals to construct a combined loss function, data-driven learning, self-supervised signals, and geometric prior knowledge are combined, enabling the model to mine supervisory signals from the point cloud data itself, significantly reducing the dependence on manually labeled data and enhancing the model's generalization ability.
[0018] 3. Through the synergistic effect of geometric constraint losses such as normal consistency, projection consistency and curvature smoothing, the sharp features and edge structures of the point cloud are effectively protected, avoiding the over-smoothing problem common in traditional deep learning methods, and significantly improving the geometric fidelity of the denoising results.
[0019] 4. By constructing a sample set through a combination of multi-scale dynamic neighborhood sampling and noise injection, the model can learn and adapt to denoising strategies for different point cloud density regions. By integrating geometric prior knowledge and combining supervised learning and self-supervised learning training processes, the model's generalization ability and robustness to non-uniformly distributed point clouds are significantly improved, achieving a technical breakthrough in obtaining high-precision and high-fidelity denoising effects under limited labeled data conditions.
[0020] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description
[0021] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Figure 1 This is a flowchart of an end-to-end point cloud denoising model construction method in Embodiment 1 of the present invention. Detailed Implementation
[0022] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.
[0023] Example 1 A specific embodiment of the present invention discloses an end-to-end point cloud denoising model construction method, such as... Figure 1 As shown, it includes the following steps: S1. After preprocessing the collected clean point cloud, point cloud samples are generated by injecting noise to construct a sample set.
[0024] It should be noted that when preprocessing clean point clouds, a fixed neighborhood radius and number of nearest neighbors may fail to accurately describe geometric features in sparse regions due to insufficient points, or reduce computational efficiency and feature discriminative power in dense regions due to excessive noise and irrelevant details. Therefore, a multi-scale strategy is introduced to dynamically adjust the neighborhood size. This allows for the selection of a smaller neighborhood in high-density regions to capture fine geometric details and avoid over-smoothing; while in low-density regions, the neighborhood is gradually expanded to ensure a sufficient number of effective neighboring points for stable feature computation. This dynamic adjustment mechanism effectively overcomes the limitations of fixed neighborhood parameters, making the feature extraction process more adaptable to the inherent non-uniformity of point clouds, thereby significantly improving subsequent geometric feature extraction.
[0025] Specifically, the multi-scale strategy obtains the adaptive neighborhood parameters of each point cloud based on the ratio of local density to global density, and then performs dynamic neighborhood sampling to obtain the neighborhood point set of each point cloud.
[0026] The adaptive neighborhood allocation parameters for each point cloud include: the neighborhood search radius and the number of nearest neighbors, which are calculated using the following formulas: , in, and They represent point clouds respectively. The neighborhood search radius and the number of nearest neighbors, and These represent the preset baseline values for the neighborhood search radius and the number of nearest neighbors, respectively, which are preset initial values based on the overall scale of the entire point cloud; Point cloud The local density is calculated from the point cloud. It is obtained by counting the number of points per unit volume within a local cube centered on the cube. The global density is represented by the ratio of the bounding box volume of the entire point cloud to the total number of points. and These represent the sensitivity indices for the neighborhood search radius and the number of nearest neighbors, respectively. , The SALib library was used for global sensitivity analysis and optimization to determine the optimal response. This represents the preset minimum number of nearest neighbor points, used to ensure that there are enough points for basic calculations even in the sparsest regions.
[0027] Furthermore, for each point cloud's neighborhood point set By calculating the mean and standard deviation of the neighborhood point set, the coordinates of each point cloud are normalized, scaling the data to a uniform scale and eliminating dimensional differences. Each point cloud and its normalized neighborhood information will be fixed and used in subsequent operations.
[0028] Furthermore, the preprocessed clean point cloud Injecting Gaussian noise, impulse noise, or a mixture of these noises generates a noisy point cloud. This forms a "noisy-clean" point cloud pair for supervised learning; noisy point clouds Each point cloud is represented as However, the index order of the points remains unchanged from the neighborhood relationship, ensuring that each noisy point cloud is consistent with its corresponding clean point cloud in terms of geometric features.
[0029] Furthermore, for the generated noisy point cloud computing geometric prior values: estimate the normal vector and the moving least squares fitted value.
[0030] Specifically, the distribution of each neighborhood point set in three-dimensional space constitutes a tiny approximate plane for each point cloud. The covariance matrix of the neighborhood point set is calculated using principal component analysis (PCA), and the eigenvector corresponding to the smallest eigenvalue is used as the point cloud. Estimated normal vector .
[0031] For each point cloud The neighborhood point set is used to fit a local surface by weighting the neighborhood points, and then the positions of the points projected onto the fitted local surface are calculated to obtain the moving least squares fitted value. .
[0032] The generated noisy point cloud is used as a point cloud sample, and the corresponding clean point cloud is used as the true label of the point cloud sample. The point cloud sample, its estimated normal vector and moving least squares fitted value, and the clean point cloud are put into the sample set.
[0033] S2. Construct a deep learning model and build a combined loss function by fusing multiple supervision signals. The deep learning model includes, in sequence, a feature extraction module, a denoising-guided prediction module, and a point cloud attribute prediction module, which are used to output the predicted coordinates and predicted normal vectors of the denoised point cloud.
[0034] It should be noted that the feature extraction module in the deep learning model uses a dynamic graph convolutional network (DGCNN) to extract local geometric features from the input point cloud samples.
[0035] The denoising guidance prediction module uses a first-level multilayer perceptron to output the denoising guidance vector of each point cloud based on the local geometric features of the point cloud, according to the fractional matching theory.
[0036] The point cloud attribute prediction module uses a second multilayer perceptron to output coordinate displacement vectors and prediction normal vectors based on the local geometric features of the point cloud and the denoising steering vector. The coordinate displacement vector is added to the coordinates of the input point cloud sample to obtain the predicted coordinates.
[0037] Specifically, the Dynamic Graph Convolutional Network (DGCNN) in the feature extraction module extracts features through multiple edge convolution operations and aggregates neighborhood information by concatenating the features, thus achieving a single message passing, as shown in the formula below: , in, Indicates the first Mid-layer point cloud The feature representation starts with the three-dimensional coordinates of each point cloud and uses a multi-layered stacked neural network. In each layer, the current features of the point are combined with the features and relationships of its neighboring points to gradually calculate and refine new and higher-level feature representations, encoding the geometric and semantic information of the points. It represents the feature difference between the nearest neighbor and the center point, captures the curvature of local set changes, and reflects the geometric and semantic differences between the two points; This represents a multilayer perceptron function that learns the weight assignments of local geometry. This represents the max pooling operation, which performs max pooling on all points in the neighborhood. The output should be the maximum value.
[0038] The first multilayer perceptron in the denoising-guided prediction module is based on fractional matching theory. It receives local geometric features extracted by the dynamic graph convolutional network and encodes the gradient information of the point cloud probability density function, providing precise guidance with a statistical theoretical basis for subsequent coordinate prediction. The formula is shown below: , in, Point cloud The denoising steering vector describes the gradient field of the noise point cloud distribution and points in the direction of probability density growth, which is also the denoising direction. The probability density distribution function represents a noisy point cloud; Represents computational point cloud The gradient is used to indicate the direction of point cloud displacement optimization.
[0039] The second multilayer perceptron in the point cloud attribute prediction module includes a displacement field prediction unit and a coordinate reconstruction unit. The displacement field prediction unit accurately estimates the offset direction and magnitude by splicing local geometric features with denoised steering vectors, combining local geometric priors and global context information, and predicts the coordinate displacement vector and normal vector of each point cloud. Then, the coordinate reconstruction unit adds the coordinate displacement vector to the coordinates of the input point cloud sample to obtain the predicted coordinates, and finally outputs the predicted coordinates and predicted normal vector.
[0040] The point cloud attribute prediction module achieves coordinated optimization of coordinate denoising and geometric attribute estimation, which not only improves the denoising quality but also provides high-quality normal information, facilitating subsequent 3D processing tasks.
[0041] Furthermore, this embodiment constructs a combined loss function by fusing multiple supervision signals, including: A supervised loss is constructed based on manually labeled signals, where the manually labeled signals are derived from the real labels of clean point clouds. A self-supervised loss is constructed based on data-generated signals, where the data-generated signals originate from pseudo-labels generated by the geometric transformation of the data itself. Geometric constraint loss is constructed based on knowledge-driven signals, where the knowledge-driven signals originate from geometric prior knowledge.
[0042] Specifically, the supervised loss is constructed by calculating the spatial distance difference between the predicted point cloud and the real point cloud output by the deep learning model for point cloud samples, as shown in the following formula: , in, Indicates monitoring losses; Indicates real point cloud collection A point cloud (i.e., a clean point cloud); This represents the predicted point cloud set generated by the deep learning model from point cloud samples. A point cloud; and Predict the number of point clouds and the number of real point clouds respectively.
[0043] Self-supervised loss is constructed by calculating the difference between the output of the deep learning model on geometrically transformed samples and the output of the deep learning model on point cloud samples after the same transformation, as shown in the formula below: , in, Indicates self-monitoring loss, Represents the geometric transformation function. This indicates the denoising network processing operation.
[0044] Geometric constraint loss is constructed by incorporating prior geometric knowledge of the point cloud to constrain the predicted point cloud output by the deep learning model to conform to geometric structural features. Geometric constraint loss includes: normal consistency loss, projection consistency loss, and curvature smoothing loss.
[0045] The normal consistency loss is used to minimize the difference between the predicted normal vector of the point cloud and the estimated normal vector of the point cloud sample, constraining the predicted normal vector to be consistent with the geometrically calculated normal vector, thus preserving edge features; the formula is shown below: , in: This indicates the loss of normal uniformity. Point cloud The estimated normal vector, This indicates a deep learning model for point clouds. The output predicted normal vector; This represents the L2 norm.
[0046] The projection consistency loss is used to minimize the distance between the predicted point cloud and its projected points on the local tangent plane, constraining the denoised point cloud to fall on the local tangent plane and preventing the point cloud from deviating excessively from the geometric surface; the formula is shown below: , in, This indicates the loss of projection consistency. Point cloud Local tangent plane The coordinates of the projection point on the surface; It is formed by the normal line Defined tangent plane; This indicates the operation of projecting onto the tangent plane, which offsets the point cloud. Project onto the tangent plane and remove the normal direction component; Point cloud The center point of the neighborhood point set, This indicates taking the absolute value.
[0047] The curvature smoothing loss is used to minimize the curvature difference between the predicted point cloud and its nearest neighbors in its neighborhood point set, constraining the curvature change of neighboring points to be gradual and suppressing abrupt curvature changes caused by noise; the formula is shown below: , in, Indicates curvature smoothing loss; Point cloud The curvature; Point cloud neighborhood point set Nearest neighbor in The curvature.
[0048] Preferably, the combined loss function further includes: manifold reconstruction loss; the manifold reconstruction loss is constructed by calculating the distance between the predicted point cloud output by the deep learning model and the local surface fitted by the point cloud samples using the moving least squares method; this loss embeds the potential smooth manifold prior knowledge of the point cloud into the network training process, effectively avoiding non-physical distortions in the denoising results. The formula is shown below: , in, Indicates the manifold reconstruction loss. Point cloud The moving least squares fitted value.
[0049] Finally, the weighted sum of the supervised loss, self-supervised loss, geometric constraint loss, and manifold reconstruction loss is taken as the total loss of the combined loss function, as shown in the following formula: , in, Indicates the total loss. , , , , and These represent the weight coefficients of the corresponding loss, which are preset based on experience before training and fine-tuned on the validation set.
[0050] S3. Based on the sample set and combined loss function, perform end-to-end training of the deep learning model, and use the trained deep learning model as the point cloud denoising model.
[0051] It should be noted that end-to-end training includes: ① Obtain point cloud samples from the current training batch from the sample set, apply random geometric transformations to the point cloud samples, and obtain geometrically transformed samples.
[0052] It should be noted that three types of data are prepared in each training batch: Point cloud samples, i.e. noisy point clouds, are used as input for supervised learning; Clean point clouds serve as a true label for supervised learning; Geometric transformation samples apply a random geometric transformation, such as rotating 30° around the Z-axis, to noisy point cloud samples for self-supervised learning; the random geometric transformation is a reversible rigid transformation.
[0053] ②During training, two forward propagations are performed: point cloud samples are input into the deep learning model for supervised learning to obtain the first predicted point cloud; geometric transformation samples are input into the deep learning model for self-supervised learning to obtain the second predicted point cloud; Specifically, supervised learning involves receiving noisy point cloud samples and using clean point clouds as the true labels for supervised learning, which is the core task of denoising; self-supervised learning involves receiving geometrically transformed samples and generating signals based on the output of supervised learning.
[0054] ③ In backpropagation, the supervised loss and geometric constraint loss are calculated based on the first prediction point cloud; the self-supervised loss is calculated based on the first prediction point cloud and the second prediction point cloud; and then the network parameters of the deep learning model are optimized based on the calculated total loss.
[0055] In backpropagation, the loss value is calculated based on the calculation formulas for each loss in step S2.
[0056] Specifically, the supervised loss is obtained by calculating the difference between the first predicted point cloud and the corresponding clean point cloud.
[0057] The normal consistency loss in the geometric constraint loss is obtained by calculating the difference between the normal vector of the first predicted point cloud and the normal vector estimated by principal component analysis; the projection consistency loss is obtained by calculating the distance between the first predicted point cloud and its projection points on the local tangent plane. The curvature smoothing loss is obtained by calculating the curvature difference between the first predicted point cloud and its nearest neighbor points in the neighborhood point set.
[0058] The manifold reconstruction loss is obtained by calculating the distance between the first predicted point cloud and the local surface fitted by the point cloud sample using the moving least squares method.
[0059] Self-supervised loss is obtained by applying the same geometric transformation to the first predicted point cloud, using the transformed first predicted point cloud as a pseudo-label for the second predicted point cloud, and calculating the difference between the second predicted point cloud and its pseudo-label.
[0060] In other words, the predicted point cloud output from noisy point cloud samples in supervised learning is subjected to the same geometric transformation as the current training batch and used as the prediction target for self-supervised learning. The self-supervised loss is calculated during backpropagation. The self-supervised loss cleverly uses random geometric transformations as a powerful source of supervisory signals. It does not rely on a large amount of labeled clean data, but rather mines supervisory information from the point cloud data itself. This not only ensures geometric consistency but also greatly improves the applicability and robustness of the method.
[0061] This embodiment combines multiple losses to balance data-driven and prior knowledge; it employs momentum gradient descent, combined with learning rate decay and early stopping strategies, to stabilize the training process. By minimizing the total loss of the combined loss function, it learns the mapping relationship from noisy point clouds to clean point clouds, optimizes the network parameters in the deep learning model, and finally trains a point cloud denoising model that satisfies both point cloud denoising and normal vector estimation.
[0062] Compared with existing technologies, this embodiment provides an end-to-end point cloud denoising model construction method. Employing an end-to-end training paradigm, it unifies feature extraction, denoising-guided prediction, and displacement field learning within a single optimization framework. This avoids the error accumulation problem of traditional multi-stage methods and achieves globally optimal denoising performance. Furthermore, through denoising-guided prediction based on score matching theory, it provides a statistically grounded optimization direction for point cloud displacement, enabling the deep learning model to more accurately learn noise distribution characteristics and improving the accuracy and reliability of denoising. By fusing multiple supervisory signals to construct a combined loss function, it combines data-driven learning, self-supervised signals, and geometric prior knowledge, allowing the model to mine supervisory signals from the point cloud data itself. This significantly reduces reliance on manually labeled data and enhances the model's generalization ability. Through the synergistic effect of geometric constraint losses such as normal consistency, projection consistency, and curvature smoothing, it effectively protects the sharp features and edge structures of the point cloud, avoiding the over-smoothing problem common in traditional deep learning methods and significantly improving the geometric fidelity of the denoising results. By constructing a sample set through a combination of multi-scale dynamic neighborhood sampling and noise injection, the model can learn denoising strategies to adapt to different point cloud density regions. By integrating geometric prior knowledge and combining supervised learning and self-supervised learning training processes, the model's generalization ability and robustness to non-uniformly distributed point clouds are significantly improved, achieving a technological breakthrough in obtaining high-precision and high-fidelity denoising effects even under limited labeled data conditions.
[0063] Example 2 Another embodiment of the present invention discloses an end-to-end point cloud denoising model construction system, thereby implementing the end-to-end point cloud denoising model construction method in Embodiment 1. The specific implementation of each module is described in the corresponding description in Embodiment 1. The system includes: The sample set construction module is used to preprocess the collected clean point cloud and then generate point cloud samples by injecting noise to construct the sample set. The model building module is used to build a deep learning model and construct a combined loss function by fusing multiple supervision signals. The deep learning model includes, in turn, a feature extraction module, a denoising-guided prediction module, and a point cloud attribute prediction module, which are used to output the predicted coordinates and predicted normal vectors of the denoised point cloud. The model training module is used to perform end-to-end training of the deep learning model based on the sample set and the combined loss function, and to use the trained deep learning model as a point cloud denoising model.
[0064] Since the end-to-end point cloud denoising model construction system of this embodiment and the aforementioned end-to-end point cloud denoising model construction method are related and can be mutually referenced, this description is redundant and will not be repeated here. Because the principle of this system embodiment is the same as that of the above method embodiment, this system embodiment also has the corresponding technical effects of the above method embodiment.
[0065] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.
[0066] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.
Claims
1. An end-to-end point cloud denoising model construction method, characterized in that, Includes the following steps: After preprocessing the collected clean point cloud, point cloud samples are generated by injecting noise to construct a sample set; A deep learning model is constructed, and a combined loss function is built by fusing multiple supervision signals, including: a supervised loss based on manually labeled signals, a self-supervised loss based on data-generated signals, and a geometric constraint loss based on knowledge-driven signals. The knowledge-driven signals originate from geometric prior knowledge calculated from point cloud samples. The geometric constraint losses include: normal consistency loss, projection consistency loss, and curvature smoothing loss. The deep learning model sequentially includes: a feature extraction module, a denoising-guided prediction module, and a point cloud attribute prediction module, used to output the predicted coordinates and predicted normal vectors of the denoised point cloud. The feature extraction module is used to extract local geometric features from the input point cloud samples. The denoising-guided prediction module is used to output the denoising-guided vectors of each point cloud based on the local geometric features of the point cloud according to the score matching theory. The point cloud attribute prediction module includes a displacement field prediction unit and a coordinate reconstruction unit. The displacement field prediction unit predicts the coordinate displacement vector and normal vector of each point cloud by concatenating local geometric features and the denoising-guided vector. Then, the coordinate reconstruction unit adds the coordinate displacement vector to the coordinates of the input point cloud sample to obtain the predicted coordinates. The deep learning model is trained end-to-end based on the sample set and the combined loss function, including: obtaining point cloud samples of the current training batch from the sample set, applying random geometric transformations to the point cloud samples to obtain geometrically transformed samples; performing two forward propagations during training: inputting the point cloud samples into the deep learning model for supervised learning to obtain a first predicted point cloud; inputting the geometrically transformed samples into the deep learning model for self-supervised learning to obtain a second predicted point cloud; in the backpropagation, the self-supervised loss is obtained by applying the same geometric transformation to the first predicted point cloud, using the transformed first predicted point cloud as a pseudo-label for the second predicted point cloud, and calculating the difference between the second predicted point cloud and its pseudo-label; the geometric constraint loss is constructed by introducing geometric prior knowledge to constrain the first predicted point cloud output by the deep learning model to conform to geometric structural features; and using the trained deep learning model as a point cloud denoising model.
2. The end-to-end point cloud denoising model construction method according to claim 1, characterized in that, The manually labeled signal comes from the real labels of the clean point cloud, while the data-generated signal comes from the pseudo labels generated by the geometric transformation of the data itself.
3. The end-to-end point cloud denoising model construction method according to claim 1, characterized in that, The feature extraction module in the deep learning model uses a dynamic graph convolutional network; the denoising and guided prediction module uses a first multilayer perceptron; and the point cloud attribute prediction module uses a second multilayer perceptron.
4. The end-to-end point cloud denoising model construction method according to claim 1, characterized in that, The normal consistency loss is obtained by calculating the difference between the normal vector of the first predicted point cloud and the normal vector estimated by principal component analysis; The projection consistency loss is obtained by calculating the distance between the first predicted point cloud and its projection points on the local tangent plane; The curvature smoothing loss is obtained by calculating the curvature difference between the first predicted point cloud and its nearest neighbor points in the neighborhood point set.
5. The end-to-end point cloud denoising model construction method according to claim 1, characterized in that, The supervised loss is obtained by calculating the difference between the first predicted point cloud and the corresponding clean point cloud.
6. The end-to-end point cloud denoising model construction method according to claim 1, characterized in that, The combined loss function also includes a manifold reconstruction loss, which is obtained by calculating the distance between the first predicted point cloud and the local surface fitted by the point cloud sample using the moving least squares method.
7. The end-to-end point cloud denoising model construction method according to claim 1, characterized in that, The preprocessing of the collected clean point cloud includes: Based on the ratio of local density to global density of each point cloud, adaptive neighborhood parameters for each clean point cloud are obtained, and then dynamic neighborhood sampling is performed to obtain the neighborhood point set of each point cloud. The coordinates of each point cloud are normalized based on the neighborhood point set of each point cloud.
8. An end-to-end point cloud denoising model construction system, used to implement the end-to-end point cloud denoising model construction method according to any one of claims 1-7, characterized in that, include: The sample set construction module is used to preprocess the collected clean point cloud and then generate point cloud samples by injecting noise to construct the sample set. The model building module is used to build a deep learning model and construct a combined loss function by fusing multiple supervision signals. The deep learning model includes, in sequence, a feature extraction module, a denoising-guided prediction module, and a point cloud attribute prediction module, which are used to output the predicted coordinates and predicted normal vectors of the denoised point cloud. The model training module is used to perform end-to-end training of the deep learning model based on the sample set and the combined loss function, and to use the trained deep learning model as a point cloud denoising model.