A heterogeneous noise point cloud denoising method based on contrast learning and intensity perception

Abstract

The application belongs to the field of three-dimensional point cloud denoising, and discloses a heterogeneous noise point cloud denoising method based on contrast learning and intensity perception. Firstly, a contrast learning feature extraction module is used to pre-train the point cloud, and robust high-dimensional feature representation is learned from the point cloud segments under different noise conditions. A spatial noise intensity prediction module uses a multi-layer perception machine to estimate the noise intensity of each point. Subsequently, a directional displacement prediction module calculates the directional displacement according to the local geometric features of the points, and the position of the points is optimized through step-by-step iteration. A dynamic iterative denoising strategy is adopted, and the number of iterations is adaptively adjusted according to the denoising progress, so that high-intensity noise can be removed while the local details of the object surface are preserved. The application shows significant application value in processing complex heterogeneous noise point clouds, and can be widely used in the fields of three-dimensional reconstruction, automatic driving, remote sensing surveying and mapping, etc.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of three-dimensional point cloud denoising, and particularly relates to a heterogeneous noise point cloud denoising method based on contrast learning and intensity perception. BACKGROUND

[0002] Laser sensors obtain object surface information through reflected signals, and intuitively represent three-dimensional space in the form of point clouds, and are widely used in automatic driving, robotics and remote sensing fields. However, the real-world scene is complex, and the point cloud data collected by the sensor usually contains a large amount of noise. The sources of these noises include internal factors such as device resolution and external factors such as weather conditions and the complexity of the object surface structure. In addition, the material and characteristics of the object surface will also significantly affect the generation of noise. Noise point clouds will interfere with downstream tasks such as 3D reconstruction, scene registration, semantic segmentation and target detection, so that point cloud denoising becomes a research focus.

[0003] The point cloud data in the real environment usually presents complex heterogeneous noise, rather than a single noise distribution. This complexity is due to various objects and materials in the scene. According to the Lambertian reflection principle, when the laser beam of the sensor scans the surface composed of different materials, due to the different reflection responses and reflection distribution functions of each material, the intensity of the reflected signal and the noise distribution will also change. Therefore, the laser reflection response of different objects and their noise characteristics differ significantly. In addition, factors such as scanning angle, object color and material characteristics will also affect the signal intensity and noise level of the point cloud. Even on the same object, due to the difference in material and color, the noise pattern of the point cloud data in different areas will also be different, so that the point cloud data essentially has high complexity. Especially in complex scenes, mutual occlusion between objects, multiple reflections and environmental scattering effects will further exacerbate the complexity of noise, so that the noise not only presents as local random errors, but also may present as systematic bias or structural noise. This multi-source and non-uniformly distributed noise feature makes it difficult for traditional rule-based denoising methods to be applicable, and how to adaptively model and remove heterogeneous noise in different scenes and object surfaces is still a challenging problem.

[0004] Despite these challenges, most existing point cloud denoising models are still primarily designed and evaluated based on a single type of synthetic noise with fixed intensity, failing to effectively address the heterogeneous noise problem widely existing in real-world scenarios. Currently, many methods mainly focus on outlier removal and surface smoothing, often only targeting simple noise distributions such as uniform noise or Gaussian noise, while ignoring the complexity of heterogeneous noise. In addition, objects are usually composed of different materials and colors, resulting in a large deviation from the assumptions in synthetic data, and different regions of the same object may have different noise intensities, making it difficult for existing methods to adapt to point cloud denoising in complex environments, and the denoised point cloud may still contain a lot of heterogeneous noise, affecting the surface quality of the 3D reconstruction task. In addition, many methods rely on globally fixed denoising parameters or filtering based on local geometric features, making it difficult to adaptively adjust to different object surfaces, resulting in over-smoothing in low-noise areas and insufficient denoising in high-noise areas. SUMMARY

[0005] To solve the above technical problems, the present application provides a heterogeneous noise point cloud denoising method based on contrast learning and intensity perception, which can have stronger adaptability in complex scenes, reducing deformation and detail loss, so that the denoised point cloud is both clean and retains the original geometric information.

[0006] To achieve the above purpose, the present application is realized by the following technical scheme:

[0007] The present application is a heterogeneous noise point cloud denoising method based on contrast learning and intensity perception, which specifically includes the following steps:

[0008] Step 1, samples are extracted from a noise-free point cloud dataset, and various data augmentation strategies are used to simulate heterogeneous noise in real scanning; the data augmentation strategies include adding Gaussian noise strategy and mixed disturbance strategy of Gaussian noise and uniform noise to more realistically reproduce the noise characteristics in different devices and scenarios;

[0009] Step 2, use noise-free point cloud and its noise point cloud to form a point cloud pair, use a contrast learning network to extract robust features from the point cloud pair, and optimize through a contrast loss to make the contrast learning network identify and separate noise information;

[0010] Step 3, use a multi-layer perception machine to combine the local geometric features of noise points, i.e. normal information and neighborhood distribution, to predict the displacement of each noise point relative to its true position, i.e. noise intensity, and obtain a noise intensity prediction module;

[0011] Step 4: Combining the local geometric features of the noise point, namely the relative relationship between the noise point and its neighboring noise points and the noise distribution, calculate the displacement vector of each noise point, so that the noise point moves closer to the true value (GT), that is, moves towards the ideal (ground truth) position.

[0012] Step 5: Dynamically adjust the number of iterations using the evaluation metric chamfer distance CD;

[0013] Step 6: Use a truncation function to limit the adjustment range of high noise points, avoid excessive movement of outliers, and prevent excessive adjustment of their positions during the denoising process, which would cause them to deviate from the true value (GT).

[0014] Step 7: Output the point cloud data after denoising.

[0015] A further improvement of the present invention is that step 1 specifically includes the following steps:

[0016] Step 1.1: Select the CLPCF dataset and the PU-Net dataset as test samples;

[0017] Step 1.2: Design a homogeneous noise dataset. Gaussian noise with different standard deviations was added to the CLPCF dataset, and the standard deviations of the noise added to the PU-Net dataset were 1%, 2%, and 3% of the radius of the boundary sphere, respectively.

[0018] Step 1.3: Design a heterogeneous noise dataset. Add Gaussian noise of different intensities to the CLPCF dataset, and add Gaussian noise and uniform noise to the PU-Net dataset.

[0019] A further improvement of the present invention is that step 2 specifically includes the following steps:

[0020] Step 2.1: Generate training samples by mapping the noise-free point cloud block X. i Add noise Φ i The noise set P is obtained. i :

[0021] P i ={X i X i +Φ i}

[0022] Where, Φ i =ψ(x, μ, σ), where x represents a point in the point cloud, μ represents the mean of the noise, controlling the center position of the noise, and σ represents the standard deviation of the noise, controlling the intensity or dispersion of the noise.

[0023] Step 2.2, from the noise set P i Randomly select patches S and T to form a point cloud pair (S, T):

[0024] S∈P i , T∈P i

[0025] wherein the patch S and the patch T adopt different sampling radii;

[0026] Step 2.3, constructing a patch set based on the point cloud pair (S, T) of step 2.2, the patch set containing positive point cloud pairs and negative point cloud pairs, the positive point cloud pairs being two point cloud blocks from the same point cloud manifold, and the negative point cloud pairs being point cloud blocks from different manifolds or different regions;

[0027] Step 2.4, optimizing the feature encoder of the contrastive learning network by minimizing the similarity between the positive point cloud pairs and maximizing the similarity between the negative point cloud pairs, using a normalized temperature scaling cross-entropy loss function for training, the normalized temperature scaling cross-entropy loss function being:

[0028]

[0029] wherein L S，T is the loss of the contrastive learning network, measuring the difference in similarity between the positive point cloud pairs and the negative point cloud pairs, |X i | is the number of noise-free noise-free point cloud blocks X i , a j，S , a j，T is the feature representation of the patch S and the patch T, sim(a j，S , a j，T ) is the cosine similarity of the positive point cloud pairs, representing the similarity of the patch S and the patch T in the feature space, ω is a temperature scaling parameter for adjusting the sensitivity of the similarity, is the weighted sum of the negative point cloud pairs, representing the similarity distribution of the negative point cloud pairs.

[0030] A further improvement of the present application is that the step 3 specifically comprises the following steps:

[0031] Step 3.1, inputting the coordinate information of each noise point and its k nearest neighbors, and passing the coordinate information through the robust feature extraction contrastive learning module trained in step 2 to obtain local geometric features, i.e. normal information and neighborhood distribution;

[0032] Step 3.2, sending the local geometric features obtained in step 3.1 into a five-layer multilayer perceptron (MLP) network for processing to predict the noise intensity wherein, is the noise point;

[0033] Step 3.3, calculating the true noise intensity by comparing the noise point with its corresponding clean point xj The position difference between the noise point and the clean point is projected onto the normal vector of the noise point The real noise intensity is calculated by projecting the position difference between the noise point and the clean point onto the normal vector of the noise point

[0034]

[0035] where n j is the normal vector of the noise point, Δ j is the offset of the noise point relative to its clean point x j ;

[0036] Step 3.4, to minimize the difference between the predicted noise intensity and the real value, an optimized loss function is designed:

[0037]

[0038] where M is the total number of all point cloud blocks, |X i | is the number of points of the point cloud block X i , and α is a hyperparameter for adjusting the sensitivity of the prediction error.

[0039] Further improvements of the present application are that the step 4 specifically comprises the following steps:

[0040] Step 4.1, combine the local geometric features of the noise points extracted in step 3 with the noise intensity, and input them into a displacement prediction network composed of five layers of MLP;

[0041] Step 4.2, introduce a position loss L p to minimize the Euclidean distance L j between the predicted point and its corresponding clean point x p :

[0042]

[0043] Step 4.3, introduce a distribution uniformity loss L DU by maximizing the distance between the predicted point and the farthest point in the real point cloud block:

[0044]

[0045] Step 4.4, the total loss L total is obtained by weighted sum of the position loss L p and the distribution uniformity loss L DU :

[0046] L total = (1-β)·L p + β·L DU

[0047] where β is a hyper-parameter, controlling the position loss L p and the distribution uniformity loss L DU with a weight between them.

[0048] Further improvement of the present application is that the step 5 specifically comprises the following steps:

[0049] Step 5.1, estimating the noise intensity of the noise point by using the noise intensity prediction module trained in step 3

[0050] Step 5.2, estimating the directional displacement of the noise point by using the displacement prediction network trained in step 4

[0051] Step 5.3, dynamically adjusting the offset of the noise point according to the noise intensity obtained in step 5.1 and the directional displacement obtained in step 5.2

[0052]

[0053] wherein, is the noise point displacement that needs to be adjusted in the denoising process;

[0054] Step 5.4, calculating the evaluation index chamfer distance (CD) after all noise points in the point cloud complete each round of denoising:

[0055]

[0056] wherein, X i is the noise-free point cloud block, is the denoised point cloud set;

[0057] Further improvement of the present application is that step 6 specifically comprises the following steps:

[0058] Step 6.1, limiting the adjustment range of high noise points by using a truncation function:

[0059]

[0060] wherein, high noise points are judged according to ρ(I), and θ(I) is an adaptive intensity threshold, which is set as the median of the noise intensity of all points in the noise point cloud, and when the noise intensity is greater than the median, it is set as a high noise point;

[0061] Step 6.2, after each round of denoising is completed, the high noise points judged by the truncation function are used to repeat the execution of steps 6.1 and 6.2 for denoising processing. ​

[0062] The further improvement of the present application is that the step 7 specifically comprises the following steps: when the index chamfer distance CD value reaches the minimum, the new position of each noise point in the point cloud is saved, and the denoised point cloud is saved as a PLY file.

[0063] The beneficial effects of the present application are:

[0064] The present application can simultaneously process different types and intensities of noise, including intensity heterogeneous noise (same type but different intensity) and type heterogeneous noise (different noise distribution mixed), making it more adaptable in complex scenes.

[0065] The present application estimates the noise level point by point through noise intensity prediction, dynamically adjusts the denoising intensity and the number of iterations, and performs enhanced denoising in high-noise areas, while avoiding over-smoothing in low-noise areas. In addition, directional displacement prediction optimizes the adjustment direction of the points in combination with geometric features, so that the denoised point cloud is more consistent with the real object surface, reducing deformation and detail loss.

[0066] The present application adopts an adaptive update strategy to specially optimize high-noise points, avoiding excessive deviation of the position of the points or structural damage during the denoising process, and ensuring that the denoised point cloud is both clean and maintains the original geometric information.

[0067] The present application significantly improves the quality of point cloud data in 3D reconstruction, target detection and semantic segmentation tasks, and improves the application reliability. BRIEF DESCRIPTION OF DRAWINGS

[0068] Figure 1 is a flowchart of the heterogeneous noise point cloud denoising method of the present application.

[0069] Figure 2 is a structural schematic diagram of the heterogeneous noise denoising network of the present application.

[0070] Figure 3 is a visual comparison diagram of the denoising effect of the present application and other algorithms on the PU-Net data set interfered by 1%, 2%, and 3% Gaussian noise and uniform noise mixed heterogeneous noise.

[0071] Figure 4 is a visual comparison diagram of the denoising effect of the present application and other algorithms on the CLPCF data set interfered by 0.8% Gaussian noise.

[0072] Figure 5 is a visual comparison diagram of the denoising effect of the present application and other algorithms on the Paris-rue-Madame data set. DETAILED DESCRIPTION

[0073] Embodiments of the present application will be described below with reference to drawings. Many practical details will be described in the following description in order to provide a thorough understanding of the present application. However, it will be apparent to those skilled in the art that the present application is not limited to these practical details. That is, these practical details are not necessary in some embodiments of the present application.

[0074] As shown in Figure 1 The present application is a heterogeneous noise point cloud denoising method based on contrastive learning and intensity perception. First, a contrastive learning feature extraction module is used to pre-train the point cloud, learning robust high-dimensional feature representations from point cloud segments under different noise conditions to enhance the model's perception of noise distribution. On this basis, a spatial noise intensity prediction module uses a multi-layer perceptron to estimate the noise intensity of each point, allowing the denoising process to adapt to different noise levels in different regions. Subsequently, a directional displacement prediction module calculates the directional displacement based on the local geometric features of the points, optimizing the position of the points through iterative methods to reduce the impact of noise while maintaining the structural integrity of the point cloud. To prevent over-smoothing and loss of detail, a dynamic iterative denoising strategy is used to adaptively adjust the number of iterations based on the denoising progress, ensuring that both high-intensity noise and local details on the object surface are preserved. In addition, the model introduces an adaptive update mechanism based on noise intensity, adjusting high-intensity noise points using a truncation function to avoid feature loss that may occur with traditional denoising methods. The present application has significant application value in handling complex heterogeneous noise point clouds and can be widely used in fields such as three-dimensional reconstruction, autonomous driving, and remote sensing mapping.

[0075] Specifically, the method specifically includes the following steps:

[0076] Step 1, extract samples from a noise-free point cloud dataset, and simulate heterogeneous noise in real scanning through various data augmentation strategies, wherein the data augmentation strategies include adding Gaussian noise strategy and mixed disturbance strategy of Gaussian noise and uniform noise to more realistically reproduce the noise characteristics under different devices and scenes.

[0077] Specifically, step 1 includes the following steps:

[0078] Step 1.1, select dataset CLPCF and dataset PU-Net as test samples;

[0079] Step 1.2, design a homogeneous noise dataset, add Gaussian noise with different standard deviations to dataset CLPCF, the specific standard deviations are 0.6%, 0.8%, 1.1%, 1.5%, and 2.0%, and add noise to dataset PU-Net with standard deviations of 1%, 2%, and 3% of the boundary sphere radius;

[0080] Step 1.3, Designing heterogeneous noise dataset, adding Gaussian noise with different intensities to the dataset CLPCF, the combination of standard deviations is (0.6%, 0.8%), (0.6%, 1.1%) and (0.6%, 2.0%); adding Gaussian noise and uniform noise to the dataset PU-Net. The standard deviation is set to 1%, 2%, and 3%.

[0081] Step 2, using the point cloud pair composed of the noise-free point cloud and its noise point cloud, extracting robust features from the point cloud pair using the contrast learning network; improving the perception and differentiation ability of the contrast learning network model to different noise modes, and optimizing through the contrast loss, so that the contrast learning network can better identify and separate noise information. Specifically, the following steps are included:

[0082] Step 2.1, generating training samples, adding noise i to the noise-free point cloud block X i to obtain the noise set P i :

[0083] P i ={X i ,X i +Φ i}

[0084] where Φ i =ψ(x, μ, σ), x represents a point in the point cloud block, μ represents the mean of the noise, which controls the center position of the noise, and σ represents the standard deviation of the noise, which controls the intensity or dispersion degree of the noise;

[0085] Step 2.2, randomly selecting patches S and T from the noise set P i to form a point cloud pair (S, T):

[0086] S∈P i , T∈P i

[0087] wherein the patch S and the patch T use different sampling radii;

[0088] Step 2.3, constructing a patch set based on the point cloud pair (S, T) of step 2.2, the patch set contains positive point cloud pairs and negative point cloud pairs, the positive point cloud pair is two point cloud blocks from the same point cloud manifold, and the negative point cloud pair is a point cloud block from different manifolds or different regions;

[0089] Step 2.4, optimizing the feature encoder of the contrast learning network by minimizing the similarity between positive point cloud pairs and maximizing the similarity between negative point cloud pairs, using a normalized temperature scaling cross-entropy loss function for training, the normalized temperature scaling cross-entropy loss function is:

[0090]

[0091] where L S，T is the loss of the contrastive learning network, measuring the similarity difference between positive point cloud pairs and negative point cloud pairs, |X i | is the noise-free noise point cloud block X i The number of points in the middle, a j，s , a j，T is the feature representation of patch S and patch T, sim(a j，s , a j，T ) is the cosine similarity of the positive point cloud pair, indicating the similarity of patch S and patch T in the feature space, ω is a temperature scaling parameter, used to adjust the sensitivity of the similarity, is the weighted sum of the negative point cloud pair, indicating the similarity distribution of the negative point cloud pair.

[0092] Step 3, use a multi-layer perception to train the intensity prediction network, combine the local geometric features of the noise points, i.e. normal information, neighborhood distribution, and predict the offset of each noise point relative to its true position, i.e. noise intensity. Specifically, it includes the following steps:

[0093] Step 3.1, input the coordinate information of each noise point and its k nearest neighbors, and obtain the local geometric features, i.e. normal information and neighborhood distribution, through the robust feature extraction contrast learning module trained in step 2;

[0094] Step 3.2, input the local geometric features obtained in step 3.1 into a five-layer multi-layer perception (MLP) network for processing to predict the noise intensity where, is the noise point;

[0095] Step 3.3, calculate the true noise intensity by projecting the position difference between the noise point and its corresponding clean point x j onto the normal vector of the noise point to calculate the true noise intensity

[0096]

[0097] where n j is the normal vector of the noise point, Δ j is the offset of the noise point relative to its clean point x j ;

[0098] Step 3.4, to minimize the difference between the predicted noise intensity and the true value, an optimized loss function is designed:

[0099]

[0100] where M is the total number of point cloud blocks, |X i is the number of points of point cloud block X i , and a is a hyperparameter used to adjust the sensitivity of the prediction error.

[0101] Step 4, combine the local geometric features of the noise points, i.e. the relative relationship between the noise points and the noise points in their neighborhood, and the noise distribution, to calculate the displacement vector of each noise point, so that the noise point moves towards a more reasonable geometric position. Specifically, it includes the following steps:

[0102] Step 4.1, in order to predict the directional displacement of the point cloud, combine the local geometric features of the noise points extracted in step 3 with the noise intensity and input them into a displacement prediction network composed of five layers of MLP; this process aims to optimize the spatial position of each point to make it as close as possible to the real point cloud, thereby improving the denoising effect.

[0103] Step 4.2, to optimize the spatial position of each point, introduce a position loss L p to minimize the Euclidean distance L between the predicted point j and its corresponding clean point x p :

[0104]

[0105] Step 4.3, to prevent the point cloud from being unevenly distributed or excessively clustered after denoising, introduce a distribution uniformity loss L DU by maximizing the distance between the predicted point and the farthest point in the real point cloud block:

[0106]

[0107] Step 4.4, to ensure that both the position of the point and the uniformity of the point cloud can be accurately restored during the optimization process, the total loss L p is obtained by weighted sum of the position loss L DU and the distribution uniformity loss L total :

[0108] L total = (1 - β) · L p + β · L DU

[0109] where β is a hyperparameter that controls the weight between the position loss L p and the distribution uniformity loss L DU .

[0110] Step 5, dynamically adjust the number of iterations using the evaluation index chamfer distance CD; specifically, it includes the following steps:

[0111] Step 5.1, estimate the noise intensity of the noise point using the noise intensity prediction module trained in step 3

[0112] Step 5.2, estimate the directional displacement of the noise point using the displacement prediction network trained in step 4

[0113] Step 5.3, according to the noise intensity obtained in step 5.1 and the directional displacement obtained in step 5.2 Adjust the offset of the noise point dynamically:

[0114]

[0115] where, is the noise point displacement that needs to be adjusted in the denoising process;

[0116] Step 5.4, when all noise points in the point cloud complete each round of denoising, calculate the evaluation index chamfer distance (CD):

[0117]

[0118] where, X i is the noise-free point cloud block, is the denoised point cloud set, if the index chamfer distance CD shows a downward trend, continue to execute the denoising step; if the index chamfer distance CD rises, stop denoising in the previous round;

[0119] Step 6, use the clipping function to limit the adjustment range of high noise points, avoid excessive movement of abnormal points, and prevent excessive adjustment of their positions in the denoising process, leading to deviation from the true value (GT). Excessive movement refers to the movement after moving to the ideal (ground truth) position, which leads to the increase of noise. Specifically, the following steps are included:

[0120] Step 6.1, to prevent over-denoising in the denoising process, use the clipping function to limit the adjustment range of high noise points:

[0121]

[0122] where, according to ρ(I) to judge high noise points, θ(I) is the adaptive intensity threshold, set as the median of the noise intensity of all points in the noise point cloud, when the noise intensity is greater than the median, set as high noise points;

[0123] Step 6.2, after each round of denoising is completed, according to the high noise points judged by the clipping function, repeat steps 6.1 and 6.2 to perform denoising processing.

[0124] Step 7, through the optimization of multiple steps such as feature extraction, noise intensity prediction, direction adjustment, adaptive denoising and updating, the noise of the point cloud is significantly reduced, while the integrity of the original geometric structure is preserved, and finally the denoised point cloud is saved. Specifically, when the index chamfer distance CD value reaches the minimum, the new position of each noise point in the point cloud is saved, and the denoised point cloud is saved as a PLY file.

[0125] In order to verify the advancement of the method in the present application, experiments are carried out and qualitative and quantitative comparisons are made with existing denoising methods.

[0126] For quantitative testing, for heterogeneous noise, the present application uses the CLPCF test set, which contains 23 shapes, each containing 100K points, and adds Gaussian noise of different intensities to each point cloud to evaluate model performance. The final denoising results are shown in Table 1, CD value multiplied by 10 -5 .

[0127] Table 1

[0128]

[0129] In addition, the PU-Net test set contains 20 shapes, and the present application down-samples each point cloud at resolutions of 10K and 50K, and adds a mixture of Gaussian noise and uniform noise with the same noise intensity to evaluate the model's performance under different noise types. The final denoising results are shown in Table 2, CD value multiplied by 10 -4 . Table 2 is a quantitative comparison of the denoising effect of the present application and other algorithms on the heterogeneous noise PU-Net data set.

[0130] Table 2

[0131]

[0132] In order to verify the denoising effect of the model under homogeneous noise, the present application again uses the CLPCF test set and only adds Gaussian noise to evaluate the model. The final denoising results are shown in Table 3.

[0133] Table 3

[0134]

[0135] For qualitative testing, the present application demonstrates its excellent denoising effect by visual evaluation on the CLPCF heterogeneous noise data set, the PU-Net homogeneous noise data set, and the Paris street data set. Figure 3 and Figure 4The visual comparison of the application and the existing competitive algorithm under the condition of heterogeneous noise and homogeneous noise is shown. Compared with the existing algorithm, the result of the application is clearer, which not only can better preserve the details, but also can more thoroughly remove the abnormal points, thereby significantly improving the visual effect and the sense of reality. In real-world applications, Figure 5 The denoising effect on the Paris-rue-Madame dataset is shown. Compared with the SC and CL methods, the application successfully retains more detailed features while removing noise, achieves more complete noise removal, and has more attractive visual effects.

[0136] In order to test the robustness of the model, different degrees of noise are added, and the specific results are shown in Table 4.

[0137] Table 4

[0138]

[0139] The application has significant application value in processing complex heterogeneous noise point clouds, and can be widely used in the fields of three-dimensional reconstruction, automatic driving, remote sensing surveying and mapping and the like.

[0140] The above only describes the embodiments of the application and is not used to limit the application. The application can have various changes and variations for those skilled in the art. Any modification, equivalent replacement, improvement and the like made within the spirit and principle of the application shall be included in the scope of claims of the application.

Claims

1. A method for denoising heterogeneous noisy point clouds based on contrastive learning and intensity perception, characterized in that: The heterogeneous noise point cloud denoising method specifically includes the following steps: Step 1: Extract samples from the noise-free point cloud dataset and simulate heterogeneous noise in real scanning through various data augmentation strategies. The data augmentation strategies include adding Gaussian noise and a mixed perturbation strategy of Gaussian noise and uniform noise. Step 2: Use noise-free point clouds and noisy point clouds to form point cloud pairs, and use a contrastive learning network to extract robust features from the point cloud pairs. Furthermore, optimization is achieved through contrastive loss, enabling the contrastive learning network to identify and separate noise information, with each noise point and its corresponding input noise level being considered. The coordinates of the nearest neighbors are used to obtain local geometric features, namely normal information and neighborhood distribution, through a robust feature extraction and contrastive learning network. Step 3: Using a multilayer perceptron, the local geometric features of the noise points, i.e., normal information, neighborhood distribution, and noise intensity are combined and input into a displacement prediction network composed of five layers of MLP to predict the offset of each noise point relative to its true position, i.e., the noise intensity, thus obtaining the noise intensity prediction module. Step 4: Combining the local geometric features of the noise point, namely the relative relationship between the noise point and its neighboring noise points and the noise distribution, calculate the displacement vector of each noise point, so that the noise point moves closer to the true value; Step 5: Use the evaluation metric of chamfer distance Dynamically adjust the number of iterations; Step 6: Use a truncation function to limit the adjustment range of high-noise points, avoid excessive movement of outliers, and prevent excessive adjustment of their positions during the denoising process, which could lead to deviation from the true value. Step 7: Output the point cloud data after denoising.

2. The heterogeneous noise point cloud denoising method based on contrastive learning and intensity perception according to claim 1, characterized in that: Step 1 specifically includes the following steps: Step 1.1: Select the CLPCF dataset and the PU-Net dataset as test samples; Step 1.2: Design a homogeneous noise dataset. Gaussian noise with different standard deviations was added to the CLPCF dataset, and the standard deviations of the noise added to the PU-Net dataset were 1%, 2%, and 3% of the radius of the boundary sphere, respectively. Step 1.3: Design a heterogeneous noise dataset. Add Gaussian noise of different intensities to the CLPCF dataset, and add Gaussian noise and uniform noise to the PU-Net dataset.

3. The heterogeneous noise point cloud denoising method based on contrastive learning and intensity perception according to claim 1, characterized in that: Step 2 specifically includes the following steps: Step 2.1: Generate training samples by feeding them into noise-free point cloud blocks. Add noise The noise set is obtained. : , in, , This represents a point in a point cloud block. This represents the mean value of the noise and controls the center location of the noise. It represents the standard deviation of noise and controls the intensity or dispersion of noise. Step 2.2, from the noise set Randomly select patch and patches , forming point cloud pairs : ， Among them, the patch and the patch Use different sampling radii; Step 2.3: Point cloud pairs based on Step 2.2 Construct a patch set, which contains positive point cloud pairs and negative point cloud pairs. The positive point cloud pairs are two point cloud blocks from the same point cloud manifold, and the negative point cloud pairs are point cloud blocks from different manifolds or different regions. Step 2.4: Optimize the feature encoder of the contrastive learning network by minimizing the similarity between positive point cloud pairs and maximizing the similarity between negative point cloud pairs. Train the network using the normalized temperature-scaled cross-entropy loss function, which is: ， in, To compare the loss of the learning network and measure the difference in similarity between positive and negative point cloud pairs, noise-free point cloud blocks The number of midpoints It's a patch. and patches Feature representation, Let cosine similarity be the sum of the positive point cloud pairs, representing the patch. and patches Similarity in feature space This is a temperature scaling parameter used to adjust the sensitivity of similarity. Let be the weighted sum of negative point cloud pairs, and let represent the similarity distribution of negative point cloud pairs.

4. The heterogeneous noise point cloud denoising method based on contrastive learning and intensity perception according to claim 1, characterized in that: Step 3 specifically includes the following steps: Step 3.1: The obtained local geometric features are fed into a five-layer multilayer perceptron network for processing to predict noise intensity. ,in, Noise point; Step 3.2: Calculate the actual noise intensity by analyzing the noise points. and its corresponding clean point The positional difference between them is projected onto the noise point. The true noise intensity is calculated on the normal vector. : ， in, The normal vector of the noise point. Noise point It is cleaner than that The offset; Step 3.3: To minimize the difference between the predicted noise intensity and the true value, an optimized loss function is designed: ， in, It is the total number of all point cloud blocks. It is a noise-free point cloud patch The number of points, α is a hyperparameter used to adjust the sensitivity to prediction error.

5. The heterogeneous noise point cloud denoising method based on contrastive learning and intensity perception according to claim 1, characterized in that: Step 4 specifically includes the following steps: Step 4.1: Introduce position loss To minimize the prediction points Its corresponding clean point Euclidean distance : ， in, It is the total number of all point cloud blocks. It is a noise-free point cloud patch The number of points, To predict noise intensity; Step 4.2: Introduce distribution uniformity loss By maximizing the distance between the predicted point and the farthest point in the actual point cloud block: ， Step 4.3, by analyzing the position loss and distribution uniformity loss The total loss is obtained by weighted summation. : ， in, It's a hyperparameter, controlling the position loss. and distribution uniformity loss The weights between the two.

6. The heterogeneous noise point cloud denoising method based on contrastive learning and intensity perception according to claim 1, characterized in that: Step 5 specifically includes the following steps: Step 5.1: Estimate the noise intensity at the noise point using the noise intensity prediction module trained in Step 3. ; Step 5.2: Use the displacement prediction network trained in Step 4 to estimate the directional displacement of the noise points. ; Step 5.3: Based on the noise intensity obtained in Step 5.1 and the directional displacement obtained in step 5.2 The offset of noise points is dynamically adjusted: ， in, Noise point The displacement that needs to be adjusted during the noise reduction process; Step 5.4: After all noisy points in the point cloud have completed each round of denoising, calculate the evaluation metric, chamfer distance CD. ， in, For noise-free point cloud blocks, This is the denoised point cloud set. It is a noise-free point cloud patch The number of points; 7. A heterogeneous noise point cloud denoising method based on contrastive learning and intensity perception according to claim 6, characterized in that: Step 6 specifically includes the following steps: Step 6.1: Use a truncation function to limit the adjustment range of high-noise points: ， Among them, according to Determine high noise points, The adaptive intensity threshold is set to the median of the noise intensity of all points in the noise point cloud. When the noise intensity is greater than the median, it is set as a high-noise point. Step 6.2: After each round of denoising is completed, based on the high noise points identified by the truncation function, repeat steps 6.1 and 6.2 for denoising processing.

8. A heterogeneous noise point cloud denoising method based on contrastive learning and intensity perception according to claim 7, characterized in that: Step 7 specifically includes the following steps: when the indicator chamfer distance When the value reaches its lowest point, save the new position of each noise point in the point cloud and save the denoised point cloud as a PLY file.

Images