Heterogeneous noise point cloud denoising method based on comparative learning and intensity perception

Abstract

The invention belongs to the field of three-dimensional point cloud denoising, and discloses a heterogeneous noise point cloud denoising method based on comparative learning and intensity perception, which comprises the following steps: pre-training a point cloud through a comparative learning feature extraction module, and learning robust high-dimensional feature representation from point cloud fragments under different noise conditions; the spatial noise intensity prediction module estimates the noise intensity of each point by using a multi-layer perceptron, and then the directional displacement prediction module calculates directional displacement according to the local geometric features of the points, optimizes the positions of the points in a step-by-step iteration mode, and adaptively adjusts the number of iterations according to the denoising progress by adopting a dynamic iteration denoising strategy, so that the directional displacement of the points is predicted. Therefore, high-intensity noise can be removed, and local details on the surface of the object can be reserved. The method shows remarkable application value in the aspect of processing complex heterogeneous noise point cloud, and can be widely applied to the fields of three-dimensional reconstruction, automatic driving, remote sensing mapping and the like.

Description

Technical Field

[0001] The present invention belongs to the field of 3D point cloud denoising, and specifically relates to a heterogeneous noise point cloud denoising method based on contrast learning and intensity perception. Background Art

[0002] Laser sensors obtain the surface information of objects through reflected signals and intuitively represent the three-dimensional space in the form of point clouds, which are widely used in fields such as autonomous driving, robotics, and remote sensing. However, the real-world scene is complex, and the point cloud data collected by sensors 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. Noisy point clouds will interfere with downstream tasks such as 3D reconstruction, scene registration, semantic segmentation, and target detection, making point cloud denoising a research focus.

[0003] The point cloud data in the real environment usually presents complex heterogeneous noises rather than a single noise distribution. This complexity stems from various different 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 and noise distribution of the reflected signal will also change accordingly. Therefore, there are significant differences in the laser reflection responses and their noise characteristics of different objects. In addition, factors such as the 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 different materials and colors, the noise patterns of the point cloud data in different regions will also be different, making the point cloud data inherently highly complex. Especially in complex scenes, the mutual occlusion, multiple reflections, and environmental scattering effects between objects will further exacerbate the complexity of the noise, making the noise not only appear as local random errors but also may present systematic biases or structural noises. This multi-source and non-uniform distribution of noise characteristics makes traditional rule-based denoising methods difficult to apply, and how to adaptively model and remove heterogeneous noises on different scenes and object surfaces remains a challenging problem.

[0004] Despite these challenges, most existing point cloud denoising models are still mainly designed and evaluated based on a single type of synthetic noise with a fixed intensity, failing to effectively address the heterogeneous noise problems widely existing in real-world scenarios. Many current methods mainly focus on outlier removal and surface smoothing, usually 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 between their noise distributions and the assumptions in synthetic data. Different regions of the same object may have different noise intensities, making it difficult for existing methods to be applicable to point cloud denoising in complex environments. The denoised point cloud may still have a lot of heterogeneous noise remaining, thus affecting the surface quality of the 3D reconstruction task. Moreover, many methods rely on globally fixed denoising parameters or filtering based on local geometric features, making it difficult to adaptively adjust for different object surfaces, resulting in over-smoothing in low-noise regions and insufficient denoising in high-noise regions. Summary of the Invention

[0005] To solve the above technical problems, the present invention provides a heterogeneous noise point cloud denoising method based on contrast learning and intensity perception. This method can have stronger adaptability in complex scenarios, reduce deformation and detail loss, making the denoised point cloud both clean and able to retain the original geometric information.

[0006] To achieve the above object, the present invention is implemented through the following technical solutions:

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

[0008] Step 1: Extract samples from the noise-free point cloud dataset and simulate heterogeneous noise in real scans through various data augmentation strategies; the data augmentation strategies include the Gaussian noise addition strategy and the mixed perturbation strategy of Gaussian noise and uniform noise to more realistically reproduce the noise characteristics under different devices and scenarios;

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

[0010] Step 3: Use a multi-layer perceptron, combined with the local geometric features of the noisy points, namely normal information and neighborhood distribution, to predict the offset of each noisy point relative to its true position, that is, the noise intensity, to obtain a noise intensity prediction module;

[0011] 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 neighborhoods, and the noise distribution, and calculate the displacement vector of each noise point to make the noise point approach the ground truth (GT), that is, move it to 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, resulting in deviation from the ground truth (GT).

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

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

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

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

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

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

[0020] Step 2.1: Generate training samples, add noise Φ i to the noise-free point cloud patch X i to obtain a noise set P i :

[0021] P i ={X i , X i +Φ i}

[0022] where Φ i =ψ(x, μ, σ), x represents a certain point in the point cloud patch, μ represents the mean of the noise, controlling the central position of the noise, and σ represents the standard deviation of the noise, controlling the intensity or spread of the noise;

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

[0024] S ∈ P i ,T ∈ P i

[0025] Among them, the patch S and the patch T adopt different sampling radii;

[0026] Step 2.3: Construct a patch set based on the point cloud pair (S, T) in Step 2.2. The patch set includes positive point cloud pairs and negative point cloud pairs. The positive point cloud pair is two point cloud patches from the same point cloud manifold, and the negative point cloud pair is a point cloud patch from different manifolds or different regions;

[0027] Step 2.4: Optimize 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, and use the normalized temperature-scaled cross-entropy loss function for training. The normalized temperature-scaled cross-entropy loss function is:

[0028]

[0029] where L S，T is the loss of the contrast learning network, measuring the similarity difference between positive point cloud pairs and negative point cloud pairs. |X i | is the number of points in the noise-free point cloud patch X i without noise. a j，S , a j，T are the feature representations of the patches S and T. sim(a j，S , a j，T ) is the cosine similarity of the positive point cloud pair, representing the similarity between the patches S and T in the feature space. ω is the temperature-scaling parameter, used to adjust the sensitivity of the similarity. is the weighted sum of negative point cloud pairs, representing the similarity distribution of negative point cloud pairs.

[0030] A further improvement of the present invention is that: Step 3 specifically includes the following steps:

[0031] Step 3.1: Input the coordinate information of each noise point and its k nearest neighbor points, and pass through the robust feature extraction contrast learning module trained in Step 2 to obtain local geometric features, namely normal information and neighborhood distribution;

[0032] Step 3.2: Feed the local geometric features obtained in Step 3.1 into a five-layer multi-layer perceptron (MLP) network for processing to predict the noise intensity where is the noise point;

[0033] Step 3.3: Calculate the true noise intensity by subtracting the noise point from its corresponding clean point xj The positional difference between them is projected onto the normal vector of this noise point to calculate the true noise intensity

[0034]

[0035] where n j is the normal vector of the noise point, and 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 true value, an optimized loss function is designed:

[0037]

[0038] where M is the total number of all point cloud patches, |X i | is the number of points in the point cloud patch X i , and α is a hyperparameter used to adjust the sensitivity of the prediction error

[0039] A further improvement of the present invention lies in that: Step 4 specifically includes 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 a five-layer MLP

[0041] Step 4.2: Introduce a position loss L p to minimize the Euclidean distance L between the predicted point j 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 true point cloud patch:

[0044]

[0045] Step 4.4: Obtain the total loss L p by performing a weighted sum of the position loss L DU and the distribution uniformity loss L total :

[0046] L total =(1 - β)·L p +β·L DU

[0047] Among them, β is a hyperparameter that controls the position loss L p and the distribution uniformity loss L DU The weight between the two.

[0048] A further improvement of the present invention lies in that: step 5 specifically includes the following steps:

[0049] Step 5.1: Use the noise intensity prediction module trained in step 3 to estimate the noise intensity of the noise points

[0050] Step 5.2: Use the displacement prediction network trained in step 4 to estimate the directional displacement of the noise points

[0051] Step 5.3: According to the noise intensity obtained in step 5.1 And the directional displacement obtained in step 5.2 Dynamically adjust the offset of the noise points:

[0052]

[0053] Among them, Is the displacement that needs to be adjusted for the noise points During the denoising process;

[0054] Step 5.4: When all the noise points in the point cloud complete each round of denoising, calculate the evaluation index Chamfer Distance (CD):

[0055]

[0056] Among them, X i Is the noise-free point cloud block, Is the set of point clouds after denoising;

[0057] A further improvement of the present invention lies in that: step 6 specifically includes the following steps:

[0058] Step 6.1: Use a truncation function to limit the adjustment range of the high-noise points:

[0059]

[0060] Among them, high-noise points are judged according to ρ(I), θ(I) is an adaptive intensity threshold, set to the median of the noise intensities 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, according to the high-noise points judged by the truncation function, repeat steps 6.1 and 6.2 for denoising processing.

[0062] A further improvement of the present invention lies in that: Step 7 specifically includes the following steps: When the index chamfer distance CD value reaches the lowest, save the new positions of each noise point in the point cloud, and save the denoised point cloud as a PLY file.

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

[0064] The present invention can simultaneously process different types and intensities of noise, including intensity heterogeneous noise (i.e., the same type but different intensities) and type heterogeneous noise (i.e., a mixture of different noise distributions), making it more adaptable in complex scenarios.

[0065] The present invention estimates the noise level point by point through noise intensity prediction, dynamically adjusts the denoising intensity and the number of iterations, performs enhanced denoising in high-noise areas, and avoids over-smoothing in low-noise areas. In addition, the directional displacement prediction combines geometric features to optimize the adjustment direction of points, making the denoised point cloud more conform to the real object surface, reducing deformation and detail loss.

[0066] The present invention adopts an adaptive update strategy to specially optimize high-noise points, avoid excessive displacement or structural damage of points during the denoising process, and ensure that the denoised point cloud is both clean and can maintain the original geometric information.

[0067] The present invention significantly improves the quality of point cloud data in tasks such as 3D reconstruction, object detection, and semantic segmentation, and improves the application reliability. Brief Description of the Drawings

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

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

[0070] Figure 3 is a schematic visual comparison diagram of the denoising effects of the method of the present invention and other algorithms on the PU-Net dataset under heterogeneous noise interference mixed with 1%, 2%, and 3% Gaussian noise and uniform noise.

[0071] Figure 4 is a schematic visual comparison diagram of the denoising effects of the present invention and other algorithms on the CLPCF dataset under 0.8% Gaussian noise interference.

[0072] Figure 5 is a schematic visual comparison diagram of the denoising effects of the present invention and other algorithms on the Paris Street dataset Paris-rue-Madame. Detailed Embodiments

[0073] The embodiments of the present invention will be disclosed below with reference to the drawings. For the sake of clarity, many practical details will be described together in the following description. However, it should be understood that these practical details are not used to limit the present invention. That is to say, in some embodiments of the present invention, these practical details are not necessary.

[0074] As Figure 1 shown, the present invention is a heterogeneous noise point cloud denoising method based on contrast learning and intensity perception. First, the point cloud is pre-trained through a contrast learning feature extraction module to learn a robust high-dimensional feature representation from point cloud segments under different noise conditions, so as to enhance the model's perception ability of the noise distribution. On this basis, the spatial noise intensity prediction module uses a multi-layer perceptron to estimate the noise intensity of each point, enabling the denoising process to adaptively adjust to the noise levels in different regions. Subsequently, the directional displacement prediction module calculates the directional displacement according to the local geometric features of the points, and optimizes the positions of the points in a step-by-step iterative manner to reduce the influence of noise while maintaining the structural integrity of the point cloud. In order to prevent over-smoothing and detail loss, a dynamic iterative denoising strategy is adopted to adaptively adjust the number of iterations according to the denoising progress, so as to ensure that both high-intensity noise can be removed and the local details of the object surface can be retained. In addition, the model also introduces an adaptive update mechanism based on noise intensity, and a truncation function is used to adjust the high-intensity noise points to avoid the feature loss problem that may be caused by traditional denoising methods. The present invention shows significant application value in dealing with complex heterogeneous noise point clouds and can be widely used in fields such as 3D reconstruction, autonomous driving, and remote sensing mapping.

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

[0076] Step 1: Extract samples from the noise-free point cloud dataset and simulate heterogeneous noise in real scans through various data augmentation strategies. Among them, the data augmentation strategies include the Gaussian noise addition strategy and the mixed perturbation strategy of Gaussian noise and uniform noise, so as to more realistically reproduce the noise characteristics under different devices and scenarios.

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

[0078] Step 1.1: Select the datasets CLPCF and PU-Net as test samples;

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

[0080] Step 1.3. Design a heterogeneous noise dataset, adding Gaussian noise with different intensities to the dataset CLPCF. The specific combinations are the standard deviations of (0.6%, 0.8%), (0.6%, 1.1%), and (0.6%, 2.0%); adding two types of noise, Gaussian noise and uniform noise, to the dataset PU-Net. The standard deviations are set to 1%, 2%, and 3%.

[0081] Step 2. Use the noiseless point cloud and its noisy point cloud to form a point cloud pair, and adopt a contrastive learning network to extract robust features from the point cloud pair; improve the perception and discrimination ability of the contrastive learning network model for different noise patterns, and optimize it through contrastive loss, so that the contrastive learning network can better identify and separate noise information. The specific steps are as follows:

[0082] Step 2.1. Generate training samples, adding noise i to the noiseless 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 certain point in the point cloud block, μ represents the mean of the noise, controlling the central position of the noise, and σ represents the standard deviation of the noise, controlling the intensity or spread of the noise;

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

[0086] S∈P i , T∈P i

[0087] where the patch S and the patch T adopt different sampling radii;

[0088] Step 2.3. Based on the point cloud pair (S, T) in Step 2.2, construct a patch set. 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 point cloud blocks from different manifolds or different regions;

[0089] 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, and use the normalized temperature-scaled cross-entropy loss function for training. The normalized temperature-scaled cross-entropy loss function is:

[0090]

[0091] Among them, L S，T is the loss of the contrastive learning network, which measures the similarity difference between the positive point cloud pairs and the negative point cloud pairs. |X i | is the number of points in the noise-free point cloud patch X i . a j，s , a j，T are the feature representations of patch S and patch T. sim(a j，s , a j，T ) is the cosine similarity of the positive point cloud pair, representing the similarity between patch S and patch T in the feature space. ω is the temperature scaling parameter used to adjust 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.

[0092] Step 3: Use a multi-layer perceptron to train the intensity prediction network, and combine the local geometric features of the noisy points, i.e., normal information and neighborhood distribution, to predict the offset of each noisy point relative to its true position, i.e., the noise intensity. The specific steps are as follows:

[0093] Step 3.1: Input the coordinate information of each noisy point and its k nearest neighbor points, and pass through the trained robust feature extraction and contrastive learning module in Step 2 to obtain the local geometric features, i.e., normal information and neighborhood distribution;

[0094] Step 3.2: Feed the local geometric features obtained in Step 3.1 into a five-layer multi-layer perceptron (MLP) network for processing to predict the noise intensity Among them, is the noisy point;

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

[0096]

[0097] Among them, n j is the normal vector of the noisy point, and Δ j is the offset of the noisy 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, design an optimized loss function:

[0099]

[0100] Among them, M is the total number of all point cloud patches, |X i | is the number of points in the point cloud patch X i , and α is a hyperparameter used to adjust the sensitivity of the prediction error.

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

[0102] Step 4.1: 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 a five-layer 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 uneven distribution or over-aggregation of the point cloud 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 patch:

[0106]

[0107] Step 4.4: To ensure that the position of the point can be accurately restored and the uniformity of the point cloud can be maintained during the optimization process, obtain the total loss L p by weighted summing the position loss L DU and the distribution uniformity loss L total :

[0108] L total =(1 - β)·L p +β·L DU

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

[0110] Step 5: Use the evaluation metric Chamfer Distance CD to dynamically adjust the number of iterations; specifically, it includes the following steps:

[0111] Step 5.1: Estimate the noise intensity of the noise points using the noise intensity prediction module trained in Step 3

[0112] Step 5.2: Estimate the directional displacement of the noise points using the displacement prediction network trained in Step 4

[0113] Step 5.3: Dynamically adjust the offset of the noise points according to the noise intensity obtained in Step 5.1 and the directional displacement obtained in Step 5.2 The dynamic adjustment of the offset of the noise points is as follows:

[0114]

[0115] where, is the displacement that needs to be adjusted for the noise points during the denoising process;

[0116] Step 5.4: When all the noise points in the point cloud have completed each round of denoising, calculate the evaluation metric Chamfer Distance (CD):

[0117]

[0118] where, X i is the noise-free point cloud block, is the set of point clouds after denoising. If the metric Chamfer Distance CD shows a downward trend, continue to execute the denoising steps; if the metric Chamfer Distance CD rises, stop denoising in the previous round;

[0119] 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, resulting in deviation from the true value (GT). Excessive movement means that after moving to the ideal (ground truth) position, continuing to move causes the noise to rise instead of fall. The specific steps are as follows:

[0120] Step 6.1: To prevent excessive denoising during the denoising process, use a truncation function to limit the adjustment range of high-noise points:

[0121]

[0122] where, high-noise points are judged according to ρ(I), θ(I) is the adaptive intensity threshold, set to the median of the noise intensities 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;

[0123] Step 6.2: After each round of denoising is completed, according to the high-noise points judged by the truncation function, repeat Steps 6.1 and 6.2 for denoising processing.

[0124] Step 7: After optimization of multiple steps such as feature extraction, noise intensity prediction, directional adjustment, adaptive denoising and updating, the noise of the point cloud is significantly reduced, while retaining the integrity of the original geometric structure, and finally the point cloud after denoising is output and saved. Specifically, the following steps are included: When the indicator 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 advancedness of the method in this application, experiments were conducted and qualitative and quantitative comparisons were made with existing denoising methods.

[0126] For quantitative testing, for heterogeneous noise, this 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 the model performance. The final denoising results are shown in the quantitative comparison data of the denoising effect of the present invention and other algorithms on heterogeneous noise and homogeneous noise CLPCF datasets, see Table 1 for details, CD value multiplied by 10 -5 .

[0127] Table 1

[0128]

[0129] In addition, using the PU-Net test set containing 20 shapes, this application downsampled each point cloud to 10K and 50K resolution, and added a mixture of Gaussian noise and uniform noise with the same noise intensity to evaluate the performance of the model under different noise types. The final denoising effect is shown in Table 2, and the CD value is multiplied by 10 -4 Table 2 is a quantitative comparison of the denoising effects of the present invention and other algorithms on the heterogeneous noise PU-Net dataset.

[0130] Table 2

[0131]

[0132] In order to verify the denoising effect of the model under homogeneous noise, this application uses the CLPCF test set again 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, this application demonstrates its excellent denoising effect through visual evaluation on the CLPCF heterogeneous noise dataset, the PU-Net homogeneous noise dataset, and the Paris street dataset. Figure 3 and Figure 4Shows a visual comparison between the present application and existing competing algorithms under heterogeneous noise and homogeneous noise conditions. In contrast, the results of the present application are clearer, not only being able to better preserve details but also more thoroughly removing outliers, thus significantly improving the visual effect and realism. In real-world applications, Figure 5 Shows the denoising effect on the Paris-rue-Madame dataset. Compared with the SC and CL methods, the present application successfully retains more detailed features while removing noise, achieving more complete noise removal and a more appealing visual effect.

[0136] To test the robustness of the model, different degrees of noise were added to the present application, and the specific results are shown in Table 4 below.

[0137] Table 4

[0138]

[0139] The present invention demonstrates significant application value in dealing with complex heterogeneous noise point clouds and can be widely used in fields such as 3D reconstruction, autonomous driving, and remote sensing mapping.

[0140] The above are only the embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, various changes and modifications can be made to the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the scope of the claims of the present invention.

Claims

1. A method for denoising heterogeneous noise point cloud based on contrastive learning and intensity perception, characterized by: The heterogeneous noise point cloud denoising method specifically comprises the following steps: Step 1: extract samples from the noise-free point cloud dataset and simulate heterogeneous noise in real scans through multiple data enhancement strategies, including adding Gaussian noise strategy and mixed perturbation strategy of Gaussian noise and uniform noise; Step 2: Use the noise-free point cloud and its noisy point cloud to form a point cloud pair, and use a contrastive learning network to extract robust features from the point cloud pair; And through contrast loss optimization, the contrastive learning network can identify and separate noise information; Step 3: Using a multi-layer perceptron, combined with the local geometric features of the noise points, i.e., normal information and neighborhood distribution, the offset of each noise point relative to its true position, i.e., noise intensity, is predicted to obtain a noise intensity prediction module; Step 4: Combine the local geometric features of the noise point, i.e., the relative relationship between the noise point and the noise points in its neighborhood and the noise distribution, and calculate the displacement vector of each noise point so that the noise point moves closer to the true value (GT); Step 5: Use the evaluation indicator chamfer distance CD to dynamically adjust the number of iterations; Step 6: Use the truncation function to limit the adjustment range of high noise points to avoid excessive movement of abnormal points and prevent excessive adjustment of their positions during the denoising process, resulting in deviation from the true value (GT); Step 7: Output the point cloud data after denoising.

2. The method for denoising heterogeneous noise point cloud based on contrastive learning and intensity perception according to claim 1, characterized in that: The 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, add Gaussian noise with different standard deviations to the CLPCF dataset, and add noise with standard deviations of 1%, 2%, and 3% of the bounding sphere radius to the PU-Net dataset; Step 1.3: Design a heterogeneous noise dataset, add Gaussian noise of different intensities to the CLPCF dataset, and add two types of noise, Gaussian noise and uniform noise, to the PU-Net dataset.

3. The method for denoising heterogeneous noise point cloud based on contrastive learning and intensity perception according to claim 1, characterized in that: The step 2 specifically includes the following steps: Step 2.1: Generate training samples and generate noise-free point cloud blocks X i Add noise Φ i , and get the noise set P i : P i ={X i ,X i +Φ i } Among them, Φ i =ψ(x, μ, σ), where 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 of the noise; Step 2.2: From the noise set P i Randomly select patches S and patches T from , forming a point cloud pair (S, T): S∈P i ,T∈P i Wherein, the patch S and the patch T use different sampling radii; Step 2.3, constructing a patch set based on the point cloud pair (S, T) of step 2.2, wherein the patch set includes a positive point cloud pair and a negative point cloud pair, wherein 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; 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. Use the normalized temperature scaled cross entropy loss function for training. The normalized temperature scaled cross entropy loss function is: Among them, L S，T To compare the loss of the learning network, measure the similarity difference between the positive point cloud pair and the negative point cloud pair, |X i | is the noise-free point cloud block X i The number of midpoints, 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, which indicates the similarity between patch S and patch T in the feature space. ω is the temperature scaling parameter used to adjust the sensitivity of the similarity. is the weighted sum of negative point cloud pairs, indicating the similarity distribution of negative point cloud pairs.

4. The method for denoising heterogeneous noise point cloud based on contrastive learning and intensity perception according to claim 1, characterized in that: The step 3 specifically includes the following steps: Step 3.1, input the coordinate information of each noise point and its k nearest neighbor points, 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; Step 3.2: Send the local geometric features obtained in step 3.1 to a five-layer multi-layer perceptron (MLP) network for processing to predict the noise intensity. in, is the noise point; Step 3.3, calculate the actual noise intensity by and its corresponding clean point x j The position difference between the two is projected onto the noise point The actual noise intensity is calculated on the normal vector of Among them, n j is the normal vector of the noise point, Δ j Noise point Relative to its clean point x j The offset of Step 3.4: To minimize the difference between the predicted noise intensity and the true value, design an optimized loss function: Where M is the total number of all point cloud blocks, |X i | is point cloud block X i The number of points, α is a hyperparameter used to adjust the sensitivity of the prediction error.

5. The method for denoising heterogeneous noise point cloud based on contrastive learning and intensity perception according to claim 1, characterized in that: The step 4 specifically includes the following steps: Step 4.1, the local geometric features of the noise points extracted in step 3 are combined with the noise intensity and input into the displacement prediction network composed of five layers of MLP; Step 4.2: Introduce position loss L p To minimize the predicted point The corresponding clean point x j The Euclidean distance L p : Step 4.3: Introduce distribution uniformity loss L DU By maximizing the distance between the predicted point and the farthest point in the real point cloud block: Step 4.4: By adjusting the position loss L p and distribution uniformity loss L DU Perform weighted summation to get the total loss L total : L total =(1-β)·L p +β·L DU Among them, β is a hyperparameter that controls the position loss L p and distribution uniformity loss L DU The weight between the two.

6. The method for denoising heterogeneous noise point cloud based on contrastive learning and intensity perception according to claim 1, characterized in that: The step 5 specifically includes the following steps: Step 5.1: Use the noise intensity prediction module trained in step 3 to estimate the noise intensity of the noise point Step 5.2: Use the displacement prediction network trained in step 4 to estimate the directional displacement of the noise point Step 5.3: The noise intensity obtained according to step 5.1 and the directional displacement obtained in step 5.2 Dynamically adjust the offset of the noise point: in, Noise point The displacement that needs to be adjusted during the denoising process; Step 5.4: After all noise points in the point cloud have completed each round of denoising, the evaluation indicator chamfer distance (CD) is calculated: Among them, X i is a noise-free point cloud block, is the denoised point cloud set.

7. The method for denoising heterogeneous noise point cloud based on contrastive learning and intensity perception according to claim 1, characterized in that: Step 6 specifically includes the following steps: Step 6.1: Use the truncation function to limit the adjustment range of high noise points: Among them, high noise points are judged according to ρ(I), θ(I) is an adaptive intensity threshold, which 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, repeat steps 6.1 and 6.2 to perform denoising according to the high noise points determined by the truncation function.

8. The method for denoising heterogeneous noise point cloud based on contrastive learning and intensity perception according to claim 7, characterized in that: The step 7 specifically includes 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.

Images