Point cloud denoising method, device and storage medium based on feature analysis and scale selection

Abstract

The present invention provides a point cloud denoising method, device and storage medium based on multi-scale feature analysis and scale selection, and constructs a point cloud denoising network model based on multi-scale feature analysis and scale selection; the point cloud denoising network model uses a plurality of local neighborhoods of different scales selected from the original noise point cloud as input; performs local feature extraction and enhancement on point cloud data at each scale, connects feature information at different scales in series, performs weight information regression through an expert mechanism module, and selects the best inverse displacement through the obtained weight information; inputs a training set into the constructed point cloud denoising network model for training, and optimizes the denoising network model; inputs a test set into the denoising network model generated after training for denoising, and outputs denoised point cloud data. The present invention can better retain sharp detail features through multi-scale feature analysis and scale selection, and solves the limitation of ignoring local geometric features at a single scale.

Description

Technical Field

[0001] The present invention relates to the technical field of three-dimensional data processing, and in particular to a point cloud denoising method, device and storage medium based on feature analysis and scale selection. Background Art

[0002] With the rise of autonomous driving technology, 3D data acquisition technology has gradually developed. 3D data can usually be represented in different formats, such as depth images, point clouds, meshes, and volume meshes. Among them, point clouds are unordered collections of three-dimensional points sampled from the two-dimensional surface of an object or scene, making it possible to directly represent the three-dimensional information extracted from stereo vision cameras and depth maps generated by RGB-D.

[0003] However, due to the inherent limitations of scanning devices or matching ambiguities in image reconstruction, they are often contaminated by noise, which severely affects the performance of downstream tasks such as shape matching, surface reconstruction, object segmentation, etc. Therefore, point cloud denoising is crucial for related 3D vision applications.

[0004] At present, point cloud denoising methods can be roughly divided into traditional methods and deep learning-based methods. Traditional methods use certain surface assumptions to remove noise, such as sparsity or non-local similarity. However, they usually over-sharpen and smooth, or fail to preserve sharp features well.

[0005] Deep learning-based methods usually use local neighborhoods (patches) as input to encode geometric information and predict the displacement of each point. However, most of them are learned at a single scale, and the feature information of the neighborhood is not learned sufficiently, and sharp detail features cannot be retained. For example, the point cloud denoising network modeled by the encoder-decoder extracts and analyzes the features of the point cloud data through the encoder, aggregates and collects the extracted features through the decoder, and finally predicts the inverse displacement of each point in the noisy point cloud and applies the inverse displacement to each point. This method takes point cloud data of a single scale as input and focuses on the loss function that can retain sharp information.

[0006] The feature information that existing technologies focus on is limited to neighborhood feature information at a single scale, and does not fully consider the specific local geometric characteristics of the points to adjust the selection range of neighboring points. As a result, the network does not fully learn the detailed information of the model, resulting in a certain deviation in the displacement prediction results, resulting in over-smoothing, and distortion of some thin and fine geometric details. Summary of the invention

[0007] In view of this, the present invention provides a point cloud denoising method, device and storage medium based on feature analysis and scale selection, which can denoise 3D irregular point cloud data and generate high-quality point cloud data.

[0008] To this end, the present invention provides the following technical solutions:

[0009] The present invention provides a point cloud denoising method based on multi-scale feature analysis and scale selection, the method comprising:

[0010] Obtain point cloud data with Gaussian noise added to form a training set and a test set of the noise point cloud;

[0011] A point cloud denoising network model based on multi-scale feature analysis and scale selection is constructed; the point cloud denoising network model takes a plurality of local neighborhoods of different scales selected from the original noise point cloud as input, and performs correction operations on each local neighborhood in the alignment space; extracts and enhances local features of point cloud data at each scale, connects feature information at different scales in series, and performs weight information regression through an expert mechanism module, and selects the best inverse displacement based on the obtained weight information; and outputs denoised point cloud data;

[0012] Input the training set into the constructed point cloud denoising network model for training and optimize the denoising network model;

[0013] The test set is input into the denoising network model generated after training for denoising, and the denoised point cloud data is output.

[0014] Furthermore, the point cloud denoising network model includes: an input and alignment module, a feature extraction module and an expert mechanism module;

[0015] The input and alignment module includes a plurality of input channels, each of which includes an alignment submodule for inputting a plurality of local neighborhoods of different scales selected from the original noise point cloud, and performing correction operations on each neighborhood in the alignment space respectively;

[0016] The feature extraction module includes a plurality of feature extraction submodules for extracting features of different scales, each feature extraction submodule including a plurality of groups of feature extractors and point pooling layers;

[0017] The expert mechanism module includes a series of fully connected layers and scale management networks, which are used to weight and optimally select the inverse displacement of regression at different scales; the fully connected layer connects the feature information output of the feature extraction module at different scales in series after pooling operation, and uses the weight information as the input of the scale management network to calculate the weight information, and selects the best inverse displacement based on the obtained weight information; the scale management network obtains the weight of each branch single scale, and selects the best inverse displacement according to the size of the weight.

[0018] Furthermore, the number of input channels is 3.

[0019] Furthermore, the alignment submodule is calibrated using principal component analysis.

[0020] Furthermore, the feature extractor is Conv-BN-ReLU.

[0021] Furthermore, the input of the point pooling layer is the feature information and coordinate information of all points in the local neighborhood with an initial scale of s. First, the feature tensor f representing the global information of the current local neighborhood is obtained by the feature extractor of the previous layer. s Then, based on the K-NN algorithm, the points are selected and the points far from the center are discarded, so as to retain the most important shape of the local neighborhood, reduce the neighborhood scale to s1, and then convert the global information f of the initial scale into s The neighborhood information with the neighborhood scale s1 continues to be transmitted to the feature extractor of the next layer.

[0022] Furthermore, training the point cloud denoising network model includes:

[0023] A branch loss function is set under each single-scale branch to ensure that the inverse displacement obtained under each single-scale branch can make the denoised point cloud as close to the true value as possible;

[0024] During the training process, it is determined whether the total loss has reached the minimum value. If not, the network parameters are optimized by back propagation until the total loss function reaches the minimum value and the training ends.

[0025] The present invention also provides a point cloud denoising device based on multi-scale feature analysis and scale selection, the device comprising:

[0026] A data acquisition unit, used for acquiring point cloud data with Gaussian noise added to form a training set and a test set of the noise point cloud;

[0027] A model building unit is used to build a point cloud denoising network model based on multi-scale feature analysis and scale selection; the point cloud denoising network model takes a plurality of local neighborhoods of different scales selected from the original noise point cloud as input, performs correction operations on each local neighborhood in the alignment space; extracts and enhances local features of point cloud data at each scale, connects feature information at different scales in series, performs weight information regression through an expert mechanism module, and selects the best inverse displacement based on the obtained weight information; and outputs denoised point cloud data;

[0028] A training unit, used for inputting the training set obtained by the data acquisition unit into the point cloud denoising network model constructed by the model construction unit for training, so as to optimize the denoising network model;

[0029] The testing unit is used to input the test set obtained by the data acquisition unit into the denoising network model generated after training by the training unit for denoising, and output the denoised point cloud data.

[0030] The present invention also provides a computer-readable storage medium, which stores a computer instruction set. When the computer instruction set is executed by a processor, the point cloud denoising method based on multi-scale feature analysis and scale selection is implemented.

[0031] Advantages and positive effects of the present invention: Through multi-scale feature analysis and scale selection, sharp detail features can be better preserved, solving the limitation of ignoring local geometric features at a single scale. In addition, the above technology will also promote related 3D model processing and applications, such as model matching, surface reconstruction, object segmentation, and object recognition in large scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0033] Figure 1 is a flow chart of a point cloud denoising method based on feature analysis and scale selection in an embodiment of the present invention;

[0034] Figure 2 is a network structure diagram of a point cloud denoising method based on feature analysis and scale selection in an embodiment of the present invention;

[0035] Figure 3 Schematic diagram of a feature extraction module in a denoising network structure in an embodiment of the present invention;

[0036] Figure 4 Schematic diagram of the algorithm of the expert mechanism module in the denoising network structure in an embodiment of the present invention;

[0037] Figure 5 3 is a visualization effect diagram before and after denoising in an embodiment of the present invention. DETAILED DESCRIPTION

[0038] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0039] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0040] like Figure 1 As shown, a point cloud denoising method based on multi-scale feature analysis and scale selection in an embodiment of the present invention includes the following specific steps:

[0041] S1. Obtain point cloud data with Gaussian noise added to form a training set and a test set of the noise point cloud.

[0042] Among them, the training set: 22 clean point cloud models with precise coordinates and normals are obtained from the existing point cloud dataset PointFilter, including 11 CAD models and 11 non-CAD models, and each point cloud model randomly extracts 100k points from its original surface. Afterwards, 5 different levels of Gaussian noise are added to each model (the added standard deviation is 0.25%, 0.5%, 1%, 1.5% and 2.5% of the diagonal length of the clean model bounding box, respectively), and the corresponding noise model is synthesized. Therefore, the training dataset in the embodiment of the present invention consists of 110 noisy point cloud models and 22 clean point cloud models with precise coordinates and normals.

[0043] Test set: For quantitative evaluation, three benchmarks are used in the embodiments of the present invention: PU-Net test set (20 models), PointCleanNet test set (10 models), and PointFilter test set (20 models). PU-Net, PointCleanNet, and PointFilter are all existing point cloud datasets. Similarly, in the embodiment of the present invention, after the point cloud model is normalized, Gaussian noise is added at a ratio of 0.5%, 1%, and 1.5% to synthesize the corresponding noisy point cloud. Therefore, the test dataset in the embodiment of the present invention consists of 150 noisy point cloud models and 50 clean point cloud models with precise coordinates and normals.

[0044] S2. Construct a point cloud denoising network model based on multi-scale feature analysis and scale selection.

[0045] In the initial stage of the network, multiple local patches of different scales are selected as input for the original noise point cloud. Since the point cloud data itself is disordered, it needs to be corrected in the alignment space. Subsequently, the point cloud data at each scale will pass through the feature extraction module to learn the local information and analyze and enhance the corresponding local features. The feature information is then fused through Max-pooling, and at the same time, the additive noise vectors at different scales are regressed through the corresponding Loss function constraints. Since the feature information learned at different scales is biased, specifically, if a large scale is selected for a smooth surface, the feature will be closer to smoothness, and for sharp details, a small scale will retain more fine details. Therefore, in the embodiment of the present invention, the feature information at different scales is connected in series and weight information regression is performed through the expert mechanism module. The optimal inverse displacement is selected through the obtained weight information, so as to achieve a denoising effect that better maintains the detail features.

[0046] The specific network structure of the point cloud denoising network model is as follows Figure 2 As shown, it mainly includes an input and alignment module, a feature extraction module and an expert mechanism module. Among them, the input and alignment module includes multiple input channels, each input channel includes an alignment submodule, which is used to input multiple local patches of different scales selected from the original noise point cloud, and perform correction operations on each patch in the alignment space; preferably, the number of input channels is 3, and the alignment submodule adopts principal component analysis calibration (PCA Alignment). The feature extraction module includes a feature extractor (MLP) and a point pooling layer (CSA), and the MLP is specifically Conv-BN-ReLU, such as Figure 3 As shown, the feature extraction module in the embodiment of the present invention includes three feature extraction submodules, each of which includes multiple groups of MLPs and CSAs. Multiple MLPs are used to extract features of different scales, and multiple CSAs are used to enhance the features extracted by the previous layer of MLPs. The input of CSA is the feature information and coordinate information of all points in the local neighborhood with an initial scale of s. First, the feature tensor f representing the global information of the current local neighborhood is obtained through the MLP of the previous layer. s Then, based on the K-NN algorithm, the points are selected and the points far from the center are discarded, so as to retain the most important shape of the local neighborhood, reduce the neighborhood scale to s1, and then convert the global information f of the initial scale into s The neighborhood information with the neighborhood scale s1 is then transmitted to the next layer of MLP.

[0047] The expert mechanism is an architecture proposed in the field of deep learning, but it has not been applied to point cloud denoising tasks. The specific functions of the expert mechanism include:

[0048] (1) Neighborhoods of different scales contain different feature information. Small-scale receptive fields tend to retain local detail information, while large-scale receptive fields tend to retain global information. In order to fully consider the specific local geometric characteristics of the points to adjust the selection range of neighboring points so that the network can learn the detailed information of the model more fully, an expert mechanism architecture is designed in the embodiment of the present invention to assign different weights to each branch single-scale network so that the network focuses on more important features.

[0049] (2) There are also multi-scale point cloud denoising tasks, such as the extension of the 3D point cloud denoising network PointCleanNet, which performs denoising at multiple scales. However, the architecture it uses is more inclined to average multiple branches during training, and adopts an expert mechanism architecture to make each single branch more accurate.

[0050] The expert mechanism module designed in the embodiment of the present invention is as follows Figure 4 As shown. In the embodiment of the present invention, the expert mechanism module includes a series of fully connected layers and scale management networks, which are used to weight and optimally select the inverse displacement di of regression at different scales. Among them, the fully connected layer is specifically Linear-BN-ReLU, which connects the feature information output of the feature extraction module at different scales in series after pooling operation, as the scale management network (that is, Figure 2 The weight information is calculated based on the input of the Wi part in the graph, and the optimal inverse displacement di is selected based on the obtained weight information. Specifically, the corresponding inverse displacement di will be learned at each branch single scale, and the scale management network will obtain the weight Wi of each branch single scale, and select the optimal inverse displacement di according to the size of the weight Wi (select the inverse displacement di required at the branch single scale corresponding to the maximum value of Wi).

[0051] In the network architecture of the embodiment of the present invention, the network of n branches learning inverse displacement at a single scale is regarded as n independent "expert" networks. The "expert" network of each scale independently extracts, learns and analyzes features, fully explores the geometric feature dependencies at different scales, obtains more valuable feature information, and obtains the corresponding inverse displacement di. The scale management network is equivalent to a "manager". By empowering each branch, adjusting the network parameters of each branch, and optimally selecting the inverse displacement according to the size of the Wi value, a denoising effect that better maintains the detail features is achieved.

[0052] The above point cloud denoising network model can fully explore the multi-scale geometric feature dependencies through multi-scale feature analysis and scale selection, thereby obtaining more valuable feature information and better maintaining the denoising effect of detail features. The specific formula is as follows:

[0053] (ω1,ω2,...,ω n )=ρ(γ(f1,f2,...,fn )).

[0054] S3. Input the training set into the constructed point cloud denoising network model for training and optimize the denoising network model.

[0055] In the embodiment of the present invention, a branch loss function is set under each single-scale branch to ensure that the inverse displacement obtained under each single-scale branch can make the denoised point cloud as close to the true value as possible. The specific formula is as follows:

[0056]

[0057] in,

[0058] Make the denoised point cloud as close to the true surface as possible; is an exclusion term that allows the denoised points to be evenly distributed and avoids point aggregation: η is a trade-off parameter that controls the exclusion term during the denoising process, and η is set to 0.97 during the training phase.

[0059] p gt It's a real value point. is the denoised point, It's point p gt The true normal direction of is a Gaussian function. Nearby points are given greater weight.

[0060] The expert mechanism module sets the total loss function: Loss = ∑Loss k *ω k .

[0061] Among them, Loss k is the loss function for each single-scale branch, ω k is the weight value of each single-scale branch.

[0062] During the training process, it is determined whether the total loss has reached the minimum value. If not, the network parameters are optimized by back propagation until the total loss function reaches the minimum value and the training ends.

[0063] S4. Input the noisy point cloud test set into the denoising network model generated after training for denoising, and output the denoised point cloud data.

[0064] The point cloud denoising method in the above embodiment is an end-to-end network architecture, which directly takes each noise point and its original neighbor points as a set input, and regresses through network deep learning to obtain an inverse displacement vector, thereby moving the noise point to its original true position.

[0065] Since most of the current denoising methods perform feature learning at a single scale, the feature information of the neighborhood is not sufficiently learned, and the sharp detail features cannot be retained, which weakens their performance in practical applications. Therefore, the point cloud denoising network model in the embodiment of the present invention uses a multi-scale neighborhood as input, fully analyzes and learns the feature information at different scales, and enhances the local features to solve the problem of insufficient neighborhood learning and ensure the effectiveness of the obtained point cloud features. On this basis, due to the deviation of the feature information learned at different scales, the small-scale receptive field tends to retain local detail information, while the large scale tends to retain global information. Specifically, if a large scale is selected for a smooth surface, the feature will be closer to smoothness. For sharp detail parts, selecting a small scale will retain more fine details. Therefore, the embodiment of the present invention also adopts an expert mechanism to select weighted constraints and inverse displacement selection for features of different scales, fully consider the specific local geometric characteristics of the point cloud, and solve the problem of neighborhood selection, so as to achieve a better denoising effect of retaining sharp detail features.

[0066] In addition, although there are also solutions for feature extraction based on multi-scale input and multiple scales in the prior art, usually multiple scale feature fusion is finally performed, and the architecture adopted tends to sum multiple branches during training. In the embodiment of the present invention, an expert mechanism architecture is adopted to make each single branch more specialized, and finally take the best instead of summing, so that denoising is more accurate and sharp details can be better maintained. In the embodiment of the present invention, CSA is also added to the feature extraction module to reduce the number of points, which can reduce the amount of calculation and time consumption.

[0067] Corresponding to the point cloud denoising method based on multi-scale feature analysis and scale selection in the present invention, the present invention also provides a point cloud denoising device based on multi-scale feature analysis and scale selection, comprising:

[0068] A data acquisition unit, used for acquiring point cloud data with Gaussian noise added to form a training set and a test set of the noise point cloud;

[0069] A model building unit is used to build a point cloud denoising network model based on multi-scale feature analysis and scale selection; the point cloud denoising network model takes a plurality of local patches of different scales selected from the original noise point cloud as input, performs correction operations on each local patch in the alignment space; extracts and enhances local features of point cloud data at each scale, connects feature information at different scales in series, performs weight information regression through an expert mechanism module, and selects the best inverse displacement based on the obtained weight information; and outputs denoised point cloud data;

[0070] A training unit, used for inputting the training set obtained by the data acquisition unit into the point cloud denoising network model constructed by the model construction unit for training, so as to optimize the denoising network model;

[0071] The testing unit is used to input the test set obtained by the data acquisition unit into the denoising network model generated after training by the training unit for denoising, and output the denoised point cloud data.

[0072] For a point cloud denoising device based on multi-scale feature analysis and scale selection in an embodiment of the present invention, since it corresponds to a point cloud denoising method based on multi-scale feature analysis and scale selection in the above embodiment, the description is relatively simple. For relevant similarities, please refer to the description of a point cloud denoising method based on multi-scale feature analysis and scale selection in the above embodiment, which will not be described in detail here.

[0073] An embodiment of the present invention further discloses a computer-readable storage medium, which stores a computer instruction set. When the computer instruction set is executed by a processor, a point cloud denoising method based on multi-scale feature analysis and scale selection as provided in any of the above embodiments is implemented.

[0074] In the several embodiments provided by the present invention, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of the units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0075] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0076] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0077] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk and other media that can store program codes.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A point cloud denoising method based on multi-scale feature analysis and scale selection, characterized in that: The method comprises: Obtain point cloud data with Gaussian noise added to form a training set and a test set of the noise point cloud; A point cloud denoising network model based on multi-scale feature analysis and scale selection is constructed; the point cloud denoising network model takes a plurality of local neighborhoods of different scales selected from the original noise point cloud as input, and performs correction operations on each local neighborhood in the alignment space; extracts and enhances local features of point cloud data at each scale, connects feature information at different scales in series, and performs weight information regression through an expert mechanism module, and selects the best inverse displacement based on the obtained weight information; and outputs denoised point cloud data; Input the training set into the constructed point cloud denoising network model for training and optimize the denoising network model; The test set is input into the denoising network model generated after training for denoising, and the denoised point cloud data is output.

2. The point cloud denoising method based on multi-scale feature analysis and scale selection according to claim 1, characterized in that: The point cloud denoising network model includes: input and alignment module, feature extraction module and expert mechanism module; The input and alignment module includes a plurality of input channels, each of which includes an alignment submodule for inputting a plurality of local neighborhoods of different scales selected from the original noise point cloud, and performing correction operations on each neighborhood in the alignment space respectively; The feature extraction module includes a plurality of feature extraction submodules for extracting features of different scales, each feature extraction submodule including a plurality of groups of feature extractors and point pooling layers; The expert mechanism module includes a series of fully connected layers and scale management networks, which are used to weight and optimally select the inverse displacement of regression at different scales; the fully connected layer connects the feature information output of the feature extraction module at different scales in series after pooling operation, and uses the weight information as the input of the scale management network to calculate the weight information, and selects the best inverse displacement based on the obtained weight information; the scale management network obtains the weight of each branch single scale, and selects the best inverse displacement according to the size of the weight.

3. The point cloud denoising method based on multi-scale feature analysis and scale selection according to claim 2, characterized in that: The number of input channels is 3.

4. The point cloud denoising method based on multi-scale feature analysis and scale selection according to claim 2, characterized in that: The alignment submodule is calibrated using principal component analysis.

5. The point cloud denoising method based on multi-scale feature analysis and scale selection according to claim 2, characterized in that: The feature extractor is Conv-BN-ReLU.

6. The point cloud denoising method based on multi-scale feature analysis and scale selection according to claim 2, characterized in that: The input of the point pooling layer is the feature information and coordinate information of all points in the local neighborhood with an initial scale of s. First, the feature tensor f representing the global information of the current local neighborhood is obtained by the feature extractor of the previous layer. s Then, based on the K-NN algorithm, the points are selected and the points far from the center are discarded, so as to retain the most important shape of the local neighborhood, reduce the neighborhood scale to s1, and then convert the global information f of the initial scale into s The neighborhood information with the neighborhood scale s1 continues to be transmitted to the feature extractor of the next layer.

7. The point cloud denoising method based on multi-scale feature analysis and scale selection according to claim 1, characterized in that: Training the point cloud denoising network model includes: A branch loss function is set under each single-scale branch to ensure that the inverse displacement obtained under each single-scale branch can make the denoised point cloud as close to the true value as possible; During the training process, it is determined whether the total loss has reached the minimum value. If not, the network parameters are optimized by back propagation until the total loss function reaches the minimum value and the training ends.

8. A point cloud denoising device based on multi-scale feature analysis and scale selection, characterized in that: The device comprises: A data acquisition unit, used for acquiring point cloud data with Gaussian noise added to form a training set and a test set of the noise point cloud; A model building unit is used to build a point cloud denoising network model based on multi-scale feature analysis and scale selection; the point cloud denoising network model takes a plurality of local neighborhoods of different scales selected from the original noise point cloud as input, performs correction operations on each local neighborhood in the alignment space; extracts and enhances local features of point cloud data at each scale, connects feature information at different scales in series, performs weight information regression through an expert mechanism module, and selects the best inverse displacement based on the obtained weight information; and outputs denoised point cloud data; A training unit, used for inputting the training set obtained by the data acquisition unit into the point cloud denoising network model constructed by the model construction unit for training, so as to optimize the denoising network model; The testing unit is used to input the test set obtained by the data acquisition unit into the denoising network model generated after training by the training unit for denoising, and output the denoised point cloud data.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer instruction set, and when the computer instruction set is executed by a processor, the point cloud denoising method based on multi-scale feature analysis and scale selection as described in any one of claims 1 to 7 is implemented.

Images