Multi-scale fusion and linear regularization denoising method and device based on iterative filtering

Abstract

The invention provides a multi-scale fusion and linear regularization denoising method and device based on iterative filtering, and relates to the technical field of point cloud denoising, and the method comprises the steps: obtaining point cloud data; preprocessing the point cloud data, and obtaining point cloud preprocessing data based on normalization, noise synthesis and down-sampling operation; performing feature coding on the point cloud preprocessing data, and generating coding features based on multi-scale feature fusion and linear regularization; and constructing an iterative filter based on the coding features, and denoising the point cloud data through the iterative filter to obtain denoised point cloud data. According to the method, the problems that the feature extraction capability is insufficient and the influence of different noise types on the point cloud data cannot be effectively handled due to the fact that the local features and the global features of the point cloud are not fully considered in the existing denoising method are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of point cloud denoising. Specifically, it relates to a multi-scale fusion and linear regularization denoising method and device based on iterative filtering. Background Art

[0002] With the rapid development of three-dimensional data acquisition devices such as depth cameras and laser scanners, the digitalization of three-dimensional scenes has become a reality. As a common three-dimensional data representation format, three-dimensional point clouds can retain the original geometric information of the acquired scene or object in three-dimensional space. However, due to the presence of noise during the acquisition process, which causes point coordinate offsets, it limits the accurate application of three-dimensional point clouds in fields such as computer graphics and computer vision. Therefore, three-dimensional point cloud denoising has become a research hotspot.

[0003] Current deep learning methods in point cloud denoising mainly focus on the extraction of local features, and the local feature extraction methods are relatively single, while ignoring the extraction of global features. This results in the model being unable to comprehensively capture the overall structural information of the point cloud when dealing with complex scenes, thus affecting the denoising effect and the accuracy of reconstruction. And in order to improve the denoising accuracy of point clouds, existing denoising networks often introduce networks with quadratic complexity such as Transformers or attention mechanisms, resulting in too long network training and prediction times. Summary of the Invention

[0004] The purpose of the present invention is to provide a multi-scale fusion and linear regularization denoising method and device based on iterative filtering to improve the above problems. To achieve the above purpose, the technical solutions adopted by the present invention are as follows:

[0005] In a first aspect, the present application provides a multi-scale fusion and linear regularization denoising method based on iterative filtering, including:

[0006] Obtain point cloud data;

[0007] Preprocess the point cloud data to obtain point cloud preprocessing data based on normalization, noise synthesis, and downsampling operations;

[0008] Perform feature encoding on the point cloud preprocessing data to generate encoded features based on multi-scale feature fusion and linear regularization;

[0009] Construct an iterative filter based on the encoded features, and use the iterative filter to denoise the point cloud data to obtain denoised point cloud data.

[0010] In a second aspect, the present application also provides a multi-scale fusion and linear regularization denoising device based on iterative filtering, including:

[0011] An acquisition unit for acquiring point cloud data;

[0012] A preprocessing unit for preprocessing the point cloud data to obtain preprocessed point cloud data based on normalization, noise synthesis, and downsampling operations;

[0013] An encoding unit for feature encoding the preprocessed point cloud data to generate encoded features based on multi-scale feature fusion and linear regularization;

[0014] A denoising unit for constructing an iterative filter based on the encoded features and denoising the point cloud data through the iterative filter to obtain denoised point cloud data.

[0015] The beneficial effects of the present invention are as follows: The multi-scale dynamic graph convolution provided by the present invention has the ability to extract local geometry and local features, and at the same time combines linear regularization to output a global perception graph, thereby improving the denoising accuracy. Compared with the ability of the attention mechanism to extract global features, which requires quadratic computational complexity, the output of the global perception graph only requires linear complexity, thereby improving the denoising performance. And an iterative filter is introduced for step-by-step denoising, enabling the point cloud to gradually approach the true geometric surface, significantly improving the denoising effect, removing noise faster through linear regularization, and at the same time maintaining the detailed structure of the point cloud, ensuring the accuracy and robustness of the denoising effect.

[0016] Other features and advantages of the present invention will be described in the subsequent description, and, in part, will be obvious from the description, or can be understood by implementing the embodiments of the present invention. The objectives and other advantages of the present invention can be realized and obtained by the structures specifically pointed out in the written description, claims, and drawings. Description of the Drawings

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for use in the embodiments. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts.

[0018] Figure 1 Schematic diagram of the multi-scale fusion and linear regularization denoising method based on iterative filtering described in the embodiments of the present invention;

[0019] Figure 2 Schematic diagram of the structure of the multi-scale dynamic graph convolution sub-module in the embodiments of the present invention;

[0020] Figure 3 Schematic diagram of the structure of the global perception graph generation sub-module in the embodiments of the present invention;

[0021] Figure 4 Schematic diagram of the iterative process of the iterative filter in the embodiment of the present invention;

[0022] Figure 5 Schematic diagram of the structure of the iterative filter in the embodiment of the present invention. Specific implementation manners

[0023] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Components of the embodiments of the present invention generally described and illustrated in the figures herein may be arranged and designed in a variety of different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but is merely representative of selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0024] It should be noted that like reference numerals and letters denote like items in the following figures, and thus, once an item is defined in one figure, it need not be further defined and explained in subsequent figures. At the same time, in the description of the present invention, the terms "first", "second", etc. are only used for descriptive distinction and cannot be construed as indicating or implying relative importance.

[0025] Embodiment 1:

[0026] This embodiment provides a multi-scale fusion and linear regularization denoising method based on iterative filtering.

[0027] See Figure 1 , which shows that this method includes step S100, step S200, step S300, and step S400.

[0028] Step S100: Obtain point cloud data;

[0029] Step S200: Preprocess the point cloud data to obtain point cloud preprocessing data based on normalization, noise synthesis, and downsampling operations;

[0030] The step S200 includes:

[0031] Step S201: Construct a point cloud preprocessing module based on the point cloud data, where the point cloud preprocessing module includes a normalization sub-module, a noise synthesis sub-module, and a downsampling sub-module connected in sequence;

[0032] In this embodiment, the point cloud preprocessing module is responsible for normalizing, noise synthesizing, and downsampling the point cloud data to improve the quality of the point cloud data.

[0033] Step S202: Normalize the point cloud data through the normalization sub-module to unify the scale and central position of the point cloud data, obtaining normalized point cloud data;

[0034] In this embodiment, normalization is an essential step in point cloud data processing. By unifying the scale and central position of the point cloud data, it enables point clouds from different sources and sizes to have consistent performance in the model. This process not only improves the stability and efficiency of model training, preventing gradient explosion or vanishing problems, but also enhances the model's robustness to noise and outliers. Therefore, the normalized point cloud data normalizes the norm of each point to within the unit sphere, facilitating subsequent batch processing and geometric operations.

[0035] In this embodiment, the specific steps for performing normalization processing are as follows:

[0036] Step A100: Calculate the center point of the point cloud data;

[0037] Step A200: Subtract the center point from each point in the point cloud data to obtain the first point cloud data;

[0038] Step A300: Calculate the Euclidean norm of each point in the point cloud data, compare the magnitudes of all Euclidean norms, and obtain the maximum norm;

[0039] Step A400: Divide the first point cloud by the maximum norm to obtain the normalized point cloud data.

[0040] Step S203: Synthesize noise for the normalized point cloud data through the noise synthesis sub-module to obtain synthesized point cloud data, and the noise synthesis sub-module uses Gaussian noise for noise synthesis;

[0041] In this embodiment, random noise sampled from a Gaussian distribution is added to each coordinate in the normalized point cloud data to simulate sensor measurement errors.

[0042] Step S204: Perform downsampling on the synthesized point cloud data through the downsampling sub-module to obtain point cloud preprocessing data.

[0043] In this embodiment, the farthest point sampling algorithm is used for downsampling.

[0044] Step S300: Perform feature encoding on the point cloud preprocessing data, and generate encoded features based on multi-scale feature fusion and linear regularization;

[0045] The step S300 includes:

[0046] Step S301: Construct an encoding module based on the preprocessed point cloud data. The encoding module includes a multi-scale dynamic graph convolutional sub-module and a global perception graph generation sub-module;

[0047] Step S302: Perform multi-scale feature fusion on the preprocessed point cloud data through the multi-scale dynamic graph convolutional sub-module to obtain multi-scale fusion features;

[0048] In this embodiment, the multi-scale dynamic graph convolutional sub-module introduces a multi-scale feature fusion strategy and proposes a multi-scale dynamic graph convolutional method, enabling the full extraction and fusion of the local geometric information of each point. By combining neighborhood information in the point cloud graph, it better captures the multi-scale geometric structure and semantic features.

[0049] The specific structure of the multi-scale dynamic graph convolutional sub-module is as Figure 2 shown, where MLP represents a shared multi-layer perceptron, MaxPooling represents a max-pooling operation, N represents the total number of points, and C represents the feature dimension.

[0050] The step S302 includes:

[0051] Step B100: Calculate the neighbor set of each point in the preprocessed point cloud data through the nearest neighbor algorithm and the 3D Euclidean distance algorithm;

[0052] Step B200: Extract features for each point and the neighbor set of each point based on the local geometric encoder to obtain a local geometric feature map;

[0053] In this embodiment, the local geometric feature map is composed of multiple local geometric features. The calculation formula for the local geometric feature is:

[0054]

[0055] In the formula, represents the local geometric feature corresponding to the i-th point in the preprocessed point cloud data, p i represents the i-th point in the preprocessed point cloud data, p ik represents the k-th neighborhood point of p i , K represents the number of neighborhood points, represents a point cloud set of size k×9.

[0056] Step B300: Aggregate the features of the local geometric feature map through a shared multi-layer perceptron and a max-pooling operation in sequence to obtain a local geometric graph;

[0057] In this embodiment, the local geometric feature map is encoded, and a max-pooling operation is applied to k neighborhoods to aggregate the local geometric feature map. The formula for feature aggregation is:

[0058]

[0059] In the formula, represents the local geometry map, max k (·) represents the max pooling operation of taking k neighborhood points, and MLP(·) represents the shared multi-layer perceptron. represents the local geometry feature map. represents a point cloud set of size, N represents the total number of points, and C represents the feature dimension.

[0060] Step B400: Perform feature encoding on the preprocessed point cloud data through a local feature encoder to obtain a local feature encoding map.

[0061] In this embodiment, a feature map in the C-dimensional space is extracted through a local feature encoder to obtain a local feature encoding map composed of multiple local feature encodings. Among them, the local feature encoding represents a point cloud set of size k×3C, C represents the feature dimension, and k represents the number of neighborhood points.

[0062] Step B500: Aggregate the features of the local feature encoding map through a shared multi-layer perceptron and a max pooling operation in sequence to obtain a local feature map.

[0063] In this embodiment, the calculation formula of the local feature map is:

[0064]

[0065] In the formula, represents the local feature map, max k (·) represents the max pooling operation of taking k neighborhood points, and MLP(·) represents the shared multi-layer perceptron. represents the local feature encoding map. represents a point cloud set of size, N represents the total number of points, and C represents the feature dimension.

[0066] Step B600: Expand the receptive field of the preprocessed point cloud data through an edge convolution block and calculate the edge based on a non-linear function.

[0067] In this embodiment, the receptive field of the local map is further expanded, and the edge convolution is introduced to define the edge. Among them, the definition formula of the edge is:

[0068] e ik = ReLU(θ m ·(p ik - p i ) + φ m·p i )

[0069] where e ik represents the edge between p i and p ik ; p i represents the i-th point in the preprocessed point cloud data; p ik represents the k-th neighboring point of p i ; θ m and φ m both represent learnable non-linear functions, and ReLU(·) represents the activation function.

[0070] Step B700: Aggregate the features of the edges through a shared multi-layer perceptron and a max-pooling operation in sequence to obtain a local graph;

[0071] In this embodiment, after defining the edges, the point cloud can be represented as a graph. Among them, the expression of the local graph is:

[0072]

[0073] where represents the local graph, max(·) represents the max-pooling operation, MLP(·) represents the shared multi-layer perceptron, e ij represents the set of edges with respect to p i ; p i represents the i-th point in the preprocessed point cloud data, represents a point cloud set of size, N represents the total number of points, and C represents the feature dimension.

[0074] Step B800: Concatenate the features of the local geometric graph, the local feature graph, and the local graph to obtain a multi-scale fusion feature.

[0075] In this embodiment, the formula for concatenating the features to obtain the multi-scale fusion feature is;

[0076]

[0077] where F L represents the multi-scale fusion feature, concat(·) represents feature concatenation, represents the local geometric graph, represents the local feature graph, represents the local graph, represents a point cloud set of size N×C, N represents the total number of points, and C represents the feature dimension.

[0078] Step S303: Perform global perception on the multi-scale fusion feature through the global perception graph generation sub-module to obtain an encoded feature.

[0079] In this embodiment, the global perception graph generation sub-module introduces global multi-scale linear features to refine the point cloud feature representation, improving the accuracy of point cloud analysis. In traditional global bilinear regularization, the global channel descriptor and the global point descriptor do not describe the features of the graph structure. Therefore, the global perception graph generation sub-module enhances the representation ability of topological information by improving the input method of global bilinear regularization and introducing local graph structures into the input.

[0080] The structure of the global perception graph generation sub-module is as Figure 3 shown, where Linear represents a linear function, ReLU represents an activation function, AvgPooling represents an average pooling operation, MLP represents a shared multi-layer perceptron, Sub represents a subtraction operation, represents taking the outer product, and ⊕ represents an addition operation.

[0081] The step S303 includes:

[0082] Step C100: Reducing the dimension of the multi-scale fusion feature through a first weight matrix to obtain a first point cloud feature;

[0083] In this embodiment, the first weight matrix W c is used to reduce the dimension of the multi-scale fusion feature, where represents a point cloud set of size, C represents the feature dimension, and r represents the reduction factor.

[0084] Step C200: Sequentially calculating the first point cloud feature through an activation function and performing an average pooling operation to obtain a global channel descriptor;

[0085] In this embodiment, non-linearity is provided through the activation function, and at the same time, an average pooling operation is performed on the elements of the spatial axis to learn the local graph structure and obtain compressed spatial information, that is, the global channel descriptor. The calculation formula of the global channel descriptor is:

[0086] g c = avg N (ReLU(F L W c ))

[0087] In the formula, g c represents the global channel descriptor, avg N (·) represents the average pooling operation with respect to N, N represents the total number of points, ReLU(·) represents the activation function, F L represents the multi-scale fusion feature, and W cdenotes the first weight matrix, denotes a point cloud set of size, C denotes the feature dimension, and r denotes the reduction factor.

[0088] Step C300: Reduce the dimension of the multi-scale fusion feature through the second weight matrix to obtain the second point cloud feature;

[0089] In this embodiment, through the second weight matrix W p obtain the second point cloud feature, where, denotes a point cloud set of size, C denotes the feature dimension, and r denotes the reduction factor.

[0090] Step B400: Pass the second point cloud feature through the activation function for calculation and average pooling operation in sequence to obtain the global point descriptor encoding;

[0091] In this embodiment, the calculation formula of the global point descriptor encoding is:

[0092]

[0093] In the formula, g p denotes the global point descriptor encoding, denotes the average pooling operation with respect to , C denotes the feature dimension, r denotes the reduction factor, F L denotes the multi-scale fusion feature, W p denotes the second weight matrix, denotes a point cloud set of size N, and N denotes the total number of points.

[0094] Step C500: After calculating the outer product of the global channel descriptor and the global point descriptor encoding, take the square root of the outer product to generate the global bilinear response;

[0095] In this embodiment, the calculation formula of the global bilinear response is:

[0096]

[0097] In the formula, G denotes the global bilinear response, g c denotes the global channel descriptor, g p denotes the global point descriptor encoding, denotes taking the outer product, denotes a point cloud set of size, N denotes the total number of points, C denotes the feature dimension, r denotes the reduction factor, η ab denotes the element in the a-th row and b-th column of G, λ a denotes the mean response of the global point descriptor encoding, μb Represents the mean response of the global channel descriptor.

[0098] Step C600: Concatenate the first point cloud feature, the second point cloud feature, and the global bilinear response, and then input them into a shared multi-layer perceptron to restore the channel dimension and generate a global perception map.

[0099] In this embodiment, through a shared multi-layer perceptron and two shortcut connections, the channel dimension is restored to generate a complete global perception map. The calculation formula of the global perception map is:

[0100] F G = MLP(G + F L W c + F L W p )

[0101] In the formula, F G represents the global perception map, MLP(·) represents the shared multi-layer perceptron, G represents the global bilinear response, F L represents the multi-scale fusion feature, W c represents the first weight matrix, and W p represents the second weight matrix.

[0102] Step C700: Subtract the global perception map from the multi-scale fusion feature to obtain a third point cloud feature, and increase the non-linearity of the third point cloud feature through an activation function to obtain an encoded feature.

[0103] In this embodiment, the Mish function is used as the activation function to increase the non-linearity and generate an encoded feature. The calculation formula of the encoded feature is:

[0104] F out = σ(F L - F G )

[0105] In the formula, F out represents the encoded feature, σ(·) represents the Mish function, F L represents the multi-scale fusion feature, and F G represents the global perception map.

[0106] Step S400: Construct an iterative filter based on the encoded feature, and use the iterative filter to denoise the point cloud data to obtain denoised point cloud data.

[0107] In this embodiment, the iterative filter is used to denoise the noisy point cloud to obtain denoised point cloud data, that is, the iterative process of the clean point cloud is as Figure 4 shown.

[0108] The step S400 includes:

[0109] Step S401: Construct a linear decoding module based on the encoded features. The input of the linear decoding module is the encoded features, and the output of the linear decoding module is the decoded features.

[0110] Step S402: Connect the point cloud preprocessing module, the encoding module, and the linear decoding module in sequence to form an iterative filtering layer.

[0111] Step S403: Connect a preset number of iterative filtering layers in sequence to form an iterative filter, and use the decoded features output by the last iterative filtering layer as the output features of the iterative filter.

[0112] Step S404: Denoise the point cloud data through the iterative filter to obtain denoised point cloud data.

[0113] In this embodiment, as Figure 5 shown, it is a schematic structural diagram of the iterative filter, which is composed of M iterative filtering layers. The denoising process is iterated multiple times through the iterative filter to optimize the denoising accuracy and make the point cloud approach the real geometric surface.

[0114] The iterative formula of the iterative filter is:

[0115]

[0116] In the formula, represents the loss function, represents the displacement after the τ-th iteration, Y (τ) represents the target position after the τ-th iteration, represents the square of the L2 norm, represents the displacement of the nearest neighbor point in Y (τ) , and NN(·) represents the iterative filtering layer.

[0117] In summary, through multi-scale feature fusion, the present invention effectively captures the geometric and feature context information of the point cloud, realizes the accurate extraction of local features. At the same time, linear regularization is introduced to improve the extraction ability of global features, and local features and global features are combined to improve the accuracy of overall point cloud denoising and the generalization of the denoising model. Therefore, the present invention better preserves the detail features of the point cloud and enhances the adaptability in complex scenarios.

[0118] At the same time, the present invention introduces iterative filtering for step-by-step denoising, making the point cloud gradually approach the real geometric surface, significantly improving the denoising effect. Using the linear regularization method, the linearization of the denoising performance is realized, and the denoising speed is faster than that using the attention mechanism. Therefore, the present invention can also remove noise at a faster speed, while maintaining the detail structure of the point cloud, ensuring the accuracy and robustness of the denoising effect.

[0119] Example 2:

[0120] This embodiment provides a multi-scale fusion and linear regularization denoising device based on iterative filtering. The device includes:

[0121] An acquisition unit for acquiring point cloud data;

[0122] A preprocessing unit for preprocessing the point cloud data to obtain preprocessed point cloud data based on normalization, noise synthesis, and downsampling operations;

[0123] An encoding unit for feature encoding the preprocessed point cloud data to generate encoded features based on multi-scale feature fusion and linear regularization;

[0124] A denoising unit for constructing an iterative filter based on the encoded features and denoising the point cloud data through the iterative filter to obtain denoised point cloud data.

[0125] The preprocessing unit includes:

[0126] A first construction subunit for constructing a point cloud preprocessing module based on the point cloud data. The point cloud preprocessing module includes a normalization submodule, a noise synthesis submodule, and a downsampling submodule connected in sequence;

[0127] A normalization subunit for normalizing the point cloud data through the normalization submodule to unify the scale and central position of the point cloud data and obtain normalized point cloud data;

[0128] A noise synthesis subunit for synthesizing noise for the normalized point cloud data through the noise synthesis submodule to obtain synthesized point cloud data. The noise synthesis submodule uses Gaussian noise for noise synthesis;

[0129] A downsampling subunit for performing a downsampling operation on the synthesized point cloud data through the downsampling submodule to obtain preprocessed point cloud data.

[0130] The encoding unit includes:

[0131] A second construction subunit for constructing an encoding module based on the preprocessed point cloud data. The encoding module includes a multi-scale dynamic graph convolution submodule and a global perception graph generation submodule;

[0132] A feature fusion subunit for performing multi-scale feature fusion on the preprocessed point cloud data through the multi-scale dynamic graph convolution submodule to obtain multi-scale fusion features;

[0133] The global perception subunit is used to globally perceive the multi-scale fusion features through the global perception graph generation sub-module to obtain encoded features.

[0134] The denoising unit includes:

[0135] The third construction subunit is used to construct a linear decoding module based on the encoded features. The input of the linear decoding module is the encoded features, and the output of the linear decoding module is the decoded features;

[0136] The fourth construction subunit is used to connect the point cloud preprocessing module, the encoding module, and the linear decoding module in sequence to form an iterative filtering layer;

[0137] The fifth construction subunit is used to connect a preset number of iterative filtering layers in sequence to form an iterative filter, and use the decoded features output by the last iterative filtering layer as the output features of the iterative filter;

[0138] The denoising subunit is used to denoise the point cloud data through the iterative filter to obtain denoised point cloud data.

[0139] It should be noted that regarding the device in the above embodiments, the specific manners in which each module performs operations have been described in detail in the embodiments related to the method, and will not be elaborated here.

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

[0141] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present invention, and all should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A multi-scale fusion and linear regularization denoising method based on iterative filtering, characterized in that Including: Obtain point cloud data; Preprocess the point cloud data, and obtain preprocessed point cloud data based on normalization, noise synthesis, and downsampling operations; Perform feature encoding on the preprocessed point cloud data, and generate encoded features based on multi-scale feature fusion and linear regularization; Construct an iterative filter based on the encoded features, and denoise the point cloud data through the iterative filter to obtain denoised point cloud data.

2. The multi-scale fusion and linear regularization denoising method based on iterative filtering according to claim 1, wherein , Preprocess the point cloud data, and obtain preprocessed point cloud data based on normalization, noise synthesis, and downsampling operations, including: Construct a point cloud preprocessing module based on the point cloud data, where the point cloud preprocessing module includes a normalization sub-module, a noise synthesis sub-module, and a downsampling sub-module connected in sequence; Perform normalization processing on the point cloud data through the normalization sub-module to unify the scale and central position of the point cloud data, and obtain normalized point cloud data; Perform noise synthesis on the normalized point cloud data through the noise synthesis sub-module to obtain synthesized point cloud data, and the noise synthesis sub-module uses Gaussian noise for noise synthesis; Perform downsampling operations on the synthesized point cloud data through the downsampling sub-module to obtain preprocessed point cloud data.

3. The multi-scale fusion and linear regularization denoising method based on iterative filtering according to claim 2, characterized in that , Perform feature encoding on the preprocessed point cloud data, and generate encoded features based on multi-scale feature fusion and linear regularization, including: Construct an encoding module based on the preprocessed point cloud data, where the encoding module includes a multi-scale dynamic graph convolution sub-module and a global perception graph generation sub-module; Perform multi-scale feature fusion on the preprocessed point cloud data through the multi-scale dynamic graph convolution sub-module to obtain multi-scale fusion features; Perform global perception on the multi-scale fusion features through the global perception graph generation sub-module to obtain encoded features.

4. The multi-scale fusion and linear regularization denoising method based on iterative filtering according to claim 3, wherein , Perform multi-scale feature fusion on the preprocessed point cloud data through the multi-scale dynamic graph convolution sub-module to obtain multi-scale fusion features, including: Calculate the neighbor set of each point in the preprocessed point cloud data through the nearest neighbor algorithm and the 3D Euclidean distance algorithm; Extract features of each point and the neighbor set of each point based on the local geometry encoder to obtain a local geometry feature map; Aggregate the features of the local geometry feature map through a shared multi-layer perceptron and a max pooling operation in sequence to obtain a local geometry map; Perform feature encoding on the preprocessed point cloud data through the local feature encoder to obtain a local feature encoding map; Aggregate the features of the local feature encoding map through a shared multi-layer perceptron and a max pooling operation in sequence to obtain a local feature map; Enlarge the receptive field of the preprocessed point cloud data through an edge convolution block, and calculate the edge based on a non-linear function; Aggregate the features of the edge through a shared multi-layer perceptron and a max pooling operation in sequence to obtain a local map; Perform feature stitching on the local geometry map, the local feature map, and the local map to obtain multi-scale fusion features.

5. The multi-scale fusion and linear regularization denoising method based on iterative filtering according to claim 3, characterized in that , Perform global perception on the multi-scale fusion features through the global perception graph generation sub-module to obtain encoded features, including: Reduce the dimension of the multi-scale fusion features through a first weight matrix to obtain first point cloud features; The first point cloud feature is successively calculated through an activation function and subjected to average pooling operation to obtain a global channel descriptor; The dimension of the multi-scale fusion feature is reduced through a second weight matrix to obtain a second point cloud feature; The second point cloud feature is successively calculated through an activation function and subjected to average pooling operation to obtain a global point descriptor encoding; After calculating the outer product of the global channel descriptor and the global point descriptor encoding, the square root of the outer product is taken to generate a global bilinear response; The first point cloud feature, the second point cloud feature, and the global bilinear response are concatenated and then input into a shared multi-layer perceptron to restore the channel dimension and generate a global perception map; The multi-scale fusion feature is subtracted from the global perception map to obtain a third point cloud feature, and the nonlinearity of the third point cloud feature is increased through an activation function to obtain an encoded feature.

6. The multi-scale fusion and linear regularization denoising method based on iterative filtering according to claim 3, characterized in that ,An iterative filter is constructed based on the encoded feature, and the point cloud data is denoised through the iterative filter to obtain denoised point cloud data, including: A linear decoding module is constructed based on the encoded feature. The input of the linear decoding module is the encoded feature, and the output of the linear decoding module is the decoded feature; The point cloud preprocessing module, the encoding module, and the linear decoding module are connected in sequence to form an iterative filtering layer; A preset number of iterative filtering layers are connected in sequence to form an iterative filter, and the decoded feature output by the last iterative filtering layer is used as the output feature of the iterative filter; The point cloud data is denoised through the iterative filter to obtain denoised point cloud data.

7. A multi-scale fusion and linear regularization denoising device based on iterative filtering, characterized in that, Including: An acquisition unit for acquiring point cloud data; A preprocessing unit for preprocessing the point cloud data to obtain point cloud preprocessing data based on normalization, noise synthesis, and downsampling operations; An encoding unit for performing feature encoding on the point cloud preprocessing data to generate an encoded feature based on multi-scale feature fusion and linear regularization; A denoising unit for constructing an iterative filter based on the encoded feature and denoising the point cloud data through the iterative filter to obtain denoised point cloud data.

8. The multi-scale fusion and linear regularization denoising device based on iterative filtering according to claim 7, wherein The preprocessing unit includes: A first construction subunit for constructing a point cloud preprocessing module based on the point cloud data. The point cloud preprocessing module includes a normalization submodule, a noise synthesis submodule, and a downsampling submodule connected in sequence; A normalization subunit for normalizing the point cloud data through the normalization submodule to unify the scale and central position of the point cloud data and obtain normalized point cloud data; A noise synthesis subunit for performing noise synthesis on the normalized point cloud data through the noise synthesis submodule to obtain synthesized point cloud data. The noise synthesis submodule uses Gaussian noise for noise synthesis; A downsampling subunit for performing downsampling operation on the synthesized point cloud data through the downsampling submodule to obtain point cloud preprocessing data.

9. The multi-scale fusion and linear regularization denoising device based on iterative filtering according to claim 8, wherein, The encoding unit includes: A second construction subunit for constructing an encoding module based on the point cloud preprocessing data. The encoding module includes a multi-scale dynamic graph convolution submodule and a global perception map generation submodule; A feature fusion subunit, configured to perform multi-scale feature fusion on the preprocessed point cloud data through the multi-scale dynamic graph convolution sub-module to obtain multi-scale fusion features; A global perception subunit, configured to perform global perception on the multi-scale fusion features through the global perception graph generation sub-module to obtain encoded features.

10. The multi-scale fusion and linear regularization denoising device based on iterative filtering according to claim 9, characterized in that, The denoising unit includes: A third construction subunit, configured to construct a linear decoding module based on the encoded features, where the input of the linear decoding module is the encoded features and the output of the linear decoding module is the decoded features; A fourth construction subunit, configured to connect the point cloud preprocessing module, the encoding module, and the linear decoding module in sequence to form an iterative filtering layer; A fifth construction subunit, configured to connect a preset number of iterative filtering layers in sequence to form an iterative filter, and use the decoded features output by the last iterative filtering layer as the output features of the iterative filter; A denoising subunit, configured to denoise the point cloud data through the iterative filter to obtain denoised point cloud data.

Images