Noise point cloud-oriented three-dimensional wireframe generation method and device

Abstract

The invention discloses a noise point cloud-oriented three-dimensional wireframe generation method and device, and belongs to the technical field of computer assistance and design. The method comprises the following steps: processing a noise point cloud data set to obtain a noise point cloud coordinate set; constructing an edge extraction network model which comprises a multi-scale feature extraction module, an edge shape classification module, a point cloud classification module and an offset prediction module; inputting the noise point cloud coordinate set into a multi-scale feature extraction module for feature extraction to obtain global features; inputting the global features into an edge shape classification module, a point cloud classification module and an offset prediction module to obtain an edge shape classification prediction result, a noise point cloud classification prediction result and an offset prediction result; and generating a three-dimensional wireframe according to the edge shape classification prediction result, the noise point cloud classification prediction result and the offset prediction result. By collecting a noise point cloud data set and combining an edge extraction network model output result, a high-quality three-dimensional wireframe is generated from noise point cloud.

Description

Technical Field

[0001] The present invention relates to the field of computer aided design technology, and more specifically, to a method and device for generating a three-dimensional wireframe for noise point clouds. Background Art

[0002] In the field of computer-aided and design technology, 3D point clouds and 3D wireframes are widely used. Obtaining point clouds by 3D scanning real objects and then editing them in design software is a common technique in reverse engineering. However, point clouds obtained by 3D scanning usually contain noise and lack clear edge features, which makes them incapable of direct application. Wireframes have more significant advantages over point clouds, especially in CAD software, where wireframes can be more conveniently edited and operated interactively. Therefore, generating wireframes from noisy point clouds has become an important and critical task in industrial design and CAD.

[0003] In traditional geometric analysis methods, edge detection usually relies on local geometric features. However, since the quality of 3D scanning point clouds is difficult to effectively support the extraction of accurate geometric features in local areas, the application effect of traditional methods is limited to a certain extent. At present, a large number of studies have been devoted to feature extraction and edge detection of point cloud deep learning, but most methods are still based on noise-free point clouds for training, and have certain limitations in noise resistance. When performing edge detection on noisy point clouds, noise will inevitably affect the results, resulting in a decrease in the quality of edge points.

[0004] In the prior art, additional post-processing steps are usually required to suppress noise interference, thereby reducing its impact on subsequent wireframe generation. In related methods, the wireframe generation step usually treats edge points as a whole for fitting, or adopts a simple segmented fitting strategy, which can easily lead to abnormal geometric distortion, especially under noise interference, edges that should be straight lines may be mistakenly fitted as curves. Therefore, generating 3D wireframes from noisy point clouds is an urgent problem to be solved. Summary of the invention

[0005] 1. Technical problems to be solved

[0006] In view of the problems of insufficient robustness and poor generation quality in the prior art when processing noisy point clouds, the present invention provides a method and device for generating a three-dimensional wireframe for noisy point clouds, which can generate high-quality three-dimensional wireframes from noisy point clouds.

[0007] 2. Technical solution

[0008] The purpose of the present invention is achieved through the following technical solutions.

[0009] A method for generating a three-dimensional wireframe from a noise point cloud comprises the following steps:

[0010] Collecting a noise point cloud data set, and processing the noise point cloud data set to obtain a noise point cloud coordinate set;

[0011] Constructing an edge extraction network model; the edge extraction network model includes a multi-scale feature extraction module, an edge shape classification module, a point cloud classification module and an offset prediction module;

[0012] Input the noise point cloud coordinate set into the multi-scale feature extraction module for feature extraction to obtain global features;

[0013] Input the global features into the edge shape classification module, the point cloud classification module and the offset prediction module respectively, and obtain the edge shape classification prediction results, the noise point cloud classification prediction results and the offset prediction results;

[0014] A three-dimensional wireframe is generated based on the edge shape classification prediction results, the noise point cloud classification prediction results and the offset prediction results.

[0015] As a further improvement of the present invention, the processing of the noise point cloud data set to obtain a noise point cloud coordinate set includes:

[0016] Original three-dimensional point cloud information is obtained from a noise point cloud data set, wherein the original three-dimensional point cloud information includes an original three-dimensional point cloud coordinate set, and random noise is added to the original three-dimensional point cloud coordinate set to obtain a noise point cloud coordinate set.

[0017] As a further improvement of the present invention, the multi-scale feature extraction module includes a pre-convolution module, a plurality of feature convolution modules and a feature processing module;

[0018] The pre-convolution module includes a graph feature extraction unit and a basic unit; the graph feature extraction unit searches for a noise point cloud coordinate set to obtain a neighbor feature, calculates the difference between the neighbor feature and the noise point cloud coordinate set, concatenates the difference with the noise point cloud coordinate set, and inputs the concatenated difference into the basic unit to output a primary feature.

[0019] As a further improvement of the present invention, the feature convolution module includes a graph feature extraction unit, a basic unit and a residual connection unit;

[0020] The residual connection unit is used to splice the input and output of the feature convolution module.

[0021] As a further improvement of the present invention, the feature processing module includes a convolutional feature splicing unit, a convolutional feature pooling unit and a global feature splicing unit;

[0022] The convolution feature splicing unit splices the convolution features obtained by multiple feature convolution modules to generate local features;

[0023] The convolution feature pooling unit performs pooling processing on the convolution features obtained by multiple feature convolution modules to generate pooled features;

[0024] The global feature concatenation unit concatenates the local features and the pooled features to obtain the global features.

[0025] As a further improvement of the present invention, global features are input into the edge shape classification module, and the global features are processed by the one-dimensional convolution layer, normalization layer, ReLU activation function layer, one-dimensional convolution layer and SoftMax activation function layer in the edge shape classification module to output the edge shape classification prediction result.

[0026] As a further improvement of the present invention, global features are input into the noise point cloud classification module, and after the global features are processed by the one-dimensional convolution layer, the normalization layer and the one-dimensional convolution layer in the noise point cloud classification module, the noise point cloud classification prediction result is output.

[0027] As a further improvement of the present invention, global features are input into the offset prediction module, and the global features are processed by the one-dimensional convolution layer, the normalization layer and the ReLU activation function layer in the offset prediction module to output the offset prediction result.

[0028] As a further improvement of the present invention, a three-dimensional wireframe is generated according to the edge shape classification prediction result, the noise point cloud classification prediction result and the offset prediction result, including:

[0029] Extract the index of the edge point based on the noise point cloud classification prediction result, combine the index of the edge point with the noise point cloud coordinate set to generate a noise edge point coordinate set, and add the offset prediction result to the noise edge point coordinate set to obtain the edge point coordinate set;

[0030] Based on the edge shape classification prediction results, the indexes of straight line segments, circles and third-order Bezier curves are extracted respectively, and the edge point coordinate set is divided into a straight line segment coordinate set, a circle coordinate set and a third-order Bezier curve coordinate set according to the indexes of the straight line segment, the circle and the third-order Bezier curve;

[0031] Classify and fit the straight line segment coordinate set, circle coordinate set and third-order Bezier curve coordinate set;

[0032] Merge the classification fitting results to generate a three-dimensional wireframe.

[0033] A three-dimensional wireframe generation device for noise point cloud, comprising:

[0034] A data processing module collects a noise point cloud data set, and processes the noise point cloud data set to obtain a noise point cloud coordinate set;

[0035] A model building module, constructing an edge extraction network model; the edge extraction network model includes a multi-scale feature extraction module, an edge shape classification module, a point cloud classification module and an offset prediction module;

[0036] Feature extraction module, inputs the noise point cloud coordinate set into the multi-scale feature extraction module for feature extraction to obtain global features;

[0037] The result prediction module inputs the global features into the edge shape classification module, the point cloud classification module and the offset prediction module respectively, and obtains the edge shape classification prediction result, the noise point cloud classification prediction result and the offset prediction result;

[0038] The wireframe generation module generates a three-dimensional wireframe based on the edge shape classification prediction results, the noise point cloud classification prediction results and the offset prediction results.

[0039] 3. Beneficial effects

[0040] Compared with the prior art, the advantages of the present invention are:

[0041] (1) The present invention provides a 3D wireframe generation method and device for noise point clouds. By collecting a noise point cloud data set, an edge extraction network model is used to extract edge points and predict noise offsets, and finally a classification fitting method is used to generate a 3D wireframe. This effectively solves the problem that existing methods cannot generate clean and complete wireframes on noise point clouds.

[0042] (2) The present invention provides a 3D wireframe generation method and device for noisy point clouds, which constructs an edge extraction network model, including a multi-scale feature extraction module, an edge shape classification module, a point cloud classification module and an offset prediction module. Based on the multi-scale feature extraction module, the local and global features of the point cloud are fully mined, and the features of edge shape, point cloud classification and offset prediction are efficiently integrated, thereby being able to effectively process various forms of features, having the flexibility to process complex structures, and having strong practicality and wide applicability.

[0043] (3) The present invention provides a 3D wireframe generation method and device for noisy point clouds, which combines the output results of the edge extraction network model and classifies them into a straight line segment coordinate set, a circle coordinate set and a third-order Bezier curve coordinate set. Targeted fitting methods are used according to different categories, which effectively reduces the negative impact of noise and significantly improves the accuracy and efficiency of 3D wireframe generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a schematic diagram of the edge extraction network model structure of an embodiment of the present invention;

[0045] Figure 2 This is a schematic diagram of the structure of a multi-scale feature extraction module according to an embodiment of the present invention;

[0046] Figure 3 Schematic diagram of the structure of a pre-convolution module according to an embodiment of the present invention;

[0047] Figure 4 Schematic diagram of the structure of a feature convolution module according to an embodiment of the present invention;

[0048] Figure 5 This is a schematic diagram of the structure of a feature processing module according to an embodiment of the present invention;

[0049] Figure 6 A schematic diagram of classification fitting of a straight line segment coordinate set according to an embodiment of the present invention;

[0050] Figure 7 It is a schematic diagram of classification fitting of a circular coordinate set according to an embodiment of the present invention;

[0051] Figure 8 It is a schematic diagram of classification fitting of a third-order Bezier curve coordinate set according to an embodiment of the present invention;

[0052] Fig. 9 A graph a showing a three-dimensional wireframe visualization result generated from a noise point cloud according to an embodiment of the present invention;

[0053] Fig.10 A three-dimensional wireframe visualization result b is generated for the noise point cloud of the embodiment of the present invention;

[0054] Fig.11 A three-dimensional wireframe visualization result c is generated for the noise point cloud of the embodiment of the present invention;

[0055] Fig.12 Figure d is a three-dimensional wireframe visualization result generated by the noise point cloud in an embodiment of the present invention. DETAILED DESCRIPTION

[0056] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0057] Example

[0058] The present embodiment provides a three-dimensional wireframe generation method for noise point clouds, comprising the following steps: collecting a noise point cloud data set, processing the noise point cloud data set to obtain a noise point cloud coordinate set; constructing an edge extraction network model; the edge extraction network model comprises a multi-scale feature extraction module, an edge shape classification module, a point cloud classification module and an offset prediction module; inputting the noise point cloud coordinate set into the multi-scale feature extraction module for feature extraction to obtain global features; inputting the global features into the edge shape classification module, the point cloud classification module and the offset prediction module respectively to obtain edge shape classification prediction results, noise point cloud classification prediction results and offset prediction results; generating a three-dimensional wireframe according to the edge shape classification prediction results, the noise point cloud classification prediction results and the offset prediction results.

[0059] Specifically in this embodiment, Figure 1 As shown, a noise point cloud data set is collected, and original three-dimensional point cloud information is obtained from the noise point cloud data set, and the original three-dimensional point cloud information includes an original three-dimensional point cloud coordinate set. It should be noted that in this embodiment, the original three-dimensional point cloud coordinate set is a noise-free three-dimensional point cloud coordinate set in the original three-dimensional point cloud data set. Random noise is added to the original three-dimensional point cloud coordinate set to obtain a noise point cloud coordinate set P, and the noise point cloud coordinate set P is used to generate a noise point cloud data set for training the edge extraction network model in the subsequent steps.

[0060] It should be noted that, in this embodiment, the original three-dimensional point cloud information also includes an original edge point coordinate set, an original three-dimensional point cloud classification label C, and an edge shape classification label T.

[0061] The original three-dimensional point cloud classification label C divides the original three-dimensional point cloud coordinate set into two categories, namely the original non-edge point coordinate set and the original edge point coordinate set, that is, the value range of the original three-dimensional point cloud classification label C is {0,1}, where 0 represents the original non-edge point coordinate set and 1 represents the original edge point coordinate set.

[0062] The edge shape classification label T divides the edge shape types into three categories, namely, straight line segments, circles, and third-order Bezier curves. That is, the value range of the edge shape classification label T is {0,1,2}, where 0 represents a straight line segment, 1 represents a circle, and 2 represents a third-order Bezier curve.

[0063] In this embodiment, random noise is added to the original three-dimensional point cloud coordinate set to obtain a noise edge point coordinate set, and then the distance difference between the noise edge point coordinate set and the original edge point coordinate set is calculated to obtain a distance offset O. The distance offset O is used for iterative training of the edge extraction network model in subsequent steps.

[0064] Further, an edge extraction network model is constructed. In this embodiment, Figure 1 As shown, the edge extraction network model includes a multi-scale feature extraction module, an edge shape classification module, a noise point cloud classification module and an offset prediction module.

[0065] Specifically, Figure 2 As shown, the multi-scale feature extraction module includes a pre-convolution module, multiple feature convolution modules and a feature processing module.

[0066] like Figure 3 As shown, the pre-convolution module includes a graph feature extraction unit and a basic unit.

[0067] The graph feature extraction unit searches the noise point cloud coordinate set P to obtain neighbor features, calculates the difference between the neighbor features and the noise point cloud coordinate set P, thereby capturing the relative geometric relationship and spatial structure in the noise point cloud coordinate set P, and then splices the difference with the noise point cloud coordinate set P to ensure that more local information and input information are retained in the feature space. It should be noted that in this embodiment, the noise point cloud coordinate set P is searched by the existing K nearest neighbor search method. In this embodiment, K=32.

[0068] The basic unit is composed of a one-dimensional convolution layer, a normalization layer and a ReLU activation function layer connected in sequence. In this embodiment, the convolution kernel size of the one-dimensional convolution layer is 1×1 and the step size is 1.

[0069] Therefore, in this embodiment, the noise point cloud coordinate set P is input into the pre-convolution module, processed by the graph feature extraction unit and the basic unit in sequence, and the primary features are output.

[0070] like Figure 4 As shown, the feature convolution module includes a graph feature extraction unit, a basic unit and a residual connection unit.

[0071] The residual connection unit is used to concatenate the input of the feature convolution module with the output processed by the basic unit.

[0072] Therefore, in this embodiment, the initial features are input into the first feature convolution module, and the input of each subsequent feature convolution module is the output of the feature convolution module of the previous layer, and finally the convolution features are output.

[0073] It should be noted that, in this embodiment, preferably, 12 feature convolution modules with the same structure are selected.

[0074] like Figure 5 As shown, the feature processing module includes a convolutional feature splicing unit, a convolutional feature pooling unit and a global feature splicing unit.

[0075] The convolution feature splicing unit splices the convolution features obtained by multiple feature convolution modules to generate local features.

[0076] The convolution feature pooling unit is composed of a one-dimensional convolution layer, a normalization layer, a ReLU activation function layer and a maximum pooling layer connected in sequence. The convolution feature pooling unit performs pooling processing on the convolution features obtained by multiple feature convolution modules to generate pooled features. It should be noted that in this embodiment, the maximum pooling layer needs to select the maximum value in the pooling result as the pooling feature, so as to achieve feature compression.

[0077] The global feature concatenation unit concatenates local features and pooled features to obtain global features.

[0078] Therefore, in this embodiment, the noise point cloud coordinate set P is input into the multi-scale feature extraction module, and is processed by the pre-convolution module, multiple feature convolution modules and feature processing module to obtain the global features.

[0079] Furthermore, the global features are input into the edge shape classification module to obtain the edge shape classification prediction results. Specifically, the edge shape classification module includes a one-dimensional convolution layer, a normalization layer, a ReLU activation function layer, a one-dimensional convolution layer, and a SoftMax activation function layer. The one-dimensional convolution layer extracts local information of the global features, the normalization layer standardizes the one-dimensional convolution layer results, the ReLU activation function layer performs a nonlinear transformation on the standardized one-dimensional convolution layer results, and then, a one-dimensional convolution layer is used to further extract the feature information of the global features, and the SoftMax activation function layer outputs the probability that the point cloud belongs to each category. Thus, the global features are processed in sequence by the one-dimensional convolution layer, the normalization layer, the ReLU activation function layer, the one-dimensional convolution layer, and the SoftMax activation function layer in the edge shape classification module, and the edge shape classification prediction results are output.

[0080] The global features are input into the noise point cloud classification module to obtain the noise point cloud classification prediction results. Specifically, the noise point cloud classification module includes a one-dimensional convolution layer, a normalization layer, and a one-dimensional convolution layer. Thus, the global features are processed by the one-dimensional convolution layer, the normalization layer, and the one-dimensional convolution layer in the noise point cloud classification module in sequence, and the noise point cloud classification prediction results are output.

[0081] The global features are input into the offset prediction module to obtain the offset prediction result. Specifically, the offset prediction module includes a one-dimensional convolution layer, a normalization layer, and a ReLU activation function layer. Thus, the global features are processed in sequence by the one-dimensional convolution layer, the normalization layer, and the ReLU activation function layer in the offset prediction module to output the offset prediction result.

[0082] Therefore, the present embodiment provides a three-dimensional wireframe generation method for noisy point clouds, which constructs an edge extraction network model, including a multi-scale feature extraction module, an edge shape classification module, a point cloud classification module and an offset prediction module. Based on the multi-scale feature extraction module, the local and global features of the point cloud are fully mined, and the features of edge shape, point cloud classification and offset prediction are efficiently integrated, thereby being able to effectively process various forms of features, having the flexibility to process complex structures, and having strong practicality and wide applicability.

[0083] Furthermore, the prediction results are classified according to the edge shape Noise point cloud classification prediction results And the offset prediction results Generate a 3D wireframe. The specific steps are as follows:

[0084] S1: Classification prediction results based on noise point cloud Extract the index of the edge point, combine the index of the edge point with the noise point cloud coordinate set P, generate the noise edge point coordinate set, and use the offset prediction result Add the noise edge point coordinate set to obtain the edge point coordinate set.

[0085] S2: Prediction results based on edge shape classification The indexes of the straight line segment, circle and third-order Bezier curve are extracted respectively, and the edge point coordinate set is divided into a straight line segment coordinate set, a circle coordinate set and a third-order Bezier curve coordinate set according to the indexes of the straight line segment, circle and third-order Bezier curve.

[0086] S3: Classify and fit the straight line segment coordinate set, circle coordinate set and third-order Bezier curve coordinate set.

[0087] Specifically, Figure 6 As shown, for a straight line segment coordinate set, first, the centroid of the straight line segment coordinate set is calculated; second, the direction of the straight line segment coordinate set is extracted using the existing singular value decomposition method as the direction vector of the fitted straight line segment, and the straight line segment coordinate set is projected onto the direction vector to determine the range of the fitted straight line segment; finally, discrete points are generated within the range, and the discrete points are sequentially connected to generate the fitted straight line segment.

[0088] like Figure 7 As shown, for the circular coordinate set, first, the plane normal vector of the circular coordinate set is extracted using the existing singular value decomposition method, so as to fit the plane where the circular coordinate set is located; secondly, the circular coordinate set is projected onto the plane, a local two-dimensional coordinate system is established, and the projected coordinates are converted into two-dimensional coordinates; thirdly, in the local two-dimensional coordinate system, the two-dimensional circle is fit using the existing least squares method to obtain the center and radius; finally, the center of the two-dimensional circle is mapped back to the three-dimensional space, and then combined with the radius to generate a three-dimensional circle.

[0089] like Figure 8 As shown, for a third-order Bezier curve coordinate set, first, a KDTree is established to perform spatial division of the third-order Bezier curve coordinate set to obtain multiple regions; second, a straight line segment and a direction vector are fitted to each individual region according to the method of a straight line segment coordinate set; finally, the straight line segments of each region are arranged in order based on the direction vector, and are merged and connected in sequence to generate a fitted third-order Bezier curve. It should be noted that in this embodiment, the KDTree used is a tree data structure that stores points in a K-dimensional space for rapid retrieval.

[0090] S4: Merge the classification fitting results to generate a three-dimensional wireframe.

[0091] Therefore, this embodiment provides a three-dimensional wireframe generation method for noisy point clouds, which is combined with the output results of the edge extraction network model, classified into a straight line segment coordinate set, a circle coordinate set, and a third-order Bezier curve coordinate set, and adopts targeted fitting methods according to different categories, which effectively reduces the negative impact of noise and significantly improves the accuracy and efficiency of three-dimensional wireframe generation.

[0092] It is worth noting that if Figure 1 As shown, in this embodiment, the edge extraction network model also needs to be iteratively trained, including cross entropy loss, weighted cross entropy loss and mean square error loss.

[0093] Specifically, the cross entropy loss is used to calculate the edge shape classification label T and the edge shape classification prediction result The error between them is calculated as:

[0094]

[0095] Among them, L shp represents the cross entropy loss, i represents a natural number, N represents the number of three-dimensional point clouds, in this embodiment, N = 10000, H C represents the cross entropy loss function, t i ∈T, T represents the edge shape classification label, Represents the edge shape classification prediction result.

[0096] Weighted cross entropy loss is used to calculate the original 3D point cloud classification label C and the noise point cloud classification prediction result The error between them is calculated as:

[0097]

[0098] Among them, L cls represents the weighted cross entropy loss, i and j both represent natural numbers, Q represents the index set of all the original edge point coordinate sets in the original 3D point cloud classification label C, |Q| represents the set length, c i ∈C, c j ∈C, C represents the original 3D point cloud classification label, Represents the classification prediction result of the noise point cloud.

[0099] The mean square error loss is used to calculate the distance offset O and the offset prediction result The error between them is calculated as:

[0100]

[0101] Among them, L off represents the mean square error loss, o i∈O, O represents the offset distance, Indicates the offset prediction result.

[0102] Therefore, in this embodiment, the edge extraction network model is iteratively trained, and the total loss is expressed as:

[0103] L=L shp +L cls +L off

[0104] Among them, L represents comprehensive loss.

[0105] The noise point cloud coordinate set P is input into the edge extraction network model for training, and the overall loss function L is calculated. When the overall loss function L tends to fit, the optimal edge extraction network model is obtained, which is used to generate the optimal edge shape classification prediction results, the optimal noise point cloud classification prediction results, and the optimal noise point cloud classification prediction results.

[0106] The 3D wireframe generation method for noise point cloud provided in this embodiment is compared with the existing 3D wireframe generation method, and the evaluation criteria are Hausdorff distance and chamfer distance. Both Hausdorff distance and chamfer distance are effective measurement indicators for evaluating the quality of 3D wireframe generation. The smaller the Hausdorff distance and chamfer distance, the higher the quality of 3D wireframe generation. The comparison results are shown in Table 1.

[0107] Table 1

[0108] RFEPS FASCC PcqI Nerve This embodiment Hausdorff distance 0.451 0.534 0.629 0.247 0.142 Chamfer distance 0.091 0.117 0.130 0.031 0.017

[0109] It can be seen from Table 1 that, compared with the existing 3D wireframe generation method, the 3D wireframe generation method for noise point cloud provided by this embodiment has the smallest Hausdorff distance and chamfer distance. Therefore, the 3D wireframe generated by the method of this embodiment has the highest quality.

[0110] Among them, the titles of the papers mentioned in Table 1 above are:

[0111] RFEPS: Reconstructing Feature-line Equipped Polygonal Surface;

[0112] FASCC: Feature-aligned segmentation using correlation clustering;

[0113] Pc2wf: 3d wireframe reconstruction from raw point clouds;

[0114] Nerve: Neural volumetric edges for parametric curve extraction from point cloud.

[0115] like Figure 9-12 As shown, the 3D wireframe generation method for noise point clouds provided in this embodiment can generate clean and complete 3D wireframes from noise point clouds. Therefore, the 3D wireframe generation method for noise point clouds provided in this embodiment collects noise point cloud data sets, uses edge extraction network models to extract edge points and predict noise offsets, and finally uses classification fitting methods to generate 3D wireframes, effectively solving the problem that existing methods cannot generate clean and complete wireframes on noise point clouds.

[0116] This embodiment also provides a three-dimensional wireframe generation device for noise point clouds, including a data processing module, a model building module, a feature extraction module, a result prediction module and a wireframe generation module. The data processing module collects a noise point cloud data set, and processes the noise point cloud data set to obtain a noise point cloud coordinate set. The model building module constructs an edge extraction network model; the edge extraction network model includes a multi-scale feature extraction module, an edge shape classification module, a point cloud classification module and an offset prediction module. The feature extraction module inputs the noise point cloud coordinate set into the multi-scale feature extraction module for feature extraction to obtain global features. The result prediction module inputs the global features into the edge shape classification module, the point cloud classification module and the offset prediction module respectively to obtain edge shape classification prediction results, noise point cloud classification prediction results and offset prediction results. The wireframe generation module generates a three-dimensional wireframe according to the edge shape classification prediction results, the noise point cloud classification prediction results and the offset prediction results. The three-dimensional wireframe generation device for noise point clouds provided in this embodiment can implement any method of the three-dimensional wireframe generation method for noise point clouds, and the specific working process of the three-dimensional wireframe generation device for noise point clouds can refer to the corresponding process in the embodiment of the three-dimensional wireframe generation method for noise point clouds. The method and device provided in this embodiment can be implemented in other ways. For example, the device embodiment described above is only schematic; for example, the division of a certain module is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point, the mutual connection or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, or it can be an electrical, mechanical or other form of connection.

[0117] This embodiment also provides a computer device. A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the three-dimensional wireframe generation method for noise point clouds when executing the computer program.

[0118] This embodiment also provides a computer-readable storage medium. A computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, a three-dimensional wireframe generation method for a noise point cloud described in this embodiment is executed. Among them, the computer-readable storage medium can be any tangible medium containing or storing a program, and the program can be used by or in combination with an instruction execution system, device or device; the program code contained on the computer-readable medium can be transmitted using any appropriate medium, including but not limited to wireless, wires, optical cables, RF, etc., or any suitable combination of the above.

[0119] The above schematically describes the invention and its implementation methods, which is not restrictive. Without departing from the spirit or basic features of the invention, the invention can be implemented in other specific forms. What is shown in the accompanying drawings is only one of the implementation methods of the invention. The actual structure is not limited thereto, and any figure mark in the claims should not limit the claims involved. Therefore, if a person of ordinary skill in the art is inspired by it, without departing from the purpose of the invention, a structural method and an embodiment similar to the technical solution are designed without creativity, which should all belong to the protection scope of the present invention. In addition, the word "including" does not exclude other elements or steps, and the word "one" before the element does not exclude the inclusion of "multiple" elements. The multiple elements stated in the product claim can also be implemented by one element through software or hardware. The words first, second, etc. are used to indicate the name, and do not indicate any specific order.

Claims

1. A method for generating a three-dimensional wireframe for a noisy point cloud, comprising the following steps: Collecting a noise point cloud data set, and processing the noise point cloud data set to obtain a noise point cloud coordinate set; Constructing an edge extraction network model; the edge extraction network model includes a multi-scale feature extraction module, an edge shape classification module, a point cloud classification module and an offset prediction module; Input the noise point cloud coordinate set into the multi-scale feature extraction module for feature extraction to obtain global features; Input the global features into the edge shape classification module, the point cloud classification module and the offset prediction module respectively, and obtain the edge shape classification prediction results, the noise point cloud classification prediction results and the offset prediction results; A three-dimensional wireframe is generated based on the edge shape classification prediction results, the noise point cloud classification prediction results and the offset prediction results.

2. A method for generating a three-dimensional wireframe for a noise point cloud according to claim 1, characterized in that: The processing of the noise point cloud data set to obtain a noise point cloud coordinate set includes: Original three-dimensional point cloud information is obtained from a noise point cloud data set, wherein the original three-dimensional point cloud information includes an original three-dimensional point cloud coordinate set, and random noise is added to the original three-dimensional point cloud coordinate set to obtain a noise point cloud coordinate set.

3. The method for generating a 3D wireframe for a noise point cloud according to claim 2, characterized in that: The multi-scale feature extraction module includes a pre-convolution module, a plurality of feature convolution modules and a feature processing module; The pre-convolution module includes a graph feature extraction unit and a basic unit; the graph feature extraction unit searches for a noise point cloud coordinate set to obtain a neighbor feature, calculates the difference between the neighbor feature and the noise point cloud coordinate set, concatenates the difference with the noise point cloud coordinate set, and inputs the concatenated difference into the basic unit to output a primary feature.

4. The method for generating a 3D wireframe for a noise point cloud according to claim 3, characterized in that: The feature convolution module includes a graph feature extraction unit, a basic unit and a residual connection unit; The residual connection unit is used to splice the input and output of the feature convolution module.

5. The method for generating a 3D wireframe for a noise point cloud according to claim 4, characterized in that: The feature processing module includes a convolutional feature splicing unit, a convolutional feature pooling unit and a global feature splicing unit; The convolution feature concatenation unit concatenates the convolution features obtained by multiple feature convolution modules to generate local features; The convolution feature pooling unit performs pooling processing on the convolution features obtained by multiple feature convolution modules to generate pooled features; The global feature concatenation unit concatenates the local features and the pooled features to obtain the global features.

6. A method for generating a three-dimensional wireframe for a noise point cloud according to claim 5, characterized in that: The global features are input into the edge shape classification module. After the global features are processed by the one-dimensional convolution layer, the normalization layer, the ReLU activation function layer, the one-dimensional convolution layer and the SoftMax activation function layer in the edge shape classification module, the edge shape classification prediction result is output.

7. The method for generating a 3D wireframe for a noise point cloud according to claim 5, characterized in that: The global features are input into the noise point cloud classification module. After the global features are processed by the one-dimensional convolution layer, the normalization layer and the one-dimensional convolution layer in the noise point cloud classification module, the noise point cloud classification prediction result is output.

8. The method for generating a three-dimensional wireframe for a noise point cloud according to claim 5, characterized in that: The global features are input into the offset prediction module. After the global features are processed by the one-dimensional convolution layer, the normalization layer and the ReLU activation function layer in the offset prediction module, the offset prediction result is output.

9. A method for generating a 3D wireframe for a noise point cloud according to claims 6-8, characterized in that: Generate a 3D wireframe based on the edge shape classification prediction results, the noise point cloud classification prediction results and the offset prediction results, including: Extract the index of the edge point based on the noise point cloud classification prediction result, combine the index of the edge point with the noise point cloud coordinate set to generate a noise edge point coordinate set, and add the offset prediction result to the noise edge point coordinate set to obtain the edge point coordinate set; Based on the edge shape classification prediction results, the indexes of straight line segments, circles and third-order Bezier curves are extracted respectively, and the edge point coordinate set is divided into a straight line segment coordinate set, a circle coordinate set and a third-order Bezier curve coordinate set according to the indexes of the straight line segment, the circle and the third-order Bezier curve; Classify and fit the straight line segment coordinate set, circle coordinate set and third-order Bezier curve coordinate set; Merge the classification fitting results to generate a three-dimensional wireframe.

10. A 3D wireframe generation device for noise point cloud, characterized in that: include: A data processing module collects a noise point cloud data set, and processes the noise point cloud data set to obtain a noise point cloud coordinate set; A model building module, constructing an edge extraction network model; the edge extraction network model includes a multi-scale feature extraction module, an edge shape classification module, a point cloud classification module and an offset prediction module; Feature extraction module, inputs the noise point cloud coordinate set into the multi-scale feature extraction module for feature extraction to obtain global features; The result prediction module inputs the global features into the edge shape classification module, the point cloud classification module and the offset prediction module respectively, and obtains the edge shape classification prediction result, the noise point cloud classification prediction result and the offset prediction result; The wireframe generation module generates a three-dimensional wireframe based on the edge shape classification prediction results, the noise point cloud classification prediction results and the offset prediction results.

Images