Fracture surface identification method and device based on neural network

Abstract

The invention discloses a fracture surface recognition method and device based on a neural network, and the method comprises the steps: obtaining a fragment point cloud, taking any one data point as a query point in a first feature dimension raising layer, and determining a neighborhood and adjacent points of the query point; performing feature extraction on the query points according to the coordinate information of the query points and the adjacent points to obtain local features and edge features of the query points as input of a next feature dimension raising layer so as to output new local features and edge features; inputting the obtained local features and edge features into a first feature dimension reduction layer to obtain dimension reduction features and taking the dimension reduction features as input of a next feature dimension reduction layer; calculating a characteristic value of a query point according to each dimension reduction characteristic, and if the characteristic value of the query point is in a set range, determining that the query point belongs to a fracture surface; and acquiring all data points belonging to the fracture surface in the fragment point cloud to determine an area belonging to the fracture surface in the fragment point cloud. According to the method and the device, the area representing the fracture surface in the fragment point cloud is automatically identified.

Description

Technical Field

[0001] The present application belongs to the field of point cloud data processing, and specifically relates to a fracture surface recognition method and device based on a neural network. Background Art

[0002] With the development of science and technology, computer technology and digital technology, the three-dimensional reconstruction of cultural relics has also entered the digital stage. Digital modeling can not only restore the appearance of cultural relics buildings, but also provide important assistance for restoration and protection work. In the process of cultural relic restoration, early cultural relic experts and researchers manually restored cultural relics based on historical materials and experience. This method is not only laborious and time-consuming, but also likely to cause secondary wear and tear on cultural relics during the splicing process. With the development of computer technology, cultural relics can be input into specific software for simulated splicing using laser scanning and other methods with digital information as the carrier, and the first step of splicing is the identification and extraction of fracture surfaces. The determination of fracture surfaces currently still relies on manual recognition by the naked eye.

[0003] Therefore, how to automatically determine the fracture surface area of ​​the fragment point cloud of the object fragments is a technical problem that needs to be solved urgently. Summary of the invention

[0004] The purpose of the present application is to automatically determine the area where the fracture surface is located for the fragment point cloud of the object fragment.

[0005] Other features and advantages of the present application will become apparent from the following detailed description, or may be learned in part by the practice of the present application.

[0006] According to an embodiment of the present application, the present application provides a fracture surface recognition method based on a neural network, wherein the neural network includes K layers of feature dimension-increasing layers and K layers of feature dimension-reducing layers, wherein the feature dimension-increasing layers are connected in sequence, and the feature dimension-reducing layers are connected to each other, and the last feature dimension-increasing layer is connected to the first feature dimension-reducing layer, and K is a natural number;

[0007] The method comprises:

[0008] Performing point cloud collection on the object fragments to obtain a fragment point cloud, wherein the fragment point cloud is composed of a plurality of data points;

[0009] In the first feature dimension-raising layer, any data point is used as a query point, and the neighborhood of the query point and the neighboring points of the query point distributed in the neighborhood are determined with the query point as the center and the distance as the radius;

[0010] Extracting features of the query point according to the coordinate information of the query point and the neighboring points to obtain local features and edge features of the query point, and using them as inputs of the next feature dimension-enhancing layer to output new local features and edge features;

[0011] The obtained local features and edge features are input into the first feature dimension reduction layer to obtain the dimension reduction features and serve as the input of the next feature dimension reduction layer;

[0012] Calculating the characteristic value of the query point according to the dimension reduction features output by each of the feature dimension reduction layers, and if the characteristic value of the query point is within a set range, determining that the query point belongs to a fracture surface;

[0013] All data points belonging to the fracture surface in the fragment point cloud are obtained to determine the area belonging to the fracture surface in the fragment point cloud.

[0014] According to one embodiment of the present application, calculating the local features and edge features of the query point in the feature dimension-raising layer includes:

[0015] In the first feature dimension-raising layer, according to a neighborhood of the query point formed by taking any data point as the query point and the first set distance as the radius, the first-order local feature of the query point is calculated, and the neighboring points of the query point are determined with the second set distance as the radius, and the vector difference between the feature vector of the query point and the feature vectors of each of the neighboring points is calculated as the first-order edge feature;

[0016] In the Nth feature dimensionality increase layer, the N-1 order neighborhood features and the N-1 order edge features are combined and dimensionality increased convolution is performed to obtain N-order local features, and the vector differences in the N-1 order edge features are subtracted pairwise to obtain N-order edge features, where N belongs to K and N is greater than 1.

[0017] According to one embodiment of the present application, calculating the dimension reduction feature in the feature dimension reduction layer includes:

[0018] In the first feature dimension reduction layer, global feature extraction is performed on the K-order edge features, the K-order local features, and the first-order target local features to obtain first-order dimension reduction features, and the feature dimension increase layer where the first-order target local features are located and the K-th feature dimension increase layer are sequentially connected;

[0019] In the Mth feature dimensionality reduction layer, global feature extraction is performed on the M-1th order dimensionality reduction features and the Mth order target local features to obtain the Mth order dimensionality reduction features. The feature dimensionality increase layer where the Mth order target local features and the M-1th order target local features are located has a sequential connection relationship, and M belongs to K, and M is greater than 1.

[0020] According to an embodiment of the present application, the local features include neighborhood features and neighborhood partition information. At the first feature dimension-raising layer, the first-order local features of the query point are calculated based on a neighborhood of the query point formed with any data point as the query point and a first set distance as the radius, including:

[0021] Taking any data point in the fragment point cloud as a query point;

[0022] Taking the query point as the center and a first set distance as the radius, determine a neighborhood of the query point, wherein the data points in the neighborhood are first neighboring points;

[0023] The coordinate information of the query point and the first neighboring point, and the feature vectors of the query point and the first neighboring point are convolved to obtain the first-order neighborhood features of the neighborhood, and the farthest point sampling is performed on the data points in the neighborhood to determine the number of sampling points in the neighborhood and the sampled coordinate information to obtain the first-order neighborhood division information.

[0024] According to an embodiment of the present application, convolving the coordinate information of the query point and the first neighboring point, and the feature vectors of the query point and the first neighboring point, includes:

[0025] The first set distance includes a plurality of radius values, and a sub-neighborhood is determined according to each of the radius values, and the largest sub-neighborhood is the neighborhood of the query point;

[0026] Calculating sub-neighborhood features of the sub-neighborhoods respectively according to the coordinate information of the query point and the data points in each sub-neighborhood;

[0027] Aggregate the sub-neighborhood features to obtain the neighborhood feature of the query point.

[0028] According to one embodiment of the present application, performing farthest point sampling on the data points in the neighborhood includes:

[0029] In the neighborhood of the query point, a first-order sampling area is formed with the query point as the center and a third set distance as the radius, where the third set distance is smaller than the first set distance;

[0030] The data point farthest from the query point in the first-order sampling area is the first-order sampling point;

[0031] In a new first-order sampling area formed with the latest first-order sampling point as the center and the third set distance as the radius, the data point farthest from the latest sampling point is used as the first-order sampling point until no more first-order sampling points can be obtained.

[0032] According to an embodiment of the present application, at the Nth feature dimension-raising layer, the N-1th order neighborhood feature and the N-1th order edge feature are combined and dimension-raising convolution is performed to obtain the Nth order local feature, including:

[0033] Aggregating the N-1 order neighborhood feature and the N-1 order edge feature to obtain an N-order initial neighborhood feature;

[0034] According to the N-1 order neighborhood division information, the sampling area characteristics of each N-1 order sampling area are calculated;

[0035] The N-order initial neighborhood features and the sampling area features of each of the N-1-order sampling areas are convolved to obtain the N-order neighborhood features, and the data points in the neighborhood are sampled at the farthest point to determine the number of N-order sampling points in the neighborhood and the coordinate information of the N-order sampling points to obtain the N-order neighborhood division information.

[0036] According to an embodiment of the present application, the farthest point sampling is performed on the data points in the neighborhood, the number of N-order sampling points in the neighborhood and the coordinate information of the N-order sampling points are determined, and the N-order neighborhood partition information is obtained, including:

[0037] In the neighborhood of the query point, an N-1-order sampling area is formed with the N-1-order sampling point as the center and a third set distance as the radius, wherein the third set distance has different values ​​in different feature dimension-raising layers, and the size of the third set distance is proportional to the size of N;

[0038] The N-1th order sampling point farthest from the query point in the Nth order sampling area is the Nth order sampling point;

[0039] In a new N-order sampling area formed with the latest N-order sampling point as the center and the third set distance as the radius, the N-1-order sampling point farthest from the latest N-order sampling point is used as a sampling point;

[0040] The number and coordinate information of the N-th order sampling points are used as neighborhood partitioning information of the N-th feature dimension-raising layer.

[0041] According to one embodiment of the present application, calculating the feature value of the query point according to all dimensionality reduction features includes:

[0042] Perform feature aggregation on all reduced-dimensional features and add residuals to obtain the target feature matrix;

[0043] The target feature matrix is ​​pooled to obtain the feature value of the query point.

[0044] According to an embodiment of the present application, the present application provides a fracture surface identification device based on a neural network, wherein the neural network includes K layers of feature dimension-increasing layers and K layers of feature dimension-reducing layers, wherein each of the feature dimension-increasing layers is connected in sequence, each of the feature dimension-reducing layers is connected to each other, and the last feature dimension-increasing layer is connected to the first feature dimension-reducing layer, and K is a natural number;

[0045] The device comprises:

[0046] A collection module is configured to collect point clouds of object fragments to obtain a fragment point cloud, wherein the fragment point cloud is composed of a plurality of data points;

[0047] The feature dimension-raising module is configured to, in the first feature dimension-raising layer, take any data point as a query point, take the query point as the center, set a distance as a radius to determine the neighborhood of the query point and the neighboring points of the query point distributed in the neighborhood;

[0048] A feature updating module is configured to extract features of the query point according to the coordinate information of the query point and the neighboring points, obtain local features and edge features of the query point, and use them as inputs of the next feature dimension-increasing layer to output new local features and edge features;

[0049] A feature dimension reduction module is configured to input the obtained local features and edge features into a first feature dimension reduction layer, obtain dimension reduction features and use them as inputs for a next feature dimension reduction layer;

[0050] A judgment module is configured to calculate a feature value of the query point according to the dimension reduction features output by each feature dimension reduction layer, and if the feature value of the query point is within a set range, determine that the query point belongs to a fracture surface;

[0051] The determination module is configured to obtain all data points belonging to the fracture surface in the fragment point cloud to determine the area belonging to the fracture surface in the fragment point cloud.

[0052] In the present application, point clouds of object fragments are collected to obtain fragment point clouds, which are composed of multiple data points. In the first feature dimension-upgrading layer, any data point is taken as a query point, and the query point is taken as the center and the distance is set as the radius to determine the neighborhood of the query point and the neighboring points of the query point in the neighborhood. Feature dimension-upgrading is performed based on the coordinate information of the query point and the neighboring points to obtain local features and edge features of the query point, which are used as inputs of the next feature dimension-upgrading layer to output new local features and edge features. The obtained local features and edge features are input into the first feature dimension-downgrading layer to obtain reduced-dimensional features and used as inputs of the next feature dimension-upgrading layer. The feature value of the query point is calculated based on each reduced-dimensional feature. If the feature value of the query point is within the set range, it is determined that the query point belongs to a fracture surface, and all data points belonging to the fracture surface in the fragment point cloud are obtained to determine the area belonging to the fracture surface in the fragment point cloud. The present application can more accurately calculate the characteristic value of each data point by performing feature dimensionality increase and feature dimensionality reduction calculations on each point in the fragment point cloud. Based on the obvious differences in the features of the data points belonging to the fracture surface and the non-fracture surface, it is further determined whether the data point in the fragment point cloud belongs to the fracture surface, thereby achieving a more accurate judgment of whether the data point belongs to the fracture surface.

[0053] Other features and advantages of the present application will become apparent from the following detailed description, or may be learned in part by the practice of the present application.

[0054] It should be understood that the foregoing general description and the following detailed description are exemplary only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The drawings herein are incorporated into the specification and constitute a part of the specification, illustrate embodiments consistent with the present application, and together with the specification are used to explain the principles of the present application. Obviously, the drawings described below are only some embodiments of the present application, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0056] Figure 1 A flowchart of a fracture surface identification method based on a neural network according to an embodiment of the present application is shown.

[0057] Figure 2 A flow chart of calculating local features and edge features of a query point in a feature dimensionality enhancement layer according to an embodiment of the present application is shown.

[0058] Figure 3 A flow chart of calculating dimensionality reduction features in a feature dimensionality reduction layer according to an embodiment of the present application is shown.

[0059] Figure 4 A flowchart is shown for calculating the first-order local features of a query point at the first feature dimension-raising layer according to an embodiment of the present application, based on a neighborhood of the query point formed with any data point as the query point and a first set distance as the radius.

[0060] Figure 5 A flowchart of calculating neighborhood features at the first feature dimensionality enhancement layer according to an embodiment of the present application is shown.

[0061] Figure 6 A flowchart of performing farthest point sampling on data points in a neighborhood according to an embodiment of the present application is shown.

[0062] Figure 7 A flowchart is shown of combining N-1 order neighborhood features and N-1 order edge features and performing dimension-up convolution at the Nth feature dimension-upgrading layer to obtain N-order local features according to an embodiment of the present application.

[0063] Figure 8 A flowchart is shown for sampling the farthest points of data points in a neighborhood, determining the number of N-order sampling points in the neighborhood and the coordinate information of the N-order sampling points, and obtaining N-order neighborhood partition information according to an embodiment of the present application.

[0064] Fig. 9 A flow chart of calculating the feature value of a query point based on all dimensionality reduction features according to an embodiment of the present application is shown.

[0065] Fig.10 A schematic diagram of a neural network according to an embodiment of the present application is shown.

[0066] Fig.11 A fracture surface identification device based on a neural network according to an embodiment of the present application is shown.

[0067] Fig.12 A hardware structure diagram of a fracture surface identification method based on a neural network according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0068] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that this application will be more comprehensive and complete and fully convey the concept of the example embodiments to those skilled in the art.

[0069] In addition, described feature, structure or characteristic can be combined in one or more embodiments in any suitable manner. In the following description, many specific details are provided to provide a full understanding of the embodiments of the present application. However, those skilled in the art will appreciate that the technical scheme of the present application can be put into practice without one or more of the specific details, or other methods, components, devices, steps, etc. can be adopted. In other cases, known methods, devices, realizations or operations are not shown or described in detail to avoid blurring the various aspects of the application.

[0070] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities may be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0071] The flowcharts shown in the accompanying drawings are only exemplary and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps can be decomposed, and some operations / steps can be combined or partially combined, so the actual execution order may change according to actual conditions.

[0072] In the embodiments of the present application, the term "module" or "unit" refers to a computer program or a part of a computer program with a predetermined function, and works together with other related parts to achieve a predetermined goal, and can be implemented in whole or in part by using software, hardware (such as processing circuits or memories) or a combination thereof. Similarly, a processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be part of an overall module or unit that includes the function of the module or unit.

[0073] See also Figure 1 , Figure 1 A flowchart of a fracture surface recognition method based on a neural network according to an embodiment of the present application is shown. The neural network includes K layers of feature dimension-increasing layers and K layers of feature dimension-reducing layers. Each feature dimension-increasing layer is connected in sequence, each feature dimension-reducing layer is connected to each other, and the last feature dimension-increasing layer is connected to the first feature dimension-reducing layer. K is a natural number. The embodiment of the application provides the execution steps of the fracture surface recognition method based on a neural network, including:

[0074] Step S110, collecting point clouds of object fragments to obtain fragment point clouds, where the fragment point clouds consist of multiple data points;

[0075] Step S120, in the first feature dimension-raising layer, any data point is used as a query point, and the neighborhood of the query point and the neighboring points of the query point distributed in the neighborhood are determined with the query point as the center and the distance as the radius;

[0076] Step S130, extracting features of the query point according to the coordinate information of the query point and its neighboring points, obtaining local features and edge features of the query point, and using them as inputs of the next feature dimension-enhancing layer to output new local features and edge features;

[0077] Step S140, inputting the obtained local features and edge features into the first feature dimension reduction layer, obtaining the reduced dimension features and using them as inputs for the next feature dimension reduction layer;

[0078] Step S150, calculating the feature value of the query point according to the dimension reduction features output by each feature dimension reduction layer, and if the feature value of the query point is within a set range, determining that the query point belongs to the fracture surface;

[0079] Step S160, obtaining all data points belonging to the fracture surface in the fragment point cloud to determine the area belonging to the fracture surface in the fragment point cloud.

[0080] The above six steps are described in detail below.

[0081] It should be noted that the neural network includes the K-layer feature dimension-raising layer and the K-layer feature dimension-reducing layer, and the number of feature dimension-raising layers and feature dimension-reducing layers is the same. Each feature dimension-raising layer is connected in sequence, each feature dimension-reducing layer is connected to each other, and the last feature dimension-raising layer is connected to the first feature dimension-reducing layer, that is, except for the first feature dimension-raising layer, the output of the previous feature dimension-raising layer is used as input for the other feature dimension-raising layers. Except for the first feature dimension-reducing layer, the output of the previous feature dimension-reducing layer is used as input for the other feature dimension-reducing layers.

[0082] In step S110, point cloud collection is performed on the fragments of the object to obtain a point cloud of the fragments. The point cloud collection of the fragments of the object refers to determining multiple measurement points on the fragments of the object, and then using a measuring device to obtain the position information of these measurement points in a preset coordinate system. It should be clear that the preset coordinate system can be constructed with any point as the origin. Then, based on the position information of each measurement point, the corresponding measurement point is found in the virtual space, and the fragment point cloud can be obtained by materialization.

[0083] The following methods can be used to find the data points corresponding to the measurement points in the virtual space.

[0084] In some embodiments, a data point having the same coordinate value as the measurement point may be searched in the virtual space as the data point corresponding to the measurement point in the virtual space.

[0085] In some embodiments, a measurement point can be arbitrarily selected as a selected point, and a point can be arbitrarily selected in the virtual space as a data point corresponding to the selected point. Based on the relative positional relationship between the selected point and other measurement points, the corresponding data point is determined in the virtual space, and is materialized to obtain a fragment point cloud. The relative positional relationship between each data point in the fragment point cloud is the same as the relative positional relationship between the measurement points corresponding to the data points.

[0086] In step S120, the fragment point cloud is input into the neural network to calculate the characteristic value of each data point in the fragment point cloud. It should be noted that multiple fragment point clouds can be input into the neural network at the same time to calculate the characteristic value of each data point of each fragment point cloud.

[0087] Now, we take a data point in a fragment point cloud as an example to explain how to calculate the feature value of the data point in the neural network. First, the features of the selected data point are calculated through the neural network, and then the feature value of the selected data point is calculated based on the features of the selected data point. The selected data point is also the query point below. It should be clear that edge features, local features and dimensionality reduction features are all features of the query point at different stages.

[0088] In a fragment point cloud, any data point is used as a query point, and a spherical space with the query point as the center and the set distance as the radius is the neighborhood of the query point. The data points in the neighborhood are the neighboring points of the query point.

[0089] In step S130, it should be noted that each feature dimension enhancement layer includes two feature extraction parts. When the two feature extraction parts are used for feature extraction, it is necessary to first determine the feature vectors of each neighboring point in the neighborhood and the feature vector of the query point.

[0090] In some embodiments, the feature vectors of the query point and each neighboring point can be determined by the normal vectors on the coordinate axes of each dimension of the query point and the neighboring point.

[0091] In some embodiments, the feature vectors of the query point and each neighboring point may also be obtained by convolving the coordinate information of the query point and each neighboring point.

[0092] The first feature extraction part uses the feature vectors of each neighboring point in the neighborhood and the feature vector of the query point as input, and is used to subtract the feature vector of the query point from the feature vector of the neighboring point to obtain the difference as the edge feature. Based on the great difference in the smoothness of the fracture surface and the non-fracture surface of the object fragments, when part of the query point and the neighboring point are located on the fracture surface and the other part are located on the non-fracture surface, the difference in the feature vectors of the query point and the neighboring point will be relatively large. Therefore, it can be determined whether there is an edge of the fracture surface and the non-fracture surface in the current neighborhood.

[0093] The following formula can be used to obtain edge features in the first feature dimension-raising layer:

[0094]

[0095] Among them, h i is the feature vector of point i (query point). j represents the neighboring points of the query point, and ε (neighboring point set) is the set of adjacency relations, which represents the adjacency relations of all point pairs (i, j). θ is a learnable function that processes the features and feature differences of points. In the specific operation of this article, first for each point i, find its k nearest neighbor points j in the feature space. Then, calculate the feature difference h of each point pair (i, j) j -h i Then, the feature h of point i is transformed into i and feature difference h j -h i Combined, a new feature representation is obtained. Finally, the feature representations of all adjacent points are pooled to obtain the new feature h of point i i ′ , which is the first-order edge feature.

[0096] The eigenvector difference can be used to quickly determine which data points have drastic changes in their eigenvectors. The query point can be used as a reference to determine which data points have drastic changes in their eigenvectors. The direction in which the eigenvector changes can be determined, and then whether there is a fracture surface edge in the current neighborhood and the location of the fracture surface edge can be determined.

[0097] The second feature extraction part calculates the spatial distribution characteristics of the neighborhood, that is, the local characteristics, based on the feature vectors and coordinate information of the query point and the neighboring points. This is because the fracture surface generally has more potholes and many irregular areas compared to the non-fracture surface. The local characteristics of the neighborhood can be used to determine whether the data point is in the fracture surface area or the non-fracture surface area.

[0098] In the feature dimensionality upgrading layer, the obtained edge features and local features are continuously upgraded to analyze the local features and edge features in a more detailed manner, and finally the significant parts of the edge features and local features are made more significant, and the insignificant parts are made less significant, so as to highlight the features of the query point and determine whether the query point is located on the fracture surface.

[0099] In step S140, the obtained local features and edge features are input into the first feature dimensionality reduction layer, and the dimensionality reduction features are obtained and used as the input of the next feature dimensionality reduction layer. The obtained local features and edge features are continuously reduced in dimension to aggregate the edge features and local features, and only the more significant features are retained through dimensionality reduction.

[0100] In step S150, the features of the query point are accurately calculated through the feature dimension enhancement layer and the feature dimension reduction layer to obtain the feature value of the query point. Based on the training process of the neural network, the feature value range of the query point when the query point belongs to the fracture surface can be determined. If the feature value of the query point is within the set range, it means that the query point is on the fracture surface. If the feature value of the query point is not within the set range, it means that the query point is not on the fracture surface.

[0101] In some embodiments of the present application, it is determined that the query point is located on the fracture surface in the following manner, and whether the neighborhood is simultaneously on the fracture surface and the non-fracture surface can be determined by the first feature extraction part.

[0102] If the first feature extraction part determines that the neighborhood is not at the fracture surface or the non-fracture surface at the same time, it means that there is no boundary line of the fracture surface in the neighborhood, that is, the neighborhood belongs entirely to the fracture surface or the non-fracture surface. At this time, according to the local features obtained by the second feature extraction part, it is determined whether the neighborhood is at the fracture surface or the non-fracture surface. If the neighborhood is at the fracture surface, it means that the query point is at the fracture surface, and if the neighborhood is at the non-fracture surface, it means that the query point is at the non-fracture surface.

[0103] It should be clear that if the eigenvector difference is large and there is a negative value, then the eigenvector difference refers to the absolute value of the difference between the two eigenvectors. The eigenvector difference is used to indicate the difference between the two eigenvectors.

[0104] If the first feature extraction part determines that the neighborhood is at both the fracture surface and the non-fracture surface, it means that there is a fracture surface in the neighborhood. According to the spatial distribution characteristics of the area with large feature vector difference and the area with small feature vector difference, it is determined which area belongs to the fracture surface. Then if the query point is in the area where the fracture surface is located, it means that the query point is on the non-fracture surface. In this way, it can be determined whether all data points in the neighborhood are on the fracture surface, and the area with large feature vector difference between the query point and the neighboring point and the area with small feature vector difference can be determined, that is, the junction of the fracture surface and the fracture surface can be determined, and the regional division of the neighborhood can be realized.

[0105] In the above-mentioned various situations where the query point is located on the fracture surface, the characteristic value of the query point has a set range. By calculating the range where the characteristic value of the query point is located, it is determined whether the query point is located on the fracture surface.

[0106] See also Figure 2 , Figure 2 A flowchart of calculating the local features and edge features of a query point in a feature dimension-raising layer according to an embodiment of the present application is shown. The embodiment of the present application provides steps for calculating the local features and edge features of a query point in a feature dimension-raising layer, including:

[0107] Step S201, at the first feature dimension-raising layer, based on a neighborhood of the query point formed by taking any data point as the query point and the first set distance as the radius, calculate the first-order local feature of the query point, determine the neighboring points of the query point with the second set distance as the radius, and calculate the vector difference between the feature vector of the query point and the feature vectors of each neighboring point as the first-order edge feature;

[0108] Step S202, in the Nth feature dimension-raising layer, the N-1 order neighborhood features and the N-1 order edge features are combined and dimension-raising convolution is performed to obtain N-order local features, and the vector differences in the N-1 order edge features are subtracted pairwise to obtain N-order edge features, where N belongs to K and N is greater than 1.

[0109] The above two steps are described in detail below.

[0110] In step S201, it should be clear that the local features obtained in the first feature dimension-raising layer are first-order local features, and the edge features are first-order edge features. That is, the local features obtained in the Nth feature dimension-raising layer are Nth-order local features and Nth-order edge features.

[0111] In the first feature dimension-enhancing layer, firstly, a data point is randomly selected in a fragment point cloud as a query point, and a spherical space is determined as the neighborhood of the query point with the first distance as the radius, and the data points in the neighborhood are the first neighboring points of the query point. The feature vectors of the query point and the first neighboring point are obtained, and the feature vectors of the query point and the first neighboring point and the coordinate information of the query point and the first neighboring point are convolved to obtain the first-order local features.

[0112] The second neighboring points of the query point are determined with the second set distance as the radius, and the vector difference between the feature vector of the query point and the feature vectors of each second neighboring point is calculated as the first-order edge feature. In some embodiments, the edge feature exists in the form of a matrix. For example, the feature vectors a (the feature vector of the query point), b, and c are obtained as vector differences, ab, ac, ba, and ca, that is, in the first-level vector difference calculation process, the feature vector of the query point must be involved.

[0113] In step S202, in the Nth feature dimension-raising layer, the N-1th order neighborhood feature and the N-1th order edge feature are combined and dimension-raising convolution is performed to obtain the Nth order local feature. That is, the local features are continuously dimension-raised and polarized, so that the significant parts of the local features become more significant and the insignificant parts become less significant, so that the local features are more representative and can more clearly express the features of the query point.

[0114] At the same time, the vector differences in the N-1 order edge features are subtracted from each other to obtain an N order edge feature, where N belongs to K and is greater than 1. For example, for the vector differences A, B, and C, vector differences AB, AC, BC, BA, CA, and CB are obtained.

[0115] See also Figure 3 , Figure 3 A flowchart of calculating dimensionality reduction features in a feature dimensionality reduction layer according to an embodiment of the present application is shown. The embodiment of the present application provides steps for calculating dimensionality reduction features in a feature dimensionality reduction layer, including:

[0116] Step S301, in the first feature dimension reduction layer, global feature extraction is performed on the K-order edge features, the K-order local features, and the first-order target local features to obtain first-order dimension reduction features, and the feature dimension increase layer where the first-order target local features are located and the K-th feature dimension increase layer are sequentially connected;

[0117] Step S302, in the Mth feature dimensionality reduction layer, global feature extraction is performed on the M-1th order dimensionality reduction features and the Mth order target local features to obtain the Mth order dimensionality reduction features, the feature dimensionality increase layer where the Mth order target local features and the M-1th order target local features are located has a sequential connection relationship, so M belongs to K, and M is greater than 1.

[0118] The above two steps are described in detail below.

[0119] In step S301, in the first feature dimensionality reduction layer, global feature extraction is performed based on the K-order local features and K-order edge features output by the K-th feature dimensionality increase layer, as well as the first-order target local features. By combining and reducing the K-order local features, K-order edge features, and the first-order target local features, more significant features are obtained.

[0120] It should be made clear that the existence of the target local features is to prevent the feature dimensionality reduction layer from causing distortion of the dimensionality reduction features during dimensionality reduction, that is, the dimensionality reduction features cannot well represent the real features of the query points. Therefore, the target local features are added as references when calculating the dimensionality reduction features to avoid distortion of the dimensionality reduction features. The first feature dimensionality reduction layer corresponds to the first-order target local features, and the Mth feature dimensionality reduction layer corresponds to the Mth-order target local features.

[0121] It should be further clarified that the feature dimension-raising layer where the first-order target local features are located has a sequential connection relationship, and in the first feature dimension-reduction layer, the feature dimension-raising layer where the first-order target local features are located has a sequential connection relationship with the Kth feature dimension-raising layer.

[0122] In the Mth feature dimension-raising layer, the feature dimension-raising layer where the M-order target local features are located has a sequential connection relationship, and has a sequential connection relationship with the target local features of the previous M-1 feature dimension-raising layers. And the feature dimension-raising layer where the M-order target local features are located can overlap with the feature dimension-raising layer where the M-1-order target local features are located.

[0123] For example, if the second-order target local features are located in the feature dimension-raising layers: K-1th feature dimension-raising layer and K-2th feature dimension-raising layer, the third-order target local features are located in the feature dimension-raising layers: K-2th feature dimension-raising layer and K-3th feature three-dimensional layer.

[0124] It should be clear that the target local features of the Kth feature dimensionality reduction layer include the coordinate information and feature vectors of the query point and its neighboring points, that is, the dimensionality reduction features are calculated with reference to the most original information of the query point, so that the final dimensionality reduction features can avoid distortion.

[0125] In step S302, at the Mth feature dimension reduction layer, global feature extraction is performed on the M-1th order dimension reduction features and the Mth order target local features to obtain the Mth order dimension reduction features.

[0126] It should be made clear that the reason why the feature dimension reduction layer has multiple layers is that if the dimension reduction is completed in one time, the accuracy of the dimension reduction features output by the feature dimension reduction layer cannot be guaranteed. The feature dimension increase layer is divided into multiple layers, and the local features of the target are used as a reference to combine the local features and edge features or the dimension reduction features with the target local features for dimension reduction, which ensures that the dimension reduction features will not be distorted, because the features that have been dimensionally increased multiple times may not be able to retain the characteristics of some features due to multiple dimension increases. Therefore, by gradually reducing the dimensions of the features of the query point, the dimension reduction features converge, which not only ensures the accuracy of the dimension reduction features, but also ensures that the dimension reduction features will not be distorted to determine the feature values ​​of the query points more efficiently and accurately. First increase the dimension and then reduce the dimension, which is equivalent to first gradually and carefully analyzing the features of the query point (upgrading the dimension), and then gradually refining and combining the analysis results (reducing the dimension), and finally obtaining the feature values ​​of the query point.

[0127] See also Figure 4 , Figure 4 A flowchart is shown of calculating the first-order local features of a query point at the first feature dimension-raising layer according to an embodiment of the present application, based on a neighborhood of the query point formed with any data point as the query point and a first set distance as a radius. Local features include neighborhood features and neighborhood partitioning information. The embodiment of the present application provides steps for calculating the first-order local features of a query point based on a neighborhood of the query point formed with any data point as the query point and a first set distance as a radius, including:

[0128] Step S401, taking any data point in the fragment point cloud as a query point;

[0129] Step S402, taking the query point as the center and the first set distance as the radius, determining the neighborhood of the query point, and the data points in the neighborhood are first neighboring points;

[0130] Step S403, convolve the coordinate information of the query point and the first neighboring point, and the feature vectors of the query point and the first neighboring point to obtain the first-order neighborhood features of the neighborhood, and perform farthest point sampling on the data points in the neighborhood to determine the number of sampling points in the neighborhood and the sampled coordinate information to obtain the first-order neighborhood division information.

[0131] The above three steps are described in detail below.

[0132] It should be clear that the neighborhood partition information is used to re-divide the neighborhood and provide a reference for calculating the neighborhood features in the next layer of feature dimensionality upgrade.

[0133] In step S401, any data point in the fragment point cloud is used as a query point.

[0134] In step S402, a neighborhood of the query point is determined with the query point as the center and the first set distance as the radius, and the data points in the neighborhood are first neighboring points.

[0135] In step S403, the coordinate information of the query point and the first neighboring point, as well as the feature vectors of the query point and the first neighboring point are convolved to obtain the first-order neighborhood features of the neighborhood to represent the spatial distribution characteristics of the neighborhood, and the data points in the neighborhood are sampled at the farthest point to determine the number of sampling points in the neighborhood and the sampled coordinate information to obtain the first-order neighborhood division information.

[0136] In some embodiments, the target local feature only includes the neighborhood partition information of the feature dimensionality-enhanced layer where the target local feature is located, and does not include the neighborhood feature.

[0137] See also Figure 5 , Figure 5 A flowchart of calculating neighborhood features at the first feature dimension-raising layer according to an embodiment of the present application is shown. The present application embodiment provides step S403 of calculating neighborhood features at the first feature dimension-raising layer, including:

[0138] Step S501, the first set distance includes a plurality of radius values, and a sub-neighborhood is determined according to each radius value, and the largest sub-neighborhood is the neighborhood of the query point;

[0139] Step S502, calculating sub-neighborhood features of the sub-neighborhoods respectively according to the coordinate information of the query point and the data points in each sub-neighborhood;

[0140] Step S503: Aggregate the sub-neighborhood features to obtain the neighborhood features of the query point.

[0141] The above three steps are described in detail below.

[0142] In step S501, the first set distance includes a plurality of radius values, and with different radius values ​​as radii and the query point as the center, a plurality of areas are respectively determined as sub-neighborhoods, wherein the largest sub-neighborhood is the neighborhood of the query point.

[0143] In step S502, the sub-neighborhood features of the sub-neighborhood are calculated based on the coordinate information of the query point and the data points in each sub-neighborhood. That is, in addition to extracting the sub-neighborhood features of the sub-domain as the neighborhood, feature extraction is also performed on multiple sub-neighborhoods within the neighborhood to obtain each sub-neighborhood feature.

[0144] In step S503, each sub-neighborhood feature is aggregated to obtain the neighborhood feature of the query point. In this way, the neighborhood features can be extracted more comprehensively.

[0145] See also Figure 6 , Figure 6A flowchart of sampling the farthest point of a data point in a neighborhood according to an embodiment of the present application is shown. The present application embodiment provides a step S403 of sampling the farthest point of a data point in a neighborhood, including:

[0146] Step S601, in the neighborhood of the query point, a first-order sampling area is formed with the query point as the center and a third set distance as the radius, where the third set distance is smaller than the first set distance;

[0147] Step S602, the data point farthest from the query point in the first-order sampling area is the first-order sampling point;

[0148] Step S603: In a new first-order sampling area formed with the latest first-order sampling point as the center and the third set distance as the radius, the data point farthest from the latest sampling point is used as the first-order sampling point until no more first-order sampling points can be obtained.

[0149] The above three steps are described in detail below.

[0150] In step S601, in the neighborhood of the query point, a first-order sampling area is formed with the query point as the center and the third set distance as the radius, and the third set distance is less than the first set distance. It should be clear that the sampling area determined in the first feature dimension-raising layer is a first-order sampling area, and the determined sampling points are first-order sampling points, that is, the sampling area determined in the Nth feature dimension-raising layer is an N-order sampling area, and the determined sampling points are N-order sampling points.

[0151] In step S602, the data point farthest from the query point in the first-order sampling area is the first-order sampling point.

[0152] In step S603, in a new first-order sampling area formed with the latest first-order sampling point as the center and the third set distance as the radius, the data point farthest from the latest acquired sampling point is used as the first-order sampling point, until no more first-order sampling points can be obtained, and all first-order sampling points are obtained.

[0153] See also Figure 7 , Figure 7 A flowchart is shown of combining the N-1 order neighborhood features and the N-1 order edge features and performing dimension-up convolution at the N-th feature dimension-up layer to obtain the N-th order local features according to an embodiment of the present application. The embodiment of the present application provides a step S202 of combining the N-1 order neighborhood features and the N-1 order edge features and performing dimension-up convolution at the N-th feature dimension-up layer to obtain the N-th order local features, including:

[0154] Step S701, aggregating the N-1 order neighborhood features and the N-1 order edge features to obtain the N order;

[0155] Step S702, calculating the sampling area characteristics of each N-1 order sampling area according to the N-1 order neighborhood division information;

[0156] Step S703, convolve the N-1 order initial neighborhood features and the sampling area features of each N-1 order sampling area to obtain the N-order neighborhood features, and perform farthest point sampling on the data points in the neighborhood to determine the number of N-order sampling points in the neighborhood and the coordinate information of the N-order sampling points to obtain the N-order neighborhood division information.

[0157] The above three steps are described in detail below.

[0158] In step S701, the N-1 order neighborhood feature and the N-1 order edge feature are aggregated to obtain the N-1 order initial neighborhood feature;

[0159] In step S702, the sampling area features of each N-1 order sampling area are calculated according to the N-1 order neighborhood division information, and new features of the neighborhood are obtained from another perspective.

[0160] Exemplarily, feature extraction is performed on the sampling region features of each N-1 order sampling region to obtain the sampling region features of each N-1 order sampling region.

[0161] In step S703, the N-1 order initial neighborhood features and the sampling area features of each N-1 order sampling area are convolved to obtain the N-order neighborhood features.

[0162] In another embodiment, the sampling region features of each N-1 order sampling region are aggregated or convolved to obtain the N-1 order sampling region features, and are aggregated or convolved with the N-1 order initial neighborhood features to obtain the N order neighborhood features.

[0163] At the same time, the farthest point sampling is performed on the data points in the neighborhood, the number of N-order sampling points in the neighborhood and the coordinate information of the N-order sampling points are determined, and the N-order neighborhood division information is obtained.

[0164] In the embodiment of the present application, more features in the neighborhood are further added to the N-order neighborhood features based on the neighborhood features and edge features obtained from the N-1th feature dimension-raising layer, and the sampling area features recalculated based on the N-1th order sampling points, so that the N-order neighborhood features are more perfect in representing the features of the neighborhood than the N-1th order features.

[0165] See also Figure 8 , Figure 8A flowchart of performing farthest point sampling on data points in a neighborhood, determining the number of N-order sampling points in the neighborhood and the coordinate information of the N-order sampling points, and obtaining N-order neighborhood partition information according to an embodiment of the present application is shown. The embodiment of the present application provides step S302 of performing farthest point sampling on data points in a neighborhood, determining the number of N-order sampling points in the neighborhood and the coordinate information of the N-order sampling points, and obtaining N-order neighborhood partition information, including:

[0166] Step S801, in the neighborhood of the query point, with the N-1 order sampling point as the center and the third set distance as the radius, an N order sampling area is formed, the third set distance has different values ​​in different feature dimension-raising layers, and the size of the third set distance is proportional to the size of N;

[0167] Step S802, the N-1th order sampling point farthest from the query point in the Nth order sampling area is the Nth order sampling point;

[0168] Step S803, in a new N-order sampling area formed with the latest N-order sampling point as the center and the third set distance as the radius, the N-1-order sampling point farthest from the latest N-order sampling point is used as a sampling point;

[0169] Step S804: Use the number and coordinate information of the N-th order sampling points as the neighborhood partitioning information of the N-th feature dimension enhancement layer.

[0170] The above four steps are described in detail below.

[0171] In step S801, in the neighborhood of the query point, an N-1 order sampling area is formed with the N-1 order sampling point as the center and the third set distance as the radius. The third set distance has different values ​​in different feature dimensionality layers, and the size of the third set distance is proportional to the size of N, that is, as N increases, the number of N-order sampling points will continue to decrease, so that new features of the neighborhood can be obtained from a new angle at each layer.

[0172] In step S802, the N-1th order sampling point farthest from the query point in the Nth order sampling area is the Nth order sampling point.

[0173] In step S803, in a new N-order sampling area formed with the latest N-order sampling point as the center and the third set distance as the radius, the N-1-order sampling point farthest from the latest N-order sampling point is used as a sampling point.

[0174] In step S804, the number and coordinate information of the N-th order sampling points are used as the neighborhood partitioning information of the N-th feature dimension raising layer.

[0175] See also Fig. 9 , Fig. 9A flowchart of calculating the feature value of a query point according to all dimensionality reduction features according to an embodiment of the present application is shown. The present application embodiment provides a step S150 of calculating the feature value of a query point according to all dimensionality reduction features, including:

[0176] Step S901, perform feature aggregation on all dimensionality reduction features and add residuals to obtain a target feature matrix;

[0177] Step S902: Perform pooling processing on the target feature matrix to obtain the eigenvalues ​​of the query points.

[0178] The above two steps are described in detail below.

[0179] In step S901, all dimension reduction features are aggregated and residuals are added to obtain a target feature matrix. The target feature matrix is ​​used to describe the point features of the query.

[0180] In the embodiment of the present application, all dimensionality reduction features are fully connected, and the dimensionality reduction features output by each feature dimensionality reduction layer represent the features of the query point in different dimensions. By extracting the dimensionality reduction features of each dimension, the features of the query point are more comprehensively represented.

[0181] In step S902, the target feature matrix is ​​pooled to obtain the eigenvalue of the query point. In some embodiments, the target feature matrix is ​​pooled into a 1x1 matrix, and the only value in the matrix is ​​the eigenvalue of the query point.

[0182] In some embodiments, the target feature matrix may be calculated as follows:

[0183]

[0184] Among them, Leaky Rectified Linear Unit (Leaky ReLU) is an activation function, which introduces the residual structure into the neural network and adds a direct shortcut connection between the input and output, so that information can be transmitted to the deep network more directly, preventing information from being lost during the transmission process, thereby improving the network's expressive power. In the embodiment of the present application, the input feature x (all dimensionality reduction features) is first convolved Conv 1 Operation and Batch Normalization B 1 ,y 1 =LeakReLU(B 1 (Conv 1 (x))), and y 1 After the second convolution Conv 2 and batch normalization B 2 Get y 2 , that is: y 2 =B2 (Conv 2 (y 1 )), replace y 2 After the residual connection with x, the Leaky ReLU activation function is applied again to get the final output y out (target feature matrix).

[0185] In some embodiments, each value in the target feature matrix is ​​substituted into a set formula to obtain the feature value of the query point.

[0186] In some embodiments, before using the neural network, the neural network is first trained, and the data points with labels are input into the neural network, and the parameters used in the convolution or feature extraction of each feature dimensionality increase layer and feature dimensionality reduction layer are adjusted according to the judgment results output by the neural network and the labels of the data points. The label indicates whether the data point belongs to a fracture surface or a non-fracture surface.

[0187] The error of the neural network is determined based on the output results of multiple data points in the neural network and the labels of multiple data points. If the error of the neural network is greater than the error threshold, the parameters used for convolution or feature extraction of each feature dimensionality increase layer and feature dimensionality reduction layer are adjusted according to the error of the neural network until the error of the neural network is less than or equal to the error threshold, and the training of the neural network is completed.

[0188] The error calculation of the neural network can be performed using the following formula:

[0189]

[0190] Where B is the batch size (the number of data points in the patch point cloud). p is the output of the pth sample (data point with a label). I is the identity matrix. ||…|| F represents the Frobenius norm, and the formula is Where A is an m×n matrix, a so is the element in the i-th row and j-th column of matrix A, that is, the square root of the sum of the squares of the matrix elements.

[0191] See also Fig.10 , Fig.10 A schematic diagram of a neural network according to an embodiment of the present application is shown. The present application embodiment provides a schematic diagram of a neural network. The neural network has three feature dimension-increasing layers and three feature dimension-reducing layers, as shown in the figure:

[0192] In the first feature dimension-increasing layer, the feature vectors of the query point and the neighboring points, as well as the coordinate information of the query point and the neighboring points are input into the second feature extraction part, and the feature vectors of the query point and the neighboring points are input into the first feature extraction part. Then the first feature extraction part outputs the first-order edge features, and the second feature extraction part outputs the first-order neighborhood features and the first-order neighborhood partition information.

[0193] In the second feature dimensionality upgrading layer, the first-order edge features and the first-order neighborhood features are combined and input into the second feature extraction part of the second feature dimensionality upgrading layer, and at the same time, the second feature extraction part of the second feature dimensionality upgrading layer of the first-order neighborhood division information outputs the second-order neighborhood features and the second-order neighborhood division information.

[0194] The first-order edge features are input to the first feature extraction part of the second feature dimension enhancement layer, and the second-order edge features are output.

[0195] In the third feature dimension upgrading layer, the second-order edge features and the second-order neighborhood features are combined and input into the second feature extraction part of the third feature dimension upgrading layer, and at the same time, the second-order neighborhood division information is input into the second feature extraction part of the third feature dimension upgrading layer, and the third-order neighborhood features and the third-order neighborhood division information are output.

[0196] The second-order edge features are input to the first feature extraction part of the third feature dimension enhancement layer, and the third-order edge features are output.

[0197] In the first feature dimension reduction layer, the third-order edge features and third-order neighborhood features are combined and input into the feature extraction part of the first feature dimension reduction layer, and the second-order neighborhood division information is input into the feature extraction part of the first feature dimension reduction layer as the target local feature, and feature extraction is performed together to obtain the first-order dimension reduction feature.

[0198] In the second feature dimension reduction layer, the first-order dimension reduction feature is input into the feature extraction part of the second feature dimension reduction layer, and the second-order neighborhood partition information and the first-order neighborhood partition information are input into the feature extraction part of the second feature dimension reduction layer as the target local feature, and feature extraction is performed together to obtain the second-order dimension reduction feature.

[0199] In the third feature dimension reduction layer, the second-order dimension reduction features are input into the feature extraction part of the third feature dimension reduction layer, and the first-order neighborhood partition information and the feature vectors of the query point and the neighboring points, as well as the coordinate information of the query point and the neighboring points are input into the feature extraction part of the third feature dimension reduction layer as the target local features, and feature extraction is performed together to obtain the third-order dimension reduction features.

[0200] Finally, the first-order dimensionality reduction features, second-order dimensionality reduction features, third-order dimensionality reduction features and residuals are fully connected to obtain the target feature matrix.

[0201] See also Fig.11 , Fig.11A fracture surface identification device based on a neural network according to an embodiment of the present application is shown, comprising:

[0202] A collection module is configured to collect point clouds of object fragments to obtain a point cloud of the fragments, where the point cloud of the fragments is composed of multiple data points;

[0203] The feature dimension upgrading module is configured to, in the first feature dimension upgrading layer, take any data point as a query point, take the query point as the center, set the distance as the radius to determine the neighborhood of the query point and the neighboring points of the query point distributed in the neighborhood;

[0204] The feature updating module is configured to extract features of the query point according to the coordinate information of the query point and the neighboring points, obtain the local features and edge features of the query point, and use them as inputs of the next feature dimension-enhancing layer to output new local features and edge features;

[0205] A feature dimension reduction module is configured to input the obtained local features and edge features into a first feature dimension reduction layer, obtain dimension reduction features and use them as inputs for a next feature dimension reduction layer;

[0206] A judgment module is configured to calculate a feature value of a query point according to the dimension reduction features output by each feature dimension reduction layer, and if the feature value of the query point is within a set range, it is determined that the query point belongs to a fracture surface;

[0207] The determination module is configured to obtain all data points belonging to the fracture surface in the fragment point cloud to determine the area belonging to the fracture surface in the fragment point cloud.

[0208] The fracture surface recognition based on the neural network in the embodiment of the present application can be performed by Fig.12 The hardware device shown is implemented. Fig.12 To describe the hardware device according to the embodiment of the present application. Fig.12 The hardware device shown is only an example and should not bring any limitation to the functions and scope of application of the embodiments of the present application.

[0209] like Fig.12 As shown, the hardware device may be in the form of a general computing device. The components of the hardware device may include, but are not limited to: at least one processing unit 810, at least one storage unit 820, and a bus 830 connecting different system components (including the storage unit 820 and the processing unit 810).

[0210] The storage unit stores program codes, which can be executed by the processing unit 810, so that the processing unit 810 performs the steps according to various exemplary embodiments of the present invention described in the description of the exemplary method described above in this specification. For example, the processing unit 810 can perform the following steps: Figure 1 The steps shown in .

[0211] The storage unit 820 may include a readable medium in the form of a volatile storage unit, such as a random access memory unit (RAM) 8201 and / or a cache memory unit 8202 , and may further include a read-only memory unit (ROM) 8203 .

[0212] The storage unit 820 may also include a program / utility 8204 having a set (at least one) of program modules 8205, such program modules 8205 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each of which or some combination may include an implementation of a network environment.

[0213] Bus 830 may represent one or more of several types of bus structures, including a memory unit bus or memory unit controller, a peripheral bus, an accelerated graphics port, a processing unit, or a local bus using any of a variety of bus architectures.

[0214] The hardware device may also communicate with one or more external devices 700 (e.g., keyboards, pointing devices, Bluetooth devices, etc.), may communicate with one or more devices that enable a user to interact with the hardware device, and / or may communicate with any device that enables the hardware device to communicate with one or more other computing devices (e.g., routers, modems, etc.). Such communication may be performed via an input / output (I / O) interface 850. Furthermore, the hardware device may also communicate with one or more networks (e.g., local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) via a network adapter 860. As shown, the network adapter 860 communicates with other modules via a bus 830. It should be understood that, although not shown in the figure, other hardware and / or software modules may be used in conjunction with the hardware device, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0215] Through the description of the above implementation methods, it is easy for those skilled in the art to understand that the example implementation methods described here can be implemented by software or by combining software with necessary hardware. Therefore, the technical solution according to the implementation methods of the present application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, and includes several instructions to enable a computing device (which can be a personal computer, a server, a terminal device, or a network device, etc.) to execute the method according to the implementation methods of the present application.

[0216] In one implementation of the present application, a storage medium for carrying the methods described in the various embodiments described above is also provided. The storage medium shown in the embodiment of the present application may be a readable signal medium or a readable storage medium or any combination of the above two. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM), a flash memory, an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present application, a readable storage medium may be any tangible medium containing or storing a program, which may be used by an instruction execution system, device or device or used in combination with it. In the present application, a readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, wherein a readable program code is carried. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. The readable signal medium may also be any storage medium other than a readable storage medium, which may send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device. The program code contained on the storage medium may be transmitted using any suitable medium, including but not limited to: wireless, wired, etc., or any suitable combination of the above.

[0217] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present application. In this regard, each box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the above-mentioned module, program segment or a part of a code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flow chart, and the combination of the boxes in the block diagram or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0218] It should be noted that, although several modules or units of the equipment for action execution are mentioned in the above detailed description, this division is not mandatory. In fact, according to the embodiments of the present application, the features and functions of two or more modules or units described above can be embodied in one module or unit. On the contrary, the features and functions of one module or unit described above can be further divided into being embodied by multiple modules or units.

[0219] Through the description of the above implementation methods, it is easy for those skilled in the art to understand that the example implementation methods described here can be implemented by software or by combining software with necessary hardware. Therefore, the technical solution according to the implementation methods of the present application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, and includes several instructions to enable a computing device (which can be a personal computer, a server, a touch terminal, or a network device, etc.) to execute the method according to the implementation methods of the present application.

[0220] Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the invention disclosed herein. The present application is intended to cover any modification, use or adaptation of the present application, which follows the general principles of the present application and includes common knowledge or customary technical means in the art that are not disclosed in the present application.

[0221] It should be understood that the present application is not limited to the precise structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present application is limited only by the appended claims.

Claims

1. A fracture surface recognition method based on neural network, characterized in that: The neural network comprises K layers of feature dimension-increasing layers and K layers of feature dimension-reducing layers, each of the feature dimension-increasing layers is connected in sequence, each of the feature dimension-reducing layers is connected to each other, the last feature dimension-increasing layer is connected to the first feature dimension-reducing layer, and K is a natural number; The method comprises: Performing point cloud collection on the object fragments to obtain a fragment point cloud, wherein the fragment point cloud is composed of a plurality of data points; In the first feature dimension-raising layer, any data point is used as a query point, and the neighborhood of the query point and the neighboring points of the query point distributed in the neighborhood are determined with the query point as the center and the distance as the radius; Extracting features of the query point according to the coordinate information of the query point and the neighboring points to obtain local features and edge features of the query point, and using them as inputs of the next feature dimension-enhancing layer to output new local features and edge features; The obtained local features and edge features are input into the first feature dimension reduction layer to obtain the dimension reduction features and serve as the input of the next feature dimension reduction layer; Calculating the characteristic value of the query point according to the dimension reduction features output by each of the feature dimension reduction layers, and if the characteristic value of the query point is within a set range, determining that the query point belongs to a fracture surface; All data points belonging to the fracture surface in the fragment point cloud are obtained to determine the area belonging to the fracture surface in the fragment point cloud.

2. The method according to claim 1, characterized in that The local features and edge features of the query point are calculated in the feature dimension-enhancing layer, including: In the first feature dimension-raising layer, according to a neighborhood of the query point formed by taking any data point as the query point and the first set distance as the radius, the first-order local feature of the query point is calculated, and the neighboring points of the query point are determined with the second set distance as the radius, and the vector difference between the feature vector of the query point and the feature vectors of each of the neighboring points is calculated as the first-order edge feature; In the Nth feature dimensionality increase layer, the N-1 order neighborhood features and the N-1 order edge features are combined and dimensionality increased convolution is performed to obtain N-order local features, and the vector differences in the N-1 order edge features are subtracted pairwise to obtain N-order edge features, where N belongs to K and N is greater than 1.

3. The method according to claim 2, characterized in that Calculate the dimension reduction features in the feature dimension reduction layer, including: In the first feature dimension reduction layer, global feature extraction is performed on the K-order edge features, the K-order local features, and the first-order target local features to obtain first-order dimension reduction features, and the feature dimension increase layer where the first-order target local features are located and the K-th feature dimension increase layer are sequentially connected; In the Mth feature dimensionality reduction layer, global feature extraction is performed on the M-1th order dimensionality reduction features and the Mth order target local features to obtain the Mth order dimensionality reduction features. The feature dimensionality increase layer where the Mth order target local features and the M-1th order target local features are located has a sequential connection relationship, and M belongs to K, and M is greater than 1.

4. The method according to claim 2, characterized in that: The local features include neighborhood features and neighborhood partition information. In the first feature dimension-raising layer, the first-order local features of the query point are calculated based on the neighborhood of the query point formed by taking any data point as the query point and the first set distance as the radius, including: Taking any data point in the fragment point cloud as a query point; Taking the query point as the center and a first set distance as the radius, determine a neighborhood of the query point, wherein the data points in the neighborhood are first neighboring points; The coordinate information of the query point and the first neighboring point, and the feature vectors of the query point and the first neighboring point are convolved to obtain the first-order neighborhood features of the neighborhood, and the farthest point sampling is performed on the data points in the neighborhood to determine the number of sampling points in the neighborhood and the sampled coordinate information to obtain the first-order neighborhood division information.

5. The method according to claim 4, characterized in that The root performs convolution on the coordinate information of the query point and the first neighboring point, and the feature vectors of the query point and the first neighboring point, including: The first set distance includes a plurality of radius values, and a sub-neighborhood is determined according to each of the radius values, and the largest sub-neighborhood is the neighborhood of the query point; Calculating sub-neighborhood features of the sub-neighborhoods respectively according to the coordinate information of the query point and the data points in each sub-neighborhood; Aggregate the sub-neighborhood features to obtain the neighborhood feature of the query point.

6. The method according to claim 4, characterized in that Performing farthest point sampling on the data points in the neighborhood includes: In the neighborhood of the query point, a first-order sampling area is formed with the query point as the center and a third set distance as the radius, where the third set distance is smaller than the first set distance; The data point farthest from the query point in the first-order sampling area is the first-order sampling point; In a new first-order sampling area formed with the latest first-order sampling point as the center and the third set distance as the radius, the data point farthest from the latest sampling point is used as the first-order sampling point until no more first-order sampling points can be obtained.

7. The method according to claim 2, characterized in that: In the Nth feature dimension-raising layer, the N-1th order neighborhood features and the N-1th order edge features are combined and convolved to obtain Nth order local features, including: Aggregating the N-1 order neighborhood feature and the N-1 order edge feature to obtain an N-order initial neighborhood feature; According to the N-1 order neighborhood division information, the sampling area characteristics of each N-1 order sampling area are calculated; The N-order initial neighborhood features and the sampling area features of each of the N-1-order sampling areas are convolved to obtain the N-order neighborhood features, and the data points in the neighborhood are sampled at the farthest point to determine the number of N-order sampling points in the neighborhood and the coordinate information of the N-order sampling points to obtain the N-order neighborhood division information.

8. The method according to claim 3, characterized in that Performing farthest point sampling on the data points in the neighborhood, determining the number of N-order sampling points in the neighborhood and the coordinate information of the N-order sampling points, and obtaining N-order neighborhood partitioning information, including: In the neighborhood of the query point, an N-1-order sampling area is formed with the N-1-order sampling point as the center and a third set distance as the radius, wherein the third set distance has different values ​​in different feature dimension-raising layers, and the size of the third set distance is proportional to the size of N; The N-1th order sampling point farthest from the query point in the Nth order sampling area is the Nth order sampling point; In a new N-order sampling area formed with the latest N-order sampling point as the center and the third set distance as the radius, the N-1-order sampling point farthest from the latest N-order sampling point is used as a sampling point; The number and coordinate information of the N-th order sampling points are used as neighborhood partition information of the N-th feature dimension enhancement layer.

9. The method according to claim 1, characterized in that: According to all the dimension reduction features, the feature value of the query point is calculated, including: Perform feature aggregation on all dimensionality reduction features and add residuals to obtain the target feature matrix; The target feature matrix is ​​pooled to obtain the feature value of the query point.

10. A fracture surface recognition device based on a neural network, characterized in that: The neural network comprises K layers of feature dimension-increasing layers and K layers of feature dimension-reducing layers, each of the feature dimension-increasing layers is connected in sequence, each of the feature dimension-reducing layers is connected to each other, the last feature dimension-increasing layer is connected to the first feature dimension-reducing layer, and K is a natural number; The device comprises: A collection module is configured to collect point clouds of object fragments to obtain a fragment point cloud, wherein the fragment point cloud is composed of a plurality of data points; The feature dimension-raising module is configured to, in the first feature dimension-raising layer, take any data point as a query point, take the query point as the center, set a distance as a radius to determine the neighborhood of the query point and the neighboring points of the query point distributed in the neighborhood; A feature updating module is configured to extract features of the query point according to the coordinate information of the query point and the neighboring points, obtain local features and edge features of the query point, and use them as inputs of the next feature dimension-increasing layer to output new local features and edge features; A feature dimension reduction module is configured to input the obtained local features and edge features into a first feature dimension reduction layer, obtain dimension reduction features and use them as inputs for a next feature dimension reduction layer; A judgment module is configured to calculate a feature value of the query point according to the dimension reduction features output by each feature dimension reduction layer, and if the feature value of the query point is within a set range, determine that the query point belongs to a fracture surface; The determination module is configured to obtain all data points belonging to the fracture surface in the fragment point cloud to determine the area belonging to the fracture surface in the fragment point cloud.

Images