Deep learning-based rock stratum surface three-dimensional feature extraction and classification method and system
By performing point cloud rasterization processing on lidar data and using CNN and improved CapsRetNet networks, the problem of traditional convolutional neural networks neglecting geometric relationships when processing three-dimensional point cloud data is solved, and three-dimensional feature extraction and classification with high accuracy and consistency are achieved.
Patent Information
- Application Number
- CN202510231640.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-28
- Publication Date
- 2025-06-17
AI Technical Summary
Traditional convolutional neural networks ignore the geometric relationship between points when processing three-dimensional point cloud data, resulting in low accuracy of rock formation feature classification, and automatic classification of three-dimensional point clouds is a very challenging problem.
The three-dimensional feature extraction and classification method of rock layer surface based on deep learning is used to perform point cloud rasterization processing on lidar data, and a feature extraction and classification model is constructed using convolutional neural network (CNN) and improved CapsRetNet network.
It achieves a high accuracy and classification consistency in three-dimensional feature extraction, and improves the recognition ability and classification accuracy of rock layer surface features.
Smart Images
Figure CN120164022A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of feature extraction and classification, and relates to a three-dimensional feature extraction and classification method and system for rock formation surfaces based on deep learning. Background Art
[0002] Using a high-frequency breaker instead of the cutter head of a shield machine for hard rock formation crushing is a new method to improve the efficiency of a shield machine. Installing multiple crushing arms on the fixed frame of the shield cutter head, with a high-frequency crusher installed at the front end of each crushing arm, and the collaborative work of multiple crushing arms on the hard rock formation surface is a method to further improve the efficiency of the shield machine. This collaborative crushing algorithm needs to be based on a mathematical model with clear rock formation characteristics.
[0003] Rock formation feature data is generally obtained through three-dimensional point cloud data scanned by a three-dimensional sensor. Deep learning training can detect objects and obtain classification categories of different rock formations. However, due to the characteristics of large data volume, irregular shape, and uneven density of the point cloud, the automatic classification of three-dimensional point clouds has always been a very challenging problem. And traditional convolutional neural networks ignore the geometric relationship between points, so they cannot recognize changes in the feature space, and the classification accuracy still needs to be improved. Summary of the Invention
[0004] The purpose of the present invention is to provide a three-dimensional feature extraction and classification method and system for rock formation surfaces based on deep learning, which performs point cloud rasterization processing on the data of a lidar (Li-DAR), and then uses a convolutional neural network (CNN) and an improved CapsRetNet network to construct a three-dimensional feature extraction and classification method for rock formation surfaces, with high accuracy and high classification consistency.
[0005] The technical solution to achieve the purpose of the present invention is as follows:
[0006] A three-dimensional feature extraction and classification method for rock formation surfaces based on deep learning, comprising the following steps:
[0007] S01: Obtain rock formation point cloud data and perform point cloud filtering;
[0008] S02: Perform point cloud raster conversion on the remaining data points after calculation, project the point cloud onto a plane, and divide it into pixels according to the resolution to obtain a point cloud raster image;
[0009] S03: Extract the rock formation surface features from the point cloud raster image through a feature extraction network, and obtain a three-dimensional feature extraction and classification result through an enhanced classification network for the extracted rock formation surface features.
[0010] In a preferred technical solution, the point cloud filtering method in step S01 includes:
[0011] Calculate each point P in the point cloud data i The k-nearest neighbors of are:
[0012] P i ={p1, p2,..., p k}
[0013] where i = 1, 2,..., N, N is the number of P in the point cloud data i p ik The k-th nearest neighbor of point P i ;
[0014] The pixel distance between each k-nearest neighbor and its corresponding P i is D i :
[0015] D i ={d i1 , d i2 ,..., d ik}
[0016] d ik The pixel distance between the k-th nearest neighbor and P i , D i The variable follows a Gaussian distribution, and the one-dimensional Gaussian function is:
[0017]
[0018] where f(x) is the probability density function of the Gaussian distribution, e is the natural constant, π is the pi, μ is the mean of D i , and σ is the standard deviation of the D i variable.
[0019] Take the pixel distance D i between each nearest neighbor and the corresponding category point as a one-dimensional variable and substitute it into the one-dimensional Gaussian function to obtain the probability density function f i (x) of each nearest neighbor, and let this value be q i , and perform normalization on it:
[0020]
[0021] where Q i is the value of the probability density function after normalizing q i , simply referred to as the weight;
[0022] Finally, sum the weighted nearest neighbors to get the updated point as:
[0023]
[0024] where P j is the data after weighted update of the nearest neighbors, pj is an unweighted neighbor point, Q j is the weight value of the neighbor point after normalization.
[0025] In the preferred technical solution, in step S01, after Gaussian filtering, SOR statistical filtering is performed. The SOR statistical filtering method includes:
[0026] Construct a k-d tree. Each node of the tree is used as a k-dimensional coordinate point, and each node of the tree is a hyperplane. Calculate the average distance d of the updated point P j to the hyperplane j , and under the condition of meeting traversability, determine whether the data is retained according to the following formula:
[0027] μ - εd ≤ d j ≤ μ + εd
[0028] where μ is the average value of the distances of N data points P j to the corresponding hyperplane, d j is the distance of the current data point P j to the hyperplane corresponding to this data point, d is the standard deviation of the distance, and ε is the set data rejection threshold;
[0029] If it is satisfied, the data point is retained; otherwise, it is determined as a noise point.
[0030] In the preferred technical solution, in step S02, after point cloud raster conversion, interpolation processing is performed. The interpolation formula is:
[0031]
[0032] where x0 is the missing value, n is the number of grid points around x0, is the estimated value for filling the gap, is the data point around the missing value, and λ i is the weight coefficient.
[0033] In the preferred technical solution, in step S03, the feature extraction network includes a ResNet network. The constructed ResNet network residual block is:
[0034] X a = H(X a-1 ) + F(X a-1 + W a-1 )
[0035] where X a is the output of the network, H(X a-1 ) is a convolution operation, F(X a-1 + W a-1 ) is a residual operation, and a represents the number of convolution residual blocks in the ResNet network;
[0036] H(X a-1 ) = W a ′ -1 X a-1
[0037] The convolutional activation function selects ReLU, and W a ′ -1 is the convolutional operation of the ResNet network, and X a-1 is the input of the convolutional operation.
[0038] In the preferred technical solution, the method for obtaining the three-dimensional feature extraction and classification result by passing the extracted rock formation surface features through the enhanced classification network in step S03 includes:
[0039] The enhanced classification network is an improved CapsNet network, which is composed of N similar capsule modules in parallel. The input of the CapsNet network is:
[0040]
[0041] Among them, is the input of the CapsNet network, b is the number of capsule modules, and a is the number of variables in each capsule module; is determined by the output X a of the ResNet network, that is: u a = X a ; w ab is the input weight coefficient of the CapsNet network;
[0042] Weighted superposition is performed on the input of each capsule module:
[0043]
[0044] Among them, is the a-th input of the b-th capsule module, and C ba is the weight value; the number I of the summation is the number of the input in the current capsule module, that is: I = 1, 2,..., i;
[0045] The Squashing function is taken as:
[0046]
[0047] v j is the three-dimensional feature.
[0048] The present invention also discloses a three-dimensional feature extraction and classification system for the rock formation surface based on deep learning, including:
[0049] Point cloud filtering module, which acquires rock formation point cloud data and performs point cloud filtering;
[0050] Point cloud raster conversion module, which performs point cloud raster conversion on the remaining data points after calculation, projects the point cloud onto a plane, and divides it into pixels according to the resolution to obtain a point cloud raster image;
[0051] 3D feature extraction and classification module, which extracts the surface features of the rock formation from the point cloud raster image through a feature extraction network, and obtains the 3D feature extraction and classification result by passing the extracted surface features of the rock formation through an enhanced classification network.
[0052] In the preferred technical solution, the feature extraction network in the 3D feature extraction and classification module includes a ResNet network, and the constructed ResNet network residual block is:
[0053] X a = H(X a-1 ) + F(X a-1 + W a-1 )(8)
[0054] Among them, X a is the output of the network, H(X a-1 ) is a convolution operation, F(X a-1 + W a-1 ) is a residual operation, and a represents the number of convolutional residual blocks in the ResNet network;
[0055] H(X a-1 ) = W a ′ -1 X a-1
[0056] The convolution activation function selects ReLU, W a ′ -1 is the convolution operation of the ResNet network, and X a-1 is the input of the convolution operation.
[0057] In the preferred technical solution, the method for obtaining the 3D feature extraction and classification result by passing the extracted surface features of the rock formation through the enhanced classification network in the 3D feature extraction and classification module includes:
[0058] The enhanced classification network is an improved CapsNet network, which is composed of N similar capsule modules in parallel. The input of the CapsNet network is:
[0059]
[0060] Among them, is the input of the CapsNet network, b is the number of capsule modules, and a is the number of variables in each capsule module; From the output X of the ResNet networka Determination, i.e.: u a = X a ; w ab is the input weight coefficient of the CapsNet network;
[0061] Weightedly superimpose the inputs of each capsule module as follows:
[0062]
[0063] Among them, is the a-th input of the b-th capsule module, and C ba is the weight value of; the number I of the summation is the number of the inputs inside the current capsule module, i.e.: I = 1, 2,..., i;
[0064] Take the Squashing function as:
[0065]
[0066] v j is a three-dimensional feature.
[0067] The present invention also discloses a computer storage medium, on which a computer program is stored, and when the computer program is executed, the above-mentioned three-dimensional feature extraction and classification method for the rock formation surface based on deep learning is implemented.
[0068] Compared with the prior art, the present invention has the following remarkable advantages:
[0069] Perform point cloud rasterization processing on the data of the lidar (Li-DAR), and then use a convolutional neural network (CNN) and an improved CapsRetNet network to construct a three-dimensional feature extraction and classification method for the rock formation surface, which has a high accuracy rate and a high classification consistency. Description of the Drawings
[0070] Figure 1 is a flowchart of the three-dimensional feature extraction and classification method for the rock formation surface based on deep learning in a preferred embodiment;
[0071] Figure 2 is a schematic block diagram of the principle of the three-dimensional feature extraction and classification system for the rock formation surface based on deep learning in a preferred embodiment;
[0072] Figure 3 is a schematic diagram of the principle of point cloud rasterization processing in a preferred embodiment;
[0073] Figure 4 is the test result of the overall accuracy rate for comparison with other algorithms;
[0074] Figure 5 The test results of the average accuracy rate for comparison with other algorithms;
[0075] Figure 6 The comparison results of the KAPPA normalization values. Detailed implementation manners
[0076] The principle of the present invention is: the present invention performs point cloud rasterization processing on the data of a lidar (Li-DAR), and then uses a convolutional neural network (CNN) and an improved CapsRetNet network to construct a three-dimensional feature extraction and classification method for the rock formation surface.
[0077] Embodiment 1:
[0078] As Figure 1 shown, a three-dimensional feature extraction and classification method for the rock formation surface based on deep learning includes the following steps:
[0079] S01: Obtain the point cloud data of the rock formation and perform point cloud filtering;
[0080] S02: Perform point cloud raster conversion on the retained data points after calculation, project the point cloud on a plane, and divide it into pixels according to the resolution to obtain a point cloud raster image;
[0081] S03: Extract the surface features of the rock formation from the point cloud raster image through a feature extraction network, and obtain a three-dimensional feature extraction and classification result through an enhanced classification network for the extracted surface features of the rock formation.
[0082] In one embodiment, the point cloud filtering method in step S01 includes:
[0083] Calculate the k nearest neighbors of each point P i in the point cloud data as:
[0084] P i ={p1, p2,..., p k}
[0085] where i = 1, 2,..., N, and N is the number of P i in the point cloud data, and p ik is the kth nearest neighbor of point P i ;
[0086] The pixel distance between each k nearest neighbor and its corresponding P i is D i :
[0087] D i ={d i1 , d i2 ,..., d ik}
[0088] dik The pixel distance between the k-th nearest neighbor point and P i , D i The variable follows a Gaussian distribution, and the one-dimensional Gaussian function is:
[0089]
[0090] where f(x) is the probability density function of the Gaussian distribution, e is the natural constant, π is the circumference ratio, and μ is the mean of D i , and σ is the standard deviation of the D i variable.
[0091] Take the pixel distance D between each nearest neighbor point and the corresponding category point i as a one-dimensional variable and substitute it into the one-dimensional Gaussian function to obtain the probability density function f i (x) of each nearest neighbor point, and let this value be q i , and perform normalization on it:
[0092]
[0093] where Q i is the value of the probability density function after normalizing q i , simply referred to as the weight;
[0094] Finally, the weighted sum of the nearest neighbor points is obtained, and the updated point is:
[0095]
[0096] where P j is the data after weighted update of the nearest neighbor points, p j is the unweighted nearest neighbor point, and Q j is the normalized weight of the nearest neighbor point.
[0097] In one embodiment, after Gaussian filtering in step S01, SOR statistical filtering is performed. The SOR statistical filtering method includes:
[0098] Construct a k-d tree. Each node of the tree is used as a k-dimensional coordinate point, and each node of the tree is a hyperplane. Calculate the average distance d j from the updated point P j to the hyperplane. Under the condition of satisfying traversability, determine whether the data is retained according to the following formula:
[0099] μ - εd ≤ d j ≤ μ + εd
[0100] where μ is the average value of the distances from N data points P j to the corresponding hyperplane, and d j is the distance from the current data point P jThe distance to the hyperplane corresponding to the data point, d is the standard deviation of the distance, and ε is the set data rejection threshold;
[0101] If it is satisfied, the data point is retained; otherwise, it is determined as a noise point.
[0102] In the preferred technical solution, after the point cloud raster conversion in step S02, interpolation processing is performed, and the interpolation formula is:
[0103]
[0104] where x0 is the missing value, n is the number of grid points around x0, is the estimated value for filling the gap, is the data point around the missing value, and λ i is the weight coefficient.
[0105] In one embodiment, the feature extraction network in step S03 includes a ResNet network, and the constructed ResNet network residual block is:
[0106] X a = H(X a-1 ) + F(X a-1 + W a-1 )
[0107] where X a is the output of the network, H(X a-1 ) is the convolution operation, F(X a-1 + W a-1 ) is the residual operation, and a represents the number of convolutional residual blocks in the ResNet network;
[0108] H(X a-1 ) = W a ′ -1 X a-1
[0109] The convolution activation function selects ReLU, W a ′ -1 is the convolution operation of the ResNet network, and X a-1 is the input of the convolution operation.
[0110] In one embodiment, the method for obtaining the three-dimensional feature extraction classification result by passing the extracted rock formation surface feature through the enhanced classification network in step S03 includes:
[0111] The enhanced classification network is an improved CapsNet network, which is composed of N similar capsule modules in parallel. The input of the CapsNet network is:
[0112]
[0113] where, is the input of the CapsNet network, b is the number of capsule modules, and a is the number of variables in each capsule module; It is determined by the output X of the ResNet network, that is: u a = a = X a ; w ab is the input weight coefficient of the CapsNet network;
[0114] Weighted superposition is performed on the input of each capsule module :
[0115]
[0116] where is the a-th input of the b-th capsule module, and C ba is 's weight value; the number I of summation is the number of inputs in the current capsule module, that is: I = 1, 2,..., i;
[0117] Take the Squashing function as:
[0118]
[0119] v j is a three-dimensional feature.
[0120] In another embodiment, a computer storage medium stores a computer program, and when the computer program is executed, the above-mentioned three-dimensional feature extraction and classification method for the rock formation surface based on deep learning is implemented.
[0121] The three-dimensional feature extraction and classification method for the rock formation surface based on deep learning can adopt any of the above-mentioned three-dimensional feature extraction and classification methods for the rock formation surface based on deep learning, and the specific implementation will not be elaborated here.
[0122] Another embodiment, as Figure 2 shown, a three-dimensional feature extraction and classification system for the rock formation surface based on deep learning includes:
[0123] A point cloud filtering module 10, which acquires rock formation point cloud data and performs point cloud filtering;
[0124] A point cloud raster conversion module 20, which performs point cloud raster conversion on the remaining data points after calculation, projects the point cloud on a plane, and divides it into pixels according to the resolution to obtain a point cloud raster image;
[0125] The three-dimensional feature extraction and classification module 30 extracts the surface features of the rock formation from the point cloud raster image through the feature extraction network, and obtains the three-dimensional feature extraction and classification result by passing the extracted surface features of the rock formation through the enhanced classification network.
[0126] Specifically, taking a preferred embodiment as an example, the working process of the three-dimensional feature extraction and classification system for the rock formation surface based on deep learning is described as follows:
[0127] The point cloud rasterization process of Li-DAR data is as Figure 3 shown.
[0128] First, perform point cloud filtering on the original data of Li-DAR, and select Gaussian filtering combined with SOR statistical filtering during the filtering process.
[0129] Let the initial point cloud data be S r , and the k-nearest neighbors (KNN) of each point P i are:
[0130] P i = {p1, p2,..., p k}, i = 1, 2,..., N(1)
[0131] where N is the number of this data set:
[0132] The pixel distance between each neighbor point and P i is:
[0133] D i = {d1, d2,..., d k}, i = 1, 2,..., N(2)
[0134] The D i variable satisfies the Gaussian distribution, and the one-dimensional Gaussian function is:
[0135]
[0136] where f(x) is the probability density function of the Gaussian distribution, e is the natural constant, π is the pi, μ is the mean of D i , and σ is the standard deviation of the D i variable.
[0137] Take the pixel distance D i between each neighbor point and the corresponding category point as a one-dimensional variable and substitute it into formula (3) to obtain the probability density function f i (x) of each neighbor point, let this value be q i , and perform normalization processing on it:
[0138]
[0139] where Qi is q i The value of the normalized probability density function, which is simply referred to as the weight here.
[0140] Finally, the weighted sum of the nearest neighbor points is calculated to obtain the updated point as:
[0141]
[0142] where, P j is the data after weighted update of the nearest neighbor points, p j is the unweighted nearest neighbor point described in formula (1), Q j is the normalized weight of the nearest neighbor points described in formula (4).
[0143] After pulling the outliers back to the dense positions through the above Gaussian filtering, statistical filtering is performed using SOR: construct a k-d tree, each node of the tree is used as a k-dimensional coordinate point, each node of the tree is a hyperplane, and calculate the average distance d of P j from the hyperplane according to the data given by formula (5). j Under the condition of satisfying traversability, determine whether the data is retained according to formula (6): j μ - εd ≤ d
[0144] ≤ μ + εd(6) j where, μ is the average value of the distances of N data points P
[0145] to the corresponding hyperplanes, d j is the distance from the current data point P j to the hyperplane corresponding to this data point, d is the standard deviation of the distances, and ε is the set data rejection threshold; if the calculated distance of the hyperplane point satisfies formula (6), then this data point is retained, otherwise this point is determined as a noise point. j
[0146] Perform point cloud raster conversion on the calculated retained data points, project the point cloud onto a plane, divide it into pixels according to the required resolution, and the value of each pixel is the maximum elevation of the point cloud in this plane area.
[0147] Interpolate the possible holes or missing values in the point cloud raster image. Here, Kriging interpolation is used. Let the missing value be x0, and there are n grid points around x0, then the interpolation formula is as shown in (7):
[0148]
[0149] where, is the estimated value that can fill the gap, is the data points around the missing value, and λ i is the weight coefficient.
[0150] The interpolated point cloud raster data is fed into the ResNet network to extract the surface features of the rock formation. ResNet34 structure is selected for ResNet. The residual block of the constructed ResNet network is shown in Equation (8):
[0151] X a =H(X a-1 )+F(X a-1 +W a-1 )(8)
[0152] where X a is the output of the network, H(X a-1 ) is the convolution operation, F(X a-1 +W a-1 ) is the residual operation, and a represents the number of convolutional residual blocks in the ResNet network;
[0153] H(X a-1 )=W a ′ -1 X a-1 (9)
[0154] The convolution activation function selects ReLU, W a ′ -1 is the convolution operation of the ResNet network, and X a-1 is the input of the convolution operation.
[0155] Taking X a as the input, it is fed into the enhanced classification network CapsNet. The CapsNet network is composed of N similar modules in parallel, and each module is called a capsule. The CapsNet network can better identify the changes in the feature space. The input of CapsNet is determined by Equation (10).
[0156]
[0157] where, is the input of the CapsNet network, b is the number of capsule modules, and a is the number of variables in each capsule module; is determined by the output X a of the ResNet network, that is: u a =X a ; w ab is the input weight coefficient of the CapsNet network;
[0158] The inputs of each capsule module are weighted and superimposed:
[0159]
[0160] Among them, is the a-th input of the b-th capsule module, and C ba is 's weight value; the number I of the summation is the number of inputs in the current capsule module, that is: I = 1, 2,..., i;
[0161] Take the Squashing function as:
[0162]
[0163] v j is a three-dimensional feature.
[0164] In order to verify the effectiveness of the method of the present invention, the feature extraction experiment was carried out under MATLAB2022b. The data training was set for 100 rounds, and the training was stopped in advance when the feature classification accuracy no longer improved. The gradient optimization algorithm selected stochastic gradient descent (SGD) and the momentum algorithm, with the momentum selected as 0.9 and the descent rate as 0.00001. As a comparison, a classification decision tree (the maximum depth was set to 100), a support vector machine (SVM, the kernel function selected the radial basis function), and a random forest were selected for comparison. Randomly selected from the dataset: 25, 50, 75, 100, 125, 150, 175, 200, 225, and 250 samples as the training set, and a fixed number of test sets were randomly selected from the remaining part. The test results of the overall accuracy are as Figure 4 shown. The abscissa is the number of samples, and the ordinate is the overall accuracy (OA), which is the quotient of the number of correctly tested samples and the number of tested samples. It can be seen that: the number of training samples is positively correlated with the overall accuracy. As the number of samples increases, the random forest and the method adopted by the present invention perform better than SVM and the classification tree, and the overall test accuracy improves significantly. Among them, the overall test accuracy of the method of the present invention increases from 87.21% to 96%. From the perspective of the overall test accuracy, the method provided by the present invention is better.
[0165] The test results of the average accuracy are as Figure 5 shown. The abscissa is the number of samples, and the ordinate is the average accuracy (AA), which is the average value of the number of correctly tested samples in multiple rounds of testing. It can be seen that: the average accuracy shown by the classification tree, SVM, random forest, and the method of the present invention is positively correlated with the training samples. The random forest and the method adopted by the present invention perform better than SVM and the classification tree, and the average test accuracy improves significantly. Among them, the overall test accuracy of the method of the present invention increases from 87.01% to 97%. From the perspective of the average accuracy, the method provided by the present invention is better.
[0166] The comparison of the KAPPA normalization values is as Figure 6As shown in the figure, the abscissa is the number of samples and the ordinate is the KAPPA normalization value. It can be seen that when the number of training samples is lower than a certain value, the classification consistency of the classification tree and SVM is relatively low, while the classification consistency of the random forest and the method of the present invention is relatively high. When the number of test training samples exceeds 100, the consistency of the method provided by the present invention reaches more than 0.8. It can be seen that the advantages of the inventive method are obvious.
[0167] The above embodiments are the preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principle of the present invention shall be equivalent replacement methods and are all included in the protection scope of the present invention.
Claims
1. A deep learning-based three-dimensional feature extraction and classification method for rock surface, characterized in that: The following steps are involved: S01: Obtain rock formation point cloud data and perform point cloud filtering; S02: convert the data points retained after calculation into point cloud raster, project the point cloud onto a plane, divide it into pixels according to the resolution, and obtain a point cloud raster image; S03: Extract rock surface features from the point cloud raster image through a feature extraction network, and obtain three-dimensional feature extraction and classification results through an enhanced classification network.
2. The rock surface three-dimensional feature extraction and classification method based on deep learning according to claim 1 is characterized in that: The point cloud filtering method in step S01 includes: Calculate each point P in the point cloud data i The k nearest neighbors of are: P i ={p i1 ,p i2 ,...,p ik } Among them, i = 1, 2, ..., N, N is the number of point cloud data points. i The number of ik Point P i The kth nearest neighbor of ; Each k nearest neighbor point and its corresponding P i The pixel distance is D i : D i ={d i1 ,d i2 ,...,d ik } d ik The kth nearest neighbor point and P i The pixel distance, D i The variable satisfies the Gaussian distribution, and the one-dimensional Gaussian function is: Among them, f(x) is the probability density function of Gaussian distribution, e is a natural constant, π is the circumference of a circle, and μ is D i The mean of i The standard deviation of the variable; The pixel distance D between each neighbor point and the corresponding category point i Substitute it into the one-dimensional Gaussian function as a one-dimensional variable to obtain the probability density function f of each neighboring point i (x), let the value be q i , and normalize it: Among them, Q i for q i The normalized probability density function value is referred to as the weight; Finally, the updated point is obtained by weighted summation of the neighboring points: Among them, P j is the weighted updated data of the neighboring points, p j is the unweighted neighbor point, Q j is the normalized weight of the neighboring points.
3. The rock surface three-dimensional feature extraction and classification method based on deep learning according to claim 2 is characterized in that: In step S01, after Gaussian filtering, SOR statistical filtering is performed. The SOR statistical filtering method includes: Construct a kd tree, with each node of the tree as a k-dimensional coordinate point, and each node of the tree as a hyperplane, and calculate the updated point P j The average distance to the hyperplane d j , under the condition of satisfying ergodicity, determine whether the data is retained according to the following formula: μ-εd≤d j ≤μ+εd Where μ is the number of N data points P j The average distance to the corresponding hyperplane, d j is the current data point P j The distance to the hyperplane corresponding to the data point, d is the distance standard deviation, and ε is the set data removal threshold; If the condition is met, the data point is retained, otherwise it is considered as a noise point.
4. The rock surface three-dimensional feature extraction and classification method based on deep learning according to claim 1 is characterized in that: In step S02, the point cloud raster is converted and then interpolated. The interpolation formula is: Where x0 is the missing value, n is the number of grid points around x0, To fill in the gaps in the estimate, are the data points around the missing value, λ i is the weight coefficient.
5. The rock surface three-dimensional feature extraction and classification method based on deep learning according to claim 1 is characterized in that: In step S03, the feature extraction network includes a ResNet network, and the constructed ResNet network residual block is: X a =H(X a-1 )+F(X a-1 +W a-1 ) Among them, X a is the output of the network, H(X a-1 ) is the convolution operation, F(X a-1 +W a-1 ) is the residual operation, and a represents the number of convolutional residual blocks in the ResNet network; H(X a-1 )=W a ′ -1 X a-1 The convolution activation function selects ReLU, W a ′ -1 is the convolution operation of the ResNet network, X a-1 is the input of the convolution operation.
6. The rock surface three-dimensional feature extraction and classification method based on deep learning according to claim 5 is characterized in that: In step S03, the method of obtaining a three-dimensional feature extraction and classification result by using an enhanced classification network to extract the rock layer surface features includes: The enhanced classification network is an improved CapsNet network, which is composed of N similar capsule modules in parallel. The input of the CapsNet network is: in, is the input of the CapsNet network, b is the number of capsule modules, and a is the number of variables in each capsule module; The output of the ResNet network is X a Decision, that is: a =X a ;w ab is the input weight coefficient of the CapsNet network; The input of each capsule module To perform weighted overlay: in, is the ath input of the bth capsule module, C ba for The weight of the sum is I, which is the input of the current capsule module. the number of Take the Squashing function as: v j It is a three-dimensional feature.
7. A deep learning-based three-dimensional feature extraction and classification system for rock surface, characterized in that: include: Point cloud filtering module, which obtains rock formation point cloud data and performs point cloud filtering; The point cloud raster conversion module converts the data points retained after calculation into point cloud raster, projects the point cloud onto a plane, divides it into pixels according to the resolution, and obtains a point cloud raster image; The three-dimensional feature extraction and classification module extracts the rock surface features from the point cloud raster image through the feature extraction network, and obtains the three-dimensional feature extraction and classification results by passing the extracted rock surface features through the enhanced classification network.
8. The deep learning-based rock surface three-dimensional feature extraction and classification system according to claim 7, characterized in that: The feature extraction network in the three-dimensional feature extraction and classification module includes a ResNet network, and the constructed ResNet network residual block is: X a =H(X a-1 )+F(X a-1 +W a-1 ) Among them, X a is the output of the network, H(X a-1 ) is the convolution operation, F(X a-1 +W a-1 ) is the residual operation, and a represents the number of convolutional residual blocks in the ResNet network; H(X a-1 )=W a ′ -1 X a-1 The convolution activation function selects ReLU, W a ′ -1 is the convolution operation of the ResNet network, X a-1 is the input of the convolution operation.
9. The deep learning-based rock surface three-dimensional feature extraction and classification system according to claim 8, characterized in that: The method of obtaining the three-dimensional feature extraction and classification result by using the extracted rock layer surface features through the enhanced classification network in the three-dimensional feature extraction and classification module includes: The enhanced classification network is an improved CapsNet network, which is composed of N similar capsule modules in parallel. The input of the CapsNet network is: in, is the input of the CapsNet network, b is the number of capsule modules, and a is the number of variables in each capsule module; The output of the ResNet network is X a Decision, that is: a =X a ;w ab is the input weight coefficient of the CapsNet network; The input of each capsule module To perform weighted overlay: in, is the ath input of the bth capsule module, C ba for The weight of the sum is I, which is the input of the current capsule module. the number of Take the Squashing function as: v j It is a three-dimensional feature.
10. A computer storage medium having a computer program stored thereon, characterized in that: When the computer program is executed, the deep learning-based rock surface three-dimensional feature extraction and classification method described in any one of claims 1 to 6 is implemented.