Deep learning-based rock weathering degree classification method

Abstract

The present invention discloses a method for classifying rock weathering degree based on deep learning, including: inputting the acquired rock image into a constructed decision forest model for edge detection to obtain a structured forest edge detection map; extracting a rock feature image based on the backbone network of the rock image ， extracting a rock edge feature image based on the structured forest edge detection map; after the global attention module fuses the rock feature image and the rock edge feature image, predicting the type of weathering degree to which it belongs. The present invention uses a structured random forest algorithm to perform edge detection on rock images, fuses rock images and rock edge images, accurately predicts and outputs the weathering degree of rocks, and has high engineering application value.

Description

Technical Field

[0001] The present invention relates to a method for classifying rock weathering degree based on deep learning, and belongs to the technical field of rock weathering degree classification. Background Technique

[0002] With the rapid development of social economy, the electricity demands of residents and enterprises are increasing continuously, and the scale of the power grid in China is also expanding continuously. Due to the lack of economic peak shaving means in the power grid, the contradiction of power grid peak shaving is becoming increasingly prominent, and the power shortage situation has changed from the lack of electricity quantity to the lack of peak shaving capacity. There is a growing consensus on building pumped-storage power stations to solve the peak shaving problem of the power grid mainly based on thermal power. With the requirements of economic operation of the power grid and adjustment of the power source structure, some power grids mainly based on hydropower have also begun to study and build pumped-storage power stations of a certain scale, effectively alleviating the power shortage situation during peak electricity consumption, and achieving good social and economic benefits.

[0003] Most pumped-storage power stations are built in mountains where suitable reservoir capacities can be constructed and the distances from the station sites to the power grid are economically reasonable. Natural reservoirs are formed by using mountain ridges to reduce construction costs. A pumped-storage power station is divided into upper and lower reservoirs, and there are three construction modes: both the upper and lower reservoirs utilize similar natural rivers or lakes; when the upper reservoir is artificially enclosed, the lower reservoir utilizes natural rivers, lakes, bays or existing reservoirs; both the upper and lower reservoirs are artificially enclosed, and only a pure pumped-storage power station can be built; the lower reservoir is artificially enclosed, and the upper reservoir is an existing reservoir. When building a pumped-storage power station, it is necessary to observe the rock state around the power station to prevent rock collapses and cause safety accidents; it is also necessary to observe the rock state during the construction of the power station to ensure that the weathering degree of the rock meets the safety construction indicators of the power station. The division of rock weathering degree and the research on engineering characteristics play a key role in the selection of the construction base surface of large-scale hydropower projects, high-rise buildings, road bridges and other projects and the determination of the foundation design and construction plan, and are also of great significance for evaluating the stability of surrounding rocks and slope engineering.

[0004] According to the characteristics and depths of rock weathering, the degree of weathering is divided into five levels: unweathered, slightly weathered, moderately weathered, highly weathered, and completely weathered. The unweathered rock has fresh rock quality, occasional weathering traces, and unchanged rock tissue structure. The slightly weathered rock has fresh rock quality, some traces of ferromanganese staining or slight color change along the joint surface, a small amount of weathering traces, no loose material, and the mineral and rock tissue structure remains basically unchanged. The moderately weathered rock has clear structural bedding, but is cut into rock blocks by joint fissures, and the fissures are filled with a small amount of weathered material; the structure is partially damaged, the mineral composition remains basically unchanged, and only secondary minerals appear along the joint surface; the hammering sound is crisp, the rock mass is not easy to break, it is difficult to dig with a pickaxe, and a core drill can be used to drill. The highly weathered rock mass is divided into fragmented blocks by joint fissures; most of the rock mass structure is damaged, the structural bedding is not clear, and the mineral composition changes significantly; the hammering sound is dull, the broken rock can be broken by hand, it is not easy to drill dry, and it can be dug with a pickaxe. The completely weathered rock mass is divided into scattered masses by joint fissures; the rock mass structure is basically damaged, and only the appearance remains in the original rock state; the crushed stone can be crushed by hand, it can be dug with a pickaxe, and it can be drilled dry.

[0005] The state of rocks is variable and there are many types. The degree of rock weathering requires professional personnel to accurately judge. If only relying on manual judgment of the degree of rock weathering, not only the cost is high, but also the efficiency is low and it is easy to make mistakes. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to overcome the defects of the prior art and provide a method for classifying the degree of rock weathering based on deep learning, which can efficiently and intelligently identify the degree of rock weathering, effectively solve the disadvantages of low efficiency and high cost of manual work, so as to judge the state of rocks during the construction or operation of a pumped-storage power station, and further ensure the safety during the construction process and operation process of the pumped-storage power station.

[0007] To achieve the above object, the present invention provides a method for classifying the degree of rock weathering based on deep learning, including:

[0008] Step 1), input the obtained rock image I into the constructed decision forest model for edge detection to obtain the structured forest edge detection map I edge ;

[0009] Step 2), based on the rock image I, the backbone network extracts the rock feature image;

[0010] Based on the structured forest edge detection map I edge , the backbone network extracts the rock edge feature image;

[0011] Step 3), after the global attention module fuses the rock feature image and the rock edge feature image, it predicts the type of weathering degree to which it belongs.

[0012] Preferably, a decision forest model is constructed, including:

[0013] (21) Each decision tree of the random decision forest is independently trained in a recursive manner;

[0014] (22) A second-order mapping is used to discretize the similarity of the structure label y into a discretized label c;

[0015] (23) All decision trees are combined into a decision forest to obtain a decision forest model.

[0016] Preferably, step (21) includes:

[0017] According to the information gain The maximum principle determines the node separation function:

[0018]

[0019] The maximum information gain among them, j is the given node, represents the training set for training the left node, represents the training set for training the right node, and the training set S j ∈X×Y, X is an image composed of pixel blocks x, y∈Y is the structure label corresponding to the pixel block x, and Y is the label set composed of structure labels y;

[0020]

[0021] In the formula, h(x,θ j ) is the node separation function. When h(x,θ j ) is 0, the pixel block x is classified to the right of node j to generate a right node; when h(x,θ j ) is 1, the pixel block x is classified to the left of node j to generate a left node; k is a quantization feature value of x, and θ j =(k,γ) is the gain parameter that maximizes I j , and γ is the threshold of the quantization feature value k;

[0022] Use Recursively train the left node, use Recursively train the right node until the set tree depth or the threshold of the set information gain is reached.

[0023] Preferably, the threshold of the set information gain is obtained using the standard information gain:

[0024]

[0025]

[0026] Among them, I j ′ is the standard information gain, H(S) is the Shannon entropy, and p y is the probability of the training set S j with the structure label y, and H(S j ) is the Shannon entropy of S j . is the Shannon entropy when the parameter m ∈ (L, R). .

[0027] Preferably, step (22) includes:

[0028] Define the mapping Π: y → z, encode the pixel block x with the structure label y into a binary vector z, calculate the Euclidean distance between the binary vectors z in the two-dimensional space Z to distinguish whether the pixel blocks x with similar structure labels y belong to the same segmentation; take m-dimensional features in the two-dimensional space Z to form a low-dimensional mapping Π: Φ: Y → Z;

[0029] Define the low-dimensional mapping Π: Φ: Z → C, and use PCA dimensionality reduction quantization to give specific discrete labels C(1, 2,..., k) according to the specific quadrant where the binary vector z is located.

[0030] Preferably, the backbone network is a dual-path ResNeXt-50 network, and the dual-path ResNeXt-50 network includes a first ResNeXt-50 network for extracting the rock feature image and a second ResNeXt-50 network for extracting the rock edge feature image;

[0031] The first ResNeXt-50 network includes a first convolutional layer, a batch normalization layer, a rectified linear unit layer, a max pooling layer, Layer1, Layer2, Layer3, and Layer4, and the first convolutional layer, the batch normalization layer, the rectified linear unit layer, the max pooling layer, Layer1, Layer2, Layer3, and Layer4 are connected in sequence;

[0032] The convolutional kernel of the first convolutional layer is 7×7, the number of output channels is 64, the padding is 3, and the stride is 2; Layer1 includes 1 convolutional module and 2 feature modules, and the 1 convolutional module and 2 feature modules are connected in sequence. The convolutional kernel of the convolutional module is 3×3, the padding is 1, and the stride is 2;

[0033] In the convolutional module, first perform convolution, batch normalization, and rectified linear unit operations on the input channel C in with 64 channels on the main path to obtain an output channel C out of 128 and a convolutional kernel S K of 1, and then perform operations of keeping the number of channels unchanged and a convolutional kernel S KGroup convolution, batch normalization, and ReLU operations with a group number of 3 and a branch number of 32, followed by an output channel C out with 256 and S K Convolution and batch normalization operations with 1, while performing input channel C on the shortcut in with 64, output channel C out with 256 and a convolution kernel S K Convolution and batch normalization operations with 1, and finally adding the output of the main path and the output of the shortcut, and performing a ReLU operation;

[0034] In the feature module, first perform an output channel C on the input channel C with 256 channels on the main path in to obtain an output channel C out with 128 and a convolution kernel S K Convolution, batch normalization, and ReLU operations with 1, and then perform a grouped convolution, batch normalization, and ReLU operation with a group number of 3 and a branch number of 32 with the same number of channels and a convolution kernel S K with 3 and a branch number of 32, followed by an output channel C out with 256 and S K Convolution, batch normalization, and ReLU operations with 1.

[0035] Preferably, Layer2 includes 1 convolution module and 3 feature modules,

[0036] Layer3 includes 1 convolution module and 5 feature modules,

[0037] Layer4 includes 1 convolution module and 2 feature modules;

[0038] The structure of the second ResNeXt-50 network is the same as the structure of the second ResNeXt-50 network.

[0039] Preferably, step 2) includes:

[0040] The rock image I is input into the first ResNeXt-50 network, and the rock feature image F is output;

[0041] Structured forest edge detection map I edge is input into the second ResNeXt-50 network, and the rock edge feature image F is output edge .

[0042] Preferably, step 3) includes:

[0043] Performing a concatenation operation on the channel numbers of the rock feature image F and the rock edge feature image F edge to obtain F c ;

[0044] Performing Fc Perform convolution with 4096 input channels, 1024 output channels and 3 convolution kernels, and then perform batch normalization and linear rectification to obtain F c ′;

[0045] F c ′ Four adaptive maximum pooling operations with sizes of 1, 5, 9, and 13 are used to obtain four feature maps with different receptive fields, and convolution and linear rectification operations with an input channel of 1024, an output channel of 512, and a convolution kernel of 1 are performed on the four feature maps respectively, and then the four feature maps are upsampled to the same size as F c ′ of the same size, and finally the four feature maps are combined with F c ′ performs a merge operation to obtain the output F p ;

[0046] F p A convolution with an input channel number of 3072, an output channel number of 1024 and a convolution kernel of 1 is performed, and then batch normalization and linear rectification operations are performed to obtain the reconstructed feature G that incorporates global information;

[0047] A full connection operation is performed on G with 5 output categories, corresponding to the weathering degree types of unweathered, slightly weathered, moderately weathered, strongly weathered and fully weathered. The category with the highest probability of output through Softmax is the weathering degree of the rock predicted by the network.

[0048] The beneficial effects achieved by the present invention are:

[0049] (1) The present invention combines the structured random forest algorithm to obtain a decision forest model. The decision forest model can effectively locate rocks and joints on rocks through edge detection, and obtain small fracture structures where there is no significant displacement of rock blocks on both sides after the rock mass is fractured by force, laying the foundation for later judgment of the degree of rock weathering.

[0050] (2) In the prior art, the traditional edge detection algorithms such as Canny tend to retain a lot of detail information in the image that is irrelevant to the target, interfering with the detection of joints. The present invention uses structured forest fast edge detection, which has the characteristics of less interference, fast detection and high accuracy. Each decision tree of the random decision forest is trained independently in a recursive manner. By recursively training multiple unrelated decision trees to form a forest, the problem of overfitting and instability of a single decision tree is better solved. The present invention uses the input of randomly sampled pixel blocks x or feature categories as training data to improve the accuracy of the decision forest.

[0051] (3) Based on ResNet, ResNeXt introduces the concept of the number of branches (cardinality) and uses group convolution to deepen the network while ensuring the same number of floating-point operations per second (FLOPS) and the number of parameters. In the present invention, a dual-path ResNeXt-50 network without a fully connected layer is used as the backbone, which can effectively improve the network classification effect of the backbone network.

[0052] When the present invention combines the visible light image features and the edge image features, with the fusion of spatial details, the position information of the object will be gradually diluted. At the same time, the receptive field of ResNeXt is too small to obtain sufficient semantic information. Therefore, the present invention designs a global attention module (Global Attention Module) to effectively expand the receptive field, obtain the detailed information of important targets, and effectively fuse the two features. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 is the system flowchart of the present invention;

[0054] Figure 2 is the structured forest edge detection flowchart of the present invention;

[0055] Figure 3 is the training diagram of the decision forest model;

[0056] Figure 4 is the schematic diagram of the network structure of the present invention;

[0057] Figure 5 is the convolutional module structure diagram in Layer1 of the present invention;

[0058] Figure 6 is the feature module structure diagram in Layer1 of the present invention;

[0059] Figure 7 is the global attention module structure diagram of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0060] The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be used to limit the protection scope of the present invention.

[0061] It should be noted that if there are directional indications (such as up, down, left, right, front, back...) in the embodiments of the present invention, they are only used to explain the relative position relationship and movement conditions between components in a specific posture. If the specific posture changes, the directional indications will also change accordingly.

[0062] In addition, if descriptions such as "first", "second", etc. are involved in the present invention, they are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include at least one such feature. In addition, the technical solutions between various embodiments may be combined with each other, but it must be based on the ability of those of ordinary skill in the art to implement. When the combination of technical solutions results in contradictions or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0063] A method for classifying rock weathering degree based on deep learning, comprising:

[0064] 1) Input the obtained rock image I into the constructed decision forest model for edge detection to obtain the structured forest edge detection map I edge ;

[0065] 2) Based on the rock image I, the backbone network extracts the rock feature image; based on the structured forest edge detection map I edge , the backbone network extracts the rock edge feature image;

[0066] 3) The global attention module fuses the feature information of the rock feature image and the feature information of the rock edge feature image to obtain a fused feature image;

[0067] 4) Through the fully connected layer and Softmax, predict and output the class with the highest probability of the weathering degree to which the fused feature image belongs, and obtain the predicted weathering degree of the rock.

[0068] In step 1), constructing the decision forest model includes:

[0069] (21) Each decision tree of the random decision forest is independently trained recursively: for a given node j and training set S j ∈X×Y, where X is an image composed of pixel blocks x; y∈Y is the classification result corresponding to the pixel block x, that is, the structure label; Y is the label set composed of structure labels y. According to the principle of maximum information gain determine the node separation function:

[0070]

[0071] In the formula, the maximum information gain, represents the training set for training the left node, represents the training set for training the right node; h(x,θ j) is a node separation function with values of 0 or 1. When h(x,θ j ) is 0, the pixel block x is classified as on the right side of node j, generating a right node. When h(x,θ j ) is 1, the pixel block x is classified as on the left side of node j, generating a left node; k is a certain quantization feature value of x, and θ j =(k,γ) is the gain parameter that maximizes I j , where γ is the threshold for having this quantization feature value;

[0072] Use to recursively train the left node, and use to recursively train the right node until the set tree depth or the threshold of information gain is reached. The threshold of information gain is obtained using the standard information gain:

[0073]

[0074]

[0075] where I j ′ is the standard information gain, H(S) is the Shannon entropy, and p y is the probability of having the structure label y in the training set S j , H(S j ) is the Shannon entropy of S j , and is the Shannon entropy when the parameter m ∈ (L, R). ;

[0076] (22) Discretize the similarity of the structure label y into a discretized label c using a second-order mapping:

[0077] First, define the mapping Π: y → z, encode the pixel block with the structure label y into a binary vector z, and calculate the Euclidean distance between the binary vectors z in the two-dimensional space Z to distinguish whether the pixel blocks with similar structure labels y belong to the same segmentation;

[0078] Take the m-dimensional features in the two-dimensional space Z to form a low-dimensional mapping Π: Φ: Y → Z, where Y is the label set composed of the structure labels y. Then define the low-dimensional mapping Π: Φ: Z → C, and use PCA dimensionality reduction quantization to give the specific discretized label C(1, 2,..., k) according to the specific quadrant where the binary vector z is located, with k = 2;

[0079] (23) Combine all decision trees into a decision forest to obtain a decision forest model.

[0080] In step 2), a dual-path ResNeXt-50 network without a fully connected layer is used as the backbone to extract the rock feature image and the rock edge feature image respectively. The dual-path ResNeXt-50 network includes a first ResNeXt-50 network for extracting the rock feature image and a second ResNeXt-50 network for extracting the rock edge feature image;

[0081] The first ResNeXt-50 network includes:

[0082] (31) The first convolutional layer performs convolution on the rock image I. The convolutional kernel of the first convolutional layer is 7×7, the number of output channels is 64, the padding is 3, the stride is 2. After convolution, the batch normalization layer performs batch normalization (BatchNormalize, BN), and the rectified linear layer performs rectification based on the rectified linear unit (ReLU). Then, max pooling is performed through the max pooling layer, where the convolutional kernel is 3×3, the padding is 1, and the stride is 2, to obtain the preprocessed feature map I'.

[0083] (32) The feature map I' is input into Layer1. Layer1 is composed of 1 convolutional module (ConvBlock) and 2 identity modules (IdentinyBlock), and their structures are as Figure 5 shown. Define the input channel of the convolution operation as C in , the output channel as C out , and the convolutional kernel as S K . The convolutional module first performs convolution, batch normalization, and rectification on the input with 64 channels on the main path, where C out is 128 and S K is 1;

[0084] then it performs grouped convolution, batch normalization, and rectification with the number of channels unchanged, where S K is 3 and the number of branches is 32;

[0085] then it performs convolution and batch normalization where C out is 256 and S K is 1; At the same time, on the shortcut path, it performs convolution and BN operations where C in is 64, C out is 256, and S K is 1; Finally, the output of the main path and the output of the shortcut path are added together, and a rectification operation is performed;

[0086] The input channels of the IdentinyBlock are 256, and no operation is performed on the shortcut. The other parts are the same as the ConvBlock. After passing through 2 IdentinyBlocks, the output of Layer1 is the feature map I1'.

[0087] (33) Input the feature map I1' into Layer2. Layer2 consists of 1 ConvBlock and 3 IdentinyBlocks. The ConvBlock first performs a convolution with C out being 256 and S K being 1, followed by BN and ReLU; then it performs a grouped convolution with the number of channels remaining unchanged, S K being 3 and the number of branches (cardinality) being 32, followed by BN and ReLU; then it performs a convolution with C out being 512 and S K being 1, followed by BN; at the same time, on the shortcut path, it performs a convolution with C in being 256, C out being 512, S K being 1 and BN operations; finally, the outputs on the main path and the shortcut path are added together and ReLU operation is performed.

[0088] The input channels of the IdentinyBlock are 512, and no operation is performed on the shortcut path. The other parts are the same as the ConvBlock. After passing through 3 IdentinyBlocks, the output of Layer2 is the feature map I2'.

[0089] (34) Input the feature map I2' into Layer3. Layer3 consists of 1 ConvBlock and 5 IdentinyBlocks. The ConvBlock first performs a convolution with C out being 512 and S K being 1, followed by BN and ReLU; then it performs a grouped convolution with the number of channels remaining unchanged, S K being 3 and the number of branches (cardinality) being 32, followed by BN and ReLU; then it performs a convolution with C out being 1024 and S K being 1, followed by BN; at the same time, on the shortcut path, it performs a convolution with C in being 512, C out being 1024, S K being 1 and BN operations; finally, the outputs on the main path and the shortcut path are added together and ReLU operation is performed.

[0090] The number of input channels of the IdentinyBlock is 1024, and no operation is performed on the shortcut path. The other parts are the same as the ConvBlock. After 5 IdentinyBlocks, the output of Layer3 is the feature map I3'.

[0091] (35) Input the feature map I3' into Layer4. Layer4 consists of 1 ConvBlock and 2 IdentinyBlocks. The ConvBlock first performs a convolution with C out being 1024 and S K being 1, followed by BN and ReLU; then it performs a grouped convolution with the number of channels remaining unchanged, S K being 3 and the number of branches cardinality being 32, followed by BN and ReLU; then it performs a convolution with C out being 2048 and S K being 1, followed by BN; at the same time, on the shortcut path, it performs a convolution with C in being 1024, C out being 2048, and S K being 1, followed by convolution and BN operations; finally, the outputs on the main path and the shortcut path are added together, followed by a ReLU operation.

[0092] The number of input channels of the IdentinyBlock is 2048, and no operation is performed on the shortcut path. The other parts are the same as the ConvBlock. After 2 IdentinyBlocks, the output of Layer4 is the rock feature map F.

[0093] The structure of the second ResNeXt-50 network is the same as that of the second ResNeXt-50 network. The structured forest edge detection map I edge is input into the second ResNeXt-50 network to obtain the rock edge feature image F edge .

[0094] In step 3) above, an attention mechanism is introduced, and a global attention module that fuses the feature information of the rock image and the rock edge image is designed. The specific steps for outputting the class with the highest probability as the weathering degree of the rock predicted by the network through a fully connected layer and Softmax are as follows:

[0095] (41) Concatenate the rock feature image F and the rock edge feature image F edge operation on the number of channels to obtain F c ;

[0096] (42) Input F c Perform convolution with 4096 input channels, 1024 output channels, and a convolution kernel of 3, followed by batch normalization and ReLU to obtain F c ';

[0097] (43) Apply four adaptive max-pooling operations with sizes 1, 5, 9, and 13 to F c ' to obtain four feature maps with different receptive fields. Then, perform convolution with 1024 input channels, 512 output channels, and a convolution kernel of 1, and ReLU operation on each of the four feature maps. Next, upsample the four feature maps to the same size as F c '. Finally, perform a Concatnate operation on the four feature maps and F c ' to obtain the output F p ; Concatnate is a pixel-level stacking operation;

[0098] (44) Perform convolution on F p with 3072 input channels, 1024 output channels, and a convolution kernel of 1, followed by batch normalization and ReLU operations to obtain the reconstructed feature G that integrates global information;

[0099] (45) Perform a fully connected operation on G with 5 output classes, corresponding to the weathering degree types of unweathered, slightly weathered, moderately weathered, highly weathered, and completely weathered. Then, output the class with the highest probability through Softmax (normalized exponential function), which is the weathering degree of the rock predicted by the network.

[0100] The above is only the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the technical principle of the present invention, several improvements and deformations can be made, and these improvements and deformations should also be regarded as the protection scope of the present invention.

Claims

1. A method for classifying the weathering degree of rocks based on deep learning, characterized in that, Including: Step 1), input the obtained rock image I into the constructed decision forest model for edge detection to obtain the structured forest edge detection map I edge ; Step 2), based on the rock image I, the backbone network extracts the rock feature image; Structured Forest Edge Detection Based Diagram I edge , the backbone network extracts the rock edge feature image; Step 3), after the global attention module fuses the rock feature image and the rock edge feature image, it predicts the weathering degree type to which it belongs; The backbone network is a dual-path ResNeXt-50 network, and the dual-path ResNeXt-50 network includes a first ResNeXt-50 network for extracting the rock feature image and a second ResNeXt-50 network for extracting the rock edge feature image; The first ResNeXt-50 network includes a first convolutional layer, a batch normalization layer, a rectified linear layer, a max pooling layer, Layer1, Layer2, Layer3, and Layer4, and the first convolutional layer, the batch normalization layer, the rectified linear layer, the max pooling layer, Layer1, Layer2, Layer3, and Layer4 are connected in sequence; The convolutional kernel of the first convolutional layer is 7×7, the number of output channels is 64, the padding is 3, and the stride is 2; Layer1 includes 1 convolutional module and 2 feature modules, the 1 convolutional module and the 2 feature modules are connected in sequence, the convolutional kernel of the convolutional module is 3×3, the padding is 1, and the stride is 2; In the convolutional module, first, on the main path, an input channel C with 64 channels in is subjected to a convolution, batch normalization, and rectified linear unit operation with an output channel C out of 128 and a convolution kernel S K of 1. Then, a grouped convolution, batch normalization, and rectified linear unit operation with a convolution kernel S K of 3 and 32 branches is performed while keeping the number of channels unchanged. Next, a convolution and batch normalization operation with an output channel C out of 256 and an S K of 1 is carried out. At the same time, on the shortcut path, a convolution and batch normalization operation is performed on the input channel C in with 64 channels and an output channel C out of 256 and a convolution kernel S K of 1. Finally, the output of the main path and the output of the shortcut path are added together, and a rectified linear unit operation is performed; In the feature module, first perform convolution, batch normalization, and rectified linear unit operations on the input channel C with 256 channels on the main path, where the output channel C in is 128 and the convolution kernel S out is 1. Then perform grouped convolution, batch normalization, and rectified linear unit operations with the number of channels remaining unchanged, where the convolution kernel S K is 3 and the number of branches is 32. Finally, perform convolution, batch normalization, and rectified linear unit operations where the output channel C K is 256 and S out is 1; K ​ Layer2 includes 1 convolutional module and 3 feature modules, Layer3 includes 1 convolutional module and 5 feature modules, Layer4 includes 1 convolutional module and 2 feature modules; The structure of the second ResNeXt-50 network is the same as that of the first ResNeXt-50 network; Step 2) includes: The rock image I is input into the first ResNeXt-50 network, and the rock feature image F is output; Structured forest edge detection diagram I edge Input into the second ResNeXt-50 network, and output the rock edge feature image F edge ; Step 3) includes: Concatenate the rock feature image F and the rock edge feature image F edge in terms of the number of channels to obtain F c ; Input F c Perform a convolution with 4096 input channels, 1024 output channels, and a convolution kernel of 3 on F, and then perform batch normalization and rectified linear unit to obtain F c ′; For F c ′, four adaptive max pooling operations with sizes of 1, 5, 9, and 13 are used to obtain four feature maps with different receptive fields; convolution and rectified linear operations with an input channel of 1024, an output channel of 512, and a convolution kernel of 1 are respectively performed on the four feature maps, and then the four feature maps are upsampled to the same size as F c ′, and finally the four feature maps are merged with F c ′ to obtain the output F p ; For F p perform convolution with 3072 input channels, 1024 output channels, and a convolution kernel of 1, and then perform batch normalization and ReLU operations to obtain the reconstructed feature G that integrates global information; Perform a fully connected operation with an output category of 5 on the reconstructed feature G, corresponding to the weathering degree types of unweathered, slightly weathered, moderately weathered, strongly weathered, and completely weathered respectively, and then output the category with the highest probability through the normalized exponential function as the predicted weathering degree type of the rock.

2. The method for classifying rock weathering degree based on deep learning according to claim 1, characterized in that Construct a decision forest model, including: (21) Each decision tree of the random decision forest is independently trained recursively; (22) Use a second-order mapping to discretely map the similarity of the structure label y to the discretized label c; (23) Combine all decision trees into a decision forest to obtain a decision forest model.

3. The method for classifying rock weathering degree based on deep learning according to claim 2, characterized in that Step (21) includes: According to the information gain I j = I(S j , S j L , S j R ), the node separation function is determined by the maximum principle: is the largest information gain among them, j is the given node, represents the training set for training the left node, represents the training set for training the right node, and the training set S j ∈ X × Y, where X is an image composed of pixel blocks x, y ∈ Y is the structure label corresponding to the pixel block x, and Y is the label set composed of the structure labels y; where h(x,θ j ) is the node separation function. When h(x,θ j ) is 0, the pixel block x is classified to the right of node j to generate a right node. When h(x,θ j ) is 1, the pixel block x is classified to the left of node j to generate a left node. k is a certain quantization feature value of x, and θ j =(k,γ) is the gain parameter that maximizes I j . γ is the threshold for having the quantization feature value k; Usage Recursively train the left node using Recursively train the right node until the set tree depth or the threshold of the set information gain is reached.

4. The method for classifying rock weathering degree based on deep learning according to claim 3, characterized in that The set threshold of information gain is obtained by using the standard information gain: Among them, I j ′ is the standard information gain, H(S) is the Shannon entropy, and p y is the probability of the training set S j with the structural label y, and H(S j ) is the Shannon entropy of S j . is the Shannon entropy when the parameter m ∈ (L, R) .

5. The method for classifying rock weathering degree based on deep learning according to claim 2, characterized in that Step (22) includes: Define the mapping Π: y → z, which encodes the pixel block x with the structure label y into a binary vector z, and calculates the Euclidean distance between the binary vectors z in the two-dimensional space Z to distinguish whether the pixel blocks x with similar structure labels y belong to the same segmentation; Take the m-dimensional features in the two-dimensional space Z to form the low-dimensional mapping Φ: Z → C; Define the low-dimensional mapping Φ: Z → C. According to the specific quadrant where the binary vector z is located, use PCA dimensionality reduction quantization to give the specific discretization label C(1, 2,..., k).

Images