Rock slope weak texture feature high-precision identification method fused with WTNet model

Abstract

The invention discloses a rock slope weak texture feature high-precision identification method fused with a WTNet model, and the method achieves the enhancement of weak texture features and weak texture boundary features on a weak texture rock slope through the use of a weak texture feature perception model (WTFP) and a weak texture boundary perception model (WTBA). And classifying and positioning the enhanced weak textures and boundary features thereof by adopting a TD-RCNN frame, and finally realizing high-precision recognition of the weak texture features. According to the invention, the WTFP and WTBA enhanced perception capabilities are fully utilized, so that the classification and positioning of the TD-RCNN frame are more accurate, a high-precision weak texture feature image can be obtained by using the WTNet model, the high-precision recognition of the weak texture features is realized, geological information is provided for generating a weak texture rock slope three-dimensional geological model, and the method has the advantages of being simple in structure and convenient to operate. A reliable basis is provided for judging the stability of the rock slope, and the time and labor cost of slope maintenance are reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing image recognition, and relates to a weak texture feature recognition method, and specifically to a high-precision recognition method for weak texture features of rock slopes integrated with a WTNet model. Background Art

[0002] Weakly textured rock slopes are widespread in southwestern regions, including Chongqing. These surface cracks, joints, and other non-continuous structural surfaces lack distinct geometric features. With the widespread application of remote sensing in geological exploration, weak texture features are often difficult to capture and scan accurately in remote sensing imagery. This is because existing drone mapping technology is often limited in capturing and scanning weak textures by factors such as light intensity, shooting angle, and observation distance. This results in low image resolution, making it impossible to fully recognize weak texture features. This indirectly leads to low 3D reconstruction accuracy and, consequently, model distortion. Summary of the Invention

[0003] To address the above-mentioned deficiencies in the background technology, the present invention provides a high-precision identification method for weak texture features of rock slopes that integrates the WTNet model. This method enhances the weak texture features and weak texture boundary features on weak-texture rock slopes by using a weak texture feature perception model (WTFP) and a weak texture boundary perception model (WTBA). The enhanced weak texture and its boundary features are classified and located using the TD-RCNN framework, ultimately achieving high-precision identification of weak texture features. This provides geological information for generating a three-dimensional geological model of a weak-texture rock slope, provides a reliable basis for judging the stability of the rock slope, and reduces the time and labor costs of slope maintenance.

[0004] The purpose of the present invention is achieved through the following technical solutions:

[0005] A high-precision recognition method for weak texture features of rock slopes integrated with the WTNet model includes the following steps:

[0006] Step 1: Collect remote sensing images with weak texture features and randomly divide them into training and test sets;

[0007] Step 2: Create a weak texture feature perception model WTFP, amplify the weak texture features of the remote sensing images in the training set, and obtain texture perception features;

[0008] Step 3: Create a weak texture boundary perception model WTBA. Through boundary map generation, boundary perception feature extraction and boundary guidance feature fusion, the boundary information is extracted from the image source and integrated into the texture perception feature extracted in step 2 to enhance the boundary features of weak texture and obtain boundary guidance features.

[0009] Step 4: Using the TD-RCNN framework, the weak texture feature detection, including texture-aware features and boundary-guided features, is decomposed into classification and localization tasks to achieve high-precision recognition and detection of weak texture features.

[0010] Step 5: Input the test set data into the WTNet model composed of the weak texture feature perception model (WTFP), the weak texture boundary perception model (WTBA) and the TD-RCNN framework, and use the average precision (AP) to evaluate the accuracy of the model;

[0011] Step 6: Input the remote sensing image of the textured rock slope into the WTNet model to identify weak texture features in the rock slope.

[0012] Compared with the prior art, the present invention has the following advantages:

[0013] The present invention uses WTFP and WTBA to enhance weak texture features and weak texture boundary features on weak texture rock slopes. The enhanced weak texture and its boundary features are classified and located using the TD-RCNN framework, ultimately achieving high-precision identification of weak texture features. The present invention fully utilizes the enhanced perception capabilities of WTFP and WTBA, making the classification and positioning of the TD-RCNN framework more accurate. By using the WTNet model, a high-precision weak texture feature image can be obtained, ultimately achieving high-precision identification of weak texture features. This provides geological information for generating a three-dimensional geological model of weak texture rock slopes, provides a reliable basis for judging the stability of rock slopes, and reduces the time and labor costs of slope maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 The overall implementation flow chart of the high-precision recognition method of weak texture features integrated with the WTNet model. DETAILED DESCRIPTION

[0015] The technical solution of the present invention is further described below with reference to the accompanying drawings, but is not limited thereto. Any modification or equivalent replacement of the technical solution of the present invention that does not depart from the spirit and scope of the technical solution of the present invention should be included in the scope of protection of the present invention.

[0016] The present invention provides a high-precision recognition method for weak texture features of rock slopes integrated with the WTNet model. The method addresses the problem that weak texture features on weak-textured rock slopes cannot be recognized with high precision. Remote sensing images are input into the WTNet model, and WTFP and WTBA are used to amplify the weak texture features of the training set remote sensing images and enhance the boundary features of the weak textures. The TD-RCNN framework is used to decompose the weak texture feature detection into classification and positioning tasks to achieve high-precision recognition and detection of weak texture features. Figure 1 As shown, the specific steps include:

[0017] Step 1: Since the non-penetrating structural surfaces such as cracks and joints on the surface of weak-textured rock slopes do not have obvious geometric features, 2,000 remote sensing images with weak texture features are collected. 1,000 of them are used as the training set, and the other half of the remote sensing images are used as the test set.

[0018] Step 2: Create WTFP, amplify the weak texture features of the training set remote sensing images, and obtain texture perception features. The specific steps are as follows:

[0019] Step 2.1: Generate multi-scale features by FPN algorithm C represents the dimension of the feature channel, W and H represent the width and height of the feature respectively. in Add three 1×1 convolutional layers to reduce the output dimension to Get three intermediate features f θ , f g , using the covariance matrix Σ to calculate With f θ to capture the correlation between pixels at different locations:

[0020]

[0021] Where (x, y) represents the spatial position of the pixel, represents matrix multiplication, and T represents the transposed matrix.

[0022] Step 2.2: In order to highlight the correlation between the target pixel and other pixels while suppressing the relationship matrix of background pixels, a spatial attention mechanism is introduced:

[0023] f in Perform global average pooling along the channel dimension to form the pooling factor f gap A 3×3 convolution layer is used to reduce redundant noise, the Sigmoid function is used to calculate the attention weight, and the attention weight is converted into a one-dimensional vector

[0024]

[0025] M(x,y)=Reshape(σ(Conv 3×3 (f gap (x,y))))

[0026] Where GAP represents the global average pooling operation, σ represents the Sigmoid activation function, and Reshape represents the reshaping function.

[0027] Subsequently, Hadamard products and covariance matrices are performed between the attention weights to produce the object augmentation covariance matrix

[0028] Σ′(x,y)=M(x,y)⊙Σ(x,y)

[0029] Where ⊙ represents the Hadamard product.

[0030] Step 2.3: Use SoftMax to regularize and reshape them to obtain the relationship matrix Mathematically represents the relationship between a specific pixel and all other pixels in an image. The relationship between similar pixels can indicate important texture distribution patterns. 1×1 convolution is used to integrate pixel correlations between different channels and Dimensionality is converted to At this point, a relationship map has been generated that can amplify homogeneous pixels and similar locations of the object, further enhancing beneficial texture features:

[0031]

[0032] Where, Conv 1×1 represents a 1×1 convolution, SoftMax represents a regularization function, and Reshape represents a reshaping function.

[0033] Step 2.4: With f gap Multiply to obtain valuable texture features, and use a 1×1 convolution layer to restore the feature dimension to Finally, the texture perception feature f is generated out :

[0034]

[0035] Where, Represents element-wise summation.

[0036] Step 3: Create WTBA. Through the steps of boundary map generation, boundary-aware feature extraction, and boundary-guided feature fusion, the boundary information is extracted from the image source and integrated into the extracted features to enhance the boundary features of weak textures and obtain boundary-guided features. The specific steps are as follows:

[0037] Step 3.1: Boundary map generation.

[0038] Given an input image I(x,y), use the Sobel kernel to slide I(x,y) horizontally and vertically to extract the oriented gradient map:

[0039]

[0040] The boundary map G(x,y) can be obtained by merging the horizontal and vertical gradient maps:

[0041] G(x,y)=I(x,y)*(G x +G y )

[0042] Where * represents the convolution operation.

[0043] Use the filter to reduce the noise of G(x,y) and generate the boundary map Support boundary feature perception:

[0044]

[0045] Where λ is the preset threshold, which means that when G(x,y) is less than the maximum value of λ×G(x,y) Take zero.

[0046] Step 3.2: Boundary feature extraction.

[0047] The boundary map is transferred to two convolutional layers, BN and ReLu, to extract the basic features f base And reduce the resolution, introduce paired 1×1 convolution in the dimension transformation, a pair of asymmetric convolutions are constructed into the module to capture the continuous lines in the features, which may be important high-frequency boundaries of the object. At the same time, the standard 3×3 convolution is accompanied by asymmetric convolution to make full use of the object boundary. Output feature f bou The basic features f base , and then after integration and activation through the Sigmoid function, the network obtains the boundary feature B related to the edge of the object f :

[0048]

[0049] B f =σ(Conv 1×1 (f bou ))

[0050] Where Bottleneck represents two 1×1 convolutions, followed by BN and ReLu convolutions, and f 1×3 ,f 3×1 ,f 3×3 Represent the features extracted from asymmetric convolution and symmetric convolution respectively.

[0051] Step 3.3: Guide boundary feature fusion.

[0052] The extracted boundary features are quickly fused and incorporated into the intermediate boundary features P iTo further highlight the edge expression of the feature, add channel attention moderation, modify the weight to amplify the feature channel effect, given the basic feature C i and boundary feature B f Perform element-by-element multiplication and element-by-element summation, and then perform convolution to achieve information interaction:

[0053]

[0054] Where, f i represents the fusion feature, i∈{2,3,4,5}, and D represents the downsampling operation.

[0055] A global average pooling across channel dimensions is used to aggregate feature information and a 1×1 convolution is used along the Sigmoid function to generate channel attention. The target feature is enhanced by element-wise multiplication. The enhanced feature is integrated into P i To obtain the boundary guidance feature f i b , which effectively emphasizes the location information of the object:

[0056]

[0057] Step 4: Use the TD-RCNN framework to decompose weak texture feature detection into classification and positioning tasks to achieve high-precision recognition and detection of weak texture features. The specific steps are as follows:

[0058] Step 4.1: Take 256 RoI samples (feature regions) obtained by decoupled sampling as training samples. RoI samples are samples with weak texture features obtained by texture perception features and boundary guidance features, where the target and background maintain a sample ratio of 1:3. Perform DPES classification: a useful local context region is defined with the RoI as the center, and its size is k times the RoI. Enlarging the RoI can effectively mine context information.

[0059] Step 4.2: Take 128 RoI target samples obtained by decoupling sampling and perform PSS positioning:

[0060] Assume that the reference point x to be fitted is i ,i∈{1,2,3,...,n}, find the center point y through continuous iteration so that the center point y and the reference point x i The distance between them is the smallest:

[0061]

[0062] Where ||||2 represents the L2 norm, and argmin represents the sum of the distances minimized by y on D.

[0063] Find the partial derivatives of the above equation to get the minimum:

[0064]

[0065] By solving the above equations using an iterative method, the center point obtained does not coincide with any reference point:

[0066]

[0067] Where k represents iteration, y k+1 Represents the minimum value found after successive iterations.

[0068] Based on two points (top left and bottom right), all target samples assigned to an object are synthesized using the “ground truth”. Therefore, PSS is able to provide high-quality synthesized RoIs for each object.

[0069] Step 4.3: Merge the high-quality RoI obtained by DPES classification and PSS positioning to obtain a high-precision weak texture feature image.

[0070] Step 5: Input the test set data into the WTNet model composed of the weak texture feature perception model (WTFP), the weak texture boundary perception model (WTBA) and the TD-RCNN framework, and use the average precision (AP) to evaluate the accuracy of the model.

[0071] Specifically, the number of correct answers in the output is denoted as TP, the number of incorrect answers in the output is denoted as FP, the number of correct answers not found in the input is denoted as FN, and the total number of input answers minus TP is denoted as TN. P and R can be calculated as:

[0072]

[0073] In the formula, Precision represents accuracy and Recall represents recall.

[0074] According to P and R, the AP of each category and the average AP (mAP) of all categories can be formulated as:

[0075]

[0076] Where C represents the total number of categories. mAP can more comprehensively evaluate the accuracy of the model; higher mAP values ​​indicate better model performance. AP is affected by the IoU threshold, which we set to 0.5.

[0077] Step 6: Input remote sensing images of weakly textured rock slopes in southwestern regions such as Chongqing into the WTNet model to identify weak texture features in rock slopes with high precision.

Claims

1. A high-precision recognition method for weak texture features of rock slopes fused with WTNet model, characterized by The method comprises the following steps: Step 1: Collect remote sensing images with weak texture features and randomly divide them into training sets and test sets; Step 2: Create a weak texture feature perception model WTFP, amplify the weak texture features of the remote sensing images in the training set, and obtain texture perception features; Step 3: Create a weak texture boundary perception model WTBA, extract boundary information from the image source through boundary map generation, boundary perception feature extraction and boundary guidance feature fusion, and integrate it into the texture perception feature extracted in step 2 to enhance the boundary features of weak texture and obtain boundary guidance features; Step 4: Using the TD-RCNN framework, the weak texture feature detection including texture-aware features and boundary-guided features is decomposed into classification and localization tasks to achieve high-precision recognition and detection of weak texture features; Step 5: Input the test set data into the WTNet model composed of WTFP, WTBA and TD-RCNN framework, and use the average precision AP to evaluate the accuracy of the model; Step 6: Input the remote sensing image of the textured rock slope into the WTNet model to identify weak texture features in the rock slope.

2. The high-precision recognition method for weak texture features fused with the WTNet model according to claim 1 is characterized in that The specific steps of step 2 are as follows: Step 2.1: Generate multi-scale features by FPN algorithm C represents the dimension of the feature channel, W and H represent the width and height of the feature respectively. in Add three 1×1 convolutional layers to reduce the output dimension to Get three intermediate features f θ , f g , using the covariance matrix Σ to calculate With f θ to capture the correlation between pixels at different locations: Where (x, y) represents the spatial position of the pixel. represents matrix multiplication, T represents the transposed matrix; Step 2.2: f in Perform global average pooling along the channel dimension to form the pooling factor f gap A 3×3 convolutional layer is used to reduce redundant noise, the Sigmoid function is used to calculate the attention weight, and the attention weight is converted into a one-dimensional vector M(x,y)=Reshape(σ(Conv 3×3 (f gap (x,y)))) In the formula, GAP represents the global average pooling operation, σ represents the Sigmoid activation function, and Reshape represents the reshaping function; Subsequently, Hadamard products and covariance matrices are performed between the attention weights, producing the object augmentation covariance matrix Σ′(x,y)=M(x,y)⊙Σ(x,y) Where, ⊙ represents the Hadamard product; Step 2.3: Use SoftMax to regularize and reshape them to obtain the relationship matrix Mathematically represents the relationship between a specific pixel and all other pixels in the image, using 1×1 convolution to integrate pixel correlations between different channels and Dimensions are converted to At this point, a relationship map of homogeneous pixels and similar locations that can magnify the object is generated, further enhancing the beneficial texture features: In the formula, Conv 1×1 represents 1×1 convolution, SoftMax represents regularization function, and Reshape represents reshaping function; Step 2.4: With f gap Multiply to obtain valuable texture features, and use a 1×1 convolution layer to restore the feature dimension to Finally, the texture perception feature f is generated out : In the formula, Represents element-wise sum.

3. The high-precision recognition method of weak texture features fused with the WTNet model according to claim 1 is characterized in that The specific steps of step 3 are as follows: Step 3.1: Boundary map generation Given an input image I(x,y), use the Sobel kernel to slide I(x,y) horizontally and vertically to extract the oriented gradient map: The boundary map G(x,y) can be obtained by merging the horizontal and vertical gradient maps: G(x,y)=I(x,y)*(G x +G y ) In the formula, * represents the convolution operation; Use the filter to reduce the noise of G(x,y) and generate a boundary map Support boundary feature perception: In the formula, λ is the preset threshold, which means that when G(x,y) is less than the maximum value of λ×G(x,y) Take zero; Step 3.2: Boundary feature extraction The boundary map is transferred to two convolutional layers, BN and ReLu, to extract the basic features f base And reduce the resolution, introduce paired 1×1 convolutions in the dimension transformation, a pair of asymmetric convolutions are constructed into the module to capture continuous lines in the features, and at the same time, the standard 3×3 convolution is accompanied by asymmetric convolutions to make full use of the object boundaries and output the feature f bou The basic features f base , and then after integration and activation through the Sigmoid function, the boundary feature B related to the edge of the object is obtained f : B f =σ(Conv 1×1 (f bou )) In the formula, Bottleneck represents two 1×1 convolutions, followed by BN and ReLu convolutions, and f 1×3 ,f 3×1 ,f 3×3 Represent the features extracted from asymmetric convolution and symmetric convolution respectively; Step 3.3: Guide boundary feature fusion. The extracted boundary features are quickly merged and incorporated into the intermediate boundary features P i To further highlight the edge expression of the feature, add channel attention moderation, modify the weight to amplify the feature channel effect, given the basic feature C i and boundary feature B f Perform element-by-element multiplication and element-by-element summation, and then perform convolution to achieve information interaction: In the formula, f i represents fusion features, i∈{2,3,4,5}, and D represents downsampling operation; A global average pooling across channel dimensions is used to aggregate feature information and a 1×1 convolution is used along the Sigmoid function to generate channel attention. The target feature is enhanced by element-by-element multiplication and the enhanced feature is integrated into P i To obtain the boundary guidance feature f i b , which effectively emphasizes the location information of the object:

4. The high-precision recognition method of weak texture features fused with the WTNet model according to claim 1 is characterized in that The specific steps of step 4 are as follows: Step 4.1: Take 256 RoI samples obtained by decoupling sampling as training samples. RoI samples are samples with weak texture features obtained by texture perception features and boundary guidance features. The target and background maintain a sample ratio of 1:3, and perform DPES classification; Step 4.2: Take 128 RoI target samples obtained by decoupling sampling and perform PSS positioning: Assume that the reference point x to be fitted is i , i∈{1,2,3,...,n}, find the center point y through continuous iteration so that the center point y and the reference point x i The distance between them is minimal: In the formula, || ||2 represents the L2 norm, and argmin represents the sum of the distances minimized by y on D; Find the partial derivatives of the above equation to get the minimum: By solving the above equations using an iterative method, the center point obtained does not coincide with any reference point: In the formula, k represents iteration, y k+1 represents the minimum value found after consecutive iterations; Based on the two points in the upper left and lower right corners, all target samples assigned to the object are synthesized using the "ground truth"; Step 4.3: Merge the high-quality RoI obtained by DPES classification and PSS positioning to obtain a high-precision weak texture feature image.

Images