A mine surface crack extraction method based on multi-scale global fusion

Abstract

This invention provides a method for extracting surface cracks in mining areas based on multi-scale global fusion, relating to the field of surface crack extraction. Specifically, it includes the following steps: acquiring orthophotos of the mining area from a UAV; constructing a multi-scale feature perception module to capture multi-scale crack textures and suppress complex background interference; constructing a token selection context module to perform channel linear modulation on deep features and fuse local details with global semantics; designing a feature fusion module based on Fourier transform to reduce background noise and enhance edge contrast through linear combination in the frequency domain; and using an image segmentation network model to infer from the UAV imagery and output segmentation results containing crack regions. The technical solution of this invention overcomes the problems existing in the prior art when identifying cracks in high-resolution orthophotos of complex mining scenes from UAVs, such as inaccurate segmentation due to similar textures to cracks, discontinuous edges of slender cracks, and incomplete extraction of connected structures.

Description

Technical Field

[0001] This invention relates to the field of surface fracture extraction, and specifically to a method for extracting surface fractures in mining areas based on multi-scale global fusion. Background Technology

[0002] Surface fissures in mining areas are a manifestation of the combined effects of mining disturbance, surface subsidence, and tectonic activity. Their spatial distribution is characterized by elongated, discontinuous, and weakly textured features. Rapid and accurate identification of fissures is crucial for safe production, disaster risk assessment, ecological restoration, and regulatory decisions in mining areas. In recent years, high-resolution orthophotos acquired by UAVs offer flexibility and high precision. Combined with deep learning technology, this has led to significant progress in automatic fissure extraction. However, several challenges remain in complex mining scenarios. First, fissure widths occupy only a few pixels in images, easily causing information overload and edge breaks during semantic segmentation downsampling, making it easy to miss elongated targets. Second, fissure length and orientation are easily influenced by the surrounding environment; a single receptive field or fixed-scale feature representation cannot simultaneously extract both elongated and fragmented fissure morphologies. Third, interference objects similar in morphology to fissures, such as shadows, roads, and ditches, are widespread, and variations in lighting and surface cover further exacerbate background confusion, leading to a higher false detection rate.

[0003] Therefore, there is a need for a multi-scale global fusion-based method for extracting surface fractures in mining areas that can simultaneously improve boundary clarity and fracture connectivity without significantly increasing computation. Summary of the Invention

[0004] The main objective of this invention is to provide a method for extracting surface cracks in mining areas based on multi-scale global fusion, in order to solve the problems in existing technologies when identifying cracks in high-resolution orthophotos of complex mining areas from UAVs, such as inaccurate segmentation due to similar textures to cracks, discontinuous edges of slender cracks, and incomplete extraction of connected structures.

[0005] To achieve the above objectives, this invention provides a method for extracting surface fractures in mining areas based on multi-scale global fusion, specifically including the following steps:

[0006] S1 acquires orthophotos of the mining area from UAVs, and performs preprocessing, cropping, and enhancement on the images.

[0007] S2 constructs a multi-scale feature perception module (MFS) in the shallow feature extraction stage of the image segmentation network model and introduces an efficient multi-scale attention module (EMA) to capture multi-scale crack textures and suppress complex background interference.

[0008] S3 constructs the Token Selection Context Module (TSCT) in the deep feature modeling stage of the image segmentation network model, and uses the Feature Modulation Module (FiLM) to generate scaling and offset coefficients based on shallow features, performs channel linear modulation on deep features, and fuses local details with global semantics.

[0009] S4. The feature fusion module AFF based on Fourier transform is designed for the image segmentation network model. It reduces background noise and enhances edge contrast through linear combination in the frequency domain. At the same time, it further refines the texture by combining with multi-scale feature fusion at the decoding end.

[0010] S5 uses the constructed image segmentation network model to perform inference on the preprocessed UAV imagery and outputs segmentation results containing crack regions.

[0011] Furthermore, step S1 specifically includes the following steps:

[0012] S1.1 involves planning flight routes for drones over the target mining area, setting flight overlap to ensure complete image coverage and geometric consistency; then, through aerial triangulation, geometric correction is performed on the acquired images, and finally, orthophoto data is generated through image reconstruction.

[0013] S1.2, based on the preprocessed orthophoto data, the areas with ground fissures are first selected, and the images are locally cropped using the sliding window method, with the window size set. During the cropping process, image blocks without fissures are manually interpreted or removed, retaining only images containing fissure targets. The images are then scaled, and the fissures are annotated with pixel set vectors using Labelme annotation software. Finally, the vector annotations are converted into raster data of the fissures using a script to form training samples of fissures in the mining area.

[0014] Furthermore, the multi-scale feature perception module (MFS) in step S2 includes: a parallel dual-branch convolutional structure, where each branch uses convolutional kernels of different scales to extract features from the input features to capture the texture patterns of cracks at different scales, specifically:

[0015] ;

[0016] ;

[0017] ;

[0018] in, , These represent features at different scales. This represents the activation function. For batch normalization, This represents depthwise separable convolution. Represents two asymmetric convolution operators. Represents dilated convolution. This represents a 3×3 convolution layer. The output of the parallel two-branch convolutional structure, These are batch size, number of channels, height, and width, respectively. For input features.

[0019] Furthermore, the efficient multi-scale attention module EMA in step S2 achieves multi-scale enhancement in a cross-spatial interaction and lightweight manner through feature grouping, specifically including the following steps:

[0020] S2.1, the features output by MFS are used as input features; EMA divides the input features into G groups according to channels, with each group having a size of C / G, to obtain grouped features. EMA for each grouping feature Horizontal and vertical pooling is performed to extract the contextual information of the strip-shaped cracks in the transverse and longitudinal directions:

[0021] ;

[0022] in, and Grouping features Horizontal and vertical perception features. This is a one-dimensional horizontal global pooling operation. This is a one-dimensional vertical global pooling operation.

[0023] S2.2, will and Spatial attention weights are obtained by fusing the weights through 1×1 convolutions and mapping them to the [0,1] interval. Used for grouping features Perform pixel weighting to obtain weighted grouped features. :

[0024] ;

[0025] At the same time, Perform 3×3 convolution Extracting local boundaries and high-frequency features yields :

[0026] .

[0027] S2.3, EMA for features and characteristics Two attention weights are obtained by performing global average pooling and normalization respectively. and The final multi-scale attention weights are obtained by interacting with and activating the features of the two branches. :

[0028] .

[0029] S2.4, the obtained weights will be used Grouping features Reshaping yields grouped feature output. And concatenate all subgroups to obtain the final output. :

[0030] ;

[0031] in, The total number of subgroups. This is for splicing operations.

[0032] Furthermore, step S3, which involves selecting the TSCT module to construct the token context, specifically includes the following steps:

[0033] S3.1, use the final output features of EMA as the input features of the TSCT module. ,Will Simultaneously generate queries through linear mapping ,key AND value Three sets of vectors:

[0034] ;

[0035] in, It is a linear transformation.

[0036] S3.2, represented by the characteristics of the token As input, , Indicates the first The feature vector of each token is processed through two fully connected layers and an activation function to output the token weights. Weight To each , Modulate and with Multiplication yields a feature representation that incorporates the global context. ;

[0037] ;

[0038] in, , Modulated , , For the dimension of attention weights, It is a normalized exponential function.

[0039] S3.3, Select attention output in token selection Subsequently, an improved hybrid feedforward network was used to further enhance feature representation, expanding the channel dimension to four times the original number of channels, and introducing depthwise separable convolutions. Enhance spatial awareness, while coordinating with channel attention Channel recalibration of features:

[0040] ;

[0041] in, These are the characteristics after channel recalibration.

[0042] S3.4, will Input-gated linear unit Candidate features are divided into channels. With weight features ,in After activation, it is used as a gating weight, and... Element-by-element multiplication yields the final feature. :

[0043] .

[0044] Furthermore, step S3, constructing the Feature Modulation Module (FiLM), specifically includes the following steps:

[0045] S3.5, the feature modulation module uses mid-to-low layers Extract a conditional representation vector The modulation signal is used to guide the recalibration of deep features, and channel-level feature modulation parameters are predicted from conditional features through a learnable mapping function:

[0046] ;

[0047] in, and This represents a linear layer that performs channel-wise affine modulation on deep features:

[0048] ;

[0049] in, Representing the The spatial location is at the first The token sequence on each channel Representative and The corresponding modulated output characteristics, This is the scaling factor. This is a bias term.

[0050] Furthermore, step S4 specifically includes the following steps:

[0051] S4.1, AFF performs cross-filtering of two-path features in the spatial domain to control background noise introduced by single-path activation; through 1×1 convolution... and Sigmoid activation function Pixel-level weights are generated to modulate another branch of features:

[0052] ;

[0053] ;

[0054] in, Represents the upsampling features during the decoding stage. The skip connection feature representing the sibling encoder, and These represent the enhanced features after being guided by cross-branch features.

[0055] S4.2, the filtered features are fused to obtain spatial fusion features. :

[0056] ;

[0057] in, , , , These represent the batch size, number of channels, height, and width, respectively.

[0058] S4.3, after completing feature cross-fusion, the features Perform a two-dimensional Fourier transform ,Will Mapping to the frequency domain yields the real part of the complex spectrum. With the imaginary part And splice along the channel dimension. The frequency domain feature representation is obtained. :

[0059] ;

[0060] .

[0061] In the frequency domain, AFF independently applies a shared 1×1 convolution to each frequency bin. :

[0062] .

[0063] S4.4, after completing the frequency domain enhancement section, will The data is split along the channel dimension into new real and imaginary parts, and then subjected to inverse Fourier transform. Reconstructing the frequency domain features back into the spatial domain yields the features. :

[0064] ;

[0065] ;

[0066] in, This is a splitting operation.

[0067] S4.5, further stabilize the features by using a 1×1 convolution-normalization-activation function. Feature processing yields the final AFF output. :

[0068] .

[0069] The present invention has the following beneficial effects:

[0070] This invention achieves linear edge detail extraction of cracks by introducing Multi-Scale Perceptual (MFS) and Efficient Multi-Scale Attention (EMA) in the shallow feature extraction stage. MFS utilizes asymmetric and dilated convolutions to enhance the model's sensitivity to slender target structures and expand the receptive field. EMA employs adaptive 1×1 convolutions in both horizontal and vertical directions to capture cross-spatial dependencies, thereby strengthening the directional consistency and boundary identification ability of crack strip structures. Furthermore, in the deep feature extraction stage, this invention designs a token selection Transformer to focus on key locations that may carry edge and texture information, thereby enhancing the global consistency expression of discontinuous and weakly textured cracks. Simultaneously, a FiLM feature modulation module is introduced to linearly modulate deep semantic information with shallow texture information, mitigating the tendency of deep features to smooth fine-grained edges. In the decoding stage, a Fourier transform-based feature fusion module, AFF, is introduced. This module uses cross-feature fusion, frequency domain enhancement, and spatial refinement reconstruction as steps to simultaneously improve boundary clarity and crack connectivity without significantly increasing computation. Attached Figure Description

[0071] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0072] Figure 1 The diagram shows the structure of the image segmentation network model of the present invention.

[0073] Figure 2 Comparison of segmentation details of the image segmentation network model provided by this invention in crack images.

[0074] Figure 3 Images of selected cracks in a mining area awaiting extraction.

[0075] Figure 4 for Figure 3 Image showing the results of crack extraction in the mining area.

[0076] Figure 5 for Figure 3 A screenshot overlay of cracks in the mining area. Detailed Implementation

[0077] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0078] A method for extracting surface fractures in mining areas based on multi-scale global fusion, characterized by the following steps:

[0079] S1: Acquire orthophotos of the mining area using drones, and perform preprocessing, cropping, and enhancement on the images. The drone used is a high-resolution drone.

[0080] S2 constructs a multi-scale feature perception module (MFS) in the shallow feature extraction stage of the image segmentation network model and introduces an efficient multi-scale attention module (EMA) to capture multi-scale crack textures and suppress complex background interference.

[0081] S3 constructs the Token Selection Context Module (TSCT) in the deep feature modeling stage of the image segmentation network model, and uses the Feature Modulation Module (FiLM) to generate scaling and offset coefficients based on shallow features, performs channel linear modulation on deep features, and fuses local details with global semantics.

[0082] S4. The feature fusion module AFF based on Fourier transform is designed for the image segmentation network model. It reduces background noise and enhances edge contrast through linear combination in the frequency domain. At the same time, it further refines the texture by combining with multi-scale feature fusion at the decoding end.

[0083] S5 uses the constructed image segmentation network model to perform inference on the preprocessed UAV imagery and outputs segmentation results containing crack regions.

[0084] like Figure 1 As shown, the image segmentation network model includes a shallow feature extraction stage and a deep feature extraction stage. The features extracted in the shallow feature extraction stage serve as the input to the deep feature extraction stage. The shallow feature extraction stage includes two superimposed MFS and EMA layers. The deep feature extraction stage includes two superimposed FiLM and TSCT layers. In the decoding stage, the features extracted in the deep feature extraction stage are first processed by TSCT and depthwise separable convolution (DW Conv), and then by three superimposed AFF and TSCT layers to output the segmented image.

[0085] Specifically, step S1 includes the following steps:

[0086] S1.1, the UAV plans flight routes over the target mining area, sets reasonable heading overlap, and ensures complete image coverage and geometric consistency; through aerial triangulation, the acquired images are geometrically corrected, and finally orthophoto data with high spatial resolution and geographic matching accuracy is generated through image reconstruction.

[0087] S1.2, based on the preprocessed orthophoto data, the areas with ground fissures are first selected, and the image is locally cropped using the sliding window method, with the window size set to 512×512 pixels. During the cropping process, image blocks without fissures are manually identified or removed, retaining only images containing fissure targets. The image is then scaled to 256×256 pixels, and the fissures are annotated with pixel set vectors using Labelme annotation software, setting the pixel value of the fissure area to 255 and the background area to 0. Finally, the vector annotations are converted into raster data of the fissures using a script to form training samples of fissures in the mining area.

[0088] Specifically, the Multi-Scale Feature Awareness (MFS) module in step S2 is used to perform multi-scale structured modeling of the input image features; it includes: a parallel dual-branch convolutional structure, where each branch uses convolutional kernels of different scales to extract features from the input features in order to capture the texture patterns of cracks at different scales, specifically:

[0089] By introducing 1×5 and 5×1 convolutional kernels into the multi-scale feature perception module, the network's directional response to slender and narrow cracks can be enhanced, thereby improving its sensitivity to linear structural features. Furthermore, by setting dilated convolutions with an inflation rate of 3, the network's receptive field can be effectively expanded, enabling it to integrate a wider range of contextual information and avoid missing large-scale cracks or sparsely distributed crack edge features. At the same time, introducing depthwise separable convolution (DSConv) and combining it with residual connection mechanisms in the convolutional branches reduces the number of convolutional operation parameters and overall computational complexity while ensuring the stability of network training and significantly enhancing the representation ability of shallow features.

[0090] ;

[0091] ;

[0092] ;

[0093] in, , These represent features at different scales. This represents the activation function. For batch normalization, This represents depthwise separable convolution. Represents two asymmetric convolution operators. Represents dilated convolution. This represents a 3×3 convolution layer. The output of the parallel two-branch convolutional structure, These are batch size, number of channels, height, and width, respectively. For input features.

[0094] Specifically, the efficient multi-scale attention module EMA in step S2 achieves multi-scale enhancement in a lightweight manner through feature grouping, enabling cross-spatial interaction. This includes the following steps:

[0095] S2.1, the features output by MFS are used as input features; EMA divides the input features into G groups according to channels, with each group having a size of C / G, to obtain grouped features. EMA for each grouping feature Horizontal and vertical pooling is performed to extract the contextual information of the strip-shaped cracks in the transverse and longitudinal directions:

[0096] ;

[0097] in, and Grouping features Horizontal and vertical perception features. This is a one-dimensional horizontal global pooling operation. This is a one-dimensional vertical global pooling operation.

[0098] S2.2, will and Spatial attention weights are obtained by fusing the weights through 1×1 convolutions and mapping them to the [0,1] interval. Used for grouping features Perform pixel weighting to obtain weighted grouped features. :

[0099] ;

[0100] At the same time, Perform 3×3 convolution Extracting local boundaries and high-frequency features yields :

[0101] .

[0102] S2.3, To further capture cross-space dependencies, EMA applies features... and characteristics Two attention weights are obtained by performing global average pooling and normalization respectively. and The final multi-scale attention weights are obtained by interacting with and activating the features of the two branches. :

[0103] .

[0104] S2.4, the obtained weights will be used Grouping features Reshaping yields grouped feature output. And concatenate all subgroups to obtain the final output. :

[0105] ;

[0106] in, The total number of subgroups. This is for splicing operations.

[0107] By employing a grouping strategy, EMA effectively reduces the computational burden; through spatial gating and channel interaction, it enhances the multi-scale representation capability of slender cracks and improves the clarity and connectivity of boundaries.

[0108] Specifically, step S3, which involves constructing the Token Context Selection (TSCT) module, is used to capture long-range dependencies in low-resolution, high-semantic features and to achieve global context modeling. This includes the following steps:

[0109] S3.1, use the final output features of EMA as the input features of the TSCT module. ,Will Simultaneously generate queries through linear mapping ,key AND value Three sets of vectors:

[0110] ;

[0111] in, It is a linear transformation.

[0112] S3.2, represented by the characteristics of the token As input, , Indicates the first The feature vector of each token is processed through two fully connected layers and an activation function to output the token weights. Weight To each , Modulate and with Multiplication yields a feature representation that incorporates the global context. ;

[0113] ;

[0114] in, , Modulated , , For the dimension of attention weights, It is a normalized exponential function.

[0115] S3.3, Select attention output in token selection Subsequently, an improved hybrid feedforward network was used to further enhance feature representation, expanding the channel dimension to four times the original number of channels, and introducing depthwise separable convolutions. Enhance spatial awareness, while coordinating with channel attention Channel recalibration of features:

[0116] ;

[0117] in, These are the characteristics after channel recalibration.

[0118] S3.4, will Input-gated linear unit Candidate features are divided into channels. With weight features ,in After activation, it is used as a gating weight, and... Element-by-element multiplication yields the final feature. :

[0119] .

[0120] Specifically, step S3, which involves constructing the Feature Modulation Module (FiLM), is used to achieve adaptive fusion of shallow detail features and deep semantic features and dynamically adjust the channel response. This includes the following steps:

[0121] S3.5, the feature modulation module uses mid-to-low layers Extract a conditional representation vector The modulation signal is used to guide the recalibration of deep features, and channel-level feature modulation parameters are predicted from conditional features through a learnable mapping function:

[0122] ;

[0123] in, and This represents a linear layer used to generate corresponding scaling factors and bias terms for each channel, performing channel-wise affine modulation on deep features:

[0124] ;

[0125] in, Representing the The spatial location is at the first The token sequence on each channel Representative and The corresponding modulated output characteristics, This is the scaling factor. This is a bias term.

[0126] Specifically, step S4 includes the following steps:

[0127] The decoding end introduces a feature fusion module AFF based on Fourier transform. This module includes: cross feature fusion, frequency domain enhancement, and spatial refinement reconstruction, which improves boundary clarity and crack connectivity without significantly increasing computation.

[0128] Given the upsampling features of the decoding stage and skip connection features from peer encoders The overall mapping relationship of AFF is represented as follows:

[0129] .

[0130] S4.1, before entering the frequency domain enhancement, AFF performs cross-filtering of the two features in the spatial domain to control the background noise introduced by single-path activation; specifically, through 1×1 convolution... and Sigmoid activation function Pixel-level weights are generated to modulate another branch of features:

[0131] ;

[0132] ;

[0133] in, Represents the upsampling features during the decoding stage. The skip connection feature representing the sibling encoder, and These represent the enhanced features after being guided by cross-branch features;

[0134] S4.2, the filtered features are fused to obtain spatial fusion features. :

[0135] ;

[0136] in, , , , These represent the batch size, number of channels, height, and width, respectively.

[0137] This cross-selection mechanism ensures that only spatial locations that are simultaneously activated in both feature paths are significantly enhanced, thereby achieving initial suppression of background noise in the spatial domain.

[0138] S4.3, after completing feature cross-fusion, the features Perform a two-dimensional Fourier transform ,Will Mapping to the frequency domain yields the real part of the complex spectrum. With the imaginary part And splice along the channel dimension. The frequency domain feature representation is obtained. :

[0139] ;

[0140] .

[0141] In the frequency domain, AFF independently applies a shared 1×1 convolution to each frequency bin. This step is used to linearly recombine the real and imaginary channels. It is important to emphasize that this step only linearly mixes the channel features within each frequency bin, preserving the frequency locality of the Fourier representation.

[0142] .

[0143] S4.4, after completing the frequency domain enhancement section, will The data is split along the channel dimension into new real and imaginary parts, and then subjected to inverse Fourier transform. Reconstructing the frequency domain features back into the spatial domain yields the features. :

[0144] ;

[0145] ;

[0146] in, This is a splitting operation.

[0147] S4.5, further stabilize the features by using a 1×1 convolution-normalization-activation function. Feature processing yields the final AFF output. :

[0148] .

[0149] Step S5 specifically involves: inputting training samples from the training dataset into the model, setting the initial learning rate of the model to 0.0001, and introducing a fusion method of Focal and Tversky loss (Focal-Tversky loss) to effectively address the problem of uneven distribution between the target and background. The Adam optimizer is then used to train the model, with a batch size of 8, for 100 training epochs to obtain the model's training weights.

[0150] Large-scale target mining area surface UAV imagery is fed into a trained model in the form of slices. Each slice is then stitched together after effective crack prediction to obtain complete crack extraction for the large-scale scene.

[0151] like Figure 2 As shown, the image results obtained by extracting surface cracks in mining areas using the model provided by this invention have clear outlines and complete structures, accurately representing the overall morphology and spatial orientation of the cracks. Figure 3 , Figure 4 and Figure 5 As shown, Figure 3 For the selected cracks in the mining area image to be extracted, in Figure 4 and Figure 5 In large-scale mining area surface images, the model of this invention can identify and continuously extract surface cracks that are distributed across regions and on a long scale. The cracks maintain good connectivity and consistency in the large-scale scene, and no obvious missed detections are observed, indicating that the method provided by this invention has good global perception capability and scale adaptability.

[0152] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for extracting surface fractures in mining areas based on multi-scale global fusion, characterized in that, Specifically, the steps include the following: S1: Acquire orthophotos of the mining area from UAVs, and perform preprocessing, cropping, and enhancement on the images; S2, a multi-scale feature perception module MFS is constructed in the shallow feature extraction stage of the image segmentation network model, and an efficient multi-scale attention module EMA is introduced to capture multi-scale crack textures and suppress complex background interference. S3. In the deep feature modeling stage of the image segmentation network model, the Token Selection Context Module (TSCT) is constructed, and the Feature Modulation Module (FiLM) is used to generate scaling and offset coefficients based on shallow features, perform channel linear modulation on deep features, and fuse local details with global semantics. S4. Design a Fourier transform-based feature fusion module (AFF) for the image segmentation network model. It reduces background noise and enhances edge contrast through linear combination in the frequency domain. At the same time, it further refines the texture by combining with multi-scale feature fusion at the decoding end. S5 uses the constructed image segmentation network model to perform inference on the preprocessed UAV imagery and outputs segmentation results containing crack regions. The Multi-Scale Feature Awareness (MFS) module includes: a parallel bi-branch convolutional structure that uses asymmetric convolution and dilated convolution to enhance the model's sensitivity to slender target structures and expand the receptive field; EMA that captures cross-spatial dependencies by using adaptive 1×1 convolutions in the horizontal and vertical directions through feature grouping; a TSCT module that captures long-range dependencies in low-resolution, high-semantic features and achieves global context modeling; and AFF that includes: cross-feature fusion, frequency domain enhancement, and spatial refinement reconstruction.

2. The method for extracting surface fractures in mining areas based on multi-scale global fusion according to claim 1, characterized in that, Step S1 specifically includes the following steps: S1.1, the flight mission of UAVs is planned in the target mining area, and the flight path overlap is set to ensure complete coverage and geometric consistency of the image; the acquired image is geometrically corrected by aerial triangulation, and finally orthophoto data is generated by image reconstruction. S1.2, based on the preprocessed orthophoto data, the areas with ground fissures are first selected, and the images are locally cropped using the sliding window method, with the window size set. During the cropping process, image blocks without fissures are manually interpreted or removed, retaining only images containing fissure targets. The images are then scaled, and the fissures are annotated with pixel set vectors using Labelme annotation software. Finally, the vector annotations are converted into raster data of the fissures using a script to form training samples of fissures in the mining area.

3. The method for extracting surface fractures in mining areas based on multi-scale global fusion according to claim 1, characterized in that, The multi-scale feature perception module (MFS) in step S2 includes: a parallel dual-branch convolutional structure, where each branch uses convolutional kernels of different scales to extract features from the input features to capture the texture patterns of cracks at different scales, specifically: ； ； ； in, , These represent features at different scales. This represents the activation function. For batch normalization, This represents depthwise separable convolution. Represents two asymmetric convolution operators. Represents dilated convolution. This represents a 3×3 convolution layer. The output of the parallel two-branch convolutional structure, As input features, For splicing operations, These are batch size, number of channels, height, and width, respectively.

4. The method for extracting surface fractures in mining areas based on multi-scale global fusion according to claim 1, characterized in that, The efficient multi-scale attention module EMA in step S2 achieves multi-scale enhancement in a lightweight manner through feature grouping, enabling cross-spatial interaction. Specifically, it includes the following steps: S2.1, use the features output by MFS as input features; EMA divides input features by channel. Groups, each group is of size ,in, Given the number of channels, we obtain the grouping features. EMA for each grouping feature Horizontal and vertical pooling is performed to extract the contextual information of the strip-shaped cracks in the transverse and longitudinal directions: ； in, and Grouping features Horizontal and vertical perception features. This is a one-dimensional horizontal global pooling operation. This is a one-dimensional vertical global pooling operation; S2.2, will and Spatial attention weights are obtained by fusing the weights through 1×1 convolutions and mapping them to the [0,1] interval. Used for grouping features Perform pixel weighting to obtain weighted grouped features. : ； At the same time, Perform 3×3 convolution Extracting local boundaries and high-frequency features yields : ； S2.3, EMA for features and characteristics Two attention weights are obtained by performing global average pooling and normalization respectively. and The final multi-scale attention weights are obtained by interacting with and activating the features of the two branches. : ； S2.4, the obtained weights will be used Grouping features Reshaping yields grouped feature output. And concatenate all subgroups to obtain the final output. : ； in, The total number of subgroups. For splicing operations, These are batch size, number of channels, height, and width, respectively.

5. The method for extracting surface fractures in mining areas based on multi-scale global fusion according to claim 1, characterized in that, Step S3, which involves selecting the TSCT module to build the token context, includes the following steps: S3.1, use the final output features of EMA as the input features of the TSCT module. ,Will Simultaneously generate queries through linear mapping ,key AND value Three sets of vectors: ； in, For linear transformation, These are batch size, number of channels, height, and width, respectively. S3.2, represented by the characteristics of the token As input, , Indicates the first The feature vector of each token is processed through two fully connected layers and an activation function to output the token weights. Weight To each , Modulate and with Multiplication yields a feature representation that incorporates the global context. ; ； in, , Modulated , , For the dimension of attention weights, It is a normalized exponential function; S3.3, Select attention output in token selection Subsequently, an improved hybrid feedforward network was used to further enhance feature representation, expanding the channel dimension to four times the original number of channels, and introducing depthwise separable convolutions. Enhance spatial awareness, while coordinating with channel attention Channel recalibration of features: ； in, These are the characteristics of the channel after recalibration; S3.4, will Input-gated linear unit Candidate features are divided into channels. With weight features ,in After activation, it is used as a gating weight, and... Element-by-element multiplication yields the final feature. : 。 6. The method for extracting surface fractures in mining areas based on multi-scale global fusion according to claim 1, characterized in that, Step S3, constructing the Feature Modulation Module (FiLM), specifically includes the following steps: S3.5, the feature modulation module uses mid-to-low layers Extract a conditional representation vector The modulation signal is used to guide the recalibration of deep features, and channel-level feature modulation parameters are predicted from conditional features through a learnable mapping function: ； in, and This represents a linear layer that performs channel-wise affine modulation on deep features: ； in, Representing the The spatial location is at the first The token sequence on each channel Representative and The corresponding modulated output characteristics, This is the scaling factor. This is a bias term.

7. The method for extracting surface fractures in mining areas based on multi-scale global fusion according to claim 1, characterized in that, Step S4 specifically includes the following steps: S4.1, AFF performs cross-filtering of two-path features in the spatial domain to control background noise introduced by single-path activation; through 1×1 convolution... and Sigmoid activation function Pixel-level weights are generated to modulate another branch of features: ； ； in, Represents the upsampling features during the decoding stage. The skip connection feature representing the sibling encoder, and These represent the enhanced features after being guided by cross-branch features; S4.2, the filtered features are fused to obtain spatial fusion features. : ； in, , , , These represent batch size, number of channels, height, and width, respectively. S4.3, after completing feature cross-fusion, Perform a two-dimensional Fourier transform ,Will Mapping to the frequency domain yields the real part of the complex spectrum. With the imaginary part And splice along the channel dimension. The frequency domain feature representation is obtained. : ； ； In the frequency domain, AFF independently applies a shared 1×1 convolution to each frequency bin. : ； S4.4, after completing the frequency domain enhancement section, will The data is split along the channel dimension into new real and imaginary parts, and then subjected to inverse Fourier transform. Reconstructing the frequency domain features back into the spatial domain yields the features. : ； ； in, For splitting operations; S4.5, further stabilize the features by using a 1×1 convolution-normalization-activation function. Feature processing yields the final AFF output. : 。

Images