A Smart Identification and Quantification Method for Surface Fractures in Mining Areas Based on MDE-UNet
By using the MDE-UNet network model and skeletonization algorithm, the problems of low efficiency of manual inspection and insufficient accuracy of crack extraction in surface crack monitoring in mining areas have been solved. High-precision crack segmentation and automatic acquisition of geometric parameters have been achieved, improving the efficiency and safety of mine geological disaster monitoring and ecological environment management.
Patent Information
- Application Number
- CN202611331614.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-31
- Publication Date
- 2026-11-10
AI Technical Summary
Existing methods for monitoring surface cracks in mining areas suffer from problems such as low efficiency of manual inspections, insufficient crack extraction accuracy under complex backgrounds, difficulty in automatically acquiring crack geometric parameters, and fixed receptive field of segmentation network scale, making it difficult to achieve high-precision crack segmentation and automatic acquisition of geometric parameters.
A method for intelligent identification and quantification of surface cracks in mining areas based on MDE-UNet is adopted. By constructing attention convolutional blocks (ACB), multi-scale attention interaction modules (MAIM), and attention-guided dense jump connection modules (ADSC), and combining UAV remote sensing imagery and skeletonization algorithms, high-precision automatic extraction of cracks and automatic acquisition of geometric parameters are achieved.
It improves the ability to accurately identify surface cracks in complex mining areas, realizes automated measurement of crack geometric parameters, enhances monitoring efficiency and accuracy, reduces manual inspection costs and safety risks, and provides high-precision data support.
Smart Images

Figure CN123071876A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of UAV remote sensing and mine geological disaster monitoring technology, specifically to a method for intelligent identification and quantification of surface cracks in mining areas based on MDE-UNet. Background Technology
[0002] Large-scale, high-intensity mining of underground coal resources disrupts the original stress balance of the overburden, triggering a series of geological disasters such as overburden movement, surface subsidence, and fissure development. Surface fissures, as the most direct manifestation of mining damage, not only damage land integrity and the surface ecological environment but may also cause deformation or even damage to infrastructure such as roads, buildings, power lines, and drainage facilities, posing a serious threat to mine safety and ecological environmental protection. Therefore, rapid identification, accurate extraction, and quantitative analysis of surface fissures in mining areas are of great significance for mine disaster early warning, land reclamation planning, and ecological environment restoration.
[0003] Traditional surface crack surveys primarily rely on manual field measurements, total station measurements, and GPS positioning. While these methods offer high measurement accuracy, they suffer from low efficiency, high labor intensity, limited monitoring range, and high safety risks, making them unsuitable for the rapid monitoring and dynamic assessment of surface cracks in large mining areas. With the development of UAV remote sensing technology, high-resolution aerial imagery can quickly acquire information about large areas of the surface, providing a new data source for the automatic identification of cracks in mining areas.
[0004] In recent years, deep learning technology has been widely applied in fields such as road crack detection, concrete crack recognition, and geological hazard interpretation. Semantic segmentation models, represented by FCN, U-Net, the DeepLab series, and the Transformer structure, can automatically extract crack targets. However, due to the complex surface environment of mining areas, influenced by factors such as vegetation cover, shadow occlusion, bare rock texture, and soil background interference, surface cracks typically exhibit characteristics such as being long and thin, curved, discontinuous, and having significant scale differences. Existing segmentation models still suffer from problems in complex mining scenarios, including blurred crack edges, missed detection of small cracks, insufficient multi-scale crack recognition capabilities, and a high false detection rate in the background, making it difficult to accurately characterize the crack geometry. Furthermore, existing research mostly focuses on pixel-level segmentation results of crack regions, with relatively little research on the automatic acquisition of geometric parameters such as crack length, width, and area. Traditional methods for acquiring crack geometric parameters typically rely on manual measurement or post-processing. Furthermore, width calculations often employ a uniform Euclidean distance transform half-width multiplied by an empirical coefficient, neglecting measurement errors at crack bifurcation and endpoints. Scale conversions often use a uniform ground sampling distance across the entire image, failing to account for local spatial resolution differences in UAV imagery caused by terrain undulations and flight yaws. Moreover, the scale receptive field of the segmentation network in existing methods is fixed during the inference phase, unable to adaptively adjust based on the actual scale distribution of cracks in the image to be detected. How to achieve high-precision crack segmentation, further automatically acquire crack geometric parameters, and enable the segmentation network to adaptively adjust multi-scale feature fusion strategies according to the crack geometry has become a key technical problem to be solved in the field of intelligent monitoring of surface cracks in mining areas.
[0005] Therefore, there is an urgent need to propose a new intelligent identification method for surface cracks in mining areas that can adapt to the complex background environment of mining areas and take into account both fine crack structure extraction and automatic quantitative analysis of geometric parameters, so as to improve the crack detection accuracy and automation level and provide reliable data support for mine geological disaster monitoring and early warning and ecological environment governance. Summary of the Invention
[0006] Purpose of the invention: To address the problems of low efficiency of manual inspection, insufficient accuracy of crack extraction under complex backgrounds, difficulty in automatically acquiring crack geometric parameters, and fixed receptive field of segmentation network scale in the existing monitoring of surface cracks in mining areas, this invention proposes an intelligent identification and quantification method for surface cracks in mining areas based on MDE-UNet.
[0007] Technical Solution: To achieve the above objectives, the technical solution adopted by this invention is: a method for intelligent identification and quantification of surface cracks in mining areas based on MDE-UNet, comprising the following steps: S1. Acquire UAV remote sensing images of the mining area, construct the Crack-UAV crack dataset, and train the MDE-UNet network model; the MDE-UNet network model includes attention convolutional block ACB, multi-scale attention interaction module MAIM, and attention-guided dense jump connection module ADSC. S2. Input the image to be detected into the MDE-UNet network model for the first forward propagation. MAIM fuses multi-scale features with initial weights and obtains a coarse crack segmentation probability map through block probability fusion. S3. Read the TFW file, establish the affine transformation relationship from pixel coordinates to projected coordinates, and coarsely segment the crack probability. Figure 2 After valueization, connected component analysis is performed; S4. Using the crack coarse segmentation probability map as a confidence guide, skeletonization is performed. Low-confidence short branches are pruned according to the average branch confidence to obtain the pruned crack skeleton. The skeleton pixels are then classified into endpoints, ordinary points, and bifurcation points. S5. Search the confidence boundary along the skeleton normal direction for ordinary points to calculate the normal width. Take the median width of the neighboring ordinary points for the bifurcation point. Take twice the Euclidean distance transformation value for the endpoints. Convert the width point by point to the physical width based on the affine transformation relationship. S6. Generate the scale modulation coefficients of MAIM based on the physical width distribution of the crack, load them into the MDE-UNet network model for the second forward propagation, and use the scale modulation coefficients to replace the initial weights of MAIM for multi-scale fusion to obtain the crack binary segmentation mask. S7. Based on the binary segmentation mask of the crack, re-execute S4 to S5, calculate the actual length, width and area of each crack connected domain in the projected coordinate system, and output the quantification results of crack geometric parameters.
[0008] Preferably, the MDE-UNet network model in step S1 adopts an encoder-decoder structure, where the encoder is used to extract crack features at different levels and the decoder is used to restore the crack spatial resolution. An attention convolution block ACB is embedded in the encoder, a multi-scale attention interaction module MAIM is embedded in the deep layers of the network, and an attention-guided dense jump connection module ADSC is set between the encoder and the decoder.
[0009] Preferably, the attention convolution block ACB in step S1 includes a deformable convolution module, a channel attention module, and a spatial attention module; wherein the deformable convolution module achieves adaptive perception of irregular boundaries of cracks and complex topological structures by dynamically learning the convolution sampling offset.
[0010] Preferably, the Multi-Scale Attention Interaction Module (MAIM) in step S1 constructs feature branches at the original scale, half-scale, and quarter-scale. It achieves spatial alignment of feature maps at different scales through bilinear interpolation and uses an attention mechanism to dynamically fuse multi-scale features, thereby realizing complementary expression of crack features at different scales.
[0011] Preferably, the Attention-Guided Dense Jump Connection Module (ADSC) in step S1 uses decoder features to guide encoder features to generate an attention weight map, filters encoder output features, retains only the effective information related to crack identification, and then fuses it with decoder features to reduce background noise interference.
[0012] Preferably, in S4, an independent branch is formed by tracing from one end of the skeleton to another end or a bifurcation point. The bifurcation point is a common node and does not participate in pruning. When the branch length is less than the branch length threshold and the average confidence of the branch is less than the confidence threshold, the entire branch is pruned. The iteration continues until there are no branches left to prune.
[0013] As a preferred embodiment, in S5, the tangent of a common point is determined by the sum of the direction vectors of its skeleton neighbors, and the normal is obtained by rotating it by 90 degrees. The search is performed along both sides of the normal by a unit step size until the confidence level is lower than the boundary confidence threshold. The sum of the step sizes on both sides is taken as the normal width. When the normal width exceeds a preset multiple of twice the Euclidean distance transformation value, it is replaced by twice the Euclidean distance transformation value. For a bifurcation point, the median of the normal widths of common points in its neighborhood is taken. When there are no common points in the neighborhood, twice the Euclidean distance transformation value is taken.
[0014] Preferably, the average crack width obtained in S5 is calculated in S6. mean w ,by w 0 =k·GSD Calculate the normalized width deviation using the reference width as the reference width. t=(mean w -w 0 ) / w The modulation coefficients of three scale branches are generated by the softmax function after the amplitude is limited to 0. When t is less than zero, the original scale weight is increased, and when t is greater than zero, the scale weight is increased by one-quarter.
[0015] As a preferred embodiment, in S7, adjacent pixels of the skeleton are transformed to projected coordinates through affine transformation, and the Euclidean distance is calculated segment by segment and summed to obtain the actual length; the actual area is obtained by multiplying the number of connected pixels by the square of the geometric mean of the pixel size in the x and y directions.
[0016] Beneficial effects: I. Benefits of Core Technology Innovation 1. Improve the ability to accurately identify surface cracks in complex mining areas: This invention proposes an intelligent identification and quantification method for surface cracks in mining areas based on MDE-UNet. By constructing attention convolutional blocks (ACB), multi-scale attention interaction modules (MAIM), and attention-guided dense jump connection modules (ADSC), it effectively solves the problems of blurred crack boundaries, missed detection of small cracks, difficulty in simultaneously identifying multi-scale cracks, and severe background noise interference in complex mining scenes, and achieves high-precision automatic extraction of surface cracks in complex environments.
[0017] 2. In this invention, after obtaining the coarse segmentation probability map of the crack during the first forward propagation, the crack width is estimated first, and then the multi-scale fusion weights of MAIM are adjusted based on the width estimation result for the second forward propagation. In this way, the network can automatically adjust the emphasis on different scale features according to the actual width of the crack in the current image during inference, instead of using fixed scale weights during the inference stage as in existing methods.
[0018] 3. Establish an integrated technical system for crack identification and engineering parameter quantification: Unlike existing research that only focuses on crack region segmentation results, this invention, based on high-precision crack segmentation, further integrates skeletonization algorithm and Euclidean distance transformation algorithm to achieve automatic acquisition of geometric parameters such as crack length, width and area. It establishes an integrated technical process from crack image recognition to engineering parameter calculation, realizing the direct conversion of crack monitoring results into engineering evaluation indicators.
[0019] II. Benefits of High-Precision Identification and Automatic Quantization 1. Improved accuracy of surface crack identification in mining areas: On the self-constructed Crack-UAV crack dataset, the MDE-UNet model proposed in this invention demonstrates excellent performance in terms of precision, recall, F1 score, and intersection-over-union ratio (IoU). The F1 score reaches 86.15%, and the IoU reaches 76.11%, outperforming many mainstream crack segmentation models. Especially under conditions of vegetation obstruction, shadow interference, exposed rock and soil, and complex texture backgrounds, it can still maintain the continuity and integrity of the crack structure, effectively reducing missed detections and false detections.
[0020] 2. Enhanced Multi-Scale Crack Recognition Capability and Robustness: Through a dynamic scale selection mechanism and cross-scale feature interaction strategy, this invention can simultaneously consider the feature representation of both fine and large cracks, achieving stable recognition of crack targets at different scales. In scenarios with complex topologies, intersecting cracks, and multi-branched cracks, it can accurately maintain the continuity and spatial structural integrity of the crack network, improving the model's adaptability in complex mining environments.
[0021] 3. Automated Measurement of Crack Geometric Parameters: This invention utilizes a skeletonization algorithm to extract the crack centerline and combines it with Euclidean distance transformation to calculate the local crack width, thereby automatically calculating crack length, average width, maximum width, and area parameters. Compared with traditional manual measurement methods, this significantly improves the efficiency of crack parameter acquisition, reduces human error, and achieves non-contact, non-destructive, and highly efficient measurement.
[0022] III. Engineering Application Benefits 1. Enhance the monitoring and early warning capabilities for geological disasters in mines: This invention can quickly obtain the spatial distribution and geometric parameter changes of surface cracks in mining areas, providing high-precision data support for the analysis of mining crack development patterns, slope stability evaluation, surface subsidence monitoring, and early warning of geological disasters in mines. It helps to promptly identify potential risk areas and improve the level of safe production in mining areas.
[0023] 2. Reduced Costs and Safety Risks of Manual Inspections: Traditional mine crack surveys typically rely on manual field inspections and on-site measurements, which are not only inefficient but also pose significant safety risks. This invention utilizes UAV remote sensing and intelligent identification technology to achieve large-scale automatic monitoring of mine cracks, reducing personnel input and on-site operational risks, and improving monitoring efficiency and automation levels.
[0024] 3. Serving land reclamation and ecological environment restoration: The spatial distribution and development characteristics of cracks obtained by this invention can provide a scientific basis for land reclamation planning, vegetation restoration design and ecological environment management in mining areas, and help improve the pertinence and implementation effect of ecological restoration work in mining areas.
[0025] IV. Benefits of Industry Promotion and Application 1. Constructing a new intelligent monitoring model for cracks in mining areas: This invention allows for the input of UAV images and the output of crack length, width, and area without manual intervention. It can be directly used for regular inspections and comparative analysis of cracks in mining areas, realizing the transformation of surface crack monitoring in mining areas from traditional manual investigation to intelligent and automated monitoring mode, and providing a new technical route for monitoring geological disasters in mines.
[0026] 2. Excellent cross-scenario applicability: This invention is not only applicable to surface crack monitoring in coal mining areas, but can also be extended to road crack detection, slope crack monitoring, building structure crack detection, and other linear defect identification scenarios. Validation results on public datasets such as Crack500, DeepCrack, and CrackForest demonstrate that this invention has good generalization ability and promising engineering application prospects.
[0027] 3. Promoting the intelligent and digital construction of mines: This invention combines deep learning, UAV remote sensing and mine disaster monitoring to realize the automatic acquisition and intelligent analysis of crack information in the mining area. It can provide key technical support for the construction of smart mines, digital mine management and mine safety supervision, and has high promotion value and application prospects. Attached Figure Description
[0028] Figure 1 This is a flowchart illustrating the overall process of an intelligent identification and quantification method for surface cracks in mining areas based on MDE-UNet, as described in this invention. Figure 2 This is a diagram of the MDE-UNet network structure of a method for intelligent identification and quantification of surface cracks in mining areas based on MDE-UNet, as described in this invention. Figure 3 This is a diagram of the attention convolutional block (ACB) structure of the intelligent identification and quantification method for surface cracks in mining areas based on MDE-UNet according to the present invention. Figure 4 This is a structural diagram of the multi-scale attention interaction module (MAIM) of the intelligent identification and quantification method for surface cracks in mining areas based on MDE-UNet, as described in this invention. Figure 5 This is a structural diagram of the Attention-Guided Dense Jump Connection Module (ADSC) of the intelligent identification and quantification method for surface cracks in mining areas based on MDE-UNet of the present invention. Figure 6 This is a comparison of the results of different models of the intelligent identification and quantification method for surface cracks in mining areas based on MDE-UNet according to the present invention; Figure 7 This image shows the results of crack extraction in a large-scale mining scene, based on the intelligent identification and quantification method for surface cracks in mining areas according to the present invention (MDE-UNet). Figure 8 This is a local area crack extraction and visualization detail result diagram of the intelligent identification and quantification method for surface cracks in mining areas based on MDE-UNet of the present invention; Figure 9 This is a generalization experiment result of the intelligent identification and quantification method for surface cracks in mining areas based on MDE-UNet according to the present invention. Figure 10 This is a field measurement image of the intelligent identification and quantification method for surface cracks in mining areas based on MDE-UNet, as described in this invention.
[0029] Figure 11 This image shows the crack length, width, and related statistical analysis results of the research area for the intelligent identification and quantification method of surface cracks in mining areas based on MDE-UNet, as presented in this invention.
[0030] Table 1 shows the performance comparison of different crack segmentation models of the intelligent identification and quantification method for surface cracks in mining areas based on MDE-UNet in this invention on the self-constructed Crack-UAV crack dataset. Table 2 shows the ablation experiment results of the intelligent identification and quantification method for surface cracks in mining areas based on MDE-UNet according to the present invention on the Crack-UAV crack dataset.
[0031] Table 3 shows the generalization experimental results of the intelligent identification and quantification method for surface cracks in mining areas based on MDE-UNet in this invention on the Crack500, DeepCrack, and CrackForest datasets.
[0032] Table 4 shows the verification results of the geometric parameter quantization accuracy of the intelligent identification and quantification method for surface cracks in mining areas based on MDE-UNet according to the present invention. Detailed Implementation
[0033] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0034] like Figure 1 As shown, this invention provides an intelligent identification and quantification method for surface cracks in mining areas based on MDE-UNet, comprising the following steps: acquiring UAV remote sensing images of the mining area and constructing a Crack-UAV crack dataset, and training an MDE-UNet network model; inputting the image to be detected into the model for the first forward propagation to obtain a coarse crack segmentation probability map; reading the TFW file to establish affine transformation relations and performing connected component analysis; using the coarse segmentation probability map as confidence level to guide skeletonization, branch pruning, and node classification; measuring the width according to node type differences and converting it into physical width; generating MAIM scale modulation coefficients based on the width distribution and performing a second forward propagation to obtain a binary crack segmentation mask; re-skeletonizing based on the binary crack segmentation mask and calculating the crack length, width, and area in the projected coordinate system.
[0035] 1. Acquisition of fracture images in mining areas and construction of the Crack-UAV fracture dataset In this embodiment, a high-resolution visible light camera mounted on a UAV is first used to conduct low-altitude aerial surveys of the coal mining area to acquire remote sensing images of the mining area's surface. The image acquisition area includes coal mining subsidence areas, the edges of goaf areas, fractured surface areas, and areas with obvious crack development. The acquired scenes cover different seasons, different lighting conditions, different vegetation cover levels, and different surface texture backgrounds to enhance the representativeness of the samples for the complex mining environment. Specifically, this embodiment uses Tangjiahui Coal Mine, Wangjiata Coal Mine, and Bojianghaizi Coal Mine as data source areas, and the acquired images cover typical interference conditions such as vegetation occlusion, snow cover, gravel accumulation, exposed soil, and semi-desertified surfaces. The acquired UAV remote sensing images are preprocessed, including image cropping, size unification, outlier removal, brightness normalization, and format conversion. Subsequently, surface cracks in the images are pixel-level labeled, with cracked areas marked as foreground areas and non-cracked areas marked as background areas, resulting in crack semantic segmentation labels. In this embodiment, the constructed mining area crack dataset is named Crack-UAV, containing 12,608 images. Each image contains at least one visible surface crack, and the dataset is divided into training, validation, and test sets in an 8:1:1 ratio. To reduce the risk of overfitting and improve generalization ability, data augmentation is performed on the samples before training. This data augmentation includes random rotation, random flipping, random cropping, scaling, color perturbation, brightness perturbation, and contrast adjustment.
[0036] 2. Training the MDE-UNet crack segmentation network After completing the dataset construction, build the MDE-UNet crack segmentation network model. For example... Figure 2 As shown, the MDE-UNet network model adopts an encoder-decoder structure. The encoder uses a ResNet50 pre-trained on ImageNet as its backbone network and consists of five stages, with the number of output channels for each stage being 64, 256, 512, 1024, and 2048, respectively. A 1×1 convolutional channel adaptation layer is set after each stage of the encoder to uniformly adjust the number of channels to the number required by the decoder. The decoder restores the spatial resolution of the feature map through progressive upsampling. ACB is embedded in each stage of the encoder to enhance crack boundaries and complex topological features; MAIM is set between the deepest layer of the encoder and the decoder for dynamic fusion of multi-scale crack features; ADSC is set on each skip connection path between the encoder and the decoder to filter and fuse features from deep and shallow layers.
[0037] like Figure 3As shown, the ACB includes a deformable convolution module, a channel attention module, and a spatial attention module. The deformable convolution module dynamically adjusts the convolution sampling position according to the crack boundary morphology, enabling the convolution kernel sampling points to adapt to the geometric changes of slender cracks, curved cracks, bifurcation cracks, and discontinuous cracks. The deformable convolution module employs a two-stage deformable convolution structure: the first-stage deformable convolution learns a coarse offset, and the second-stage deformable convolution further learns a fine offset based on the first stage; each stage of deformable convolution is followed by batch normalization (BN) and ReLU activation functions. The sampling position of the deformable convolution is calculated as follows: (1) in, p This indicates the sampling position in the output feature map. p k This represents the regular sampling offset in standard convolution. Δp k This represents the additional offset learned. To ensure training stability, the offset prediction layer is initialized to zero in the initial training phase, allowing the model to be trained using standard convolutions first, and then gradually learning the effective offset corresponding to the crack boundary morphology during subsequent training.
[0038] The channel attention module models the importance of different channels in the input features. Specifically, it performs global average pooling and global max pooling on the input features to obtain two channels' descriptive information. Then, it establishes the dependencies between channels through convolutional mapping to generate channel attention weights, thereby enhancing the response of crack-related channels and suppressing redundant background channels. The spatial attention module enhances the spatial location of the crack region. Specifically, it spatially aggregates the features processed by channel attention and inputs them into multiple convolutional branches with different receptive fields to generate a spatial attention weight map. The calculation formula is as follows: (2) in, F S As input features, f i (·) Represents convolution branches at different scales. a i Indicates the weight of the corresponding branch. σ(·) This represents the Sigmoid activation function. M s This is the spatial attention weight map. Through the joint processing of deformable convolution, channel attention, and spatial attention, ACB can improve the model's ability to identify crack boundaries, crack bifurcation, and fine cracks in complex backgrounds of mining areas.
[0039] like Figure 4As shown, the Multi-Scale Attention Interaction Module (MAIM) is used to address the significant variations in the width, length, and spatial morphology of surface cracks in mining areas. For the input features processed by ACB, feature branches at the original scale, half-scale, and quarter-scale are constructed. The spatial resolution of the input features is preserved in the original scale branch to capture fine cracks, crack tips, and local texture information. In the half-scale and quarter-scale branches, the input features are downsampled to obtain crack context information within a larger receptive field. Convolutional enhancement processing is applied to each scale branch, and low-resolution scale features are upsampled to the original spatial resolution to achieve spatial alignment between features at different scales. Subsequently, a weighted fusion method is used to generate multi-scale fused features, calculated as follows: (3) in, X fusion Indicates multi-scale fusion features, , and These represent feature maps aligned at different scales. w 1. w 2 and w 4 represents the corresponding scale weight; fused features X fusion An attention weight map is generated through 1×1 convolution and a sigmoid function, and then multiplied element-wise with the input features to obtain multi-scale augmented features. During the first forward propagation... w 1. w 2. w 4. Take equal initial weights; during the second forward propagation, replace the initial weights with scale modulation coefficients generated based on the crack width distribution.
[0040] like Figure 5 As shown, the Attention-Guided Dense Skip Connection Module (ADSC) is used to improve the skip connection method between the encoder and decoder. Traditional skip connections directly splice shallow features from the encoder and deep features from the decoder, easily transmitting background noise such as rock textures, vegetation edges, shadow boundaries, and broken soil textures to the decoding stage. In this embodiment, the upper-layer features of the decoder are directly spliced together. Conquer x Features corresponding to the encoder layer skip Channel mapping and spatial alignment are performed, and the encoder features are guided by the decoder features to generate an attention weight map. The calculation method is as follows: (4) Where x represents the decoder feature and skip represents the encoder feature. g(·) Represents the channel mapping function. s(·) Represents the feature transformation function. Conv 1×1This indicates convolution processing with a kernel size of 1×1. ReLU Represents the linear rectified activation function. σ(·) This represents the Sigmoid activation function. ψ This is an attention weight map.
[0041] The attention weight map is multiplied element-wise with the corresponding layer features of the encoder to obtain the filtered skip connection features. These filtered skip connection features are then concatenated with the decoder features and input into subsequent convolutional modules to complete feature reconstruction. Through ADSC, the model can preserve crack edge details in the shallow features of the encoder while suppressing background texture information unrelated to crack recognition, thus improving the fusion quality of deep semantic features and shallow spatial details.
[0042] During model training, the MDE-UNet network model is trained using a training set, and model parameters are adjusted using a validation set. A combined loss function, consisting of a binary cross-entropy loss function and a Dice loss function, is used during training. The binary cross-entropy loss optimizes the pixel-level classification results between cracks and the background, while the Dice loss optimizes the overlap between predicted and actual crack regions. This combined loss function ensures that the model simultaneously focuses on crack pixel classification accuracy and crack region integrity during training, thereby mitigating the class imbalance problem caused by the small proportion of foreground pixels in the cracks. In this embodiment, the Adam optimizer is used for training, with an initial learning rate of 0.0005, a weight decay coefficient of 0.0001, a batch size of 4, 100 training epochs, and a cosine annealing strategy to adjust the learning rate, with a minimum learning rate of 0.00001.
[0043] 3. First forward propagation and coarse segmentation The UAV remote sensing image of the mining area to be detected is divided into overlapping blocks of 256×256 pixels with an overlap of 128 pixels. These blocks are then input into the trained MDE-UNet network model for the first forward propagation. During the first forward propagation, MAIM fuses multi-scale features with equal initial weights. After each block outputs a crack probability map, the cracks are fused according to their block positions. For overlapping areas, the average of multiple predicted probabilities is taken to obtain a coarse crack segmentation probability map for the entire image.
[0044] The coarse segmentation probability map of the crack is binarized with a threshold of 0.5. Morphological opening and closing operations are performed to remove isolated noise and smooth the crack region. Then, 8-connected component analysis (i.e., diagonally adjacent pixels are also considered to belong to the same connected component) is used to extract the connected components of each crack and remove isolated noise regions with an area of less than 200 pixels.
[0045] 4. Local scale mapping based on TFW files Read the TFW world file accompanying the UAV remote sensing imagery, obtain six affine transformation parameters A, D, B, E, C, and F, and establish the affine transformation relationship from pixel coordinates (x, y) to projected coordinates (X, Y): (5) (6) Where A is the pixel size in the x-direction, D and B are rotation terms, E is the pixel size in the y-direction, and C and F are the projection coordinates of the top-left pixel center; The average ground sampling distance (GSD) of the image is taken as the arithmetic mean of the pixel size |A| in the x-direction and the pixel size |E| in the y-direction. (7) Considering that terrain undulations and flight yaws may cause inconsistencies in local spatial resolution in UAV imagery, this invention calculates the local ground sampling distance point by point for each skeleton pixel. For a pixel (x, y), the Euclidean distance between it and its x-direction neighboring pixel (x+1, y) in the projected coordinate system is calculated. gsd x and the Euclidean distance in the projected coordinate system to the y-direction neighboring pixel (x, y+1). gsd y The average of the two values is taken as the local ground sampling distance at that point: (8) (9) (10) in, GSD local (x,y) The local ground sampling distance at pixel (x,y) is the distance between the pixels.
[0046] 5. Confidence-guided skeletonization and node classification Using a coarse crack segmentation probability map as a confidence guide, the binarized crack region is skeletonized, extracting the centerline of a single-pixel-width crack while preserving the crack topology. An eight-neighbor convolution is performed on the skeleton, and the number of neighbors for each skeleton pixel is counted. Pixels with one neighbor are marked as endpoints, and pixels with three or more neighbors are marked as branch points. Starting from each unvisited endpoint, the skeleton is traced pixel by pixel until another endpoint or branch point is reached. The tracing path is recorded as an independent branch, and branch points are retained as common nodes of multiple branches and are not pruned.
[0047] For each independent branch, the number of pixels on the branch excluding the bifurcation point is counted as the branch length, and the average confidence of each pixel on the branch in the crack coarse segmentation probability map is calculated as the branch average confidence. When the branch length is less than 5 pixels and the branch average confidence is less than 0.6, all pixels on the branch excluding the bifurcation point are pruned. After pruning, the endpoints and bifurcation points are re-identified, and the above tracking and pruning process is repeated until there are no branches that meet the pruning conditions in the current iteration.
[0048] After pruning, each skeleton pixel is classified into nodes based on the number of skeleton pixels connected within its eight neighborhoods: those with 1 neighbor are marked as endpoints, those with 2 neighbors are marked as ordinary points, and those with 3 or more neighbors are marked as branching points.
[0049] 6. Differentiated width measurement and physical scale conversion A differentiated width measurement strategy is adopted for skeleton pixels based on node type. For ordinary points, all skeleton neighbor pixels within their eight neighborhoods are collected. The sum of the direction vectors of each neighbor relative to the point is used as the local tangent direction and normalized. The tangent direction is rotated 90 degrees to obtain the normal direction. The search is performed along the positive and negative normal directions with unit steps. At each step position, the confidence value of the crack coarse segmentation probability map is sampled. When the confidence value first falls below 0.7, the side boundary distance is recorded. If no position with a confidence value below the threshold is detected within the maximum search radius of 25 pixels, the maximum search radius is used as the side boundary distance. The sum of the positive and negative boundary distances is used as the normal width (in pixels) of the ordinary point. When the normal width exceeds three times twice the Euclidean distance transformation value of the point, the normal direction search is considered abnormal, and twice the Euclidean distance transformation value of the point is used as its width.
[0050] For a bifurcation point, search all ordinary points within a radius of 5 pixels centered on the bifurcation point. Sort the normal widths of these ordinary points by size and take the median value as the width of the bifurcation point. If there are no ordinary points within the range, take twice the Euclidean distance transformation value of the bifurcation point as its width.
[0051] For endpoints, the width is directly taken as twice the Euclidean distance transformation value of that point.
[0052] The Euclidean distance transformation value is the Euclidean distance from a pixel within the crack region to the nearest background pixel. This value represents the half-width of the crack at the skeleton point. Therefore, the local width of the crack is twice the Euclidean distance transformation value.
[0053] After obtaining the pixel width of each skeleton point, multiply it by the local ground sampling distance of that point to convert it into physical width (unit: millimeters). Skeleton points with a physical width greater than 2000 millimeters are marked as outliers. Outliers are excluded when calculating the average and maximum width of cracks, but their skeleton pixels are retained to maintain the topological continuity of cracks.
[0054] 7. Generation of Scale Modulation Coefficients and Second Forward Propagation Calculate the average crack width based on the physical width of all obtained skeleton points. mean w (Unit: millimeters) w 0 =k· GSD mm As the reference width, where GSD mm The average ground sampling distance is expressed in millimeters, and k is the baseline width coefficient. In this embodiment, k is set to 8, meaning the baseline width is 8 times the ground sampling distance. The normalized width deviation t is calculated as follows: (11) The amplitude of t is limited to [-2, 2] to avoid excessive weight shift due to extreme values. The modulation coefficients for the three scale branches of MAIM are generated using the softmax function: (12) (13) (14) (15) in, exp(·) This represents an exponential function with base e. Z To normalize the denominator, α In this embodiment, α is taken as 1.5, which is the sensitivity coefficient. w 1. w 2. w 4 represents the modulation coefficients of the original scale, half-scale, and quarter-scale branches, respectively. The sum of the three is 1. When t is less than zero, the weight of the original scale branch is increased, and when t is greater than zero, the weight of the quarter-scale branch is increased.
[0055] The scale modulation coefficients are loaded into the MAIM module of the MDE-UNet network model, and the image to be detected is propagated a second time in the same block-based manner. MAIM uses... w 1. w 2. w 4. Replace the equal initial weights used in the first forward propagation with multi-scale fusion. After fusing the probability maps of each block, a binary segmentation mask for the crack is obtained.
[0056] 8. Quantization and output of precise segmentation geometric parameters Based on the binary segmentation mask of the crack, skeletonization, branch pruning, node classification and differential width measurement are re-executed to obtain the fine segmentation skeleton and the physical width of each skeleton point.
[0057] The actual length of the crack is calculated in the projected coordinate system. For each pair of eight neighboring pixels on the skeleton, the coordinates of the two pixels are transformed to projected coordinates using formulas (5) and (6), respectively. The Euclidean distance between the two points in the projected coordinate system is calculated as the actual ground length of the skeleton segment. All adjacent pixel pairs are traversed and summed. (16) in, (X i ,Y i ) and (X i+1 ,Y i+1 ) These are the projected coordinates of adjacent skeleton pixels. The length L is in meters, multiplied by 100 to convert to centimeters.
[0058] The actual area of the crack is determined by counting the number of pixels within the crack's connected region. N px Multiplying by the square of the geometric mean of the pixel size in the x-direction and the pixel size in the y-direction gives: (17) The area S is in square meters, which is converted to square centimeters by multiplying by 10000.
[0059] For each crack's connected component, calculate the mean and maximum physical widths (in millimeters) of its skeleton points to obtain the crack's average and maximum widths, respectively. Output the crack number, spatial location (bounding box and centroid coordinates), actual length, average width, maximum width, and actual area for each crack.
[0060] 9. Experimental Results Table 1 shows the performance comparison of different crack segmentation models on the Crack-UAV crack dataset test set. The MDE-UNet proposed in this invention outperforms the comparative models in terms of precision, recall, F1 score, and intersection-over-union ratio (IoU), with an F1 score of 86.15% and an IoU of 76.11%.
[0061] Table 1. Performance comparison of different models on the Crack-UAV crack dataset To evaluate the contribution of each module to the surface fracture segmentation performance in the mining area, ablation experiments were conducted on the self-developed Crack-UAV fracture dataset. ResNet50-UNet++ was used as the baseline model, and three ablation experiments were designed: single-module, dual-module combination, and full-model ablation experiments, to quantitatively evaluate the independent contribution and synergistic effect of each module. The experimental results are shown in Table 2.
[0062] Table 2 Ablation Experiment Results Single-module ablation results show that each of the three modules contributes to performance improvements. Adding ACB alone improves precision by 2.52%, achieves an F1 score of 84.85%, and increases the cross-union ratio (CUNR) to 74.14%, indicating that ACB helps suppress background interference such as bare soil, coal gangue, and shadows. Adding MAIM alone improves recall by 1.02%, achieves an F1 score of 85.00%, and increases the CUNR to 74.36%, indicating that multi-scale feature interaction enhances the ability to capture cracks of different widths. Adding ADSC alone improves precision by 1.39%, achieves an F1 score of 84.80%, and increases the CUNR to 74.08%, indicating that attention gating can filter out false crack interference in skip connections.
[0063] The performance of the dual-module combination was further improved. The recall rate of the ACB and MAIM combination reached 87.29%, indicating that boundary awareness and multi-scale feature fusion have complementary effects in reducing false negatives. The precision of the ACB and ADSC combination reached 85.94%, with both working together to suppress false background detections. The cross-union ratio (CUI) of the MAIM and ADSC combination was 75.59%, the best among the dual-module combinations. When all three modules were used simultaneously, the F1 score and CUI reached 86.15% and 76.11%, respectively, both superior to any single dual-module combination. These results demonstrate that these three modules play unique and complementary roles in the network.
[0064] Table 3 presents the generalization results of MDE-UNet on the Crack500, DeepCrack, and CrackForest public crack datasets. On the Crack500 dataset, MDE-UNet outperforms other competing models overall, achieving the highest F1 score and intersection-over-union (IoU) ratio, at 83.10% and 71.71%, respectively. On the DeepCrack dataset, MDE-UNet demonstrates a significant advantage, achieving the best performance across all four metrics, including precision, recall, F1 score, and IoU, with a recall of 90.26% and an IoU ratio of 81.36%. A similar trend was observed on the CrackForest dataset, where MDE-UNet achieved the highest recall (80.56%), F1 score (72.27%), and IoU ratio (57.06%). These results demonstrate that MDE-UNet possesses strong generalization ability and robustness in cross-domain evaluation and various complex crack scenarios.
[0065] like Figures 6 to 9As shown, this invention can effectively distinguish cracks from background features such as vegetation, rocks, and roads in complex mining environments, maintaining the continuity of crack extension direction and branch structure; it can identify crack tips and small branches in local detail areas; and it can adapt to different crack morphologies such as road cracks and concrete cracks on public datasets.
[0066] Table 3 Generalization Experiment Results To verify the accuracy of the model's quantification of surface crack geometric parameters, 30 surface crack samples were selected, and the model's predictions were compared with the corresponding field measurements. Field measurements were conducted using a measuring tape: the length and width (in centimeters) of each surface crack were recorded, with a measurement accuracy of 1 centimeter each time; the length and width of each sample were measured three times, and the average of the three measurements was used as the field reference value. The field measurement process is as follows: Figure 10 As shown.
[0067] Table 4. Verification Results of Geometric Parameter Quantization Accuracy The model demonstrated high accuracy in quantifying surface crack length. The MAPE value for length estimation was only 4.71%, while the Pearson correlation coefficient and coefficient of determination reached 0.9975 and 0.9950, respectively. These results indicate that the model can effectively preserve the relative length differences among surface cracks and capture their overall trend of change. The MAE, RMSE, and MAPE for width estimation were 1.53 cm, 1.77 cm, and 10.99%, respectively. The small difference between MAE and RMSE indicates that the width error remains relatively consistent across different samples and is not significantly affected by a small number of outliers. The Pearson correlation coefficient reached 0.7565, indicating that the model can capture the overall trend of crack width change. However, the R² value was only 0.5723, indicating that the model's fitting performance for individual width changes is weaker compared to length estimation. This difference is mainly attributed to factors such as irregular crack boundaries, local occlusion, limited image resolution, and the fact that field measurements of width were recorded with an accuracy of only 1 cm. Given that the MAE for width estimation is approximately 1–2 cm, the obtained accuracy is sufficient for regional-scale width statistics and preliminary crack risk assessment. However, additional field measurements are still needed to achieve more refined width classification and high-precision deformation monitoring. Overall, MDE-UNet provides reliable length quantification results and effectively captures the overall variation pattern of crack width, indicating its application potential in large-scale automated geometric parameter extraction and rapid surface crack detection in mining areas.
[0068] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various improvements, substitutions, or modifications can be made to the present invention without departing from its technical concept, and all such improvements, substitutions, or modifications should fall within the protection scope of the present invention.
Claims
1. A method for intelligent identification and quantification of surface cracks in mining areas based on MDE-UNet, characterized in that, Includes the following steps: S1. Acquire UAV remote sensing images of the mining area and construct the Crack-UAV crack dataset; and train the MDE-UNet network model; the MDE-UNet network model includes attention convolutional block ACB, multi-scale attention interaction module MAIM and attention-guided dense jump connection module ADSC. S2. Input the image to be detected into the MDE-UNet network model for the first forward propagation. MAIM fuses multi-scale features with initial weights and obtains a coarse crack segmentation probability map through block probability fusion. S3. Read the TFW file, establish the affine transformation relationship from pixel coordinates to projected coordinates, and perform connected component analysis on the binarized coarse segmentation probability map of the crack. S4. Using the crack coarse segmentation probability map as a confidence guide, skeletonization is performed. Low-confidence short branches are pruned according to the average branch confidence to obtain the pruned crack skeleton. The skeleton pixels are classified into endpoints, ordinary points, and bifurcation points. S5. Search the confidence boundary along the skeleton normal direction for ordinary points to calculate the normal width. Take the median width of the neighboring ordinary points for the bifurcation point. Take twice the Euclidean distance transformation value for the endpoints. Convert the width point by point to the physical width based on the affine transformation relationship. S6. Generate the scale modulation coefficients of MAIM based on the physical width distribution of the crack, load them into the MDE-UNet network model for the second forward propagation, and use the scale modulation coefficients to replace the initial weights of MAIM for multi-scale fusion to obtain the crack binary segmentation mask. S7. Based on the binary segmentation mask of the crack, re-execute S4 to S5, calculate the actual length, width and area of each crack connected domain in the projected coordinate system, and finally output the quantization results of the crack geometric parameters.
2. The intelligent identification and quantification method for surface cracks in mining areas based on MDE-UNet according to claim 1, characterized in that, The attention convolutional block (ACB) includes a deformable convolutional module, a channel attention module, and a spatial attention module. The deformable convolutional module dynamically adjusts the convolution sampling position based on the crack boundary morphology to enhance the model's representation ability of irregular crack boundaries and complex topological structures. The channel attention module enhances the response of crack-related channels and suppresses background noise. The spatial attention module uses multi-scale convolutional branches to generate spatial attention weights, the expression of which is: in, F S For input features, f i (·) Represents convolution branches at different scales. a i Indicates the weight of the corresponding branch. σ(·) This represents the Sigmoid activation function. M s This is a spatial attention weight map.
3. The intelligent identification and quantification method for surface cracks in mining areas based on MDE-UNet according to claim 1, characterized in that, The Multi-Scale Attention Interaction Module (MAIM) acquires crack features at different scales by constructing feature branches at the original scale, half-scale, and quarter-scale; it uses bilinear interpolation to achieve spatial alignment between feature maps at different scales; and it employs an attention mechanism to dynamically fuse features at different scales, with the fusion expression being: in, X fusion Indicates multi-scale fusion features, , and These represent feature maps aligned at different scales. w 1. w 2 and w 4 represents the corresponding scale weight; the fused features are multiplied element-wise with the input features after attention mapping to enhance the crack target region.
4. The intelligent identification and quantification method for surface cracks in mining areas based on MDE-UNet according to claim 1, characterized in that, The Attention-Guided Dense Skip Connection (ADSC) module uses decoder features to guide encoder features in generating an attention weight map. It then filters the encoder output features, retaining only the valid information related to crack identification before fusing it with the decoder features. The expression for generating the attention weight map is: Where x represents the decoder feature and skip represents the encoder feature. g(·) Represents the channel mapping function. s(·) Represents the feature transformation function. Conv 1×1 This indicates convolution processing with a kernel size of 1×1. ReLU Represents the linear rectified activation function. σ (·) This represents the Sigmoid activation function. ψ This is an attention weight map.
5. The intelligent identification and quantification method for surface cracks in mining areas based on MDE-UNet according to claim 1, characterized in that, The TFW file described in S3 contains six affine transformation parameters A, D, B, E, C, and F. The transformation relationship from pixel coordinates (x, y) to projected coordinates (X, Y) is as follows: Where A is the pixel size in the x-direction, D and B are rotation terms, E is the pixel size in the y-direction, and C and F are the projection coordinates of the top-left pixel center; In S5, the distances between the skeleton point and its adjacent pixels in the x and y directions in the projected coordinate system are calculated respectively, and the average value is taken as the local ground sampling distance. The pixel width is multiplied by this local ground sampling distance to obtain the physical width.
6. The intelligent identification and quantification method for surface cracks in mining areas based on MDE-UNet according to claim 1, characterized in that, In S4, tracing from one end of the skeleton to another end or a bifurcation point forms an independent branch. The bifurcation point is a common node and does not participate in pruning. When the branch length is less than the branch length threshold and the average confidence of the branch is less than the confidence threshold, the entire branch is pruned. This process is repeated until no branch can be pruned.
7. The intelligent identification and quantification method for surface cracks in mining areas based on MDE-UNet according to claim 1, characterized in that, In S5, the tangent of a normal point is determined by the sum of the direction vectors of its skeleton neighbors, and the normal is obtained by rotating it by 90 degrees. The search is performed along the normal with a unit step size on both sides until the confidence level is lower than the boundary confidence threshold. The sum of the step sizes on both sides is used as the normal width. When the normal width exceeds a preset multiple of twice the Euclidean distance transformation value, it is replaced by twice the Euclidean distance transformation value. For a bifurcation point, the median of the normal widths of the normal points in its neighborhood is taken. When there are no normal points in the neighborhood, twice the Euclidean distance transformation value is taken.
8. The intelligent identification and quantification method for surface cracks in mining areas based on MDE-UNet according to claim 1, characterized in that, The average crack width obtained from S5 is calculated in S6. mean w ,by w 0 =k·GSD The reference width is defined as GSD, where GSD is the average ground sampling distance and k is the reference width coefficient; the normalized width deviation is calculated. t=(mean w -w 0 ) / w 0 and limit the amplitude to [-2,2]; generate three-scale modulation coefficients via softmax: in, exp(·) This represents an exponential function with base e. Z To normalize the denominator, α This is the sensitivity coefficient; w 1. w 2. w 4 represents the modulation coefficients of the original scale, half-scale, and quarter-scale branches, respectively. The sum of the three is 1. When t is less than zero, the weight of the original scale branch is increased, and when t is greater than zero, the weight of the quarter-scale branch is increased.
9. The intelligent identification and quantification method for surface cracks in mining areas based on MDE-UNet according to claim 5, characterized in that, The actual length mentioned in S7 is obtained by calculating the Euclidean distance segment by segment after the coordinates of adjacent pixels of the skeleton are transformed to projected coordinates through affine transformation and summing them; the actual area is obtained by multiplying the number of pixels in the connected region by the square of the geometric mean of the pixel size in the x and y directions.
