Ground fracture extraction deep learning method based on direction perception and Bayesian fusion

Abstract

The invention discloses a ground fracture extraction deep learning method based on direction perception and Bayesian fusion, which solves the problem of feature fracture caused by the fixed form of the traditional convolution kernel by constructing a direction priori knowledge extractor, strengthening the perception ability of a network to the direction and adjusting the receptive field along the fracture trend by adopting the dynamic deformation convolution kernel. And a Bayesian fusion module is designed, the weights of the global semantic features and the direction perception features are dynamically distributed based on a Bayesian probability model, the fusion effect between different features is optimized, and compared with other deep learning methods, the ground fracture extracted by the method is more complete and can adapt to a mining area environment with relatively complex geological conditions.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing technology, in particular to a ground fissure extraction deep learning method based on direction perception and Bayesian fusion. BACKGROUND

[0002] The accurate extraction of ground fissures is one of the core tasks of mine safety monitoring and ecological restoration. Traditional ground fissure extraction methods mainly rely on artificial field survey techniques such as total station and GPS positioning to indirectly infer fissure distribution by measuring ground surface displacement changes. However, these methods have significant limitations in terms of efficiency and coverage.

[0003] In recent years, the rapid development of unmanned aerial vehicle photogrammetry technology has provided a new solution for ground fissure monitoring. Unmanned aerial vehicles can flexibly obtain centimeter-level resolution images and support multi-modal data acquisition such as multispectral and thermal infrared, significantly improving the visualization capability of ground fissures. However, ground fissure extraction based on deep learning from unmanned aerial vehicle images still faces the following challenges:

[0004] 1) The fixed receptive field of traditional convolution kernels cannot capture the bending and branching characteristics of fissures, leading to incomplete extraction results.

[0005] 2) Lack of prior modeling of fissure direction, the model cannot utilize geometric topological constraints, and the reasoning ability in occluded areas (such as vegetation coverage and gravel) significantly decreases. SUMMARY

[0006] The present application aims to provide a ground fissure extraction deep learning method based on direction perception and Bayesian fusion to solve the problems raised in the background.

[0007] According to one aspect of the present application, a ground fissure extraction deep learning method based on direction perception and Bayesian fusion is provided, which comprises the following steps:

[0008] S1, obtaining unmanned aerial vehicle images of a predetermined area and performing preprocessing, the preprocessing including one or more of radiation correction, geometric correction, registration, stitching, and cropping;

[0009] S2, creating a deep learning data set using the preprocessed unmanned aerial vehicle images, the data set including labeled information of ground fissure areas;

[0010] S3. Constructing a directional prior knowledge extraction module to generate a topologically coherent directional field, wherein, based on the ground fissure prior region, a multi-scale initial directional field is generated through structural tensor analysis, and the multi-scale initial directional fields are fused using the maximum coherence criterion to generate a basic directional map. Anisotropic diffusion and fast marching methods are implemented along the fracture skeleton line to propagate high-confidence directional information to optimize the topological coherence of the directional field;

[0011] S4. Implementing multi-source feature fusion through a Bayesian probability fusion module, wherein an uncertainty propagation model of global semantic features and direction perception features is established, and the weight distribution of global semantic features and direction perception features is dynamically optimized based on the model;

[0012] S5. Construct a Prior-VDBFNet network. In the encoding stage, a VMamba global semantic perception branch is constructed based on the VSSM architecture. In the decoding stage, an input channel-aware VSS decoding block (CAVSS) is used to construct a four-stage feature decoding architecture. In the feature fusion stage, global semantic features and direction perception features are integrated through the Bayesian probability fusion module.

[0013] Preferably, in step S3, generating the basic directional pattern includes the following steps:

[0014] Construct the structure tensor based on the gradient vector field. For the grayscale image I: Its gradient vector field Constitute the first-order differential characteristics;

[0015] The structure tensor expression is:

[0016]

[0017] Among them, is a two-dimensional Gaussian kernel with a standard deviation of σ, the symbol * represents the convolution operation, and the symbol represents the tensor product; G σ

[0018] Perform eigenvalue decomposition on the structure tensor to calculate the main direction angle θ and the directional coherence coefficient C:

[0019]

[0020] Among them, λ1≥λ2 is the eigenvalue, and ò is a very small positive number to prevent the denominator from being zero;

[0021] A multi-scale spatial pyramid strategy is used for robust estimation. The scale set Σ = {16, 32, 64} is defined and the direction field θ is calculated for each scale σ∈Σ. σ and the coherence coefficient C σ , and finally generate the initial direction field through the maximum confidence criterion fusion:

[0022] σ * = argmax σ∈Σ C σ (x)

[0023] where θ init (x) is the initial direction estimate at pixel position x at the optimal scale σ * , and the optimal scale σ * is the scale at which the maximum coherence coefficient is achieved.

[0024] Preferably, the anisotropic diffusion along the crack skeleton line in step S3 and the fast marching method comprise the following steps:

[0025] An anisotropic diffusion expression is constructed by using the direction field function θ(x, t) in combination with an anisotropic diffusion and a coherence coefficient direction constraint mechanism:

[0026]

[0027] where t is a time variable, D is a diffusion tensor, and λ is a regularization parameter;

[0028] The diffusion tensor D is defined as:

[0029]

[0030] where κ is a gradient sensitivity coefficient, φ = θ init (x) is an initial direction, and η > 1 controls anisotropy strength;

[0031] A direction continuity constraint is implemented in the propagation process:

[0032]

[0033] where N(x) represents an eight-neighborhood of x, and S is a skeleton line of a crack region,

[0034] For each skeleton point x ∈ S on the crack skeleton line S, an initial direction θ init (x) is initialized,

[0035] For each x in the eight-neighborhood N(x), a direction difference is calculated, and for each y ∈ N(x) ∩ S, a direction difference constraint is checked. If |θ(x) - θ(y)| < π / 6, the direction of x is propagated to y: θ(x) → θ(y), otherwise, the original direction of y is retained or marked as needing further optimization.

[0036] Preferably, the propagation of high-confidence direction information in step S3 to optimize the topological coherence of the direction field comprises:

[0037] For the crack branch noise region in the prior position region, a direction feature vector of the mixed feature space is constructed:

[0038]

[0039] where (x i ,y i ) is the pixel coordinate,

[0040] A similarity matrix W is defined:

[0041]

[0042] where ‖f i -f j ‖ 2 is the Euclidean distance square between feature vectors, σ s is a sensitive parameter for controlling the spatial proximity, σ θ is a sensitive parameter for controlling the direction consistency, and the higher the similarity W ij ∈ [0, 1] is, the more similar i and j are in space and direction;

[0043] Spectral clustering is performed on the similarity matrix W by using K-means, and for each sub-cluster S k , a direction statistic is calculated:

[0044]

[0045] where is the average direction angle of the sub-cluster S k , and σ k is the sub-cluster direction standard deviation,

[0046] The in-class direction consistency is measured, and the direction of the abnormal sub-cluster with σ k > π / 12 is corrected, and the iteration is performed until all sub-clusters meet the direction consistency constraint.

[0047] Preferably, in step S4, the multi-source feature fusion realized by the Bayesian probability fusion module includes the following steps:

[0048] A Bayesian fusion target function is constructed;

[0049] Feature fusion is performed based on prior knowledge guidance; and

[0050] Based on the fusion result, the geometric consistency of the features is enhanced by local optimization.

[0051] Preferably, the construction of the Bayesian fusion target function includes:

[0052] The global feature of the VMamba backbone is defined as The direction-aware feature of DSCNet is defined as According to Bayes' theorem, the posterior distribution objective function is represented as:

[0053] p(F f |F g ,F d )∝p(F g |F f )p(F d |F f )p(F f )

[0054] Wherein, p(F g |F f ) and p(F d |F f ) are likelihoods, and p(F f ) is a prior term.

[0055] The global feature and the direction feature covariance matrix Λ g , Λ d is:

[0056]

[0057] Wherein is the feature of the i th channel of the global feature, is the feature of the i th channel of the direction feature.

[0058] The objective function for minimizing the comprehensive error in the optimization process is:

[0059]

[0060] Wherein, Λ g and Λ d are the covariance matrices of the two types of features respectively, and R(F f ) is a regularization term.

[0061] Preferably, the feature fusion based on prior knowledge guidance comprises:

[0062] The objective function is converted into a trainable neural network module;

[0063] According to the high-dimensional feature hypothesis, each element of the feature covariance inverse matrix is The channel statistics are calculated by global average pooling:

[0064]

[0065] Wherein, μ c is the global mean of the c th channel, is the global feature map in the VMamba branch, H and W are the height and width of the feature map;

[0066] Use two fully connected layers to generate attention weights:

[0067] W g =FC2(GELU(FC1(μ))),μ=[μ1,…,μ C ]

[0068] Among them, FC1 and FC2 are fully connected layers, and GELU is the activation function.

[0069] Generate W d , the final fusion feature is:

[0070] F f =W g ⊙F g +W d ⊙F d

[0071] Among them, F g The global features extracted from VMamba backbone, F d is the direction-aware feature extracted by DSCNet, and ⊙ is the channel-by-channel multiplication.

[0072] Preferably, the features enhanced by local optimization include:

[0073] Through lightweight network D θ Realize iterative feature fusion:

[0074]

[0075] Among them, U (l) is the Bayesian weighted fusion feature of the lth layer, is the feature enhanced by local optimization, where D θ It consists of 3×3 convolution and residual connection. The gradient update rule of the iterative algorithm is:

[0076]

[0077] X (k) is the feature estimate of the kth iteration, is the data fidelity gradient, and η is the learning rate.

[0078] According to another aspect of the present application, a computer program product is provided, comprising a computer program, wherein the computer program is executed to implement the deep learning method for ground fissure extraction based on direction perception and Bayesian fusion as described in any one of claims 1 to 8.

[0079] The method of the present application strengthens the perception ability of the network to the direction by constructing the direction prior knowledge extractor, at the same time, adopts the dynamic deformation convolution kernel to adjust the receptive field along the crack trend, solves the feature fracture problem caused by the fixed form of the traditional convolution kernel, and designs the Bayesian fusion module, dynamically allocates the weight of the global semantic feature and the direction perception feature based on the Bayesian probability model, optimizes the fusion effect between different features, and through the comparison with other deep learning methods, the ground fissure extracted by the present application is more complete, and can adapt to the mine environment with more complex geological conditions. BRIEF DESCRIPTION OF DRAWINGS

[0080] Figure 1 A flowchart based on direction perception and Bayesian fusion of the embodiment of the present application;

[0081] Figure 2 A network diagram of the ground fissure extraction deep learning method of the embodiment of the present application;

[0082] Figure 3 A structure diagram of the direction prior knowledge extraction module of the embodiment of the present application;

[0083] Figure 4 A VSS2D module diagram of the core module of the VSSB of the embodiment of the present application;

[0084] Figure 5 A core VSSB module diagram of the VMamba branch of the embodiment of the present application;

[0085] Figure 6 A structure diagram of the dynamic separable convolution of the embodiment of the present application;

[0086] Figure 7 A research area overview diagram of the embodiment of the present application;

[0087] Figure 8 A ground fissure extraction result comparison diagram of different methods of the open-pit mine of the embodiment of the present application. DETAILED DESCRIPTION

[0088] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0089] Please refer to Figure 1 , according to an embodiment of the present application, a ground fissure extraction deep learning method based on direction perception and Bayesian fusion is provided, which comprises the following steps.

[0090] S1. Acquire drone images of a predetermined area and perform preprocessing, wherein the preprocessing includes one or more of radiation correction, geometric correction, registration, stitching, and cropping.

[0091] Use drone aerial photography to obtain remote sensing images of the mining area surface, and perform the following preprocessing operations in sequence:

[0092] Radiometric correction: Eliminate uneven lighting and imaging device differences.

[0093] Geometric correction: Eliminate image distortion and make the image consistent with geographic space.

[0094] Registration and stitching: Multiple images are seamlessly aligned to form a complete coverage area.

[0095] Cropping: Focus on the study area and remove redundant background.

[0096] This generates high-quality, geometrically and spectrally consistent input images, which serve as the basis for subsequent deep learning model training and inference, improving the model's ability to perceive small-scale features (such as fine cracks).

[0097] S2. Create a deep learning dataset using the preprocessed UAV imagery, where the dataset includes labeled information about the ground fissure area.

[0098] Based on preprocessed images, ground fissure areas are manually or semi-automatically annotated at the pixel level, and training, validation, and testing subsets are constructed to form a standard supervised deep learning dataset, thereby providing reliable "input-output" pairs. This allows the model to learn the geometric morphology, texture characteristics, and contextual information of the cracks, enhance the model's generalization ability, and improve its adaptability in complex terrain backgrounds.

[0099] S3. Construct a directional prior knowledge extraction module to generate a topologically coherent directional field. Based on the ground fissure prior area, a multi-scale initial directional field is generated through structural tensor analysis. The multi-scale initial directional field is fused using the maximum coherence criterion to generate a basic directional map. Anisotropic diffusion and rapid marching methods are implemented along the crack skeleton line to propagate high-confidence directional information to optimize the topological coherence of the directional field.

[0100] Structural tensor analysis and multi-scale direction estimation: The structural tensor is constructed using the image gradient vector, the main direction angle and coherence coefficient of each pixel are extracted, the optimal scale is selected through the multi-scale pyramid strategy, and the fusion is carried out into the basic direction map.

[0101] Crack skeleton-guided directional diffusion: Anisotropic diffusion and fast marching methods are implemented along the main skeleton of the ground fissure, preferentially propagating high-confidence directions to low-confidence areas to repair fractures or disordered directions.

[0102] Spectral clustering and anomaly correction: construct a similarity matrix in the space-direction hybrid feature space, perform K-means clustering and statistical evaluation, identify and correct abnormal direction clusters.

[0103] By introducing the "geometric direction prior", the defects of convolution receptive field in direction modeling are compensated, the continuity of crack extraction in complex structure areas such as branching, turning and shielding is guaranteed, and the recognition stability of weak boundary cracks is improved.

[0104] S4, realize multi-source feature fusion through a Bayesian probability fusion module, wherein an uncertainty propagation model of global semantic features and direction perception features is established, and weight distribution of the global semantic features and the direction perception features is dynamically optimized based on the model.

[0105] The semantic features from the VMamba backbone and the DSCNet direction perception features are respectively modeled, a covariance matrix thereof is constructed, and a minimum comprehensive error objective function of weighted fusion is calculated. Through global pooling and a fully connected layer, fusion weights are dynamically generated, two feature sources are weighted and synthesized according to channel importance, and a lightweight convolution residual module is introduced to locally optimize the fused features, so as to guarantee spatial structure integrity and edge smoothness.

[0106] Thus, reliable fusion of two types of heterogeneous features of "direction" and "semantic" is realized, the discrimination and positioning accuracy of the model for cracks are enhanced in a complex background, feature conflicts are reduced, and the stability and generalization ability of extraction are improved.

[0107] S5, construct a Prior-VDBFNet (Prior-guided Vision Mamba Network with Dynamic Snake Convolution and Bayesian Fusion Network) network, wherein a VMamba global semantic perception branch is constructed based on a VSSM architecture in an encoding stage, a four-stage feature decoding architecture is constructed by using a CAVSS (Channel Attention-aware VSS Decoding Block) in a decoding stage, and global semantic features and direction perception features are integrated through the Bayesian probability fusion module in a feature fusion stage.

[0108] VMamba semantic perception branch: based on the VSSM module, a large receptive field is established by using an attention mechanism, the image context is globally understood, and macro topographic structures are captured.

[0109] DSCNet direction perception branch: dynamic separable convolution (DSConv) is used to simulate direction-adjustable convolution kernels, and accurately respond to the directionality of crack morphology.

[0110] Bayesian fusion module: use the fusion mechanism constructed in the last step to perform multi-level synthesis at different semantic levels.

[0111] CAVSS decoder: use the channel-aware structure to gradually upsample and recover the crack boundary, output the pixel-level crack probability map, and realize the final prediction.

[0112] Thus, the "structure prior and global semantic" dual-channel perception is realized, and the detection rate, positioning accuracy and structure continuity of complex morphological cracks are improved, and a more complete and reliable ground crack extraction result is output.

[0113] In one embodiment, specifically, in step S3, generating the base direction map includes the following steps:

[0114] Based on the gradient vector field, the structure tensor is constructed, and for the gray image I: Its gradient vector field Forms a first-order differential feature;

[0115] The structure tensor expression is:

[0116]

[0117] Where, G σ is a two-dimensional Gaussian kernel with a standard deviation of σ, the symbol * represents convolution operation, and the symbol represents tensor product;

[0118] Eigenvalue decomposition is performed on the structure tensor to calculate the principal direction angle θ and the direction coherence coefficient C:

[0119]

[0120] Where λ1≥λ2 are the eigenvalues, and ò is a small positive number to prevent the denominator from being zero;

[0121] Robust estimation is performed using a multi-scale spatial pyramid strategy, and a scale set Σ={16,32,64} is defined. For each scale σ∈Σ, the direction field θ σ and the coherence coefficient C σ are calculated, and finally the initial direction field is fused by the maximum confidence criterion:

[0122] σ * = argmax σ∈Σ C σ (x) (formula 4)

[0123] Where θ init (x) is the initial direction estimation value at the optimal scale σ * for the pixel position x, and the optimal scale σ* is the maximum coherence coefficient at scale s.

[0124] According to the technical solution, the structure tensor is constructed: the principal direction of each pixel is calculated through a gradient vector field (a first-order derivative of a gray scale change in an image), and the structure tensor essentially describes "direction consistency" and a "dominant direction" of a local image block.

[0125] Eigenvalue decomposition: the principal axis direction of the tensor is extracted, and a direction coherence coefficient is used to measure the reliability or confidence of the direction.

[0126] Multi-scale direction estimation and fusion: the structure tensor is constructed at multiple scales (i.e., the direction is evaluated at different blur levels), and the most reliable direction information is selected as a final estimation value through a maximum confidence criterion.

[0127] A stable and reliable direction field is generated through the structure tensor and multi-scale fusion method, the perception ability of a model to a complex crack direction feature is enhanced, and the robustness and direction continuity of the overall system in a complex environment are improved.

[0128] In one embodiment, specifically, the anisotropic diffusion and fast marching method along the crack skeleton line in step S3 includes the following steps:

[0129] An anisotropic diffusion expression is constructed by using a direction field function θ (x, t) in combination with an anisotropic diffusion and a coherence coefficient direction constraint mechanism:

[0130]

[0131] where t is a time variable, D is a diffusion tensor, and λ is a regularization parameter;

[0132] The diffusion tensor D is defined as:

[0133]

[0134] where κ is a gradient sensitivity coefficient, φ = θ init (x) is an initial direction, and η > 1 controls the anisotropy strength;

[0135] A direction continuity constraint is implemented in the propagation process:

[0136]

[0137] where N (x) represents an eight-neighborhood of x, S is a skeleton line of a crack region,

[0138] For each skeleton point x ∈ S on the crack skeleton line S, an initial direction θ init (x) is initialized,

[0139] For each x in the octree N(x), compute the directional difference, check the directional difference constraint for each y∈N(x)∩S, if |θ(x)-θ(y)|<π / 6, then propagate the x direction to y: θ(x)→θ(y), otherwise, keep the original direction of y or mark it for further optimization.

[0140] According to the technical solution, the direction field function is anisotropic diffusion: the diffusion function is designed to consider the direction consistency, so that the information is more easily propagated in the direction of the crack, and the diffusion is inhibited in the vertical direction.

[0141] Diffusion tensor design: adjust the diffusion strength according to the gradient and direction information;

[0142] Skeleton-guided propagation and directional continuity constraint: starting from the points on the crack skeleton line, propagate the direction information one by one, check the direction difference between each neighborhood pixel and the skeleton point, and only allow propagation when the direction difference is less than a set threshold, otherwise mark it for further optimization or keep the original direction.

[0143] The direction information is stably propagated on the crack skeleton with constraints, so that the constructed direction field has better topological continuity and direction consistency, thereby laying a key foundation for accurately and completely extracting complex crack structures.

[0144] In one embodiment, specifically, the propagation of high-confidence direction information in step S3 to optimize the topological coherence of the direction field includes:

[0145] For the crack branch noise region in the prior position region, construct the direction feature vector of the mixed feature space:

[0146]

[0147] where (x i ,y i ) is the pixel coordinate,

[0148] Define the similarity matrix W:

[0149]

[0150] where ‖f i -f j ‖ 2 is the Euclidean distance square between feature vectors, σ s is a sensitive parameter that controls the spatial proximity, σ θ is a sensitive parameter that controls the direction consistency, and the higher the similarity W ij ∈[0,1] indicates that i and j are more similar in space and direction.

[0151] K-means is used to perform spectral clustering on the similarity matrix W, and for each subcluster S k , calculate the directional statistics:

[0152]

[0153]

[0154] in, is a subcluster S k The average direction angle, σ k is the sub-cluster direction standard deviation,

[0155] Measuring the consistency of direction within the class, k Abnormal subclusters with >π / 12 are oriented and corrected, and the process is repeated until all subclusters satisfy the directional consistency constraint.

[0156] According to the above technical solution, a directional feature vector is constructed: the spatial coordinates of each pixel are combined with the directional angle information to form a mixed feature space (space + direction).

[0157] Similarity matrix construction: Euclidean distance and direction angle difference are used to jointly define the similarity between pixels.

[0158] Spectral clustering (K-means): Groups directional features based on the similarity matrix to form several subclusters.

[0159] Abnormal cluster identification and direction correction: Calculate the average direction and standard deviation of each sub-cluster. Clusters with excessively large direction standard deviations are considered abnormal and direction adjustments are performed. This is iterated until the directions of all sub-clusters are stable and consistent.

[0160] By identifying directional anomaly areas through spectral clustering and making iterative corrections, the entire directional field is made more spatially consistent and topologically coherent, thus effectively supporting high-precision and complete ground fissure extraction.

[0161] In one embodiment, in step S4, implementing multi-source feature fusion through the Bayesian probability fusion module includes the following steps.

[0162] Construct the Bayesian fusion objective function.

[0163] Based on the principle of Bayesian inference, the fusion problem of directional perception features and global semantic features is modeled as an optimization problem of minimizing the posterior error. Prior terms and likelihood terms are introduced to constrain the credibility of directional features and the consistency of semantic features, respectively.

[0164] Feature fusion is guided by prior knowledge.

[0165] The prior direction information is used to guide the dynamic regulation of the importance of features in the feature fusion process, and the attention weighting is realized through a neural network, so that the weight distribution process has self-adaptive ability.

[0166] And on the basis of the fusion result, the geometric consistency of the features is enhanced through local optimization.

[0167] On the basis of the fusion result, a light convolution module is further used to perform local structure optimization on the fused feature map, so as to ensure the continuity of the crack boundary and the smooth transition of the direction, and enhance the fine expression ability of the model in space.

[0168] In one embodiment, specifically, constructing the Bayesian fusion target function comprises:

[0169] The global feature of the VMamba backbone is defined as The direction perception feature of the DSCNet is defined as According to the Bayes theorem, the posterior distribution target function is represented as:

[0170] p(F f |F g ,F d )∝p(F g |F f )p(F d |F f )p(F f )

[0171] Wherein, p(F g |F f ) and p(F d |F f ) are likelihoods, and p(F f ) is a prior term.

[0172] The global feature and the direction feature covariance matrix Λ g are constructed d , Λ g is:

[0173]

[0174] Wherein is the feature of the i th channel of the global feature, is the feature of the i th channel of the direction feature.

[0175] The target function for minimizing the comprehensive error in the optimization process is:

[0176]

[0177] Wherein, Λ g and Λ d are the covariance matrices of the two types of features, and R(Ff ) is a regularization term.

[0178] According to the technical solution, Bayesian theorem is used to construct a posterior probability model between the direction feature and the global feature, covariance matrices of the two feature sources (representing uncertainty) are calculated to realize the formalization of the fusion target function, and the target function aims to minimize the comprehensive error between the two types of features to guide the weight optimization process.

[0179] In one embodiment, specifically, the feature fusion based on prior knowledge guidance includes:

[0180] transforming the target function into a trainable neural network module;

[0181] According to the high-dimensional feature assumption, each element of the feature covariance inverse matrix is The channel statistics are calculated by global average pooling:

[0182]

[0183] where μ c is the global mean of the c-th channel, is the global feature map in the VMamba branch, and H and W are the height and width of the feature map;

[0184] Two fully connected layers are used to generate attention weights:

[0185] W g = FC2(GELU(FC1(μ))), μ = [μ1,…, μ C ](Formula 15)

[0186] where FC1 and FC2 are fully connected layers, and GELU is an activation function,

[0187] W d is generated, and the final fused feature is:

[0188] F f = W g ⊙F g +W d ⊙F d (Formula 16)

[0189] where F g is the global feature extracted by the VMamba backbone, F d is the direction perception feature extracted by the DSCNet, and is the channel-wise multiplication.

[0190] According to the technical solution, the channel statistics of the feature map is calculated by using the Gaussian assumption, the attention weight is generated through a two-layer fully connected network, and the channel-by-channel weighted fusion of the direction perception feature and the semantic feature is realized.

[0191] In one embodiment, specifically, the feature enhanced by local optimization includes:

[0192] The feature enhanced by the light network D θ The feature iterative fusion is realized:

[0193]

[0194] Wherein, U (l) is the Bayesian weighted fusion feature of the lth layer, is the feature enhanced by local optimization, wherein D θ is composed of a 3*3 convolution and a residual connection, and the gradient update rule of the iterative algorithm is:

[0195]

[0196] X (k) is the feature estimation value of the kth iteration, is the data fidelity item gradient, and η is the learning rate.

[0197] According to the technical solution, the local geometric consistency of the fused feature is enhanced by using the light residual convolution module, the iterative mechanism of approximate gradient descent is introduced, and the feature is continuously corrected, so that the problems that cannot be completely processed in the fusion stage, such as local fracture and fuzzy boundary, are repaired.

[0198] Table 1 is a comparison of Prior-VDBFNet and other methods

[0199] Method Pr(％) Re (%) F1(％) mloU (%) PSPNet 92.98 92.48 92.73 86.44 DeepLabV3+ 93.66 93.77 93.72 88.18 K-Net 91.61 95.2 93.37 87.57 Mask2Former 90.26 93.74 91.96 85.12 SegNext 93.68 92.34 93.0 86.92 Prior-VDBFNet 96.32 94.39 95.33 91.44

[0200] Please refer to Figures 2 to 8 In order to verify the effectiveness of the ground fissure extraction deep learning method based on direction perception and Bayesian fusion proposed in the application, different types of typical crack regions are selected for comparative experiments, which cover complex scenes such as curved, bifurcated and weak feature regions. In the experiment, the method of the application is compared with various existing crack extraction networks, including DeepLabV3+, K-Net, SegNext, PSPNet, Mask2Former, etc.

[0201] In the complex crack morphology region (such as Figure 8In the case of p1 and p3, the method of the present application can maintain the continuity and topological integrity of the crack structure. The Prior-VDBFNet network uses a direction-aware dynamic convolution mechanism, which can dynamically adjust the convolution kernel shape along the crack direction in the feature extraction stage, thereby obtaining a more optimal directional response effect, which helps to improve the extraction accuracy of curved and branched structures. In contrast, some existing methods have problems such as crack interruption and discontinuous structure in the above-mentioned areas.

[0202] In the weak feature area where the edge of the ground crack is blurred and the contrast with the background is low (such as Figure 8 In the case of p2, p5 and p6, the present application effectively integrates direction-aware features and semantic features by introducing a Bayesian fusion module, and enhances the crack topological expression capability through the VSSB module. This mechanism helps to improve the model's ability to perceive weak cracks in complex texture backgrounds or occluded areas.

[0203] Further, in order to compare the performance of different methods, commonly used evaluation indicators including precision (Pr), recall (Re), F1-score and mean intersection over union (mIoU) are used to quantitatively evaluate the crack extraction effect of each method.

[0204] The experimental results show that the evaluation results of the method proposed in the present application on the public dataset 2 are: Pr is 96.32%, Re is 94.39%, F1-score is 95.33%, and mIoU is 91.44%. This group of indicators is better than existing similar methods, indicating that the present application has significant advantages in extraction integrity, recognition accuracy and robustness.

[0205] The parts not involved in the present application are the same as or can be implemented by the prior art. Although embodiments of the present application have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A deep learning method for ground fissure extraction based on direction perception and Bayesian fusion, characterized by: The method comprises the following steps: S1. Acquire drone images of a predetermined area and perform preprocessing, wherein the preprocessing includes one or more of radiation correction, geometric correction, registration, stitching, and cropping; S2. Creating a deep learning dataset using the preprocessed drone imagery, wherein the dataset includes labeled information about the ground fissure area; S3. Constructing a directional prior knowledge extraction module to generate a topologically coherent directional field, wherein, based on the ground fissure prior region, a multi-scale initial directional field is generated through structural tensor analysis, and the multi-scale initial directional fields are fused using the maximum coherence criterion to generate a basic directional map. Anisotropic diffusion and fast marching methods are implemented along the fracture skeleton line to propagate high-confidence directional information to optimize the topological coherence of the directional field; S4. Implementing multi-source feature fusion through a Bayesian probability fusion module, wherein an uncertainty propagation model of global semantic features and direction perception features is established, and the weight distribution of global semantic features and direction perception features is dynamically optimized based on the model; S5. Construct a Prior-VDBFNet network. In the encoding stage, a VMamba global semantic perception branch is constructed based on the VSSM architecture. In the decoding stage, an input channel-aware VSS decoding block (CAVSS) is used to construct a four-stage feature decoding architecture. In the feature fusion stage, global semantic features and direction perception features are integrated through the Bayesian probability fusion module.

2. The method according to claim 1, characterized in that In step S3, generating a basic directional pattern includes the following steps: Constructing a structure tensor based on the gradient vector field, for grayscale images Its gradient vector field ▽I=[I x , I y ] T , constituting the first-order differential characteristic; The structure tensor expression is: Among them, G σ is a two-dimensional Gaussian kernel with a standard deviation of σ, the symbol * represents the convolution operation, and the symbol represents the tensor product; Perform eigenvalue decomposition on the structure tensor to calculate the main direction angle θ and the directional coherence coefficient C: Among them, λ1≥λ2 is the eigenvalue, and ò is a very small positive number to prevent the denominator from being zero; A multi-scale spatial pyramid strategy is used for robust estimation. The scale set Σ = {16, 32, 64} is defined and the direction field θ is calculated for each scale σ∈Σ. σ and the coherence coefficient C σ , and finally generate the initial direction field through the maximum confidence criterion fusion: s * =arg max σ∈Σ C σ (x) Among them, θ init (x) is the pixel position x at the optimal scale σ * The initial direction estimate under the optimal scale σ * is the maximum coherence coefficient under scale σ.

3. The method according to claim 2, characterized in that The anisotropic diffusion and rapid marching method is implemented along the crack skeleton line in step S3, including the following steps: The anisotropic diffusion expression is constructed by combining the direction field function θ(x, t) with the direction constraint mechanism of the coherence coefficient: Where t is the time variable, D is the diffusion tensor, and λ is the regularization parameter; The diffusion tensor D is defined as: Where κ is the gradient sensitivity coefficient, φ = θ init (x) is the initial direction, and η>1 controls the anisotropy strength; Enforce directional continuity constraints during propagation: Where N(x) represents the eight-neighborhood of x, S is the skeleton line of the crack area, For each skeleton point x∈S on the crack skeleton line S, initialize the direction θ init (x), For each x in the eight-neighborhood N(x), calculate the direction difference and check the direction difference constraint for each y∈N(x)∩S. If |θ(x)-θ(y)|<π / 6, propagate the x direction to y: θ(x)→θ(y). Otherwise, retain the original direction of y or mark it as needing further optimization.

4. The method according to claim 1, wherein in, The propagation of high-confidence direction information in step S3 to optimize the direction field topology coherence includes: For the crack branch noise area in the prior position area, construct the directional feature vector of the hybrid feature space: Where (x i ,y i ) are pixel coordinates, Define the similarity matrix W: Among them, ‖f i -f j ‖ 2 is the square of the Euclidean distance between eigenvectors, σ s is a sensitive parameter controlling spatial proximity, σ θ is a sensitive parameter to control the direction consistency, similarity W ij The higher ∈[0, 1] is, the more similar i and j are in space and direction; K-means is used to perform spectral clustering on the similarity matrix W, and for each subcluster S k , calculate the directional statistics: in, is a subcluster S k The average direction angle, σ k is the sub-cluster direction standard deviation, Measuring the consistency of direction within the class, k The abnormal subclusters with >π / 12 are oriented and corrected, and the process is repeated until all subclusters meet the directional consistency constraint.

5. The method according to claim 1, wherein In step S4, the multi-source feature fusion is realized by the Bayesian probability fusion module, including the following steps: Construct Bayesian fusion objective function; Feature fusion guided by prior knowledge; and Based on the fusion results, the geometric consistency of features is enhanced through local optimization.

6. The method according to claim 5, characterized in that Constructing the Bayesian fusion objective function includes: Define the global characteristics of the VMamba backbone as The direction perception feature of DSCNet is defined as According to Bayes' theorem, the posterior distribution objective function is expressed as: p(F f |F g ,F d )∝p(F g |F f )p(F d |F f )p(F f ) Among them, p(F g |F f ) and p(F d |F f ) is the likelihood phase, p(F f ) is a priori term; Construct global features and directional feature covariance matrix Λ g , Λ d for: in is the feature of the ith channel of the global feature, is the feature of the i-th channel of the directional feature; The objective function used to minimize the comprehensive error during the optimization process is: Among them, Λ g and Λ d are the covariance matrices of the two types of features, R(F f ) is the regularization term.

7. The method according to claim 6, characterized in that Feature fusion guided by prior knowledge includes: Converting the objective function into a trainable neural network module; According to the Gaussian feature assumption, the inverse feature covariance matrix Each element of Calculate channel statistics through global average pooling: Among them, μ c is the global mean of the c-th channel, is the global feature map in the VMamba branch, H and W are the height and width of the feature map; Use two fully connected layers to generate attention weights: Among them, FC1 and FC2 are fully connected layers, and GELU is the activation function. Generate W d , the final fusion feature is: F f =W g ⊙F g +W d ⊙F d Among them, F g The global features extracted from VMamba backbone, F d is the direction-aware feature extracted by DSCNet, and ⊙ is the channel-by-channel multiplication.

8. The method according to claim 7, characterized in that Features enhanced through local optimization include: Through lightweight network D θ Realize iterative feature fusion: Among them, U (l) is the Bayesian weighted fusion feature of the lth layer, is the feature enhanced by local optimization, where D θ It consists of 3×3 convolution and residual connection. The gradient update rule of the iterative algorithm is: X (k+1) =prox f，η (X (k) -η▽g(X (k) )) X (k) is the feature estimate of the kth iteration, ▽g(X (k) is the data fidelity gradient, and η is the learning rate.

9. A computer program product comprising a computer program, characterized in that The computer program is executed to implement the deep learning method for ground fissure extraction based on direction perception and Bayesian fusion as described in any one of claims 1 to 8.

Images