A rock mass fracture sub-pixel level identification and parameter extraction method and system

Abstract

The application provides a rock mass crack sub-pixel level recognition and parameter extraction method and system, and belongs to the technical field of computer vision and digital image processing. The method synchronously generates a segmentation probability matrix and a boundary signed distance field matrix by constructing a multi-task convolutional neural network; sub-pixel level crack boundary extraction is realized by combining energy function optimization; the skeleton line is extracted by adopting constrained Delaunay triangulation and direction weight optimization to avoid topological distortion; and finally, geometric parameters are calculated based on high-precision boundary and skeleton data. The system includes image processing, boundary optimization, skeleton extraction and parameter calculation modules, and realizes the integration of recognition and parameter extraction. The application adopts the above rock mass crack sub-pixel level recognition and parameter extraction method and system, overcomes the problems of low precision, skeleton distortion and fragmented process of traditional methods, and improves the precision and reliability of rock mass crack recognition.

Description

Technical Field

[0001] This invention relates to the field of computer vision and digital image processing technology, and in particular to a method and system for subpixel-level identification and parameter extraction of rock mass fractures. Background Technology

[0002] Accurate identification and parameter extraction of rock mass fractures are of great significance for engineering geological analysis. Currently, this technical field mainly suffers from the following limitations:

[0003] Traditional image processing methods (such as edge detection and threshold segmentation) heavily rely on manual parameter adjustment, have poor adaptability to different imaging conditions, and are easily affected by noise. More importantly, these methods can only achieve pixel-level recognition accuracy and cannot meet the engineering requirements of high-precision measurement.

[0004] While recent deep learning-based methods (such as U-Net segmentation networks) have improved the automation of recognition, they are essentially still pixel-level segmentation, making it difficult to obtain smooth and accurate crack boundaries. Furthermore, these methods typically rely on subsequent independent skeleton extraction algorithms (such as thinning algorithms) to obtain the crack skeleton. However, traditional thinning algorithms are prone to topological distortion at crack intersections, destroying the true shape of the crack and leading to significant errors in subsequent parameter extraction.

[0005] In summary, existing technologies have not yet been able to effectively solve the integrated problem of sub-pixel-level precise localization of rock mass fractures, topology-preserving skeleton extraction, and accurate calculation of geometric parameters. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for subpixel-level identification and parameter extraction of rock mass fractures, which solves the problems of existing technologies that are difficult to achieve subpixel-level fracture boundary positioning, cannot effectively maintain the skeleton topology, and have a fragmented identification and parameter extraction process.

[0007] To achieve the above objectives, this invention provides a method for sub-pixel level identification and parameter extraction of rock mass fractures, comprising the following steps:

[0008] Step S1: Construct and train a multi-task convolutional neural network computing model to perform parallel computing processing on the input rock mass image data, and simultaneously generate pixel-level fracture segmentation probability matrix and boundary signed distance field matrix.

[0009] Step S2: Based on the segmentation probability matrix and the signed distance field matrix of the boundary, generate crack boundary vector data with sub-pixel precision;

[0010] Step S3: Based on the crack boundary vector data, generate crack skeleton line data with sub-pixel precision;

[0011] Step S4: Based on the crack boundary vector data and crack skeleton line data, output the set of geometric parameter data of the crack.

[0012] Preferably, step S1 includes the following steps:

[0013] Step S11: Construct a multi-task convolutional neural network computation model. This model adopts an encoder-decoder architecture, where the encoder contains five cascaded residual convolutional blocks, and each residual convolutional block contains two... Convolutional layers and cross-layer connections are used. The decoder adopts a feature pyramid structure and fuses feature maps of different scales from the encoder through bilinear interpolation upsampling and skip connections.

[0014] Step S12: Construct a cross-task feature interaction module, which uses a cross-attention mechanism to interact the deep features of the segmentation task branch with the corresponding features of the distance field regression task branch.

[0015] ;

[0016] in, This represents the fused feature map. This represents the feature map for the segmentation task. Represents the feature map of the distance field task. This represents the Sigmoid activation function. This indicates element-wise multiplication. express Convolution operation;

[0017] Step S13: Train a multi-task convolutional neural network computation model, using a dynamic weight multi-task loss function based on gradient statistics.

[0018] Preferably, step S13 includes the following steps:

[0019] Step S131: Calculate the Dice loss value for the segmentation task. Smooth L1 loss value for distance field regression task ;

[0020] Step S132: In each training batch, calculate the relative proportion of the losses for the two tasks:

[0021] ;

[0022] ;

[0023] Step S133: Dynamically adjust the weighting coefficients according to the relative proportions:

[0024] ;

[0025] ;

[0026] in, This represents the weight coefficient for the segmentation task. This represents the weight coefficients for the distance field regression task;

[0027] Step S134: Calculate the total loss function:

[0028] ;

[0029] And then propagate back to update the network parameters.

[0030] Preferably, step S2 includes the following steps:

[0031] Step S21: Based on the boundary signed distance field matrix Data items for constructing the energy function :

[0032] ;

[0033] in, This represents the boundary contour function to be optimized. Represents the x-coordinate on the image plane. Represents the ordinate on the image plane. Dirac function, Represents the image domain;

[0034] Step S22: Construct the smoothing term of the energy function Anisotropic constraints based on the geological characteristics of rock mass fractures are introduced:

[0035] ;

[0036] in, Indicates the boundary profile in arc length parameter curvature at that point Indicates based on boundary tangent direction Direction weight function, Represents the balance coefficient. Indicates the boundary outline;

[0037] Step S23: Construct the distance field gradient regularization term:

[0038] ;

[0039] in, Represents the regularization coefficient. express The gradient;

[0040] Step S24: Combine to obtain the complete energy function:

[0041] ;

[0042] in, Representing the complete energy function, and solving it numerically using the gradient descent method, specifically including:

[0043] Step S241: Use the binary conversion result of the segmentation probability matrix as the initial contour. ;

[0044] Step S242: Calculate the energy function for the boundary profile. gradient ;

[0045] Step S243: Update the contour according to the gradient direction:

[0046] ;

[0047] in, Indicates the learning rate. Indicates the first Boundary contour function of the next iteration Indicates the first The boundary contour function of the next iteration;

[0048] Step S244, repeat steps S242-S243 until... Generate sub-pixel precision crack boundary vector data ,in Indicates the first Boundary contour function at the next iteration The corresponding energy function value, This represents the convergence threshold.

[0049] Preferably, the direction weighting function in step S22 Specifically as follows:

[0050] Step S221: Calculate the tangent direction angle at each point on the boundary profile. :

[0051] ;

[0052] in, and Both represent the parametric equations of the boundary profile;

[0053] Step S222: Based on the geological characteristics of rock mass fractures, apply the directional weighting function:

[0054] .

[0055] Preferably, step S3 includes the following steps:

[0056] Step S31, for Perform constrained Delaunay triangulation calculations to ensure that boundary segments serve as constraint edges;

[0057] Step S32: Calculate the coordinates of the circumcenter of each Delaunay triangle;

[0058] Step S33: Connect the circumcenters of adjacent triangles to form the initial skeleton lines;

[0059] Step S34: Based on the direction weight function The initial skeleton lines are smoothed and optimized, specifically including:

[0060] Step S341: Calculate the length of each branch in the initial skeleton line. and direction ;

[0061] Step S342: For each branch, calculate its direction weight. ;

[0062] Step S343, when the branch length If so, then delete that branch, where Indicates a preset threshold;

[0063] Step S344: Apply direction-weighted B-spline curve fitting to the retained skeleton segments to generate the final fracture skeleton line data. .

[0064] Preferably, the crack width calculation in step S4 includes the following steps:

[0065] Step S41: For fracture skeleton line data Each point in Calculate the shortest distance from it to both sides of the boundary. and ,in, Point of coordinate, Point of coordinate, This represents the shortest distance to the left boundary. Indicates the shortest distance to the right boundary;

[0066] Step S42: Calculate the crack width at this point. :

[0067] ;

[0068] Step S43: Perform anisotropic smoothing filtering on the width sequence based on the skeleton line direction weight function, specifically including:

[0069] Step S431: Calculate the tangent angle at each point on the fracture skeleton line. ;

[0070] Step S432: Construct the anisotropic Gaussian kernel function:

[0071] ;

[0072] in, Indicates the distance along the fracture skeleton line. Indicates the scale parameter. This indicates the tangent direction of the fracture skeleton line at that point. This represents the value of the anisotropic Gaussian kernel function. The weighting function represents the direction of the fracture skeleton line;

[0073] Step S433: Apply anisotropic Gaussian kernels to perform convolution filtering on the width sequence:

[0074] ;

[0075] in, This represents the smoothed width value. and This represents the arc length parameter along the skeleton line. Indicates the first tangent direction of the skeleton line at each point Indicates the first The original width value of each point;

[0076] Step S434: Count the width values ​​of all points, and calculate the average width and maximum width.

[0077] Preferably, the crack length in step S4 is calculated as follows:

[0078] Step S401: Represent the fracture skeleton line data as an ordered vertex sequence;

[0079] Step S402: Calculate the Euclidean distance between adjacent vertices;

[0080] Step S403: Calculate the total crack length by summing the lengths of all line segments;

[0081] Step S404: Perform direction sensitivity correction on length measurement based on the direction weighting function, specifically including:

[0082] Step S4041: Calculate the direction angle of each segment of the fracture skeleton line;

[0083] Step S4042: Calculate the average direction angle of the fracture skeleton line;

[0084] Step S4043: Correct the total length according to the direction weight function value corresponding to the average direction angle, wherein the length correction amount of the horizontal direction crack is less than that of cracks in other directions.

[0085] Preferably, the fracture orientation calculation in step S4 is as follows:

[0086] Step S405: Adaptively sample and extract key turning points from the fracture skeleton line data. Specifically, the Douglas-Puk algorithm is used to adaptively select key points based on curvature changes.

[0087] Step S406: Calculate the direction vector of the line segment between adjacent key points;

[0088] Step S407: Perform weighted statistics on the direction vector based on the direction weight function, specifically including:

[0089] Step S4071: Assign weights to each line segment direction. The weight values ​​are calculated by the direction weight function based on the line segment direction.

[0090] Step S4072: Construct a weighted directional distribution histogram;

[0091] Step S408: Determine the dominant attitude parameters by means of kernel density estimation method, specifically including kernel density estimation of weighted direction distribution and finding the direction with the highest density as the dominant attitude.

[0092] This invention also provides a sub-pixel level identification and parameter extraction system for rock mass fractures, comprising:

[0093] The image processing module is used to build and train a multi-task convolutional neural network computing model, perform parallel computing processing on the input rock mass image data, and simultaneously generate pixel-level fracture segmentation probability matrix and boundary signed distance field matrix.

[0094] The boundary optimization module is used to generate sub-pixel precision crack boundary vector data based on the segmentation probability matrix and the signed distance field matrix of the boundary through energy function optimization.

[0095] The skeleton extraction module is used to perform constrained Delaunay triangulation and orientation weight optimization based on the crack boundary vector data to generate crack skeleton line data with sub-pixel accuracy.

[0096] The parameter calculation module is used to calculate and output a set of geometric parameter data of the fracture based on the fracture boundary vector data and the fracture skeleton line data.

[0097] Therefore, the present invention employs the above-described sub-pixel-level identification and parameter extraction method and system for rock mass fractures, and the beneficial technical effects are as follows:

[0098] (1) This invention generates a segmentation probability matrix and a signed distance field matrix at the boundary simultaneously through a multi-task convolutional neural network, and constructs an energy function based on these for optimization, thereby evolving the initial pixel-level segmentation result to sub-pixel accuracy. This method overcomes the limitation of traditional image processing methods and ordinary deep learning models, which can only obtain jagged pixel-level boundaries.

[0099] (2) This invention is based on sub-pixel boundary vector data. It extracts the initial skeleton by constrained Delaunay triangulation and combines it with a directional weight function based on geological characteristics for optimization and B-spline fitting. This method avoids the topological distortions (such as burrs and breaks) that are easily generated at fracture intersections and endpoints in traditional thinning algorithms, and ensures that the skeleton lines can truly reflect the extension path and connectivity of the fractures.

[0100] (3) This invention integrates boundary identification, skeleton extraction and geometric parameter calculation into a coherent framework. Based on high-precision boundary and skeleton data, key parameters such as the width, length and orientation of fractures are directly calculated. Anisotropic smoothing and direction sensitivity correction are introduced to reduce the cumulative error caused by process fragmentation and intermediate result errors. Finally, a set of geometric parameters with high reliability and good consistency is output, which improves the efficiency and accuracy of engineering geological analysis. Attached Figure Description

[0101] Figure 1 This is a flowchart of a method for sub-pixel level identification and parameter extraction of rock mass fractures according to the present invention;

[0102] Figure 2 This is a schematic diagram of a multi-task convolutional neural network structure;

[0103] Figure 3 This is an architecture diagram of a rock mass fracture subpixel-level identification and parameter extraction system according to the present invention. Detailed Implementation

[0104] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0105] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0106] Example 1

[0107] like Figure 1 As shown, a method for sub-pixel level identification and parameter extraction of rock mass fractures includes the following steps:

[0108] Step S1: Construct and train a multi-task convolutional neural network computation model to perform parallel computation processing on the input rock mass image data, and simultaneously generate pixel-level fracture segmentation probability matrix and boundary signed distance field matrix.

[0109] Step S11: Construct a multi-task convolutional neural network computation model, which adopts an encoder-decoder architecture.

[0110] like Figure 2 The encoder contains five cascaded residual convolutional blocks, each containing two... The decoder consists of convolutional layers, batch normalization layers, ReLU activation functions, and cross-layer connections (residual connections) to extract multi-level features. The decoder employs a feature pyramid structure, fusing feature maps from different scales of the encoder through bilinear interpolation upsampling and skip connections to enhance feature representation while preserving detailed information.

[0111] Step S12: Construct a cross-task feature interaction module, which uses a cross-attention mechanism to interact the deep features of the segmentation task branch with the corresponding features of the distance field regression task branch.

[0112] ;

[0113] in, This represents the fused feature map. This represents the feature map for the segmentation task. Represents the feature map of the distance field task. This represents the Sigmoid activation function. This indicates element-wise multiplication. express Convolution operation.

[0114] Step S13: Train a multi-task convolutional neural network computation model and use a dynamic weight multi-task loss function based on gradient statistics to balance the learning progress of segmentation and distance field regression tasks.

[0115] Step S131: Calculate the Dice loss value for the segmentation task. Smooth L1 loss value for distance field regression task ;

[0116] Step S132: In each training batch, calculate the relative proportion of the losses for the two tasks:

[0117] ;

[0118] ;

[0119] Step S133: Dynamically adjust the weighting coefficients according to the relative proportions:

[0120] ;

[0121] ;

[0122] in, This represents the weight coefficient for the segmentation task. This represents the weight coefficients for the distance field regression task;

[0123] Step S134: Calculate the total loss function:

[0124] ;

[0125] And then propagate back to update the network parameters.

[0126] The above technical solution achieves the following results: by employing a multi-task collaboration and feature interaction mechanism, it solves the problem of insufficient information utilization in traditional single-task segmentation networks. Existing technologies (such as U-Net) only output segmentation probability maps, resulting in coarse boundary information; while this invention innovatively introduces a signed distance field as a parallel regression task, which itself contains sub-pixel-level boundary localization information. Through a cross-task feature interaction module, the semantic features of the segmentation task and the geometric features of the distance field task are deeply integrated, enabling the network to understand the transition relationship between inside and outside the boundary. The dynamic weight loss function ensures that the two tasks co-evolve during training, rather than competing with each other, avoiding performance degradation caused by task conflicts.

[0127] Step S2: Based on the segmentation probability matrix and the signed distance field matrix of the boundary, generate sub-pixel precision crack boundary vector data.

[0128] Step S21: Based on the boundary signed distance field matrix Data items for constructing the energy function :

[0129] ;

[0130] in, This represents the boundary contour function to be optimized. Represents the x-coordinate on the image plane. Represents the ordinate on the image plane. Dirac function, Represents the image domain.

[0131] Step S22: Construct the smoothing term of the energy function Anisotropic constraints based on the geological characteristics of rock mass fractures are introduced:

[0132] ;

[0133] in, Indicates the boundary profile in arc length parameter curvature at that point Indicates based on boundary tangent direction Direction weight function, Represents the balance coefficient. Indicates the boundary outline;

[0134] Directional weighting function Specifically as follows:

[0135] Step S221: Calculate the tangent direction angle at each point on the boundary profile. :

[0136] ;

[0137] in, and Both represent the parametric equations of the boundary profile;

[0138] Step S222: Based on the geological characteristics of rock mass fractures, namely that rock mass fractures have better continuity and extension in the horizontal direction, apply the directional weighting function:

[0139] ;

[0140] This function ensures that horizontal cracks have a low smoothing constraint to maintain natural extension, while imposing a strong smoothing penalty on noise and anomalous bending in other directions.

[0141] Step S23: Construct the distance field gradient regularization term:

[0142] ;

[0143] in, Represents the regularization coefficient. express The gradient.

[0144] Step S24: Combine to obtain the complete energy function:

[0145] ;

[0146] in, Representing the complete energy function, and solving it numerically using the gradient descent method, specifically including:

[0147] Step S241: Use the binary conversion result of the segmentation probability matrix as the initial contour. ;

[0148] Step S242: Calculate the energy function for the boundary profile. gradient ;

[0149] Step S243: Update the contour according to the gradient direction:

[0150] ;

[0151] in, Indicates the learning rate. Indicates the first Boundary contour function of the next iteration Indicates the first The boundary contour function of the next iteration;

[0152] Step S244, repeat steps S242-S243 until... Generate sub-pixel boundary vector data ,in Indicates the first Boundary contour function at the next iteration The corresponding energy function value, This represents the convergence threshold.

[0153] The above technical solution achieves the following results: Traditional methods (such as active contour models) typically rely solely on image gradients, which easily get trapped in local optima on rock mass images with complex textures. This invention constructs an energy function that integrates a distance field data term, an anisotropic smoothing term, and a gradient regularization term. Its innovation lies mainly in: 1) The data term utilizes the distance field predicted by the network as the driving force. This field crosses zero at the true boundary, providing a strong and precise attraction, overcoming the shortcomings of traditional gradient information; 2) The smoothing term introduces a directional weight function based on geological priors. It acknowledges that rock mass fractures have different extension characteristics in different directions (e.g., stronger continuity in the horizontal direction), thus implementing anisotropic smoothing constraints rather than a uniform smoothing with a one-size-fits-all approach. This is more in line with geological laws, effectively preserving the natural morphology of fractures and suppressing non-geologically significant bending; 3) The regularization term ensures the stability of the optimization process. Ultimately, through the optimization of this energy function, the pixel-level, jagged initial boundary is transformed into a smooth, continuous vector boundary with sub-pixel precision, achieving a transformation from discrete to continuous and from coarse to precise.

[0154] Step S3: Generate sub-pixel precision crack skeleton line data based on crack boundary vector data.

[0155] Step S31, for Perform constrained Delaunay triangulation calculations to ensure that boundary segments serve as constraint edges.

[0156] Step S32: Calculate the coordinates of the circumcenter of each Delaunay triangle.

[0157] Step S33: Connect the circumcenters of adjacent triangles to form the initial skeleton line.

[0158] Step S34: Based on the direction weight function Smoothing optimization of skeleton lines includes:

[0159] Step S341: Calculate the length of each branch in the skeleton line. and direction ;

[0160] Step S342: For each branch, calculate its direction weight. ;

[0161] Step S343, when the branch length ,in If the threshold is set, then delete the branch.

[0162] Step S344: Apply direction-weighted B-spline curve fitting to the retained skeleton line segments to generate the final skeleton line data. .

[0163] The above technical solution achieves the following results: It proposes a new paradigm for skeleton extraction and optimization that combines geometry-driven and geological knowledge-guided approaches, completely solving the topological distortion problem inherent in traditional thinning algorithms. Existing skeleton extraction methods, when dealing with fracture intersections and endpoints, rely solely on local pixel patterns, easily generating topological errors such as burrs, breakpoints, and false branches. This invention uses constrained Delaunay triangulation (CDT) to generate the initial skeleton based on high-precision sub-pixel boundaries. This method is geometrically rigorous, preserving boundary constraints and the topological structure of the region. Subsequently, the geological prior knowledge of the direction weight function is used again to filter (remove excessively short branches) and optimize (B-spline fitting) the initial skeleton. This optimization process not only smooths the geometry but also enhances geological rationality, ensuring that the final generated skeleton lines are not only highly accurate and smooth and continuous but also that their topological structure truly reflects the extension trend and connectivity of fractures within the geological body.

[0164] Step S4: Based on the crack boundary vector data and crack skeleton line data, output the set of geometric parameter data of the crack.

[0165] Step S41: For fracture skeleton line data Each point in Calculate the shortest distance from it to both sides of the boundary. and ,in, Point of coordinate, Point of coordinate, This represents the shortest distance to the left boundary. This represents the shortest distance to the right boundary.

[0166] Step S42: Calculate the crack width at this point. :

[0167] .

[0168] Step S43: Perform anisotropic smoothing filtering on the width sequence based on the skeleton line direction weight function, specifically including:

[0169] Step S431: Calculate the tangent direction angle at each point on the skeleton line. ;

[0170] Step S432: Construct the anisotropic Gaussian kernel function:

[0171] ;

[0172] in, Indicates the distance along the skeleton line. Indicates the scale parameter. This indicates the direction of the tangent to the skeleton line at that point. This represents the value of the anisotropic Gaussian kernel function. The skeleton line direction weighting function has the same form as in step S22. same;

[0173] Step S433: Apply anisotropic Gaussian kernels to perform convolution filtering on the width sequence:

[0174] ;

[0175] in, This represents the smoothed width value. and This represents the arc length parameter along the skeleton line. Indicates the first tangent direction of the skeleton line at each point Indicates the first The original width value of each point;

[0176] Step S434: Count the width values ​​of all points, and calculate the average width and maximum width.

[0177] The crack length is calculated as follows:

[0178] Step S401: Represent the fracture skeleton line data as an ordered vertex sequence;

[0179] Step S402: Calculate the Euclidean distance between adjacent vertices;

[0180] Step S403: Calculate the total crack length by summing the lengths of all line segments;

[0181] Step S404: Perform direction sensitivity correction on length measurement based on the direction weighting function, specifically including:

[0182] Step S4041: Calculate the direction angle of each segment of the skeleton line;

[0183] Step S4042: Calculate the average direction angle of the skeleton lines;

[0184] Step S4043: Correct the total length according to the direction weight function value corresponding to the average direction angle, wherein the length correction amount of the horizontal direction crack is less than that of cracks in other directions.

[0185] The fracture orientation is calculated as follows:

[0186] Step S405: Adaptively sample and extract key turning points from the fracture skeleton line data. Specifically, the Douglas-Puk algorithm is used to adaptively select key points based on curvature changes.

[0187] Step S406: Calculate the direction vector of the line segment between adjacent key points;

[0188] Step S407: Perform weighted statistics on the direction vector based on the direction weight function, specifically including:

[0189] Step S4071: Assign weights to each line segment direction. The weight values ​​are calculated by the direction weight function based on the line segment direction.

[0190] Step S4072: Construct a weighted directional distribution histogram;

[0191] Step S408: Determine the dominant attitude parameters by means of kernel density estimation method, specifically including kernel density estimation of weighted direction distribution and finding the direction with the highest density as the dominant attitude.

[0192] The effects of the above technical solution are as follows: This step combines high-precision geometric data with a geological model, achieving an upgrade from graphic measurement to intelligent calculation of geological parameters. Traditional methods directly measure parameters from binary images, which is limited by pixel accuracy and noise, and the calculation process for each parameter is independent, resulting in serious error accumulation. This invention, based on the sub-pixel boundaries and topologically accurate skeleton generated in the preceding steps, allows parameter calculation to be based on higher-order and more reliable data. In the width calculation, anisotropic Gaussian filtering is introduced, and its filter kernel is no longer isotropic but adaptively adjusted according to the direction of the skeleton line. This protects the natural variation in width along the fracture direction while effectively smoothing abnormal fluctuations in the vertical direction, making the calculation results more consistent with geological understanding. The length calculation performs direction sensitivity correction, recognizing and compensating for the differences in projection deformation of fractures with different attitudes during imaging, thus improving the accuracy of the measurement. The attitude calculation intelligently extracts the dominant direction by performing weighted statistics and kernel density estimation on the skeleton line, rather than simple averaging, which better reflects the dominant attitude of the fracture group. The final output of this step is a set of internally consistent, high-precision geometric parameters, which improves the automation level and reliability of engineering geological analysis.

[0193] The invention will be further illustrated below with specific examples.

[0194] In the rock mass of the left bank slope of a hydropower station, the method of this invention is used to systematically identify and extract parameters of the rock mass fissures.

[0195] Eighty-five images of the rock surface were acquired using a drone equipped with a high-resolution camera (42 megapixels), covering an area of ​​approximately 1200 m². 2 During implementation, the network was fine-tuned using 320 labeled images obtained in the early stages of the project, with 50 training cycles. The final model achieved a segmentation IoU of 92.3% on the validation set and a distance field prediction MAE of 1.2 pixels. Boundary optimization was performed on 85 test images, with an average processing time of 4.2 seconds per image. The optimized boundaries, compared with the results of manual fine annotation, showed an average Hausdorff distance of 0.31 pixels. A total of 427 effective gap skeletons were extracted, including 38 complex intersecting gaps. Compared with traditional thinning algorithms, the method of this invention achieved a topological accuracy of 98.5% at intersections.

[0196] Select the same area (50m) 2 The methods were compared, and the results are shown in Table 1.

[0197] Table 1 Analysis Results

[0198]

[0199] The comparison results show that this method exhibits higher sensitivity in identifying completeness, identifying 6 to 14 more cracks than traditional methods, effectively avoiding missed detections. It also achieves a qualitative leap in measurement accuracy, with relative errors in length and width measurements significantly lower than all comparative methods, and an attitude estimation bias of only 3.5, fully demonstrating the technical advantages of sub-pixel-level processing and high-fidelity skeleton extraction. Furthermore, in terms of overall efficiency, although the fully automated process takes 35 minutes, it eliminates the large amount of manual post-processing, verification, and parameter measurement required by traditional methods, resulting in a significantly lower overall time cost. Thus, it achieves superior overall work efficiency while ensuring the highest accuracy.

[0200] Example 2

[0201] like Figure 3 As shown, a sub-pixel-level identification and parameter extraction system for rock mass fractures includes:

[0202] The image processing module is used to build and train a multi-task convolutional neural network computing model, perform parallel computing processing on the input rock mass image data, and simultaneously generate pixel-level fracture segmentation probability matrix and boundary signed distance field matrix.

[0203] The boundary optimization module is used to generate sub-pixel precision crack boundary vector data based on the segmentation probability matrix and the signed distance field matrix of the boundary through energy function optimization.

[0204] The skeleton extraction module is used to perform constrained Delaunay triangulation and orientation weight optimization based on the crack boundary vector data to generate crack skeleton line data with sub-pixel accuracy.

[0205] The parameter calculation module is used to calculate and output a set of geometric parameter data of the fracture based on the fracture boundary vector data and the fracture skeleton line data.

[0206] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0207] Therefore, the present invention employs the above-mentioned sub-pixel level identification and parameter extraction method and system for rock mass fractures, which can improve the accuracy and reliability of fracture identification and overcome the inherent defects of traditional methods in boundary positioning, skeleton extraction and parameter calculation.

[0208] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for sub-pixel-level identification and parameter extraction of rock mass fractures, characterized in that, Includes the following steps: Step S1: Construct and train a multi-task convolutional neural network computing model to perform parallel computing processing on the input rock mass image data, and simultaneously generate pixel-level fracture segmentation probability matrix and boundary signed distance field matrix. Step S2: Based on the segmentation probability matrix and the signed distance field matrix of the boundary, generate crack boundary vector data with sub-pixel precision; Step S3: Based on the crack boundary vector data, generate crack skeleton line data with sub-pixel precision; Step S4: Based on the fracture boundary vector data and fracture skeleton line data, output the set of geometric parameter data of the fracture; Step S2 includes the following steps: Step S21: Based on the boundary signed distance field matrix Data items for constructing the energy function : ； in, This represents the boundary contour function to be optimized. Represents the x-coordinate on the image plane. Represents the ordinate on the image plane. Dirac function, Represents the image domain; Step S22: Construct the smoothing term of the energy function Anisotropic constraints based on the geological characteristics of rock mass fractures are introduced: ； in, Indicates the boundary profile in arc length parameter curvature at that point Indicates based on boundary tangent direction Direction weight function, Represents the balance coefficient. Indicates the boundary outline; Step S23: Construct the distance field gradient regularization term : ； in, Represents the regularization coefficient. express The gradient; Step S24: Combine to obtain the complete energy function: ； in, Representing the complete energy function, and solving it numerically using the gradient descent method, specifically including: Step S241: Use the binary conversion result of the segmentation probability matrix as the initial contour. ; Step S242: Calculate the energy function for the boundary profile. gradient ; Step S243: Update the contour according to the gradient direction: ； in, Indicates the learning rate. Indicates the first Boundary contour function of the next iteration Indicates the first The boundary contour function of the next iteration; Step S244, repeat steps S242-S243 until... Generate sub-pixel precision crack boundary vector data ,in Indicates the first Boundary contour function at the next iteration The corresponding energy function value, This represents the convergence threshold.

2. The method for sub-pixel level identification and parameter extraction of rock mass fractures according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: Construct a multi-task convolutional neural network computation model. This model adopts an encoder-decoder architecture, where the encoder contains five cascaded residual convolutional blocks, and each residual convolutional block contains two... Convolutional layers and cross-layer connections are used. The decoder adopts a feature pyramid structure and fuses feature maps of different scales from the encoder through bilinear interpolation upsampling and skip connections. Step S12: Construct a cross-task feature interaction module, which uses a cross-attention mechanism to interact the deep features of the segmentation task branch with the corresponding features of the distance field regression task branch. ； in, This represents the fused feature map. This represents the feature map for the segmentation task. Represents the feature map of the distance field task. This represents the Sigmoid activation function. This indicates element-wise multiplication. express Convolution operation; Step S13: Train a multi-task convolutional neural network computation model, using a dynamic weight multi-task loss function based on gradient statistics.

3. The method for sub-pixel level identification and parameter extraction of rock mass fractures according to claim 2, characterized in that, Step S13 includes the following steps: Step S131: Calculate the Dice loss value for the segmentation task. Smooth L1 loss value for distance field regression task ; Step S132: In each training batch, calculate the relative proportion of the losses for the two tasks: ； ； Step S133: Dynamically adjust the weighting coefficients according to the relative proportions: ； ； in, This represents the weight coefficient for the segmentation task. This represents the weight coefficients for the distance field regression task; Step S134: Calculate the total loss function: ； And then propagate back to update the network parameters.

4. The method for sub-pixel level identification and parameter extraction of rock mass fractures according to claim 3, characterized in that, Direction weighting function in step S22 Specifically as follows: Step S221: Calculate the tangent direction angle at each point on the boundary profile. : ； in, and Both represent the parametric equations of the boundary profile; Step S222: Based on the geological characteristics of rock mass fractures, apply the directional weighting function: 。 5. The method for sub-pixel level identification and parameter extraction of rock mass fractures according to claim 4, characterized in that, Step S3 includes the following steps: Step S31, for Perform constrained Delaunay triangulation calculations to ensure that boundary segments serve as constraint edges; Step S32: Calculate the coordinates of the circumcenter of each Delaunay triangle; Step S33: Connect the circumcenters of adjacent triangles to form the initial skeleton lines; Step S34: Based on the direction weight function The initial skeleton lines are smoothed and optimized, specifically including: Step S341: Calculate the length of each branch in the initial skeleton line. and direction ; Step S342: For each branch, calculate its direction weight. ; Step S343, when the branch length If so, then delete that branch, where Indicates a preset threshold; Step S344: Apply direction-weighted B-spline curve fitting to the retained skeleton segments to generate the final fracture skeleton line data. .

6. The method for sub-pixel level identification and parameter extraction of rock mass fractures according to claim 5, characterized in that, The crack width calculation in step S4 includes the following steps: Step S41: For fracture skeleton line data Each point in Calculate the shortest distance from it to both sides of the boundary. and ,in, Point of coordinate, Point of coordinate, This represents the shortest distance to the left boundary. Indicates the shortest distance to the right boundary; Step S42: Calculate the crack width at this point. : ； Step S43: Perform anisotropic smoothing filtering on the width sequence based on the skeleton line direction weight function, specifically including: Step S431: Calculate the tangent angle at each point on the fracture skeleton line. ; Step S432: Construct the anisotropic Gaussian kernel function: ； in, Indicates the distance along the fracture skeleton line. Indicates the scale parameter. This indicates the tangent direction of the fracture skeleton line at that point. This represents the value of the anisotropic Gaussian kernel function. The weighting function represents the direction of the fracture skeleton line; Step S433: Apply anisotropic Gaussian kernels to perform convolution filtering on the width sequence: ； in, This represents the smoothed width value. and This represents the arc length parameter along the skeleton line. Indicates the first tangent direction of the skeleton line at each point Indicates the first The original width value of each point; Step S434: Count the width values ​​of all points, and calculate the average width and maximum width.

7. The method for sub-pixel level identification and parameter extraction of rock mass fractures according to claim 6, characterized in that, The crack length in step S4 is calculated as follows: Step S401: Represent the fracture skeleton line data as an ordered vertex sequence; Step S402: Calculate the Euclidean distance between adjacent vertices; Step S403: Calculate the total crack length by summing the lengths of all line segments; Step S404: Perform direction sensitivity correction on length measurement based on the direction weighting function, specifically including: Step S4041: Calculate the direction angle of each segment of the fracture skeleton line; Step S4042: Calculate the average direction angle of the fracture skeleton line; Step S4043: Correct the total length according to the direction weight function value corresponding to the average direction angle, wherein the length correction amount of the horizontal direction crack is less than that of cracks in other directions.

8. The method for sub-pixel level identification and parameter extraction of rock mass fractures according to claim 7, characterized in that, The fracture orientation calculation in step S4 is as follows: Step S405: Adaptively sample and extract key turning points from the fracture skeleton line data. Specifically, the Douglas-Puk algorithm is used to adaptively select key points based on curvature changes. Step S406: Calculate the direction vector of the line segment between adjacent key points; Step S407: Perform weighted statistics on the direction vector based on the direction weight function, specifically including: Step S4071: Assign weights to each line segment direction. The weight values ​​are calculated by the direction weight function based on the line segment direction. Step S4072: Construct a weighted directional distribution histogram; Step S408: Determine the dominant attitude parameters by means of kernel density estimation method, specifically including kernel density estimation of weighted direction distribution and finding the direction with the highest density as the dominant attitude.

9. A sub-pixel-level identification and parameter extraction system for rock mass fractures, characterized in that, A method for implementing a sub-pixel-level identification and parameter extraction of rock mass fractures as described in any one of claims 1-8 includes: The image processing module is used to build and train a multi-task convolutional neural network computing model, perform parallel computing processing on the input rock mass image data, and simultaneously generate pixel-level fracture segmentation probability matrix and boundary signed distance field matrix. The boundary optimization module is used to generate sub-pixel precision crack boundary vector data based on the segmentation probability matrix and the signed distance field matrix of the boundary through energy function optimization. The skeleton extraction module is used to perform constrained Delaunay triangulation and orientation weight optimization based on the crack boundary vector data to generate crack skeleton line data with sub-pixel accuracy. The parameter calculation module is used to calculate and output a set of geometric parameter data of the fracture based on the fracture boundary vector data and the fracture skeleton line data.

Images