Slope crack high-resolution image intelligent identification and quantification method and system

Abstract

The invention discloses a slope crack high-resolution image intelligent identification and quantification method and system, and relates to the technical field of image processing and computer vision. Comprising the steps that S1, multi-scale illumination equalization processing and multi-scale Gaussian difference filtering processing based on the Retinex theory are sequentially executed on an obtained slope high-resolution image so as to compensate non-uniform illumination and suppress background noise; s2, extracting an initial edge of the preprocessed image by using an improved Canny edge detection algorithm; according to the method, multi-scale illumination equalization and multi-scale Gaussian difference filtering based on the Retinex theory are performed on a slope high-resolution image, non-uniform illumination is effectively compensated, background noise is suppressed, and the problems of illumination distortion and noise interference in a field scene are solved; improved Canny edge detection is combined with gray gradient, direction consistency and texture continuity feature weighted fusion, effective crack segments are accurately marked, and false crack misjudgment is avoided.

Description

Technical Field

[0001] This invention relates to the fields of image processing and computer vision technology, specifically to a method and system for intelligent recognition and quantification of high-resolution images of slope cracks. Background Technology

[0002] Slope safety monitoring is a crucial link in geological disaster early warning and infrastructure protection, with its core lying in the early identification and accurate quantification of surface cracks on slopes. With the development of high-resolution imaging technology, acquiring slope images using digital cameras or drone platforms has become the mainstream approach. Combining this with computer vision algorithms enables automatic crack extraction and parameter measurement, significantly improving monitoring efficiency and coverage. These methods typically rely on image processing workflows to segment, skeletonize, and calculate geometric features of cracked areas, and under ideal conditions, can effectively support slope deformation trend analysis and risk assessment.

[0003] Among them, crack image intelligent recognition technology focuses on accurately separating weak, slender, and irregularly shaped target structures from complex natural backgrounds and reconstructing their continuous topological paths to support subsequent quantification tasks. The basic goal of this technology is to overcome the adverse effects of illumination variations, texture interference, and imaging noise while preserving the true direction of cracks, thereby providing a reliable basis for slope stability assessment.

[0004] Current technologies still face the following challenges when dealing with real-world field scenarios. Due to long-distance photography, non-uniform lighting, and sensor limitations, real cracks often appear as fragmented pixel segments or isolated clusters of points, lacking stable edge continuity. Furthermore, current mainstream algorithms, mostly based on region growing, edge connecting, or traditional morphological operations, struggle to recover the true connected structure of cracks in noisy environments, resulting in numerous breakpoints and omissions in the identification results. This not only leads to a severe underestimation of crack length but also distorts its spatial distribution, thereby weakening the scientific validity and reliability of slope evolution modeling and critical state determination based on image data. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for intelligent recognition and quantification of high-resolution images of slope cracks. This solves the problems of existing technologies, such as difficulty in recovering the true connectivity structure of cracks under strong noise backgrounds, omission of breakpoints in recognition, and underestimation of crack length and distortion of morphology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligent identification and quantification of high-resolution images of slope cracks, comprising: S1. Perform multi-scale illumination equalization and multi-scale Gaussian difference filtering based on Retinex theory on the acquired high-resolution image of the slope in sequence to compensate for non-uniform illumination and suppress background noise. S2. For the image after multi-scale illumination equalization and multi-scale Gaussian difference filtering, the improved Canny edge detection algorithm is used to extract the initial edge. Based on the initial edge, the gray-level gradient feature, orientation consistency feature and texture continuity feature of each candidate crack segment are calculated. The three types of features are weighted and fused to obtain the fused feature value. Candidate segments with fused feature values ​​greater than the preset fused threshold are marked as valid crack segments. S3. Reconstruct the topological connectivity of all valid crack segments, including: calculating the comprehensive similarity between any two adjacent valid crack segments, which is composed of weighted similarity of direction, gray-level gradient, and texture; for segment pairs with a comprehensive similarity greater than a preset threshold, dynamic constraint judgment is further performed based on angle constraints, distance constraints, and noise region exclusion rules; for segment pairs that pass the dynamic constraint judgment, curve fitting is used to generate connected paths, and pseudo-branches that do not conform to the natural crack morphology and whose angle with the main crack is less than 30° and whose length accounts for less than 20% of the length of the main crack are removed, so as to output a continuous crack topology structure. S4. The continuous crack topology is skeletonized, and the geometric parameters of the crack are calculated based on the skeleton axis. The geometric parameters include length, width and orientation angle. S5. Correct the geometric parameters according to the preset correction rules and output the final quantization result.

[0007] Furthermore, the multi-scale illumination equalization processing based on Retinex theory includes: The original image is converted to the HSV color space, and the luminance channel and color channel are separated. A Gaussian pyramid is constructed for the luminance channel to decompose it into multiple scales. Logarithmic transformation is performed on each scale and the illumination component and reflection component are separated by Gaussian filtering. The illumination component is subjected to gradient magnitude compression and then recombined with the reflection component to obtain the luminance channel after illumination equalization. The processed luminance channel and the original color channel are recombined into an RGB image.

[0008] Furthermore, the multi-scale Gaussian difference filtering process includes: Multiple Gaussian filter kernels with different standard deviations are generated. Each Gaussian filter kernel is used to convolve the image after illumination equalization to obtain multiple filtered images. The difference between the filtered images corresponding to adjacent serial numbers after sorting by standard deviation from smallest to largest is calculated to obtain a difference image representing the gray-level difference between filtered images at adjacent scales. Pixels with an absolute gray-level gradient greater than the dynamic threshold in the difference image are retained and merged to obtain a denoised image.

[0009] Furthermore, the extraction of the directional consistency feature includes: For each candidate crack segment, construct an orientation histogram and count the orientation angles of the internal edge pixels to determine the main orientation. Calculate the deviation of the orientation angles of all pixels in the segment from the main orientation. If the percentage of pixels with a deviation less than a first preset angle is greater than a first preset proportion, then the segment is determined to meet the orientation consistency feature requirement.

[0010] Furthermore, the extraction of the texture continuity features includes: The texture features of each candidate crack segment are calculated using a local binary pattern operator with a sampling radius greater than 1 pixel, and the texture features are converted into a local binary pattern histogram. The calculated local binary pattern histogram is compared with the continuous linear texture feature templates pre-stored in the feature database. If the matching degree is greater than the second preset ratio, the segment is determined to meet the texture continuity feature requirements.

[0011] Furthermore, the angle constraint in the dynamic constraint determination requires that the angle between the connecting line segment and the main direction of the two segments to be connected is less than the second preset angle; the distance constraint requires that the connection distance is less than the product of the average width of the two segments and a preset multiple; the noise region exclusion rule requires that if there is a region between the two segments that is marked as background and whose area is greater than a preset multiple of the total area of ​​the two segments, then the connection is terminated.

[0012] Furthermore, the step of generating connected paths using curve fitting includes: Multiple edge endpoints of the two segments to be connected are selected as control points, and a smooth connected path is generated by fitting a third-order Bézier curve. For blank pixels on the path, if the number of effective crack pixels in their neighborhood is greater than the preset number, a gray value is assigned by bilinear interpolation to fill the breakpoint.

[0013] Furthermore, the skeletonization process employs an improved Zhang-Suen thinning algorithm, which protects endpoint pixels during the iteration process and uses a skeleton pixel count change rate lower than a preset change rate as the iteration termination condition.

[0014] Furthermore, the step of correcting the geometric parameters according to the preset correction rules includes: If the crack has branches, the total length is equal to the sum of the length of the main crack and the length of each branch multiplied by a preset branch weight coefficient, which is pre-calibrated based on the width ratio of the branch cracks; if the local width of the crack is greater than the preset width threshold and its texture continuity feature value is lower than the third preset ratio, the width calculation result of the local area is discarded.

[0015] This invention also provides a high-resolution image intelligent recognition and quantization system for slope cracks, including: The image preprocessing module is used to sequentially perform multi-scale illumination equalization processing and multi-scale Gaussian difference filtering processing based on Retinex theory on the acquired high-resolution images of the slope in order to compensate for non-uniform illumination and suppress background noise. The multi-dimensional feature enhancement module for crack segments is used to extract gray-level gradient features, orientation consistency features, and texture continuity features of candidate crack segments, and to mark valid crack segments. The crack topology connectivity reconstruction module is used to calculate the comprehensive similarity between effective crack segments, perform dynamic constraint judgment, and reconstruct the continuous crack topology through curve fitting and pseudo-branch removal. The crack geometry parameter quantization module is used to perform skeletonization processing on the topology of continuous cracks and calculate and correct the geometric parameters of the cracks. The feature database stores continuous linear texture feature templates, branch weight coefficients, and various preset thresholds, and supports template updates based on incremental sample iterative training.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention addresses the problems of illumination distortion and noise interference in field scenes by performing multi-scale illumination equalization and multi-scale Gaussian difference filtering based on Retinex theory on high-resolution slope images. It effectively compensates for non-uniform illumination and suppresses background noise. An improved Canny edge detection method combined with weighted fusion of gray-level gradient, directional consistency, and texture continuity features accurately marks effective crack segments, avoiding false crack misjudgments. Through comprehensive similarity calculation, dynamic constraint judgment, and curve fitting, it reconstructs the topological connectivity of cracks, solving the problems of multiple breakpoints, omissions, underestimated lengths, and morphological distortions caused by crack segment breaks in existing technologies. An improved Zhang-Suen thinning algorithm is used for skeletonization, and preset rules are used to correct geometric parameters, ensuring accurate calculation of crack length, width, and orientation angle. The system achieves full-process automation, improving monitoring efficiency and providing reliable data support for slope safety monitoring and stability assessment, effectively meeting the needs of geological disaster early warning and infrastructure protection. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a flowchart of the image preprocessing process of the present invention; Figure 3 This is a flowchart of the feature extraction and fusion process of the present invention; Figure 4 This is a flowchart of the topology connectivity reconstruction process of the present invention; Figure 5 This is a flowchart of the geometric parameter quantization process of the present invention; Figure 6 This is a system structure diagram of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1 Please see Figure 1-5 This embodiment provides a method for intelligent identification and quantification of high-resolution images of slope cracks, and provides a detailed explanation in conjunction with the actual scenario of safety monitoring of large open-pit mine slopes in mountainous areas. The specific implementation includes the following steps: First, the high-resolution images of the acquired slope are subjected to multi-scale illumination equalization and multi-scale Gaussian difference filtering based on Retinex theory to compensate for non-uniform illumination and suppress background noise. In the open-pit mine slope monitoring scenario, the images are acquired by a high-resolution industrial camera mounted on a multi-rotor UAV. The camera resolution is 12 megapixels, the shooting height is 50m, and the ground resolution of the image is 0.01m / pixel. The acquired original images have obvious shadow areas and strong light reflection areas, requiring multi-scale illumination equalization based on Retinex theory. The specific operations of this processing are as follows: the original image is converted to the HSV color space, and the luminance channel and color channel are separated; a Gaussian pyramid is constructed for the luminance channel to decompose it into multiple scales; a logarithmic transformation is performed on each scale, and the illumination component and reflection component are separated by Gaussian filtering; the illumination component is gradient-compressed and then recombined with the reflection component to obtain the illumination-equalized luminance channel; the processed luminance channel and the original color channel are recombined into an RGB image. The conversion formula for converting the original image to the HSV color space is: ; ; ; In the formula, These represent the red, green, and blue channel pixel values ​​of the original RGB image, with a value range of [value range missing]. ; For hue components, For saturation components, For the luminance component, when constructing the Gaussian pyramid, the luminance channel is decomposed into 4 scales, with each scale having a downsampling factor of 2, i.e., the luminance component. The image size at the layer scale is the first Layer A logarithmic transformation is performed on the brightness components at each scale, using the following formula: ; in This represents the pixel value of the brightness channel in this layer. For light component, For the reflection component, the purpose of the logarithmic transformation is to convert the multiplicative relationship between the illumination and reflection components into an additive relationship for separation. Subsequently, Gaussian filters with different kernel sizes are used to filter the logarithmic transformation results at each scale to separate the illumination component. The Gaussian filter kernel function is: ; In the formula, x and y are the coordinate offsets of pixels within the filter kernel relative to the kernel center. The standard deviation of the Gaussian filter is taken separately for each of the four scales. The filter kernel size is This ensures that the filtering range covers an effective area of ​​3 times the standard deviation. After separating the illumination component, it is subjected to gradient amplitude compression, using the following compression formula: ,in The compression factor is taken after sample calibration. This coefficient can suppress overexposure in bright areas and enhance brightness in shadow areas while maintaining the dynamic range of brightness. It compresses the light components... With reflection component The luminance components are recombined into equalized components through exponential transformation. ,Right now Finally With the original HSV space , The components are converted back to RGB space to complete the illumination equalization process. After this step, the details in the shadow areas of the open-pit mine slope image are effectively improved, the overexposure in the strong light areas is eliminated, and the image distortion caused by non-uniform illumination is significantly improved, laying the foundation for subsequent crack identification.

[0020] After illumination equalization, multi-scale Gaussian difference filtering is performed. This step involves generating multiple Gaussian filter kernels with different standard deviations; convolving each kernel with the equalized image to obtain multiple filtered images; calculating the difference between adjacent filtered images sorted by standard deviation from smallest to largest to obtain a difference image representing the grayscale difference between adjacent scale filtered images, and retaining pixels with an absolute grayscale gradient greater than a dynamic threshold. These are then merged to obtain the denoised image. Four Gaussian filter kernels are generated, with standard deviations of various values. The size of each filter kernel remains the same. After converting the RGB image to grayscale after illumination equalization, it is convolved with four Gaussian filter kernels to obtain four filtered images. Then, the difference between filtered images at adjacent scales is calculated. Dynamic thresholding. The calculation employs an adaptive method based on the statistical characteristics of image grayscale, and the formula is as follows: ; In the formula The mean gray level of the difference image. The standard deviation of the gray levels in the difference image. The threshold adjustment coefficient was determined after calibration using ROC curves from 100 sets of open-pit mine slope image samples. For each difference image, retain those with an absolute grayscale gradient greater than 1. The pixels retained in the three difference images are merged to obtain the final denoised image. This step effectively suppresses background noise such as gravel textures and vegetation edges in the image, while preserving the faint edge features of cracks, thus solving the noise interference problem caused by complex backgrounds in open-pit mine scenes.

[0021] After image preprocessing, the next step is to extract initial edges from the preprocessed image using an improved Canny edge detection algorithm. Based on these initial edges, the grayscale gradient features, orientation consistency features, and texture continuity features of each candidate crack segment are calculated. These three features are then weighted and fused to obtain a fused feature value. Candidate segments with fused feature values ​​greater than a preset fusion threshold are marked as valid crack segments. First, the improved Canny edge detection algorithm is used to extract initial edges from the denoised image. The improvement lies in adding an endpoint pixel protection mechanism during the non-maximum suppression stage, and using an adaptive calculation method for high and low thresholds. low threshold ,in The gradient magnitude of the image is calculated using the Sobel operator, with the following formula: The gradient is in the x-direction. The gradient in the y-direction is given by the Sobel operator's kernel in the x-direction. The core in the y-direction is .

[0022] Based on the extracted initial edges, the regions where edge pixels are connected are divided into candidate crack segments. Then, three types of features are calculated for each candidate crack segment. The first type is the grayscale gradient feature, which is calculated as the average gradient magnitude of all pixels within the candidate crack segment, using the following formula: ,in The number of pixels within the segment. For the first The gradient magnitude of each pixel. The second type is the orientation consistency feature, which is extracted as follows: An orientation histogram is constructed for each candidate crack segment, and the orientation angles of its internal edge pixels are statistically analyzed to determine the main orientation; the deviation of the orientation angles of all pixels within the segment from the main orientation is calculated. If the proportion of pixels with a deviation less than a first preset angle is greater than a first preset proportion, the segment is determined to meet the orientation consistency feature requirements. In the open-pit mine scenario, the first preset angle is determined using a sample statistical method. 50 manually labeled standard crack segments are selected, and the distribution of the deviations of pixel orientation angles from the main orientation within each segment is statistically analyzed. The maximum deviation value of 95% of the samples is taken as the first preset angle, i.e., 15°. The first preset proportion is determined through a validation set, with 80% taken as the threshold to ensure that more than 90% of the valid crack segments pass the judgment. The orientation angle is calculated as follows: The range of values ​​is The orientation histogram is divided into 18 intervals, each spanning 10°. The main orientation of the segment is determined by counting the number of pixels in each interval. The third category is texture continuity features, which are extracted as follows: a local binary pattern operator with a sampling radius greater than 1 pixel is used to calculate the texture features of each candidate crack segment, and the texture features are converted into a local binary pattern histogram. The calculated local binary pattern histogram is compared with the continuous linear texture feature templates pre-stored in the feature database. If the matching degree is greater than the second preset ratio, the segment is determined to meet the texture continuity feature requirements. In this embodiment, the sampling radius is 2 pixels, the number of neighboring pixels is 8, and the calculation method of the local binary pattern operator is to compare the gray values ​​of its neighboring pixels with those of the center pixel for each pixel. If the gray value is greater than or equal to that of the center pixel, it is recorded as 1; otherwise, it is recorded as 0. An 8-bit binary number is generated and converted into a decimal number as the local binary pattern value of the pixel. The continuous linear texture feature templates in the feature database were obtained through cluster analysis of the local binary pattern histograms of 1000 standard cracks. The template matching degree was calculated using the Bach distance formula: ; in This represents the number of intervals in the histogram. The histogram interval percentage of candidate segments. The histogram interval percentage of the template has a matching degree of . The second preset ratio is set to 75% after calibration by the template matching accuracy curve.

[0023] After obtaining the three types of features, a weighted fusion method is used to obtain the fusion feature value. Candidate segments with fusion feature values ​​greater than a preset fusion threshold are marked as valid crack segments. The fusion formula is as follows: ; in To fuse feature values, These are the weight coefficients for the three types of features. This is the normalized value of the directional consistency feature. This represents the normalized value of the texture continuity feature. The weighting coefficients are determined using the analytic hierarchy process (AHP). After constructing the judgment matrix and calculating the eigenvectors, the values ​​are taken as... The fusion threshold is set to 0.6 after training with samples. When the fusion feature value of the candidate segment... When the value is greater than 0.6, it is marked as a valid crack segment. This step can effectively eliminate false crack segments such as gravel edges and vegetation textures in open-pit mining scenes, greatly improve the recognition accuracy of valid cracks, and solve the problem that traditional algorithms easily misjudge background interference as cracks.

[0024] The following steps involve reconstructing the topological connectivity of all valid crack segments, including: calculating the comprehensive similarity between any two adjacent valid crack segments, where the comprehensive similarity is weighted by directional similarity, gray-level gradient similarity, and texture similarity; for segment pairs with a comprehensive similarity greater than a preset threshold, dynamic constraint determination is further performed based on angle constraints, distance constraints, and noise region exclusion rules; for segment pairs that pass the dynamic constraint determination, curve fitting is used to generate connected paths, and pseudo-branches that do not conform to the natural crack morphology are removed to output a continuous crack topology. The comprehensive similarity between any two adjacent valid crack segments is calculated first, using the following formula: ; in, The combined similarity between two adjacent segments. The weighting coefficients are taken after calibration using the analytic hierarchy process. ; For directional similarity, the calculation method is as follows: , These represent the main directions of the two segments; The gray-level gradient similarity is calculated as follows: , These are the average gradient magnitudes of the two segments, respectively; Texture similarity refers to the local binary pattern histogram matching degree between two segments. The preset threshold for overall similarity, after connectivity verification using 100 sets of fracture crack samples, is set to 0.7. When the overall similarity between two adjacent segments... When the value is greater than 0.7, the dynamic constraint determination stage begins.

[0025] The specific rules for dynamic constraint determination are as follows: Angle constraints require that the angle between the connecting line segment and the main direction of the two segments to be connected is less than a second preset angle. The second preset angle is determined by statistically analyzing the continuity of natural crack directions. In open-pit mine scenarios, the deviation of natural crack directions is generally less than 10°, so 10° is taken as the second preset angle. Distance constraints require that the connection distance is less than the product of the average width of the two segments and a preset multiple. The average width of the two segments is the average of their respective pixel widths. The preset multiple is calibrated by statistically analyzing the relationship between crack breakage spacing and width, and is set to 3, meaning the connection distance must be less than... , The average width of the two segments; the noise region exclusion rule requires that if there is a region marked as background between the two segments and its area is greater than a preset multiple of the total area of ​​the two segments, then the connection is terminated. The preset multiple is 0.5, that is, if the area of ​​the background region between the two segments is greater than 0.5 times the total area of ​​the two segments, it indicates that there is an obvious non-crack obstacle, and the connection is terminated.

[0026] For segment pairs determined by dynamic constraints, curve fitting is used to generate connected paths. The specific operation is as follows: multiple edge endpoints of the two segments to be connected are selected as control points, and a smooth connected path is generated using a third-order Bézier curve. For blank pixels on the path, if the number of effective crack pixels in their neighborhood is greater than a preset number, a grayscale value is assigned through bilinear interpolation to fill the breakpoint. In this embodiment, the first and last three edge endpoints are selected as control points for each segment. The parametric equation of the third-order Bézier curve is: ; Where t is a parameter variable of the curve, representing the position of a point on the curve. , The combination number is the number of ways to select i control points from three control points. The coordinates of four control points correspond to the curve's starting point, two intermediate control points, and the ending point. Two Bézier curves are constructed using six endpoints to ensure path smoothness. For blank pixels on the path, a 3×3 neighborhood is used for judgment, with a preset count of 5. That is, when the number of valid crack pixels in the neighborhood is greater than 5, a grayscale value is assigned using bilinear interpolation. The bilinear interpolation formula is: ; Where x and y are the two-dimensional coordinates of the blank pixels to be interpolated. , , These are the values ​​of the four pixels in the neighborhood of the blank pixel. This represents the coordinates of the neighboring pixels to the top-left corner of the blank pixel. After path fitting, pseudo-branch removal is performed on the generated crack topology. The removal criteria are based on the branching characteristics of natural cracks, i.e., the angle between the branch and the main crack must be greater than 30° and the branch length must be less than the main crack length. Branches that do not conform to this characteristic are identified as pseudo-branches and discarded. This step can effectively restore the continuous topological structure of cracks on open-pit mine slopes, solve the problems of length underestimation and morphological distortion caused by crack segment breakage in existing technologies, and provide a complete structural basis for subsequent parameter quantification.

[0027] After completing the topological connectivity reconstruction, the continuous crack topology is skeletonized, and the geometric parameters of the cracks, including length, width, and orientation angle, are calculated based on the skeleton axis. The specific skeletonization process is as follows: an improved Zhang-Suen thinning algorithm is used, which protects endpoint pixels during iteration and uses a skeleton pixel count change rate lower than a preset change rate as the iteration termination condition. The improved Zhang-Suen thinning algorithm's iteration rule adds endpoint protection judgment to the pixel deletion condition of the traditional algorithm; that is, if a pixel is an endpoint, it is not deleted. The iteration termination condition is that the skeleton pixel count change rate of two consecutive iterations is lower than a preset change rate, which is calibrated to 0.01. The iteration terminates at time, where For the first Number of skeleton pixels in the next iteration. For the first The number of skeleton pixels in the next iteration.

[0028] Based on the extracted skeleton axis, the geometric parameters of the crack are calculated: the length is the number of pixels on the skeleton axis multiplied by the ground resolution of the image; the width is the pixel distance between the crack edges on both sides of each skeleton pixel in the direction perpendicular to the skeleton axis, calculated as the average of all pixel distances and multiplied by the ground resolution; the orientation angle is the principal orientation angle of the skeleton axis, obtained by statistically analyzing the orientation angles of the skeleton pixels and selecting the most frequent orientation angle as the orientation angle. This step can accurately extract the core axis of the crack, ensuring the reliability of the geometric parameter calculation and solving the problem of length calculation deviation caused by the loss of endpoints in traditional skeletonization algorithms.

[0029] Finally, the geometric parameters are corrected according to preset correction rules, and the final quantization result is output. The specific operation of the correction rules is as follows: If the crack has branches, the total length is equal to the sum of the length of the main crack and the length of each branch multiplied by a preset branch weight coefficient. The preset branch weight coefficient is pre-calibrated based on the width ratio of the branch cracks. If the local width of the crack is greater than a preset width threshold and its texture continuity feature value is lower than a third preset ratio, the width calculation result of that local area is discarded. The branch weight coefficient is calibrated by setting the average width of the main crack to... The average width of the branch crack is Then the branch weight coefficient ,like If a branch is identified as a micro-branch, its weight coefficient is set to 0, meaning it is not included in the total length calculation. The preset width threshold is 5 times the overall average width of the crack, determined by statistically analyzing the width distribution of cracks on open-pit mine slopes. A third preset ratio is set to 60%, meaning that when the texture continuity characteristic value of a local area is below 60%, the area is considered to have width anomalies caused by noise interference, and its width calculation results are discarded. The corrected geometric parameters more realistically reflect the actual morphology of the cracks, providing reliable data support for the stability analysis of open-pit mine slopes and solving the problem of distorted results caused by branches and noise interference in traditional parameter calculation methods.

[0030] Example 2 Please see Figure 6 The present invention also provides a high-resolution image intelligent recognition and quantization system for slope cracks, used to implement the high-resolution image intelligent recognition and quantization method for slope cracks in Embodiment 1. The system includes: The core function of the image preprocessing module is to sequentially perform multi-scale illumination equalization processing and multi-scale Gaussian difference filtering processing based on Retinex theory on the acquired high-resolution slope images to compensate for non-uniform illumination and suppress background noise. This module has built-in algorithm units such as HSV color space conversion, Gaussian pyramid decomposition, Gaussian filter convolution, and dynamic threshold determination. It can directly read high-resolution images transmitted by UAVs or digital cameras and output preprocessed denoised equalized images. Its runtime parameter configuration (such as Gaussian kernel standard deviation, gradient magnitude compression coefficient, dynamic threshold adjustment coefficient, etc.) can be retrieved and updated through the feature database.

[0031] The multi-dimensional feature enhancement module for crack segments receives the output data from the image preprocessing module. Its core function is to extract the gray-level gradient features, orientation consistency features, and texture continuity features of candidate crack segments and mark valid crack segments. This module integrates an improved Canny edge detection unit, a connected component analysis unit, a multi-dimensional feature calculation unit, and a feature weighted fusion unit. During operation, it first extracts the initial edges using the improved Canny algorithm, then analyzes the edge connected components to obtain candidate crack segments. Subsequently, it calculates the three types of features for each segment and performs weighted fusion. Finally, segments that meet the fusion threshold are marked as valid crack segments and transmitted to the next module. The weight coefficients and judgment thresholds for feature calculation are provided by the feature database.

[0032] The crack topology connectivity reconstruction module receives valid crack segment data. Its core function is to calculate the comprehensive similarity between valid crack segments, perform dynamic constraint judgment, and reconstruct the continuous crack topology through curve fitting and pseudo-branch removal. This module includes a comprehensive similarity calculation unit, a dynamic constraint judgment unit, a Bézier curve fitting unit, and a pseudo-branch removal unit. During operation, it first traverses all adjacent valid crack segments and calculates the comprehensive similarity. For segment pairs that meet the similarity threshold, it performs triple dynamic constraint judgment based on angle, distance, and noise region. For segment pairs that pass the judgment, it generates smooth connected paths. Finally, it removes pseudo-branches to obtain the continuous crack topology. The similarity weight, constraint judgment threshold, and other parameters are all retrieved from the feature database.

[0033] The core function of the crack geometry parameter quantization module is to perform skeletonization processing on the topology of continuous cracks, calculate and correct the geometric parameters of cracks. This module has built-in improved Zhang-Suen refinement unit, geometric parameter calculation unit and parameter correction unit. During operation, the crack topology is first skeletonized to obtain the skeleton axis, and then the initial geometric parameters such as length, width and orientation angle are calculated based on the skeleton axis. Finally, the parameter correction is completed according to the correction rules in the feature database, and the final quantization result is output.

[0034] The core function of the feature database is to store continuous linear texture feature templates, branch weight coefficients, and various preset thresholds. It also supports template updates based on incremental sample iterative training. The database has built-in sample management units, template clustering units, and parameter calibration units. The continuous linear texture feature templates are generated by LBP histogram clustering of a large number of manually annotated crack samples. The branch weight coefficients are calibrated by the width ratio fitting model. Various preset thresholds are determined through sample statistics and the AHP method. At the same time, the database supports the addition of new annotated samples for incremental iterative training to optimize the templates and parameters, thereby improving the system's adaptability and recognition accuracy.

[0035] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0036] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent identification and quantization of high-resolution images of slope cracks, characterized in that, include: S1. Perform multi-scale illumination equalization and multi-scale Gaussian difference filtering based on Retinex theory on the acquired high-resolution image of the slope in sequence to compensate for non-uniform illumination and suppress background noise. S2. For the image after multi-scale illumination equalization and multi-scale Gaussian difference filtering, the improved Canny edge detection algorithm is used to extract the initial edge. Based on the initial edge, the gray-level gradient feature, orientation consistency feature and texture continuity feature of each candidate crack segment are calculated. The three types of features are weighted and fused to obtain the fused feature value. Candidate segments with fused feature values ​​greater than the preset fused threshold are marked as valid crack segments. S3. Reconstruct the topological connectivity of all valid crack segments, including: calculating the comprehensive similarity between any two adjacent valid crack segments, which is composed of weighted similarity of direction, gray-level gradient, and texture; for segment pairs with a comprehensive similarity greater than a preset threshold, dynamic constraint judgment is further performed based on angle constraints, distance constraints, and noise region exclusion rules; for segment pairs that pass the dynamic constraint judgment, curve fitting is used to generate connected paths, and pseudo-branches that do not conform to the natural crack morphology and whose angle with the main crack is less than 30° and whose length accounts for less than 20% of the length of the main crack are removed, so as to output a continuous crack topology structure. S4. The continuous crack topology is skeletonized, and the geometric parameters of the crack are calculated based on the skeleton axis. The geometric parameters include length, width and orientation angle. S5. Correct the geometric parameters according to the preset correction rules and output the final quantization result.

2. The method for intelligent recognition and quantization of high-resolution images of slope cracks according to claim 1, characterized in that, The multi-scale illumination equalization process based on Retinex theory includes: The original image is converted to the HSV color space, and the luminance channel and color channel are separated. A Gaussian pyramid is constructed for the luminance channel to decompose it into multiple scales. Logarithmic transformation is performed on each scale and the illumination component and reflection component are separated by Gaussian filtering. The illumination component is subjected to gradient magnitude compression and then recombined with the reflection component to obtain the luminance channel after illumination equalization. The processed luminance channel and the original color channel are recombined into an RGB image.

3. The method for intelligent recognition and quantization of high-resolution images of slope cracks according to claim 1, characterized in that, The multi-scale Gaussian difference filtering process includes: Multiple Gaussian filter kernels with different standard deviations are generated. Each Gaussian filter kernel is used to convolve the image after illumination equalization to obtain multiple filtered images. The difference between the filtered images corresponding to adjacent serial numbers after sorting by standard deviation from smallest to largest is calculated to obtain a difference image representing the gray-level difference between filtered images at adjacent scales. Pixels with an absolute gray-level gradient greater than the dynamic threshold in the difference image are retained and merged to obtain a denoised image.

4. The method for intelligent recognition and quantization of high-resolution images of slope cracks according to claim 1, characterized in that, The extraction of the directional consistency feature includes: For each candidate crack segment, construct an orientation histogram and count the orientation angles of the internal edge pixels to determine the main orientation. Calculate the deviation of the orientation angles of all pixels in the segment from the main orientation. If the percentage of pixels with a deviation less than a first preset angle is greater than a first preset proportion, then the segment is determined to meet the orientation consistency feature requirement.

5. The method for intelligent identification and quantization of high-resolution images of slope cracks according to claim 1, characterized in that, The extraction of the texture continuity features includes: The texture features of each candidate crack segment are calculated using a local binary pattern operator with a sampling radius greater than 1 pixel, and the texture features are converted into a local binary pattern histogram. The calculated local binary pattern histogram is compared with the continuous linear texture feature templates pre-stored in the feature database. If the matching degree is greater than the second preset ratio, the segment is determined to meet the texture continuity feature requirements.

6. The method for intelligent recognition and quantization of high-resolution images of slope cracks according to claim 1, characterized in that, The dynamic constraint determination requires that the angle between the connecting line segment and the main direction of the two segments to be connected is less than the second preset angle; the distance constraint requires that the connection distance is less than the product of the average width of the two segments and a preset multiple; the noise region exclusion rule requires that if there is a region between the two segments that is marked as background and whose area is greater than a preset multiple of the total area of ​​the two segments, then the connection is terminated.

7. The method for intelligent identification and quantization of high-resolution images of slope cracks according to claim 1, characterized in that, The method of generating connected paths using curve fitting includes: Multiple edge endpoints of the two segments to be connected are selected as control points, and a smooth connected path is generated by fitting a third-order Bézier curve. For blank pixels on the path, if the number of effective crack pixels in their neighborhood is greater than the preset number, a gray value is assigned by bilinear interpolation to fill the breakpoint.

8. The method for intelligent recognition and quantization of high-resolution images of slope cracks according to claim 1, characterized in that, The skeletonization process employs an improved Zhang-Suen thinning algorithm, which protects endpoint pixels during the iteration process and uses a rate of change of skeleton pixel count below a preset rate of change as the iteration termination condition.

9. The method for intelligent recognition and quantization of high-resolution images of slope cracks according to claim 1, characterized in that, The step of correcting the geometric parameters according to the preset correction rules includes: If the crack has branches, the total length is equal to the sum of the length of the main crack and the length of each branch multiplied by a preset branch weight coefficient, which is pre-calibrated based on the width ratio of the branch cracks; if the local width of the crack is greater than the preset width threshold and its texture continuity feature value is lower than the third preset ratio, the width calculation result of the local area is discarded.

10. A high-resolution image intelligent recognition and quantization system for slope cracks, applied to the high-resolution image intelligent recognition and quantization method for slope cracks as described in any one of claims 1 to 9, characterized in that, include: The image preprocessing module is used to sequentially perform multi-scale illumination equalization processing and multi-scale Gaussian difference filtering processing based on Retinex theory on the acquired high-resolution images of the slope in order to compensate for non-uniform illumination and suppress background noise. The multi-dimensional feature enhancement module for crack segments is used to extract gray-level gradient features, orientation consistency features, and texture continuity features of candidate crack segments, and to mark valid crack segments. The crack topology connectivity reconstruction module is used to calculate the comprehensive similarity between effective crack segments, perform dynamic constraint judgment, and reconstruct the continuous crack topology through curve fitting and pseudo-branch removal. The crack geometry parameter quantization module is used to perform skeletonization processing on the topology of continuous cracks and calculate and correct the geometric parameters of the cracks. The feature database stores continuous linear texture feature templates, branch weight coefficients, and various preset thresholds, and supports template updates based on incremental sample iterative training.

Images