Concrete structure crack image detection method and system based on deep learning

Abstract

The present invention discloses a method and system for detecting cracks in concrete structures using deep learning. The method comprises: acquiring multi-view images; spatially aligning the multi-view images to obtain a matching image sequence; fusing the matching image sequence to obtain a fused image; performing semantic segmentation based on the fused image and a preset neural network model using transfer learning to obtain a crack mask; performing skeleton extraction based on the crack mask and the fused image to obtain a preliminary crack network topology; optimizing the preliminary crack network topology to obtain a crack network topology; performing finite element analysis based on preset material properties, stress state, and concrete structure to obtain a stress distribution cloud map; and spatially matching the stress distribution cloud map with the crack network topology to obtain a stress-crack correlation model. This method is capable of analyzing the dynamic changes of concrete cracks.

Description

Technical Field

[0001] The present invention relates to the field of image detection technology, and in particular to a method and system for detecting cracks in concrete structures based on deep learning. Background Art

[0002] Concrete structures are widely used in construction projects. However, cracks can develop during use due to natural factors (such as temperature, humidity, and freeze-thaw cycles) and loads (such as gravity and wind loads). These cracks not only affect the structure's appearance but can also reduce its bearing capacity and durability, posing safety risks. Therefore, timely and accurate detection and assessment of concrete crack status is crucial for ensuring building safety, extending its service life, and reducing maintenance costs.

[0003] Ultrasonic testing is a non-destructive testing method that determines the presence and depth of cracks by measuring the propagation characteristics of ultrasonic waves in concrete. This method uses the propagation speed, attenuation, and changes in the reflected signal of ultrasonic waves in concrete to identify cracks. When ultrasonic waves encounter cracks, they reflect and scatter, resulting in increased propagation time and reduced signal strength. By analyzing these changes, the location and depth of the cracks can be determined.

[0004] However, ultrasonic testing still faces numerous challenges in practical application. The dynamic nature of cracks and their correlation with internal stress distribution complicate continuous monitoring and comprehensive analysis. Consequently, ultrasonic testing cannot integrate information about cracks and stress distribution, and thus cannot analyze the dynamic changes in concrete cracks. Consequently, existing technologies lack models and methods for analyzing the dynamic changes in concrete cracks. Summary of the Invention

[0005] The present invention provides a method and system for detecting concrete structure crack images based on deep learning, so as to analyze the dynamic changes of concrete cracks.

[0006] In a first aspect, in order to solve the above technical problems, the present invention provides a method for detecting cracks in concrete structures based on deep learning, comprising:

[0007] Acquire multi-view images;

[0008] Performing spatial alignment on the multi-view images to obtain a matching image sequence;

[0009] Performing image fusion according to the matching image sequence to obtain a fused image;

[0010] Performing semantic segmentation based on the fused image and a preset neural network model based on transfer learning to obtain a crack mask;

[0011] Performing skeleton extraction based on the crack mask and the fused image to obtain a preliminary crack network topology;

[0012] Optimizing the preliminary fracture network topology to obtain a fracture network topology;

[0013] According to the preset material properties, stress state and concrete structure, finite element analysis is performed to obtain the stress distribution cloud map;

[0014] According to the stress distribution cloud map and the crack network topology, spatial matching is performed to obtain a stress-crack correlation model.

[0015] In an optional implementation, performing spatial alignment based on the multi-view images to obtain a matching image sequence includes:

[0016] Establishing an image acquisition coordinate system based on the multi-view images to obtain image coordinates;

[0017] Performing spatial transformation according to the image coordinates to obtain a spatial transformation matrix;

[0018] Resampling is performed according to the spatial transformation matrix and the multi-view images to obtain a unified view image;

[0019] Performing pixel interpolation based on the unified perspective image to obtain a complete image;

[0020] Performing image denoising according to the complete image to obtain a low-noise image;

[0021] Edge detection is performed on the low-noise image to obtain a matching image sequence.

[0022] In an optional embodiment, performing image fusion according to the matching image sequence to obtain a fused image includes:

[0023] Performing multi-scale decomposition according to the matching image sequence to obtain a multi-scale feature image;

[0024] Perform pixel fusion according to the multi-scale feature image and a preset weight rule to obtain a preliminary fused image;

[0025] performing contrast enhancement on the preliminary fused image to obtain an enhanced contrast image;

[0026] Performing unsharp mask processing on the contrast-enhanced image to obtain a clear image;

[0027] Performing edge detection based on the clear image to obtain a crack outline;

[0028] Image superposition is performed based on the crack outline and the clear image to obtain a fused image.

[0029] In an optional embodiment, skeleton extraction is performed based on the crack mask and the fused image to obtain a preliminary crack network topology, including:

[0030] Performing mask skeleton extraction based on the crack mask and the fused image to obtain an extracted image;

[0031] Performing morphological filtering on the extracted image to obtain a filtered image;

[0032] Based on the filtered image, a connection analysis is performed to obtain a preliminary crack network topology.

[0033] In an optional embodiment, optimizing the preliminary fracture network topology to obtain the fracture network topology includes:

[0034] Performing pruning optimization based on the preliminary crack network topology to obtain a de-redundant topology;

[0035] Perform skeleton segmentation according to the de-redundant topology to obtain a main skeleton line;

[0036] Perform branch detection based on the trunk skeleton line and the de-redundancy topology to obtain branch positions;

[0037] Performing geometric calculations based on the branch positions to obtain branch length angles;

[0038] Information fusion is performed based on the branch length angle and the de-redundant topology to obtain the crack network topology.

[0039] In an optional embodiment, a finite element analysis is performed based on preset material properties, stress state, and concrete structure to obtain a stress distribution cloud diagram, including:

[0040] Performing finite element meshing on the concrete structure according to the preset material properties and the stress state to obtain a finite element model;

[0041] According to the finite element model, stress calculation is performed to obtain a stress distribution cloud map.

[0042] In an optional embodiment, spatial matching is performed based on the stress distribution cloud map and the crack network topology to obtain a stress-crack correlation model, including:

[0043] Performing feature extraction based on the crack network topology to obtain crack features;

[0044] A stress-crack correlation model is constructed based on the crack characteristics and the stress distribution cloud map to obtain a stress-crack correlation model.

[0045] In a second aspect, the present invention provides a concrete structure crack image detection system based on deep learning, comprising:

[0046] An input module for acquiring multi-view images;

[0047] an alignment module, configured to perform spatial alignment based on the multi-view images to obtain a matching image sequence;

[0048] A fusion module, configured to perform image fusion according to the matching image sequence to obtain a fused image;

[0049] a mask generation module, configured to perform semantic segmentation based on the fused image and a preset neural network model and transfer learning to obtain a crack mask;

[0050] A topology generation module, configured to perform skeleton extraction based on the crack mask and the fused image to obtain a preliminary crack network topology;

[0051] A topology optimization module, configured to optimize the preliminary fracture network topology to obtain a fracture network topology;

[0052] The stress calculation module is used to perform finite element analysis based on the preset material properties, stress state and concrete structure to obtain the stress distribution cloud map;

[0053] The correlation construction module is used to perform spatial matching based on the stress distribution cloud map and the crack network topology to obtain a stress-crack correlation model.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] (1) The present invention adopts a method that combines multi-view image acquisition with a pre-trained neural network model. A multi-view industrial camera is used to obtain a sequence of concrete surface images. The spatial alignment technology of the images is used to form a consistent image sequence. The image sequence is then fused to generate a high-resolution fused image, providing a high-quality image data foundation for subsequent analysis.

[0056] (2) This paper combines a pre-trained neural network model with transfer learning technology to perform semantic segmentation on the fused image and extract crack masks, providing a reliable foundation for subsequent crack network topology extraction. On this basis, skeleton extraction is performed based on the crack mask and fused image to construct a preliminary crack network topology. This is then optimized using graph theory methods to obtain an accurate crack network topology that truly reflects the dynamic changes of the cracks.

[0057] (3) The present invention performs finite element analysis based on preset material properties and stress states to generate stress distribution cloud maps, providing an important basis for understanding the mechanical mechanism of crack dynamic changes. By spatially matching the stress distribution cloud map with the crack network topology, a stress-crack correlation model is constructed, and the correlation between crack dynamic changes and their internal stress distribution is established, achieving a comprehensive analysis of crack dynamic changes.

[0058] (4) This invention achieves efficient and accurate analysis of the dynamic changes of concrete cracks by combining multi-view image acquisition with a pre-trained neural network model. The spatial alignment and fusion processing of multi-view images improves the integrity of crack information. The semantic segmentation function of the pre-trained neural network model improves the accuracy of crack identification. Skeleton extraction and graph theory optimization further refine the topological structure of the crack network. The spatial matching of finite element analysis and crack network topology clearly presents the mechanical mechanism of the dynamic changes of cracks. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 1 is a flow chart of a method for detecting cracks in concrete structures based on deep learning according to an embodiment of the present invention;

[0060] Figure 2 It is a structural schematic diagram of a concrete structure crack image detection system based on deep learning provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0061] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0062] Reference Figure 1 , an embodiment of the present invention provides a method for detecting cracks in concrete structures based on deep learning, comprising the following steps:

[0063] S11, acquiring multi-view images;

[0064] S12, performing spatial alignment based on the multi-view images to obtain a matching image sequence;

[0065] S13, performing image fusion according to the matching image sequence to obtain a fused image;

[0066] S14, performing semantic segmentation based on the fused image and a preset neural network model based on transfer learning to obtain a crack mask;

[0067] S15, performing skeleton extraction based on the crack mask and the fused image to obtain a preliminary crack network topology;

[0068] S16, optimizing the preliminary fracture network topology to obtain a fracture network topology;

[0069] S17, performing finite element analysis based on preset material properties, stress state and concrete structure to obtain a stress distribution cloud map;

[0070] S18, performing spatial matching based on the stress distribution cloud map and the crack network topology to obtain a stress-crack correlation model.

[0071] In step S11, it is necessary to obtain multi-view images. It should be noted that the acquisition of multi-view images is completed by a high-resolution industrial camera, and the camera must have sufficient resolution and frame rate to ensure that the details of the concrete surface are clearly visible. Exemplarily, a camera that can capture thirty frames per second and has a resolution of 20 million pixels can be selected. Of course, depending on the application scenario and user needs, cameras with other parameters can also be selected. The pre-trained neural network model has achieved basic functions and good robustness through training with a large amount of data in advance. Exemplarily, the present invention adopts a deep residual network. Of course, depending on the user needs and application scenarios, other pre-trained neural network models can also be selected.

[0072] In step S12, spatial alignment is performed based on the multi-view images to obtain a matching image sequence, including:

[0073] Establishing an image acquisition coordinate system based on the multi-view images to obtain image coordinates;

[0074] Performing spatial transformation according to the image coordinates to obtain a spatial transformation matrix;

[0075] Resampling is performed according to the spatial transformation matrix and the multi-view images to obtain a unified view image;

[0076] Performing pixel interpolation based on the unified perspective image to obtain a complete image;

[0077] Performing image denoising according to the complete image to obtain a low-noise image;

[0078] Edge detection is performed on the low-noise image to obtain a matching image sequence.

[0079] It should be noted that in step S12, spatial alignment of the multi-view images is achieved through a series of image processing operations, ultimately yielding a sequence of matched images. First, the establishment of an image acquisition coordinate system serves as the foundation for spatial alignment of the multi-view images. A fixed point on the concrete surface is selected as the origin to establish a three-dimensional rectangular coordinate system. This coordinate system enables subsequent spatial transformation calculations to accurately align images from different viewpoints. For example, the origin of the coordinate system can be set to the center point of the concrete surface or a significant feature point, such as the starting point of a crack or the center of a hole. Of course, other representative locations can also be selected depending on the application scenario and user needs, and this is not limited by the present invention. Secondly, the calculation of the spatial transformation matrix is ​​achieved through feature point matching. Specifically, significant features such as corners and edges are extracted from the multi-view images, and these feature points are used to calculate the correspondence between adjacent images. For example, natural features such as cracks and holes on the concrete surface can be used as matching points, combined with the least squares method to optimize the transformation parameters, thereby achieving precise registration. This process ensures spatial alignment of images from different viewpoints, providing a foundation for subsequent image fusion and analysis.

[0080] Next, the multi-view images are resampled according to the spatial transformation matrix to obtain a unified-view image. Pixel interpolation must be considered during the resampling process to ensure that the image remains clear and complete after the transformation. Bilinear interpolation is generally suitable, improving computational efficiency while maintaining image quality. For applications requiring higher precision, bicubic interpolation can be used. While computationally more complex, it can produce smoother transitions. Of course, other interpolation methods can be chosen depending on the specific application scenario and user needs, and this is not a limitation of the present invention. Pixel interpolation is then performed on the resampled unified-view image to obtain the complete image. This process further fills in missing pixels caused by the perspective transformation, ensuring image integrity. Interpolation is performed based on the grayscale values ​​of surrounding pixels. By calculating estimated values ​​for missing pixels, the image appears more continuous and complete. Next, the complete image is denoised to obtain a low-noise image. Image denoising aims to eliminate random noise introduced during the acquisition and processing process, thereby improving image quality. Examples of methods that can be used include median filtering or Gaussian filtering. Taking median filtering as an example, a five-by-five pixel filter window is selected, and the median of the surrounding area is used to replace the original value for each pixel, which not only preserves the edge features but also suppresses noise. Of course, other denoising algorithms can be selected according to the noise type and image characteristics, and the present invention does not limit this. Finally, edge detection is performed on the low-noise image to obtain image feature points. For example, the present invention selects to extract significant features such as corners and edges in the image. Then, image matching is performed based on the detected image feature points to obtain a matching image sequence. Edge detection is a key step in identifying defects such as cracks and spalling on the concrete surface. For example, the Sobel or Canny operator can be used for edge detection. Taking crack detection as an example, the image is first grayscaled, and an appropriate detection threshold is set to extract the crack outline that forms a clear contrast with the background. Through this process, defect features such as cracks can be clearly identified and extracted, providing an important basis for subsequent crack analysis and evaluation.

[0081] In step S13, image fusion is performed according to the matching image sequence to obtain a fused image, including:

[0082] Performing multi-scale decomposition according to the matching image sequence to obtain a multi-scale feature image;

[0083] Perform pixel fusion according to the multi-scale feature image and a preset weight rule to obtain a preliminary fused image;

[0084] performing contrast enhancement on the preliminary fused image to obtain an enhanced contrast image;

[0085] Performing unsharp mask processing on the contrast-enhanced image to obtain a clear image;

[0086] Performing edge detection based on the clear image to obtain a crack outline;

[0087] Image superposition is performed based on the crack outline and the clear image to obtain a fused image.

[0088] It should be noted that in step S13, multiscale decomposition is first achieved by performing a wavelet transform on the matching image sequence. This algorithm effectively captures crack features at different scales on the concrete surface. The wavelet transform decomposes the image into high-frequency and low-frequency components. The high-frequency component contains crack details, while the low-frequency component preserves the overall crack structure. For example, fine cracks are more pronounced at higher decomposition levels, while wider cracks are better preserved in the low-frequency components. For example, for a crack with a width of 0.5 mm, the third decomposition level yields the best feature representation. Of course, other decomposition levels can be selected depending on the application scenario and user needs, and this is not limited by the present invention. Secondly, pixel fusion is performed based on the multi-scale feature images and a preset weighting rule. The design of the weighted fusion rule must consider the importance of features at different scales, assigning weights based on the crack's salience. Regions with clear edges are given higher weights, while fuzzy regions are given lower weights. For example, regions with a significant feature value exceeding 0.8 can be given a weight of 0.6, while regions with a feature value below 0.3 can be given a weight of 0.2. In this way, the preliminary fused image can better preserve the key information of the cracks while suppressing noise and irrelevant details.

[0089] Subsequently, in the present invention, contrast enhancement is achieved, for example, through histogram equalization, which can expand the image's grayscale dynamic range, thereby improving crack identification. In practical applications, increasing the contrast of the original image from 50 to 200 to 10 to 240 significantly improves the visibility of crack areas. The contrast-enhanced image more clearly demonstrates the difference between the crack and the background, providing a better foundation for subsequent processing. Next, unsharp masking is performed by subtracting the Gaussian blurred image from the original image to highlight high-frequency details in the image. Choosing the appropriate Gaussian kernel size and weighting coefficient significantly impacts the results. For concrete surface images, for example, a 5×5 Gaussian kernel and a weighting coefficient of 0.7 achieve good results. The clear image after unsharp masking more clearly displays crack details while reducing noise interference. Edge detection is then performed using the Sobel or Canny operator to extract crack outlines. The Sobel operator is effective for detecting vertical and horizontal crack edges, while the Canny operator is suitable for detecting diagonal cracks. In terms of threshold selection, a double threshold method is used, and the high threshold is set to twice the low threshold, which can effectively suppress false edges. The extracted crack contour provides accurate crack location information for subsequent image superposition. Finally, image superposition is the process of combining the crack contour with a clear image. The natural fusion of the contour and the original image is achieved by adjusting the transparency. For example, the present invention superimposes the crack contour on the optimized image with a transparency of 30%, which not only maintains the details of the original image, but also highlights the location information of the crack. In addition, by color-coding cracks of different widths, the severity of the cracks can be intuitively reflected, providing a basis for subsequent damage assessment. For example, areas with crack widths less than 0.5 mm can be marked in yellow, and areas with widths greater than 0.5 mm can be marked in red. Of course, other color coding methods can be selected according to different application scenarios and user needs, and the present invention is not limited to this.

[0090] In step S14, semantic segmentation is performed based on the fused image and a preset neural network model based on transfer learning to obtain a crack mask;

[0091] It should be noted that in step S14, semantic segmentation is performed based on transfer learning by fusing the image and a preset neural network model to obtain a crack mask. First, the fused image is obtained through a series of processes, including spatial alignment and image fusion, on the multi-view images. This process ensures the integrity and accuracy of the crack information in the image, providing a high-quality data foundation for subsequent semantic segmentation. Second, the preset neural network model is the core tool for semantic segmentation based on transfer learning. Transfer learning utilizes a pre-trained model. Exemplarily, the present invention uses a deep residual network and fine-tunes it to suit the characteristics of concrete cracks, thereby improving the model's accuracy and efficiency. Of course, other types of pre-trained models can also be selected depending on the application scenario and user needs, and this is not limited by the present invention. Next, semantic segmentation uses the transfer learning-based neural network model to perform pixel-level classification on the fused image to extract a crack region mask. This process uses the network model's forward propagation calculations to perform a classification prediction for each pixel, generating a crack mask. Exemplarily, the output of the semantic segmentation model is a mask of the same size as the input image, in which the pixel values ​​in the crack region are labeled with a specific class label, while non-crack regions are labeled as background or other categories. Finally, the resulting crack mask provides the basis for further crack analysis. Through the crack mask, the location, shape, and distribution of cracks can be clearly identified, providing important data support for subsequent steps such as skeleton extraction and crack network topology construction.

[0092] In step S15, skeleton extraction is performed based on the crack mask and the fused image to obtain a preliminary crack network topology, including:

[0093] Performing mask skeleton extraction based on the crack mask and the fused image to obtain an extracted image;

[0094] Performing morphological filtering on the extracted image to obtain a filtered image;

[0095] Based on the filtered image, a connection analysis is performed to obtain a preliminary crack network topology.

[0096] It should be noted that in step S15, mask skeleton extraction is first performed based on the crack mask and the fused image. Its purpose is to extract the crack centerline from the crack mask to characterize the spatial distribution of the crack. Skeleton extraction utilizes a distance transform method, which determines the crack centerline by calculating the shortest distance from each pixel to the boundary. In this embodiment, the distance transform threshold is set to half the crack width to ensure that the extracted skeleton accurately reflects the crack center. Of course, other skeleton extraction methods or threshold parameters can be selected depending on the application scenario and user needs, and this is not limited by this invention. Secondly, morphological filtering is a key step in processing the extracted image. Its purpose is to remove noise and irrelevant details while preserving the primary crack structure. Morphological filtering utilizes a combination of opening and closing operations, achieved by setting an appropriate structuring element size. For example, the structuring element size can be set to 3×3 pixels. By performing continuous opening and closing operations, small branches and breakpoints can be successfully removed, while preserving the primary crack structure. Of course, the size and shape of the structuring element can be adjusted based on the crack characteristics and the image noise level, and this is not limited by this invention.

[0097] Next, connectivity analysis is performed based on the filtered image, with the goal of constructing the topological structure of the crack network. This process, based on graph theory, defines crack intersections as nodes and connecting lines as edges. A complete connectivity matrix is ​​then established using a depth-first search algorithm. For example, node spacing can be distributed within a range of 5 to 20 centimeters, with orientation angles primarily concentrated in the 45-degree and 90-degree directions. This connectivity analysis method not only clearly demonstrates the distribution patterns of cracks but also provides important topological information for subsequent crack propagation analysis. Of course, the definitions of nodes and edges can be adjusted based on the complexity of the crack network and the analysis requirements, which is not limited by the present invention. Ultimately, through the acquisition and processing of the above features, step S15 achieves the conversion from the crack mask and fused image to the preliminary crack network topology. This preliminary crack network topology not only retains the crack centerline information but also constructs the topological relationships between the cracks through connectivity analysis. This topological structure provides an important data foundation for subsequent crack propagation analysis, stress correlation analysis, and the formulation of repair plans.

[0098] In step S16, the preliminary fracture network topology is optimized to obtain a fracture network topology, including:

[0099] Performing pruning optimization based on the preliminary crack network topology to obtain a de-redundant topology;

[0100] Perform skeleton segmentation according to the de-redundant topology to obtain a main skeleton line;

[0101] Perform branch detection based on the trunk skeleton line and the de-redundancy topology to obtain branch positions;

[0102] Performing geometric calculations based on the branch positions to obtain branch length angles;

[0103] Information fusion is performed based on the branch length angle and the de-redundant topology to obtain the crack network topology.

[0104] It should be noted that in step S16, pruning optimization is first performed based on the preliminary crack network topology, aiming to remove redundant connections and retain key crack paths. This process is implemented using a minimum spanning tree algorithm. For example, the present invention employs the Kruskal algorithm, which effectively removes redundant branches from the crack network while preserving the network's connectivity and integrity. Of course, other graph-theoretic optimization algorithms can also be selected depending on the application scenario and user needs, and the present invention is not limited to this. Secondly, skeleton segmentation is performed based on the de-redundant topology, aiming to extract the main skeleton lines. This process is implemented using a thinning algorithm. For example, the present invention employs the Zhang thinning method, which can simplify the crack network into skeleton lines with a single pixel width while preserving the crack topology. For example, for a crack of uneven width, the skeleton lines obtained after thinning can accurately reflect the crack's direction and morphological characteristics. For example, the number of iterations in the present invention is set to 10,000. Of course, the number of iterations of the thinning algorithm can be adjusted based on the complexity of the crack to ensure the accuracy and integrity of the skeleton lines. Next, branch detection is performed based on the main skeleton lines and the de-redundant topology, aiming to identify branch locations. This process is achieved by analyzing the connectivity of the pixels on the skeleton line. The specific steps are as follows: first, traverse each pixel on the skeleton line; second, for each pixel, count the number of connected pixels in its neighborhood; finally, when it is determined that the neighborhood connectivity number of a certain pixel point is greater than a preset threshold, the point is determined to be a branch point, and its position information is recorded. Exemplarily, the present invention uses 8 neighborhoods as the statistical range. When the neighborhood connectivity number of a pixel point is greater than 2, the point is considered to be a branch point. In this way, the positions of all branches can be accurately located, providing basic data for subsequent geometric calculations and information fusion.

[0105] Subsequently, geometric calculations are performed based on branch locations to determine branch lengths and angles. Branch lengths are calculated using Euclidean distance from the branch point to the branch end, while branch angles are calculated using vector angles. Finally, information fusion is performed based on branch lengths, angles, and redundant topology to obtain a complete fracture network topology. This process integrates geometric parameters with topological structure to construct a dataset containing features such as the number of branches, average length, and distribution of primary angles, ultimately forming a complete fracture network topology.

[0106] In step S17, finite element analysis is performed based on the preset material properties, stress state, and concrete structure to obtain a stress distribution cloud diagram, including:

[0107] Performing finite element meshing on the concrete structure according to the preset material properties and the stress state to obtain a finite element model;

[0108] According to the finite element model, stress calculation is performed to obtain a stress distribution cloud map.

[0109] It should be noted that in step S17, finite element analysis is performed based on the preset material properties and stress state to generate a stress distribution cloud map. First, the finite element network division is performed based on the preset material properties and stress state. Its purpose is to divide the concrete structure into several small units for accurate numerical calculations. The quality of the mesh division directly affects the accuracy of the finite element analysis. For example, for ordinary concrete beam structures, tetrahedron or hexahedron units can be used for division, and the unit size is controlled at the centimeter level. For example, for a concrete beam with a length of 6 meters and a cross-section of 400 mm × 400 mm, the initial mesh size can be set to 50 mm. Of course, other unit types can be selected or the unit size can be adjusted according to different application scenarios and user needs, and the present invention is not limited to this. Then, preset material properties are assigned to each unit. For example, the present invention uses elastic modulus, Poisson's ratio, compressive strength, etc., and combines the stress state of each unit to construct a mechanical model.

[0110] Subsequently, stress calculations are performed based on the finite element model, using numerical methods to determine the stress distribution within the structure. For example, stress contours generated through finite element analysis can visually display the stress distribution within the structure. For example, taking a simply supported beam as an example, the lower edge of the midspan is in tension, where stress can reach 2.5 MPa, while the upper edge is in compression, where stress can reach 12 MPa. Stress concentrations occur at load points, supports, and locations with sudden changes in cross-section. To refine the mesh in these stress concentration areas, the original element size can be halved. For example, if stress at a support exceeds 20 MPa, the mesh size in that area can be refined from 50 mm to 25 mm, and recalculation can yield a more accurate stress distribution. Finally, the generation of stress contours provides important guidance for structural optimization. For example, if excessive stress is detected in certain areas, optimization can be achieved by increasing the cross-section size, adding steel reinforcement, or installing stiffeners. For example, if the midspan bending moment stress of a beam exceeds the allowable value, the cross-section height can be increased from 400 mm to 500 mm, or the reinforcement ratio can be increased from 0.8% to 1.2%. Furthermore, in practical engineering, factors such as temperature stress and shrinkage stress must be considered when designing concrete components. For example, large concrete structures require analysis of the temperature stress caused by the heat of hydration during the pouring process. This can be achieved by adding a heat conduction analysis module to the finite element model, setting a cement hydration heat release curve, and calculating the temperature field distribution to determine the temperature stress.

[0111] In step S18, spatial matching is performed based on the stress distribution cloud map and the crack network topology to obtain a stress-crack correlation model, including:

[0112] Performing feature extraction based on the crack network topology to obtain crack features;

[0113] A stress-crack correlation model is constructed based on the crack characteristics and the stress distribution cloud map to obtain a stress-crack correlation model.

[0114] It should be noted that in step S18, a stress-crack correlation model is constructed by spatially matching the stress distribution cloud map with the crack network topology, thereby achieving quantitative analysis and prediction of the crack development law. First, crack feature extraction is based on the crack network topology, with the aim of extracting key geometric characteristics of the crack from the topological structure, such as the number of branches, average length, main angle distribution and other features. These features constitute a feature vector, which provides basic data for subsequent stress-crack correlation analysis. Of course, other crack features, such as the number of branches and angle distribution of the crack, can also be extracted depending on the application scenario and user needs, and the present invention is not limited to this. Secondly, the stress distribution cloud map is obtained through finite element analysis. For example, for concrete with a strength grade of C30, its compressive strength is 30 MPa, so the stress value in the high stress area reaches more than 21 MPa. The stress distribution cloud map intuitively reflects the stress state inside the structure and provides a visual basis for the association between crack features and stress.

[0115] Next, the stress-crack correlation model is constructed by spatially matching crack characteristics with stress distribution cloud maps. Specifically, after spatially matching the crack network topology with the stress distribution cloud map, the degree of spatial overlap between the stress zone and the crack network can be calculated. For example, areas with an overlap exceeding 80% are often high-risk areas for crack expansion. This spatial matching method can quantitatively analyze the relationship between stress concentration areas and crack distribution, providing a basis for predicting the direction and rate of crack expansion. The direction of crack expansion is determined based on the principal stress direction. When the stress component in a certain area exceeds 60% of the concrete tensile strength, that direction becomes the preferred direction for crack expansion. For example, in floor slab crack analysis, it was found that cracks tend to extend along areas where the stress component reaches 4 MPa. Calculating the crack expansion rate requires considering multiple factors, such as the angle between the crack expansion direction and the existing crack principal direction. For example, when the angle is less than 15 degrees, the expansion rate reaches twice the normal rate. For example, the prediction model established by the present invention using the support vector regression algorithm shows that under constant load, cracks expand by 0.1 mm per day. Of course, the crack propagation rate prediction model can be adjusted according to the specific structure and load conditions, and the present invention does not limit this. Finally, the dynamic update of the topological map information is based on the crack propagation rate. For example, for a specific shear wall component, when the crack propagation rate in a certain area exceeds 0.5 mm per day, the area will be marked as a high-risk area. This dynamic update mechanism can timely reflect the evolution of the structural damage state and provide an important reference for structural maintenance. By establishing a dynamic correlation model between stress and cracks, the law of crack development can be mastered.

[0116] In summary, to analyze the dynamic changes of concrete cracks, this solution combines multi-view image acquisition with a pre-trained neural network model. First, a multi-view industrial camera acquires a sequence of images of the concrete surface. Using spatial image alignment techniques, images from different viewpoints are matched to form a consistent image sequence. Next, the matched image sequence is fused to generate a high-resolution fused image. Furthermore, combined with the pre-trained neural network model, transfer learning techniques are used to perform semantic segmentation on the fused image and extract crack masks. Subsequently, skeleton extraction is performed based on the crack mask and the fused image to construct a preliminary crack network topology. This structure is then optimized using graph theory methods to ultimately obtain a precise crack network topology. Furthermore, finite element analysis is performed based on the preset material properties and stress state to generate a stress distribution cloud map. Finally, the stress distribution cloud map is spatially aligned with the crack network topology to construct a stress-crack correlation model.

[0117] In a dynamic monitoring environment, the morphology and orientation of concrete cracks change over time, with varying stress conditions and environmental factors. Therefore, continuous crack monitoring is necessary to capture their dynamic changes. Multi-view image acquisition enables comprehensive capture of crack morphology and orientation from different angles, resolving the issue of a single viewpoint hindering comprehensive representation of crack information. Spatial image alignment and fusion further enhance the integrity and accuracy of crack information. The introduction of a pre-trained neural network model enables semantic segmentation to efficiently and accurately identify crack masks, providing a reliable foundation for subsequent crack network topology extraction. Skeleton extraction and graph optimization further refine the topology of the crack network, enabling it to more realistically reflect dynamic crack changes. Stress distribution cloud maps generated by finite element analysis provide a crucial basis for understanding the mechanical mechanisms underlying crack dynamics. By spatially matching the stress distribution cloud map with the crack network topology, a correlation can be established between crack dynamics and their internal stress distribution, enabling a comprehensive analysis of crack dynamics.

[0118] This solution combines multi-view image acquisition with a pre-trained neural network model to achieve efficient and accurate analysis of the dynamic changes in concrete cracks. Spatial alignment and fusion of multi-view images improves the integrity of crack information. The semantic segmentation capabilities of the pre-trained neural network model enhance crack identification accuracy, while skeleton extraction and graph-theoretic optimization further refine the topology of the crack network. The spatial alignment of finite element analysis with the crack network topology clearly demonstrates the mechanical mechanisms underlying the dynamic changes in cracks.

[0119] Reference Figure 2 , an embodiment of the present invention provides a concrete structure crack image detection device based on deep learning, comprising:

[0120] An input module for acquiring multi-view images;

[0121] an alignment module, configured to perform spatial alignment based on the multi-view images to obtain a matching image sequence;

[0122] A fusion module, configured to perform image fusion according to the matching image sequence to obtain a fused image;

[0123] a mask generation module, configured to perform semantic segmentation based on the fused image and a preset neural network model and transfer learning to obtain a crack mask;

[0124] A topology generation module, configured to perform skeleton extraction based on the crack mask and the fused image to obtain a preliminary crack network topology;

[0125] A topology optimization module, configured to optimize the preliminary fracture network topology to obtain a fracture network topology;

[0126] The stress calculation module is used to perform finite element analysis based on the preset material properties, stress state and concrete structure to obtain the stress distribution cloud map;

[0127] The correlation construction module is used to perform spatial matching based on the stress distribution cloud map and the crack network topology to obtain a stress-crack correlation model.

[0128] It should be noted that the deep learning-based concrete structure crack image detection device provided in an embodiment of the present invention is used to execute all the process steps of the deep learning-based concrete structure crack image detection method in the above embodiment. The working principles and beneficial effects of the two correspond one to one, so they will not be repeated here.

[0129] An embodiment of the present invention further provides an electronic device. The electronic device includes: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps in the above-mentioned embodiments of the method for detecting cracks in concrete structures based on deep learning are implemented, for example: Figure 1 Alternatively, the processor implements the functions of the modules / units in the above-mentioned device embodiments when executing the computer program.

[0130] Exemplarily, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to implement the present invention. The one or more modules / units may be a series of computer program instruction segments capable of implementing specific functions, and the instruction segments are used to describe the execution process of the computer program in the electronic device.

[0131] The electronic device may be a computing device such as a desktop computer, notebook, PDA, or smart tablet. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will appreciate that the aforementioned components are merely examples of electronic devices and do not constitute a limitation of the electronic device. The electronic device may include more or fewer components than those described above, or a combination of certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, and the like.

[0132] The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor. The processor is the control center of the electronic device and connects various parts of the entire electronic device using various interfaces and lines.

[0133] The memory can be used to store the computer programs and / or modules. The processor implements the various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and accessing the data stored in the memory. The memory may primarily include a program storage area and a data storage area. The program storage area may store an operating system and at least one application required for a function (such as a sound playback function or an image playback function); the data storage area may store data generated based on the use of the mobile phone (such as audio data, a phone book, etc.). Furthermore, the memory may include high-speed random access memory and non-volatile memory, such as a hard disk, internal memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one disk storage device, a flash memory device, or other volatile solid-state storage device.

[0134] If the module / unit integrated into the electronic device is implemented as a software functional unit and sold or used as a standalone product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention can also implement all or part of the process steps in the above-mentioned method embodiments by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When executed by a processor, the computer program can implement the steps of each of the above-mentioned method embodiments. The computer program includes computer program code, which can be in source code form, object code form, executable file, or some intermediate form. The computer-readable medium can include any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a mobile hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunications signal, and a software distribution medium.

[0135] It should be noted that the device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed across multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present embodiment. In addition, in the drawings of the device embodiments provided by the present invention, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines. A person of ordinary skill in the art can understand and implement the present invention without inventive effort.

[0136] The specific embodiments described above further illustrate the objectives, technical solutions, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for detecting cracks in concrete structures based on deep learning, characterized in that: include: Acquire multi-view images; Performing spatial alignment on the multi-view images to obtain a matching image sequence; Performing image fusion according to the matching image sequence to obtain a fused image; Performing semantic segmentation based on the fused image and a preset neural network model based on transfer learning to obtain a crack mask; Performing skeleton extraction based on the crack mask and the fused image to obtain a preliminary crack network topology; Optimizing the preliminary fracture network topology to obtain a fracture network topology; According to the preset material properties, stress state and concrete structure, finite element analysis is performed to obtain the stress distribution cloud map; Performing spatial matching based on the stress distribution cloud map and the crack network topology to obtain a stress-crack correlation model; The step of optimizing the preliminary fracture network topology to obtain the fracture network topology includes: Performing pruning optimization based on the preliminary crack network topology to obtain a de-redundant topology; Perform skeleton segmentation according to the de-redundant topology to obtain a main skeleton line; Perform branch detection based on the trunk skeleton line and the de-redundancy topology to obtain branch positions; Performing geometric calculations based on the branch positions to obtain branch length angles; Performing information fusion according to the branch length angle and the de-redundant topology to obtain a crack network topology; The pruning optimization adopts a minimum spanning tree algorithm, including a Kruskal algorithm or a graph theory optimization algorithm; the skeleton segmentation is achieved by a thinning algorithm, which is a Zhang thinning algorithm, which is used to simplify the crack network into a skeleton line with a single pixel width while maintaining its topological structure unchanged. The number of iterations of the thinning algorithm can be set according to the complexity of the crack; the branch detection includes: traversing each pixel point on the skeleton line and counting the number of pixels connected to each pixel point in the neighborhood; when it is determined that the neighborhood connectivity number of a pixel point is greater than a preset threshold, the point is determined to be a branch point and its position information is recorded; The number of iterations of the refinement algorithm is set to 10,000; Wherein, in the branch detection, the neighborhood is an 8-neighborhood, the preset threshold is 2, and when the 8-neighborhood connectivity number of a pixel point is greater than 2, the point is determined to be a branch point; The geometric calculation includes: calculating the distance from the branch point to the branch end by Euclidean distance to obtain the branch length; and obtaining the branch angle by calculating the vector angle; The information fusion includes: integrating branch lengths and branch angles with the de-redundant topology to construct a fracture network topology dataset including branch number, average length, and main angle distribution; The step of performing spatial matching based on the stress distribution cloud map and the crack network topology to obtain a stress-crack correlation model includes: The spatial overlap between the stress distribution cloud map and the crack network topology was calculated, and areas with an overlap exceeding 80% were marked as high-risk areas for crack extension. Based on the analysis of the principal stress direction, when the stress component in a certain area exceeds 60% of the tensile strength of the concrete, this direction is determined to be the preferred direction of crack expansion; A prediction model was established based on the support vector regression algorithm to calculate the crack propagation rate.

2. The method for detecting cracks in concrete structures based on deep learning according to claim 1, characterized in that: The performing spatial alignment according to the multi-view images to obtain a matching image sequence includes: Establishing an image acquisition coordinate system based on the multi-view images to obtain image coordinates; Performing spatial transformation according to the image coordinates to obtain a spatial transformation matrix; Resampling is performed according to the spatial transformation matrix and the multi-view images to obtain a unified view image; Performing pixel interpolation based on the unified perspective image to obtain a complete image; Performing image denoising according to the complete image to obtain a low-noise image; Edge detection is performed on the low-noise image to obtain a matching image sequence.

3. The method for detecting cracks in concrete structures based on deep learning according to claim 1, characterized in that: The step of performing image fusion according to the matching image sequence to obtain a fused image includes: Performing multi-scale decomposition according to the matching image sequence to obtain a multi-scale feature image; Perform pixel fusion according to the multi-scale feature image and a preset weight rule to obtain a preliminary fused image; performing contrast enhancement on the preliminary fused image to obtain an enhanced contrast image; Performing unsharp mask processing on the contrast-enhanced image to obtain a clear image; Performing edge detection based on the clear image to obtain a crack outline; Image superposition is performed based on the crack outline and the clear image to obtain a fused image.

4. The method for detecting cracks in concrete structures based on deep learning according to claim 1, wherein: The step of performing skeleton extraction based on the crack mask and the fused image to obtain a preliminary crack network topology includes: Performing mask skeleton extraction based on the crack mask and the fused image to obtain an extracted image; Performing morphological filtering on the extracted image to obtain a filtered image; Based on the filtered image, a connection analysis is performed to obtain a preliminary crack network topology.

5. The method for detecting cracks in concrete structures based on deep learning according to claim 1, wherein: The finite element analysis is performed based on the preset material properties, stress state and concrete structure to obtain the stress distribution cloud map, including: Performing finite element meshing on the concrete structure according to the preset material properties and the stress state to obtain a finite element model; According to the finite element model, stress calculation is performed to obtain a stress distribution cloud map.

6. The method for detecting cracks in concrete structures based on deep learning according to claim 1, characterized in that: The method of performing spatial matching based on the stress distribution cloud map and the crack network topology to obtain a stress-crack correlation model includes: Performing feature extraction based on the crack network topology to obtain crack features; A stress-crack correlation model is constructed based on the crack characteristics and the stress distribution cloud map to obtain a stress-crack correlation model.

7. A concrete structure crack image detection system based on deep learning, characterized in that: The method for detecting cracks in concrete structures based on deep learning according to any one of claims 1 to 6 comprises: An input module for acquiring multi-view images; an alignment module, configured to perform spatial alignment based on the multi-view images to obtain a matching image sequence; A fusion module, configured to perform image fusion according to the matching image sequence to obtain a fused image; a mask generation module, configured to perform semantic segmentation based on the fused image and a preset neural network model and transfer learning to obtain a crack mask; A topology generation module, configured to perform skeleton extraction based on the crack mask and the fused image to obtain a preliminary crack network topology; A topology optimization module, configured to optimize the preliminary fracture network topology to obtain a fracture network topology; The stress calculation module is used to perform finite element analysis based on the preset material properties, stress state and concrete structure to obtain the stress distribution cloud map; The correlation construction module is used to perform spatial matching based on the stress distribution cloud map and the crack network topology to obtain a stress-crack correlation model.

8. An electronic device, characterized in that: The method comprises a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, the method for detecting cracks in concrete structures based on deep learning is implemented.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium includes a stored computer program, wherein, when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the deep learning-based concrete structure crack image detection method according to any one of claims 1 to 6.

Images