Computer vision-based automatic identification method for surface cracks of concrete structure

Abstract

The present application relates to the technical field of concrete structure detection vision, in particular to a concrete structure surface crack automatic identification method based on computer vision, which comprises the following steps: obtaining a concrete structure surface image to be analyzed, separating a background texture layer, a suspected crack feature layer and a noise interference layer through multi-scale texture perception decomposition, constructing an initial crack feature field containing pixel-level position and direction information, inputting a space-time continuous crack reasoning calculation network, coupling neighborhood feature propagation and global path optimization to infer a complete crack topology skeleton, generating crack vector lines through pixel-level width regression and edge positioning, completing credibility verification in combination with background texture statistical characteristics and noise spectrum features, fusing collected visual angle information to inverse reconstruct a crack three-dimensional morphology, and calculating crack length, width, area and volume parameters. The present method can accurately separate image features and completely infer crack topology, thereby improving crack identification authenticity and morphology integrity.

Description

Technical Field

[0001] This invention relates to the field of visual technology for concrete structure inspection, and in particular to a method for automatic identification of surface cracks in concrete structures based on computer vision. Background Technology

[0002] Current methods for identifying surface cracks in concrete structures largely rely on threshold segmentation, edge detection, and simple connected component analysis. Some solutions use conventional neural networks for pixel-level crack classification, but the image processing stage only performs noise filtering and feature extraction at a single scale, without refining the image features into fine-grained layers. These techniques depend on single-dimensional feature discrimination, with crack feature extraction and network inference using independent and singular processing mechanisms. They fail to combine pixel position and orientation information to construct a feature field, nor do they employ a neighborhood-and-global collaborative inference approach.

[0003] Conventional crack identification methods fail to effectively separate background textures, suspected crack features, and noise interference in images, easily leading to misidentification of background textures as cracks and noise obscuring true crack features. Furthermore, crack fracture regions lack coherent feature connections, resulting in missing and fragmented overall crack topology, making it difficult to form a complete crack representation. Therefore, it is necessary to implement multi-scale texture-aware decomposition of images to separate three independent feature layers, while simultaneously constructing an initial crack feature field carrying pixel-level position and orientation information. Finally, a spatiotemporal continuous network coupling neighborhood feature propagation and global path optimization is used to infer the complete crack topological skeleton. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and propose an automatic identification method for surface cracks in concrete structures based on computer vision.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: an automatic identification method for surface cracks in concrete structures based on computer vision, comprising:

[0006] Obtain the image of the surface of the target concrete structure to be analyzed;

[0007] Perform multi-scale texture-aware decomposition on the image to be analyzed to separate the background texture layer, the suspected crack feature layer, and the noise interference layer;

[0008] Based on the separated suspected crack feature layer, an initial crack feature field containing pixel-level position and orientation information is constructed.

[0009] The initial crack feature field is input into a spatiotemporally continuous crack inference computing network, which infers the complete crack topology skeleton by coupling neighborhood feature propagation and global path optimization.

[0010] Pixel-level width regression and edge localization are performed on the inferred complete crack topology skeleton to generate crack vector lines with geometric properties;

[0011] By combining the statistical characteristics of the background texture layer and the spectral characteristics of the noise interference layer, the credibility verification of the crack vector lines is performed to filter out false lines caused by background texture or noise.

[0012] The crack vector lines that have been verified for credibility are fused with the acquisition perspective information of the multi-view high-resolution original image, and a three-dimensional spatial state inversion algorithm is executed to reconstruct the three-dimensional spatial morphology of the crack on the concrete structure surface.

[0013] Based on the reconstructed three-dimensional spatial morphology of the cracks on the concrete structure surface, the length, average width, maximum width, area, and volume parameters of the cracks are calculated.

[0014] As a further aspect of the present invention, the step of acquiring the image to be analyzed of the surface of the target concrete structure includes:

[0015] Acquire multi-view high-resolution original images of the target concrete structure surface, the multi-view high-resolution original images covering different lighting conditions and observation angles of the concrete structure surface;

[0016] Adaptive illumination equalization and distortion correction are performed on the multi-view high-resolution original images to generate a set of images to be analyzed, specifically including:

[0017] The uneven illumination regions in the multi-view high-resolution original image are detected, and a color constancy algorithm based on the retina model is used to independently compensate each color channel.

[0018] Radial and tangential distortion corrections are performed on the illumination-compensated image using the camera intrinsic parameters and lens distortion coefficients obtained in advance by the calibration board.

[0019] Scale-invariant feature points are extracted from all distortion-corrected images, and feature matching and viewpoint alignment are performed on multiple images based on the extracted scale-invariant feature points.

[0020] In the aligned image space, the median brightness and contrast range of all images are unified, and the image to be analyzed is output with a consistent visual appearance.

[0021] As a further aspect of the present invention, multi-scale texture-aware decomposition is performed on the image to be analyzed to separate a background texture layer, a suspected crack feature layer, and a noise interference layer, including:

[0022] A filter bank consisting of Gaussian difference filters of different scales is constructed, and the image to be analyzed is convolved using the filter bank to obtain a multi-scale filter response map;

[0023] In the multi-scale filtered response map, pixel regions with response values ​​consistently higher than the threshold and consistent orientation are identified, and the set of pixel regions is initialized as the suspected crack feature layer.

[0024] The residual image is obtained by subtracting the pixel information corresponding to the suspected crack feature layer from the image to be analyzed;

[0025] Nonlocal mean denoising and local binary pattern analysis are performed on the residual image to separate regions with periodic or regular patterns as the background texture layer;

[0026] The background texture layer is further subtracted from the residual image, and the remaining discrete, irregular pixel responses are assigned to the noise interference layer.

[0027] As a further aspect of the present invention, the initial crack feature field is input into a spatiotemporally continuous crack inference computation network. This network infers a complete crack topological skeleton by coupling neighborhood feature propagation with global path optimization, including:

[0028] In the initial crack feature field, pixels with high confidence are marked as seed points;

[0029] Starting from the seed point, the crack inference computing network simulates the crack growth process and iteratively propagates activation signals to adjacent pixel regions.

[0030] In each propagation iteration, the probability of a pixel being activated as a crack pixel is calculated based on the orientation and intensity information of adjacent pixels, and the state of the pixels in the initial crack feature field is updated.

[0031] Set a propagation stop condition: stop propagation when continuous iterations fail to activate new valid pixels or the preset maximum number of iterations is reached.

[0032] Connect all activated pixels and use a thinning algorithm to extract the center line of a single pixel width to form the complete crack topology skeleton.

[0033] As a further aspect of the present invention, pixel-level width regression and edge localization are performed on the inferred complete crack topology skeleton to generate crack vector lines with geometric attributes, including:

[0034] For each pixel on the complete crack topology skeleton, search for the extreme points of gray-level gradient change in the suspected crack feature layer along its normal direction.

[0035] The two extreme points found in the normal direction are marked as the left edge point and the right edge point of the crack, respectively;

[0036] Calculate the Euclidean distance between the left edge point and the right edge point, and use it as the crack width at the pixel location;

[0037] Record the coordinates of the left edge point, the right edge point, and the center skeleton point, and organize the left edge point, the right edge point, and the center skeleton point into a series of short line segments connected end to end according to the skeleton connection order to form the crack vector line;

[0038] The data structure of the crack vector lines stores the coordinates of each line segment endpoint, the crack width at the line segment endpoint, and the coordinates of the corresponding edge point.

[0039] As a further aspect of the present invention, combining the statistical characteristics of the background texture layer and the spectral characteristics of the noise interference layer, a credibility verification is performed on the crack vector lines, including:

[0040] Extract the local binary pattern histogram features of the background texture layer, and calculate the similarity between the local binary pattern histogram features of the region where each crack vector line is located and the typical concrete background texture features;

[0041] Calculate the energy value of the noise interference layer in the area covered by the crack vector line;

[0042] A similarity threshold and an energy threshold are set. If the texture similarity of the region where a crack vector line is located is higher than the similarity threshold, the crack vector line is determined to be background texture interference.

[0043] If the noise energy value of the area where a crack vector line is located is lower than the energy threshold, and the changes in its length and width conform to the continuity and gradation of the crack, then the crack vector line is determined to be a real crack.

[0044] Lines identified as background texture interference are removed from the crack vector line set, while lines identified as real cracks are retained.

[0045] As a further aspect of the present invention, the step of fusing the reliability-verified crack vector lines with the acquisition perspective information of the multi-view high-resolution original image, and executing a three-dimensional spatial state inversion algorithm to reconstruct the three-dimensional spatial morphology of the crack on the concrete structure surface includes:

[0046] Using the camera intrinsic and extrinsic parameters of the multi-view high-resolution original images, stereo matching is performed on the feature points on each crack vector line, and the disparity in images from different viewpoints is calculated.

[0047] Based on the parallax information, the three-dimensional spatial coordinates of each feature point on the crack vector line in the camera coordinate system are calculated using the principle of triangulation.

[0048] Based on the calculated set of three-dimensional spatial coordinate points, a spatial curve fitting algorithm is used to reconstruct the continuous trajectory of the crack centerline in three-dimensional space.

[0049] Based on the crack edge points located in the two-dimensional image and their corresponding width information, the three-dimensional spatial surface contour of the crack is extended along the normal plane direction of the three-dimensional centerline trajectory, thus completing the three-dimensional spatial morphology reconstruction of the crack on the concrete structure surface.

[0050] As a further aspect of the present invention, based on the reconstructed three-dimensional spatial morphology of the crack on the surface of the concrete structure, the length, average width, maximum width, area, and volume parameters of the crack are calculated, including:

[0051] In three-dimensional space, the length of the three-dimensional crack is obtained by piecewise integration of the three-dimensional trajectory of the crack centerline and summing the results.

[0052] At multiple equally spaced sampling points along the center line of the crack, the corresponding crack width values ​​are read, and the arithmetic mean of the crack width values ​​is calculated as the average width.

[0053] Compare the crack width values ​​at all sampling points and find the maximum value as the maximum width;

[0054] The surface contour of the three-dimensional crack is triangularized and the sum of the areas of all triangular meshes is calculated as the approximate area of ​​the crack.

[0055] Using the three-dimensional crack surface profile as the boundary, calculate the volume of the spatial region enclosed by the crack and the concrete structure surface, and use it as the approximate volume of the crack.

[0056] As a further aspect of the present invention, the method also includes integrating the vector lines, three-dimensional spatial morphology, and geometric parameters of all cracks to generate a structured crack identification report, the specific steps of which include:

[0057] Assign a globally unique identifier to each identified crack;

[0058] The two-dimensional vector line data, three-dimensional spatial morphology mesh or point cloud data, and the calculated length, average width, maximum width, area and volume parameters of each crack are associated and stored with the crack identifier.

[0059] Sort all crack data in descending order of crack length;

[0060] The sorted crack data, including identifiers, various geometric parameters, and data indexes of three-dimensional spatial morphology, are filled into a preset report template to generate a structured crack identification report containing text and data indexes.

[0061] As a further aspect of the present invention, uneven illumination regions in the multi-view high-resolution original image are detected, and a color constancy algorithm based on a retina model is used to independently compensate each color channel, including:

[0062] For each color channel of the multi-view high-resolution original image, the local contrast of each pixel is calculated through a center-surround mechanism, and continuous areas with local contrast below a preset contrast threshold are marked as suspected uneven illumination areas.

[0063] Within the suspected uneven lighting area, the standard deviation of the pixel values ​​within the suspected uneven lighting area is calculated. If the standard deviation is lower than a preset variance threshold, the suspected uneven lighting area is finally confirmed as an uneven lighting area.

[0064] For each confirmed area of ​​uneven illumination, the pixel value distribution of the red, green, and blue color channels of the multi-view high-resolution original image within the corresponding area is extracted.

[0065] Based on the response characteristics of photoreceptor cells in the retinal model, a gain function related to the brightness distribution of surrounding pixels is independently constructed for each color channel;

[0066] The gain function is applied to the pixel values ​​of the corresponding color channel in the uneven illumination area, and the pixel values ​​are non-linearly adjusted so that the statistical distribution of each pixel value in the adjusted area is consistent with the global statistical distribution trend of the color channels in the whole image, thereby achieving independent compensation for each color channel.

[0067] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0068] Multi-scale texture-aware decomposition decomposes the image to be analyzed into features. Based on the differences in texture attributes and feature distribution within the image, it directly divides the image into a background texture layer, a suspected crack feature layer, and a noise interference layer. It performs layered isolation processing on visual features from different sources within the image, physically stripping away the natural texture of the concrete surface, irrelevant noise, and potential crack features. This blocks the transmission of background texture and noise interference to the crack feature extraction stage, ensuring that the suspected crack feature layer retains only visual response information related to cracks. This weakens the influence of non-target features on the crack identification process, maintains the feature purity of the crack feature extraction stage, reduces feature interference caused by non-crack features in subsequent processing, and achieves refined layered control of image features.

[0069] An initial crack feature field carrying pixel-level position and orientation information is constructed. This feature field is then input into a spatiotemporally continuous crack inference computation network. The network simultaneously initiates two computational mechanisms: neighborhood feature propagation and global path optimization. Neighborhood feature propagation continuously transmits pixel features in local fracture areas of the crack, filling in feature gaps at discontinuous locations. Global path optimization regularizes and connects the crack topology based on the feature distribution patterns of the overall image. The two mechanisms are coupled to form a coherent crack topology structure, eliminating fractures and discontinuities in the crack skeleton in local areas and generating a complete and intact crack topology skeleton. This ensures the continuity and integrity of the crack topology, providing a regular structural foundation for subsequent pixel-level width regression, edge localization, and 3D morphology inversion. Attached Figure Description

[0070] Figure 1 This is a flowchart of the automatic identification method for surface cracks in concrete structures based on computer vision, as described in this invention.

[0071] Figure 2 A flowchart for multi-scale texture perception decomposition;

[0072] Figure 3 A diagram illustrating the iterative process of crack topological reasoning;

[0073] Figure 4 This is a diagram showing the distribution of crack width along its length.

[0074] Figure 5 This is a multi-dimensional analysis diagram of the three-dimensional geometric parameters of concrete cracks. Detailed Implementation

[0075] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0076] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0077] See Figure 1This invention provides a method for automatic identification of surface cracks in concrete structures based on computer vision, the method comprising:

[0078] The process involves acquiring an image of the target concrete structure surface, performing multi-scale texture-aware decomposition to separate a background texture layer, a suspected crack feature layer, and a noise interference layer. Based on the separated suspected crack feature layer, an initial crack feature field containing pixel-level position and orientation information is constructed. This initial crack feature field is then input into a spatiotemporally continuous crack inference computation network, which infers a complete crack topology skeleton through coupled neighborhood feature propagation and global path optimization. Pixel-level width regression and edge localization are performed on the inferred complete crack topology skeleton to generate crack vector lines with geometric attributes. Combining the statistical characteristics of the background texture layer and the spectral characteristics of the noise interference layer, the credibility of the crack vector lines is verified, filtering out false lines caused by background texture or noise. The credibility-verified crack vector lines are fused with the acquisition viewpoint information of the multi-view high-resolution original image, and a three-dimensional spatial state inversion algorithm is executed to reconstruct the three-dimensional spatial morphology of the crack on the concrete structure surface. Based on the reconstructed three-dimensional spatial morphology of the crack, the length, average width, maximum width, area, and volume parameters of the crack are calculated.

[0079] In one embodiment of the present invention, multi-view high-resolution raw images of the target concrete structure surface are acquired, covering different lighting conditions and viewing angles of the concrete structure surface. Adaptive illumination equalization and distortion correction processing are performed on the multi-view high-resolution raw images to generate a set of images to be analyzed. To detect uneven illumination regions in multi-view high-resolution original images, a color constancy algorithm based on a retina model is employed to independently compensate each color channel. Specifically, for each color channel of the multi-view high-resolution original image, the local contrast of each pixel is calculated using a center-surround mechanism. Continuous regions with local contrast below a preset contrast threshold are marked as suspected uneven illumination regions. Within the suspected uneven illumination regions, the standard deviation of pixel values ​​within that region is calculated. If the standard deviation is below a preset variance threshold, the region is finally confirmed as an uneven illumination region. For each confirmed uneven illumination region, the pixel value distribution of the red, green, and blue color channels within the corresponding region is extracted. Based on the response characteristics of photoreceptor cells in the retina model, a gain function related to the brightness distribution of surrounding pixels is independently constructed for each color channel. This gain function is applied to the pixel values ​​of the corresponding color channel within the uneven illumination region, performing a non-linear adjustment of the pixel values ​​so that the statistical distribution of each pixel value within the adjusted region is consistent with the global statistical distribution trend of the color channels in the entire image. Using camera intrinsic parameters and lens distortion coefficients pre-obtained from a calibration board, radial and tangential distortion corrections are performed on the illumination-compensated image. Scale-invariant feature points are extracted from all distortion-corrected images. Based on these extracted feature points, feature matching and viewpoint alignment are performed on multiple images. In the aligned image space, the median brightness and contrast range of all images are unified, resulting in an output image with a consistent visual appearance for analysis.

[0080] In practice, high-resolution original images of the target concrete structure surface are acquired from multiple perspectives. Taking the inspection of a concrete bridge pier surface as an example, the acquisition process uses an industrial camera equipped with a high-resolution sensor to capture images from multiple fixed points surrounding the pier. The high-resolution original images from multiple perspectives cover different lighting conditions on the concrete structure surface under natural sunlight and partial shadow areas, as well as different viewing angles from the front to the side, to ensure sufficient visual coverage of crack areas. Adaptive illumination equalization and distortion correction are performed on the high-resolution original images from multiple perspectives to generate a set of images with geometric and visual consistency to be analyzed. This process is a prerequisite for accurate subsequent analysis.

[0081] In some embodiments, uneven illumination regions in multi-view high-resolution original images are detected, and a color constancy algorithm based on a retinal model is used to independently compensate each color channel. Specifically, the red, green, and blue color channels are processed separately. The local contrast value of each pixel is calculated by simulating the central-surround mechanism of the receptive field of retinal ganglion cells. Continuous regions in the image with local contrast values ​​lower than a preset contrast threshold are marked as suspected uneven illumination regions. Within a suspected uneven illumination region, the standard deviation of the pixel values ​​within that region is calculated. If the calculated standard deviation is lower than a preset variance threshold, the suspected uneven illumination region is finally confirmed as an uneven illumination region requiring compensation. For each finally confirmed uneven illumination region, histograms of pixel value distributions for the red, green, and blue color channels within the corresponding region are extracted. Based on the nonlinear response characteristics of photoreceptor cells to different brightness levels in the retinal model, a gain function related to the brightness distribution of surrounding pixels is independently constructed for the red color channel. This gain function is used to adjust the pixel values ​​of the red channel in areas of uneven illumination, ensuring that the statistical distribution of the adjusted red channel pixel values ​​in these areas aligns with the global statistical distribution trend of the red channel in the entire image. The green and blue channels are processed using the same principle but with independently constructed gain functions, thus achieving independent compensation for each color channel. This suppresses color shifts and brightness differences caused by illumination in the compensated image. Specifically, the gain function for a single color channel... One form of expression is:

[0082]

[0083] in: This represents the matrix of original pixel values ​​for the color channels within the unevenly lit region. This represents the local standard deviation of pixel values ​​in that region. It represents the global standard deviation of the pixel values ​​for that color channel across the entire image. and These are adjustable model parameters used to control the intensity and shape of the gain adjustment curve. This is a hyperbolic tangent function. This gain function acts on the pixel values ​​within areas of uneven illumination, achieving non-linear adjustment of the pixel values.

[0084] In some embodiments, radial and tangential distortion corrections are performed on the illumination-compensated image using camera intrinsic parameters and lens distortion coefficients pre-obtained from the calibration board. Specifically, a high-precision checkerboard calibration board of known size is used, and multiple images of the calibration board are captured at different positions and angles. The camera intrinsic parameter matrix, radial distortion coefficients, and tangential distortion coefficients are solved using the Zhang Zhengyou calibration method. Using the solved camera intrinsic parameters and distortion coefficients, a distortion correction model is applied to each illumination-compensated multi-view high-resolution original image, calculating its new position pixel-by-pixel in the corrected image to eliminate radial and tangential distortion introduced by the lens, resulting in an image with higher geometric accuracy.

[0085] In practice, scale-invariant feature points are extracted from all distortion-corrected images. Based on these extracted feature points, feature matching and viewpoint alignment are performed on multiple images. This can be understood as using SIFT or ORB algorithms to extract a large number of scale-invariant feature points and their descriptors from each distortion-corrected image. By calculating the Hamming or Euclidean distance between the descriptors, feature point matching is performed across multiple images to find a set of matching point pairs corresponding to the same physical point on the concrete surface in images from different viewpoints. Using these matching point pairs, the fundamental matrix or homography transformation matrix between images is estimated using a random sampling consensus algorithm. Based on this, all images are transformed to a common viewpoint coordinate system, completing the viewpoint alignment of multiple images.

[0086] In practice, within the aligned image space, the median brightness and contrast range of all images are unified, resulting in an output image with a consistent visual appearance. Optionally, the overall distribution of pixel brightness in all aligned images is statistically analyzed, and its global median brightness is calculated. The median brightness of each aligned image is calculated separately, and then adjusted to match the global median brightness using a linear or piecewise linear grayscale transformation function. Simultaneously, the grayscale histogram of each image is stretched or compressed to unify the grayscale value distribution range of all images to a preset standard contrast range. The final output image exhibits consistent brightness and contrast, constituting the input image for the subsequent crack detection algorithm.

[0087] In one embodiment of the present invention, see [reference] Figure 2A filter bank consisting of Gaussian difference filters at different scales is constructed. This filter bank is then used to convolve the image to be analyzed, resulting in a multi-scale filtered response map. In the multi-scale filtered response map, pixel regions with response values ​​consistently above a threshold and exhibiting directional consistency are identified, and this set of pixel regions is initialized as a suspected crack feature layer. The pixel information corresponding to the suspected crack feature layer is subtracted from the image to be analyzed, yielding a residual image. Nonlocal mean denoising and local binary pattern analysis are performed on the residual image to separate regions with periodic or regular patterns, which are then used as a background texture layer. The background texture layer is further subtracted from the residual image, and the remaining discrete and irregular pixel responses are classified as a noise interference layer. In the initial crack feature field, pixels with high confidence are marked as seed points. Starting from the seed points, a crack inference computation network simulates the crack growth process, iteratively propagating activation signals to adjacent pixel regions. In each propagation iteration, the probability of a pixel being activated as a crack pixel is calculated based on the direction and intensity information of adjacent pixels, and the state of the pixels in the initial crack feature field is updated accordingly. Set a propagation stopping condition: stop propagation when continuous iterations fail to activate new valid pixels or the preset maximum number of iterations is reached. Connect all activated pixels and use a thinning algorithm to extract the centerline of a single pixel width, forming a complete crack topology skeleton.

[0088] In the specific implementation, a filter bank composed of Gaussian difference filters of different scales is constructed. This filter bank contains multiple Gaussian kernels with different standard deviations, and its difference is used to simulate edge and texture responses at multiple scales. The filter bank is used to perform convolution operations on the preprocessed image to be analyzed, resulting in a series of multi-scale filtered response maps. In the multi-scale filtered response maps, by setting a dynamic threshold and analyzing the directional consistency within the pixel neighborhood, pixel regions with response values ​​consistently higher than the threshold and directional consistency are identified, and the set of these pixel regions is initialized as a suspected crack feature layer. The pixel information corresponding to the suspected crack feature layer is subtracted from the image to be analyzed, resulting in a residual image containing background and noise. Non-local mean denoising is performed on the residual image to smooth random noise, while local binary pattern analysis is performed to capture texture patterns, thereby separating regions with periodic or regular patterns as the background texture layer. The background texture layer is further subtracted from the residual image, and the remaining discrete and irregular pixel responses are classified as noise interference layers. This completes the multi-scale texture-perceptual decomposition of the image to be analyzed, separating the background texture layer, the suspected crack feature layer, and the noise interference layer.

[0089] In some embodiments, an initial crack feature field containing pixel-level position and orientation information is constructed based on the separated suspected crack feature layer. In the initial crack feature field, pixels with high confidence are marked as seed points. The determination of high confidence is typically based on the gradient magnitude of the pixel and its response intensity in the suspected crack feature layer. Starting from the seed points, the crack inference computation network simulates the crack growth process and iteratively propagates activation signals to neighboring pixel regions. In each propagation iteration, the probability of a pixel being activated as a crack pixel is calculated based on the orientation and intensity information of neighboring pixels, and the state of the pixels in the initial crack feature field is updated. It is understood that calculating the probability of a pixel being activated as a crack pixel involves coupled analysis of neighborhood features. One implementation method evaluates the probability of a pixel being activated using the following formula:

[0090]

[0091] in: Indicates the location at coordinates The probability that a pixel is activated as a crack pixel in this iteration. This indicates the intensity response value of the pixel in the suspected crack feature layer. This indicates the degree of consistency between the direction of this pixel and the direction of the current propagation path. and These are weighting parameters used to balance the contributions of intensity and directionality. This represents the natural exponential function. This probability calculation is the core step in the coupled neighborhood feature propagation. The crack inference computation network uses this probability value to determine whether to update the pixel state to an active state.

[0092] A propagation stopping condition is set: propagation stops when continuous iterations fail to activate new valid pixels or when the preset maximum number of iterations is reached. After the propagation process stops, all activated pixels form a connected region. A thinning algorithm is used to iteratively erode this region, extracting the centerline of a single pixel width, ultimately forming a complete crack topology skeleton.

[0093] In its implementation, the iterative propagation process of the crack inference computation network relies on the directional information of pixels in the initial crack feature field. This directional information can be obtained by calculating the gradient direction of the pixel or using a directional filter. In each iteration, the crack inference computation network examines the eight neighboring pixels of the currently active pixel, calculates the activation probability of each neighboring pixel according to a formula, and if the probability exceeds a set activation threshold, the state of that neighboring pixel is updated to active and added to the set of active pixels for the current iteration. Updating the state of pixels in the initial crack feature field includes recording the pixel's activation round and cumulative confidence. This information is used to influence the selection of propagation paths in subsequent iterations, achieving global path optimization. Optionally, global path optimization can be achieved by evaluating the cumulative probability cost of multiple potential propagation paths; the crack inference computation network tends to prioritize the path with the highest cumulative probability cost for propagation.

[0094] In some embodiments, the "inability to activate new valid pixels" condition in the propagation stopping condition means that in a number of consecutive iterations, no pixel's activation probability exceeds the activation threshold, indicating that crack growth may have reached its endpoint or encountered a non-crack region. The maximum number of iterations is a preset integer used to prevent the propagation process from getting stuck in an infinite loop or overgrowing in complex scenes. It can be understood that the thinning algorithm extracting the centerline is a repetitive process. In each thinning iteration, the pixel's eight-neighborhood pattern is used to determine whether it is a boundary point and can be deleted, until no more pixels can be deleted, ultimately obtaining a crack topology skeleton with a single pixel width. Optionally, after extracting the centerline, the skeleton can be post-processed, such as removing short, burr branches to smooth the skeleton lines, ensuring that the complete crack topology skeleton accurately reflects the main direction and connection relationships of the crack.

[0095] In one embodiment of the present invention, for each pixel on the complete crack topological skeleton, extreme points of gray-level gradient change are searched along its normal direction in the suspected crack feature layer. The two extreme points found in the normal direction are marked as the left edge point and the right edge point of the crack, respectively. The Euclidean distance between the left edge point and the right edge point is calculated as the crack width at the pixel location. The coordinates of the left edge point, the right edge point, and the central skeleton point are recorded, and these points are organized into a series of short line segments connected end to end according to the skeleton connection order to form crack vector lines. In the data structure of crack vector lines, the coordinates of the endpoints of each line segment, the crack width at the endpoints of the line segments, and the coordinates of the corresponding edge points are stored. The local binary pattern histogram features of the background texture layer are extracted, and the similarity between the local binary pattern histogram features of the region where each crack vector line is located and the typical concrete background texture features is calculated. The energy value of the noise interference layer in the region covered by the crack vector lines is calculated. A similarity threshold and an energy threshold are set. If the texture similarity of the region where a crack vector line is located is higher than the similarity threshold, the line is determined to be background texture interference. If the noise energy value of the region where a crack vector line is located is lower than the energy threshold, and its length and width changes conform to the continuity and gradation of a crack, the line is determined to be a real crack. Lines determined to be background texture interference are deleted from the crack vector line set, and lines determined to be real cracks are retained.

[0096] In practical implementation, for each pixel on the complete crack topological skeleton, the extreme points of gray-level gradient change are searched along its normal direction in the suspected crack feature layer. The normal direction is determined by the orthogonal direction of the local skeleton orientation at that skeleton point. The two extreme points found in the normal direction are marked as the left and right edge points of the crack, respectively. These two points correspond to the zero-crossing points in the suspected crack feature layer where the gradient changes from positive to negative or from negative to positive. The Euclidean distance between the left and right edge points is calculated and used as the crack width at the location of the central skeleton point. The coordinates of the left, right, and central skeleton points are recorded, and these points are organized into a series of short line segments connected end to end according to the skeleton connection order, forming crack vector lines. In the data structure of crack vector lines, the two-dimensional or three-dimensional coordinates of each line segment endpoint, the crack width value at the line segment endpoint, and the coordinates of the corresponding left and right edge points are stored, thereby generating crack vector lines with complete geometric properties.

[0097] In some embodiments, local binary pattern histogram features of the background texture layer are extracted, and the similarity between the local binary pattern histogram features of the region where each crack vector line is located and the typical concrete background texture features is calculated. It is understood that the local binary pattern histogram features reflect the texture statistical characteristics of local image regions, and the similarity can be calculated using histogram intersection or chi-square distance. Simultaneously, the energy value of the noise interference layer in the region covered by the crack vector line is calculated. The energy value can be measured by calculating the sum of squares of the pixel values ​​of the noise interference layer in that region. Similarity thresholds and energy thresholds are set. If the texture similarity of the region where a crack vector line is located is higher than the similarity threshold, then this crack vector line is determined to be background texture interference; for example, formwork joints or water stains on a concrete surface may be misjudged as cracks. If the noise energy value of the region where a crack vector line is located is lower than the energy threshold, and its length and width changes conform to the inherent continuity and gradation of cracks, then this crack vector line is determined to be a real crack. When determining a crack vector line to be a real crack, the assessment of the continuity and gradation of the length and width changes can be achieved by analyzing the geometric attribute sequence of the crack vector line. One implementation method is to evaluate the continuity score of width variation using the following formula:

[0098]

[0099] in: The score represents the continuity of the changes in the width and direction of the crack vector line. This indicates the number of points sampled sequentially along the crack vector line. Indicates the first Crack width at each sampling point Indicates the first Tangential direction of the crack skeleton at each sampling point Indicates the first Crack width at +1 sampling point Indicates the first Tangential direction of the crack skeleton at +1 sampling point It is a smoothing function of angular differences; the smaller the change in direction between adjacent points, the closer the function value is to 1. (Higher...) The value indicates that the width and direction change gradually, which conforms to the geometric characteristics of a real crack. Combining similarity, noise energy, and continuity scores for comprehensive judgment can improve the robustness of the discrimination.

[0100] In practice, lines identified as background texture interference are removed from the crack vector line set, while lines identified as genuine cracks are retained, thus verifying the credibility of the crack vector lines. Optionally, typical concrete background texture features can be obtained by calculating the average value or principal component features of the local binary pattern histogram features in crack-free concrete sample areas. Optionally, similarity thresholds and energy thresholds can be predetermined by performing statistical analysis on annotated image datasets containing known genuine cracks and known texture interference to maximize the retention rate of genuine cracks and the filtering rate of false lines. It can be understood that calculating the energy value of the noise interference layer in the area covered by crack vector lines requires accurately mapping the coordinates of the crack vector lines onto the noise interference layer image and extracting the pixels of the corresponding area for calculation.

[0101] See Figure 3 This is a graph analyzing the iterative process of crack topology inference, visually demonstrating the convergence characteristics and activation patterns of the crack growth algorithm. The number of seed points is approximately 50 at iteration 1, slowly increasing with each iteration, stabilizing at approximately 360 at iteration 10. After iteration 7, the growth slows, indicating that the seed points have largely completed their diffusion and activation. The number of activated crack pixels is approximately 120 at iteration 1, rapidly increasing with each iteration, reaching approximately 1600 at iteration 10. After iteration 8, the growth rate slows, gradually approaching saturation, indicating that the crack topology framework is essentially complete. High-confidence pixels, as seed points, are the starting point for crack growth; their increase represents the gradual expansion of the effective crack region. The rapid increase in the number of activated pixels reflects the algorithm's process of propagating crack features to the neighborhood and constructing a continuous topology framework. Both curves flatten out later, indicating that the algorithm completes crack topology inference after 8-10 iterations, and further iterations will not significantly improve the recognition performance.

[0102] In one embodiment of the present invention, the camera intrinsic and extrinsic parameters of the multi-view high-resolution original images are used to perform stereo matching on the feature points of each crack vector line, and the disparity in different view images is calculated. Based on the disparity information, the three-dimensional spatial coordinates of each feature point on the crack vector line in the camera coordinate system are calculated using the principle of triangulation. Based on the calculated set of three-dimensional spatial coordinate points, a spatial curve fitting algorithm is used to reconstruct the continuous trajectory of the crack centerline in three-dimensional space. Based on the crack edge points located in the two-dimensional image and their corresponding width information, the three-dimensional spatial surface contour of the crack is extended along the normal plane direction of the three-dimensional centerline trajectory, completing the three-dimensional spatial morphology reconstruction of the crack on the concrete structure surface. In three-dimensional space, the three-dimensional trajectory of the crack centerline is piecewise integrated, and the sum is obtained to obtain the length of the three-dimensional crack. At multiple equally spaced sampling points of the crack centerline, the corresponding crack width values ​​are read, and the arithmetic mean of these width values ​​is calculated as the average width. The crack width values ​​at all sampling points are compared, and the maximum value is found as the maximum width. The three-dimensional crack surface contour is triangulated, and the sum of the areas of all triangular meshes is calculated as the approximate area of ​​the crack. Using the three-dimensional crack surface profile as the boundary, calculate the volume of the spatial region enclosed by the crack and the concrete structure surface, and use it as the approximate volume of the crack.

[0103] In practical implementation, the camera intrinsic and extrinsic parameters of multi-view high-resolution original images are used to perform stereo matching on feature points on each verified crack vector line. The disparity of feature points in different viewpoints is calculated. Feature points are typically selected as the central skeleton points or edge points on the crack vector line. Based on the calculated disparity information, the three-dimensional spatial coordinates of each feature point on the crack vector line in the camera coordinate system are calculated using the principle of triangulation. These three-dimensional spatial coordinate points constitute a discrete set of spatial points. Based on the calculated set of three-dimensional spatial coordinate points, a spatial curve fitting algorithm, such as the B-spline curve fitting algorithm, is used to reconstruct the continuous trajectory of the crack centerline in three-dimensional space. Based on the crack edge points located in the two-dimensional image and their corresponding width information, the crack surface contour points are generated along the normal plane direction of the three-dimensional centerline trajectory, based on the width value at each center point. By connecting these contour points, the three-dimensional spatial morphology of the crack on the concrete structure surface is reconstructed. The reconstruction result can be represented as a triangular mesh model.

[0104] In some embodiments, a series of geometric parameters of the crack are calculated based on the reconstructed three-dimensional spatial morphology of the crack on the concrete structure surface. In three-dimensional space, the length of the three-dimensional crack is obtained by piecewise integration of the three-dimensional trajectory of the crack centerline and summing the lengths of these segments. Piecewise integration involves discretizing the centerline trajectory into a large number of tiny line segments and summing their lengths. At multiple equally spaced sampling points along the crack centerline, the crack width value corresponding to each sampling point is read, and the arithmetic mean of these width values ​​is calculated. The result is used as the average crack width. The crack width values ​​at all sampling points are compared, and the maximum value is identified and used as the maximum crack width. The three-dimensional crack surface contour is triangulated, and the sum of the areas of all triangular meshes is calculated. This sum of areas is used as the approximate surface area of ​​the crack. Using the three-dimensional crack surface contour as the boundary, the volume of the spatial region enclosed by the crack surface and the concrete structure surface is calculated, and this volume is used as the approximate volume of the crack. It can be understood that triangulation of the crack surface contour is the process of generating triangular patches from discrete surface contour points according to topological connections, calculating the area of ​​each triangular patch, and summing the areas to obtain the total area. The approximate volume of a crack can be calculated by discretizing the space between the crack surface and the structural surface into triangular prisms or tetrahedrons and summing their volumes. Alternatively, the formula for calculating the approximate surface area of ​​the crack can be expressed as:

[0105]

[0106] in: This represents the calculated approximate surface area of ​​the crack. This represents the total number of triangular patches obtained after triangulating the 3D surface contour of the crack. Indicates the first The area of ​​the triangular facet. This represents the summation of the areas of all triangular faces. The area of ​​a triangular face can be calculated from the three-dimensional coordinates of its three vertices.

[0107] In practical implementation, the calculation of crack geometric parameters can be illustrated using a three-dimensional crack model of a specific concrete beam member. Table 1 below shows an example of the main geometric parameter results obtained after calculating an identified crack:

[0108] Table 1: Example Table of Crack Geometric Parameter Calculation Results

[0109] Crack identifier CRACK-2024-001 none Three-dimensional length 1.245 rice Average width 0.85 millimeters Maximum width 1.62 millimeters Approximate surface area 0.0106 square meters Approximate volume 3.82e-6 cubic meter

[0110] In practical implementation, the calculation of the approximate volume of the crack is based on the three-dimensional crack surface contour as the boundary. The spatial region enclosed by the crack surface and the concrete structure surface can be regarded as an irregular thin layer. Optionally, one method for calculating this volume is to discretize the spatial region along the depth direction (i.e., the direction perpendicular to the normal to the structural surface) into a series of thin slices, with the volume of each slice approximating the volume of a prism. The total volume is obtained by summing the volumes of all slices. It can be understood that the piecewise integration of the three-dimensional trajectory of the crack centerline, along with the calculation of parameters such as width, area, and volume, together constitute a quantitative description of the three-dimensional spatial morphology of the crack, providing data for structural health assessment. In some embodiments, the number of equally spaced sampling points along the crack centerline can be adaptively determined according to the three-dimensional length of the crack, for example, setting a sampling point every one centimeter or five centimeters to ensure the representativeness of the width statistics.

[0111] See Figure 4 This is a distribution chart analyzing the crack width along its length, visually illustrating the variation in crack width along the length direction. The width from 0–700 mm is generally higher than the average, reaching a peak of 1.62 mm at approximately 300 mm, representing the most dangerous area of ​​the crack; the overall fluctuation is significant, consistent with the characteristics of real cracks: "local expansion, continuous gradual change." The width from 700–1300 mm is consistently lower than the average, gradually narrowing to the 0.3–0.6 mm range, reflecting the physical characteristic of the crack gradually closing towards its end. The width variation is not drastic, consistent with the geometric continuity of real cracks, and can serve as a direct basis for the "continuity score" in credibility verification. This curve was generated by a pixel-level width regression and edge localization algorithm, completely recording the width, edge coordinates, and other geometric information of each point along the crack centerline. The peak width (1.62 mm) is a key indicator for structural safety assessment; exceeding the standard threshold requires focused monitoring or reinforcement. The average width is used to quantify the overall degree of damage.

[0112] In one embodiment of the invention, a globally unique identifier is assigned to each identified crack. The two-dimensional vector line data, three-dimensional spatial morphology mesh or point cloud data, and calculated length, average width, maximum width, area, and volume parameters of each crack are associated and stored with the crack's identifier. All crack data are sorted in descending order of crack length. The sorted crack data, including the identifier, various geometric parameters, and a data index of the three-dimensional spatial morphology, is then filled into a preset report template to generate a structured crack identification report containing text and a data index.

[0113] In practical implementation, each identified crack is assigned a globally unique identifier. The generation of the globally unique identifier can combine the image acquisition timestamp, spatial location information, and sequence number, for example, using the encoding rule of "structure number-acquisition date-crack sequence number". The two-dimensional vector line data, three-dimensional spatial morphology mesh or point cloud data, and calculated length, average width, maximum width, area, and volume parameters of each crack are associated and stored with the globally unique identifier of the crack. The associated storage can be implemented by establishing a relational database table, using the globally unique identifier as the primary key, and storing the file paths referenced by the two-dimensional vector line data, three-dimensional morphology data, and various geometric parameters as associated fields or foreign keys.

[0114] In some embodiments, the data for all cracks is sorted in descending order of crack length. The sorting process requires reading the length parameter from each crack data record. It is understood that, to achieve stable and reproducible sorting, a secondary sorting key is needed in the rare case where two cracks have exactly equal lengths. One implementation calculates a comprehensive sorting key, as follows:

[0115]

[0116] in: Indicates the first The sorting key value of the crack. Indicates the first The length of the crack, Indicates the first The average width of the crack, This represents the maximum average width among all the cracks to be sorted. It is a positive integer used to ensure the length. The dominant bit in the key value (e.g., setting it to 6, assuming the length unit is meters, is equivalent to converting the length to the order of micrometers as the integer part). It is much smaller A positive weighting coefficient (e.g., set to 1) is used to differentiate based on average width when the lengths are the same. Calculate the average width of all cracks. After the value is set, according to By sorting the values ​​from largest to smallest, cracks can be arranged in descending order of length, and when the lengths are strictly equal, they can be further sorted by average width.

[0117] In practice, the sorted crack data, including globally unique identifiers, various geometric parameters, and data indexes of 3D spatial morphology, is filled into a pre-defined report template. The report template is a structured document framework; optionally, it can be an XML file containing predefined data field placeholders. When filling in the data, the globally unique identifier of the first sorted crack is entered into the "Identifier" field corresponding to the first crack record in the report template. The crack's length, average width, maximum width, area, and volume parameters are then entered into their respective fields. The path to the corresponding 3D mesh file or index code is entered into the "3D Morphology Data Index" field. This process is repeated for all crack data until all sorted crack information is filled into the corresponding positions in the report template, ultimately generating a structured crack identification report containing text and data indexes. It can be understood that the structured crack identification report not only contains readable text and tables, but its data index also points to specific 2D vector line data files and 3D morphology data files, achieving a link between summary information and detailed data.

[0118] In some embodiments, the associated storage of globally unique identifiers can utilize a database or a standard file format with structured header information. For example, a separate JSON file can be created for each crack, named with a globally unique identifier. The file stores all geometric parameters of the crack in key-value pairs, and stores the relative paths or binary codes of two-dimensional vector line data and three-dimensional morphological data in specific fields. Optionally, the preset report template can be a Word document template, an HTML webpage template, or a PDF form template. The process of generating the report involves batch-filling the crack data into the corresponding positions in the template and rendering and exporting the final document.

[0119] See Figure 5 This is a multi-dimensional analysis diagram of the three-dimensional geometric parameters of concrete cracks, visually presenting the relationship between crack length, average width, and volume. The longest and largest crack has a three-dimensional length of approximately 1.55 meters, an average width of approximately 1.05 millimeters, and a volume of approximately 5800 cubic millimeters, making it the most dangerous crack in the current sample. Medium-sized cracks have a three-dimensional length of approximately 1.25 meters, an average width of approximately 0.85 millimeters, and a volume of approximately 3800 cubic millimeters, indicating moderate damage. Small cracks have a three-dimensional length between 0.75 and 1.0 meters, an average width between 0.6 and 0.8 millimeters, and a volume between 1300 and 2800 cubic millimeters, indicating less damage. Three-dimensional length and average width show a positive correlation; the longer the crack, the larger its average width. Volume is positively correlated with length, width, and height; the longest and widest cracks correspond to the largest volume, consistent with the calculation logic of three-dimensional geometric volume.

[0120] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for automatic identification of surface cracks in concrete structures based on computer vision, characterized in that, include: Obtain the image of the surface of the target concrete structure to be analyzed; Perform multi-scale texture-aware decomposition on the image to be analyzed to separate the background texture layer, the suspected crack feature layer, and the noise interference layer; Based on the separated suspected crack feature layer, an initial crack feature field containing pixel-level position and orientation information is constructed. The initial crack feature field is input into a spatiotemporally continuous crack inference computing network, which infers the complete crack topology skeleton by coupling neighborhood feature propagation and global path optimization. Pixel-level width regression and edge localization are performed on the inferred complete crack topology skeleton to generate crack vector lines with geometric properties; By combining the statistical characteristics of the background texture layer and the spectral characteristics of the noise interference layer, the credibility verification of the crack vector lines is performed to filter out false lines caused by background texture or noise. The reliability-verified crack vector lines are fused with the acquisition perspective information of multi-view high-resolution original images, and a three-dimensional spatial state inversion algorithm is executed to reconstruct the three-dimensional spatial morphology of the crack on the concrete structure surface. Based on the reconstructed three-dimensional spatial morphology of the cracks on the concrete structure surface, the length, average width, maximum width, area, and volume parameters of the cracks are calculated.

2. The method for automatic identification of surface cracks in concrete structures based on computer vision as described in claim 1, characterized in that, The process of acquiring the image to be analyzed of the surface of the target concrete structure includes: Acquire multi-view high-resolution original images of the target concrete structure surface, the multi-view high-resolution original images covering different lighting conditions and observation angles of the concrete structure surface; Adaptive illumination equalization and distortion correction are performed on the multi-view high-resolution original images to generate a set of images to be analyzed, specifically including: The uneven illumination regions in the multi-view high-resolution original image are detected, and a color constancy algorithm based on the retina model is used to independently compensate each color channel. Radial and tangential distortion corrections are performed on the illumination-compensated image using the camera intrinsic parameters and lens distortion coefficients obtained in advance by the calibration board. Scale-invariant feature points are extracted from all distortion-corrected images, and feature matching and viewpoint alignment are performed on multiple images based on the extracted scale-invariant feature points. In the aligned image space, the median brightness and contrast range of all images are unified, and the image to be analyzed is output with a consistent visual appearance.

3. The method for automatic identification of surface cracks in concrete structures based on computer vision as described in claim 1, characterized in that, Multi-scale texture-aware decomposition is performed on the image to be analyzed to separate the background texture layer, the suspected crack feature layer, and the noise interference layer, including: A filter bank consisting of Gaussian difference filters of different scales is constructed, and the image to be analyzed is convolved using the filter bank to obtain a multi-scale filter response map; In the multi-scale filtered response map, pixel regions with response values ​​consistently higher than the threshold and consistent orientation are identified, and the set of pixel regions is initialized as the suspected crack feature layer. The residual image is obtained by subtracting the pixel information corresponding to the suspected crack feature layer from the image to be analyzed; Nonlocal mean denoising and local binary pattern analysis are performed on the residual image to separate regions with periodic or regular patterns as the background texture layer; The background texture layer is further subtracted from the residual image, and the remaining discrete, irregular pixel responses are assigned to the noise interference layer.

4. The method for automatic identification of surface cracks in concrete structures based on computer vision as described in claim 1, characterized in that, The initial crack feature field is input into a spatiotemporally continuous crack inference computation network. This network, through coupled neighborhood feature propagation and global path optimization, infers a complete crack topological skeleton, including: In the initial crack feature field, pixels with high confidence are marked as seed points; Starting from the seed point, the crack inference computing network simulates the crack growth process and iteratively propagates activation signals to adjacent pixel regions. In each propagation iteration, the probability of a pixel being activated as a crack pixel is calculated based on the orientation and intensity information of adjacent pixels, and the state of the pixels in the initial crack feature field is updated. Set a propagation stop condition: stop propagation when continuous iterations fail to activate new valid pixels or the preset maximum number of iterations is reached. Connect all activated pixels and use a thinning algorithm to extract the center line of a single pixel width to form the complete crack topology skeleton.

5. The method for automatic identification of surface cracks in concrete structures based on computer vision as described in claim 4, characterized in that, Pixel-level width regression and edge localization are performed on the inferred complete crack topology skeleton to generate crack vector lines with geometric properties, including: For each pixel on the complete crack topology skeleton, search for the extreme points of gray-level gradient change in the suspected crack feature layer along its normal direction. The two extreme points found in the normal direction are marked as the left edge point and the right edge point of the crack, respectively; Calculate the Euclidean distance between the left edge point and the right edge point, and use it as the crack width at the pixel location; Record the coordinates of the left edge point, the right edge point, and the center skeleton point, and organize the left edge point, the right edge point, and the center skeleton point into a series of short line segments connected end to end according to the skeleton connection order to form the crack vector line; The data structure of the crack vector lines stores the coordinates of each line segment endpoint, the crack width at the line segment endpoint, and the coordinates of the corresponding edge point.

6. The method for automatic identification of surface cracks in concrete structures based on computer vision as described in claim 1, characterized in that, Combining the statistical characteristics of the background texture layer and the spectral features of the noise interference layer, a confidence verification is performed on the crack vector lines, including: Extract the local binary pattern histogram features of the background texture layer, and calculate the similarity between the local binary pattern histogram features of the region where each crack vector line is located and the typical concrete background texture features; Calculate the energy value of the noise interference layer in the area covered by the crack vector line; A similarity threshold and an energy threshold are set. If the texture similarity of the region where a crack vector line is located is higher than the similarity threshold, the crack vector line is determined to be background texture interference. If the noise energy value of the area where a crack vector line is located is lower than the energy threshold, and the changes in its length and width conform to the continuity and gradation of the crack, then the crack vector line is determined to be a real crack. Lines identified as background texture interference are removed from the crack vector line set, while lines identified as real cracks are retained.

7. The method for automatic identification of surface cracks in concrete structures based on computer vision as described in claim 1, characterized in that, The process of fusing the verified crack vector lines with the acquired viewpoint information of the multi-view high-resolution original image, and executing a three-dimensional spatial state inversion algorithm to reconstruct the three-dimensional spatial morphology of the crack on the concrete structure surface includes: Using the camera intrinsic and extrinsic parameters of the multi-view high-resolution original images, stereo matching is performed on the feature points on each crack vector line, and the disparity in images from different viewpoints is calculated. Based on the parallax information, the three-dimensional spatial coordinates of each feature point on the crack vector line in the camera coordinate system are calculated using the principle of triangulation. Based on the calculated set of three-dimensional spatial coordinate points, a spatial curve fitting algorithm is used to reconstruct the continuous trajectory of the crack centerline in three-dimensional space. Based on the crack edge points located in the two-dimensional image and their corresponding width information, the three-dimensional spatial surface contour of the crack is extended along the normal plane direction of the three-dimensional centerline trajectory, thus completing the three-dimensional spatial morphology reconstruction of the crack on the concrete structure surface.

8. The method for automatic identification of surface cracks in concrete structures based on computer vision as described in claim 7, characterized in that, Based on the reconstructed three-dimensional spatial morphology of the cracks on the concrete structure surface, the length, average width, maximum width, area, and volume parameters of the cracks are calculated, including: In three-dimensional space, the length of the three-dimensional crack is obtained by piecewise integration of the three-dimensional trajectory of the crack centerline and summing the results. At multiple equally spaced sampling points along the center line of the crack, the corresponding crack width values ​​are read, and the arithmetic mean of the crack width values ​​is calculated as the average width. Compare the crack width values ​​at all sampling points and find the maximum value as the maximum width; The surface contour of the three-dimensional crack is triangularized and the sum of the areas of all triangular meshes is calculated as the approximate area of ​​the crack. Using the three-dimensional crack surface profile as the boundary, calculate the volume of the spatial region enclosed by the crack and the concrete structure surface, and use it as the approximate volume of the crack.

9. The method for automatic identification of surface cracks in concrete structures based on computer vision as described in claim 1, characterized in that, The method also includes integrating the vector lines, three-dimensional spatial morphology, and geometric parameters of all cracks to generate a structured crack identification report. Specific steps include: Assign a globally unique identifier to each identified crack; The two-dimensional vector line data, three-dimensional spatial morphology mesh or point cloud data, and the calculated length, average width, maximum width, area and volume parameters of each crack are associated and stored with the crack identifier. Sort all crack data in descending order of crack length; The sorted crack data, including identifiers, various geometric parameters, and data indexes of three-dimensional spatial morphology, are filled into a preset report template to generate a structured crack identification report containing text and data indexes.

10. The method for automatic identification of surface cracks in concrete structures based on computer vision as described in claim 2, characterized in that, The uneven illumination regions in the multi-view high-resolution original image are detected, and a color constancy algorithm based on the retina model is used to independently compensate each color channel, including: For each color channel of the multi-view high-resolution original image, the local contrast of each pixel is calculated through a center-surround mechanism, and continuous areas with local contrast below a preset contrast threshold are marked as suspected uneven illumination areas. Within the suspected uneven lighting area, the standard deviation of the pixel values ​​within the suspected uneven lighting area is calculated. If the standard deviation is lower than a preset variance threshold, the suspected uneven lighting area is finally confirmed as an uneven lighting area. For each confirmed area of ​​uneven illumination, the pixel value distribution of the red, green, and blue color channels of the multi-view high-resolution original image within the corresponding area is extracted. Based on the response characteristics of photoreceptor cells in the retinal model, a gain function related to the brightness distribution of surrounding pixels is independently constructed for each color channel; The gain function is applied to the pixel values ​​of the corresponding color channel in the uneven illumination area, and the pixel values ​​are non-linearly adjusted so that the statistical distribution of each pixel value in the adjusted area is consistent with the global statistical distribution trend of the color channels in the whole image, thereby achieving independent compensation for each color channel.

Images