Concrete crack precision detection system and method based on YOLO algorithm

Abstract

The application discloses a concrete crack precise detection system and method based on a YOLO algorithm, belongs to the technical field of intelligent detection, and aims to solve the problems of low efficiency, high false and missed detection, fixed unreasonable collection period and inability to timely capture crack evolution abnormalities of traditional coastal concrete crack inspection. A collection period is configured, a target detection object surface image is collected, and a surface global image is generated; after each global image is labeled, an image set is constructed, suspected crack abnormal points are extracted through multi-feature fusion, abnormal point clusters are identified in combination with density clustering, and potential abnormal areas are adaptively divided; a pre-trained YOLO algorithm model is used for preliminary inspection of the potential abnormal areas, real cracks are identified, and quantitative parameters are calculated; the collection period is dynamically corrected based on the quantitative parameters, crack timing changes are analyzed, evolution characteristics are extracted, and abnormal early warning is completed. The overall realization of the precise and efficient detection of the coastal concrete cracks and the whole-cycle dynamic management and control provides reliable support for structural safety operation and maintenance.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection technology, and more specifically to a precise detection system and method for concrete cracks based on the YOLO algorithm. Background Technology

[0002] In the safe operation and maintenance management of concrete engineering structures, crack detection is a core link in controlling the service status of the structure and preventing safety hazards. This is especially true for concrete engineering structures in coastal settings such as cross-sea bridges, port terminals, and undersea tunnels, which are subjected to harsh service environments of salt spray erosion, tidal alternation, and high humidity and corrosion. The concrete matrix is ​​prone to crack initiation, and the crack evolution rate is rapid. If crack detection and anomaly identification cannot be completed accurately and promptly, cracks can easily continue to expand, affecting the overall load-bearing capacity of the structure and causing various engineering safety problems. Currently, in the field of concrete crack detection, existing detection methods generally suffer from insufficient accuracy. Affected by interference such as rough textures on the concrete surface, salt spots, and water stains, it is difficult to effectively distinguish between actual crack areas and interference, leading to false positives and false negatives, and failing to provide accurate crack data support for structural safety assessments. Meanwhile, existing inspection and data collection methods mostly use fixed time intervals, failing to differentiate them based on the actual evolution of cracks in different parts of the concrete structure, the structural importance level, and the dynamic changes in the service environment. This easily leads to insufficient inspection frequency for high-risk crack areas and wasted inspection resources for low-risk, crack-free areas, failing to achieve precise matching between the inspection cycle and the crack evolution status. Furthermore, existing detection processes cannot promptly identify abnormal evolution trends such as rapid crack expansion and the concentrated appearance of new cracks. Cracks are often only discovered after they have developed to the point of affecting structural safety, missing the optimal opportunity for operation and maintenance, and failing to meet the accuracy, timeliness, and dynamic requirements of crack detection for concrete engineering structures, especially those in harsh coastal environments. Therefore, to overcome these limitations, this invention proposes a precise concrete crack detection system and method based on the YOLO algorithm. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the present invention aims to provide a precise concrete crack detection system and method based on the YOLO algorithm, which solves the problems of insufficient detection accuracy, inability to dynamically adapt the inspection and acquisition cycle to the crack evolution state, and difficulty in timely identifying abnormal crack evolution trends during concrete crack detection.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A precise concrete crack detection method based on the YOLO algorithm includes:

[0006] Configure the acquisition cycle, acquire and generate a global image of the surface of the target object;

[0007] Each global surface image is labeled with a collection timestamp and a target detection object ID to construct a set of surface images. Suspected crack anomaly points in the global surface images are identified by multi-feature fusion, and clusters of suspected crack anomalies are extracted based on density clustering to identify clusters of suspected crack anomalies in order to adaptively divide the potential abnormal regions of the global surface images.

[0008] For potential anomaly regions, a pre-trained YOLO algorithm model is used to perform initial crack detection, outputting crack candidate boxes, corresponding confidence scores, and class probabilities within each potential anomaly region. This determines whether cracks exist in the potential anomaly region, identifies real crack regions, and calculates crack quantization parameters for real crack regions.

[0009] Based on the crack quantization parameters of each real crack region in the global image of the surface of the target object, the acquisition period is dynamically adjusted, and the time-series dynamic change analysis of the real crack region in each global image of the surface in the surface image set is performed to extract the evolution characteristic index of each crack quantization parameter and provide anomaly warning for the target object.

[0010] Specifically, the steps for identifying a set of suspected crack anomalies in a global surface image include:

[0011] Adaptive histogram equalization preprocessing is performed on the global surface image, and a dynamic gradient threshold is set based on the global gray mean and standard deviation of the global surface image to extract edge points where pixel gray abrupt changes occur in the global surface image.

[0012] For each edge point, a neighborhood window is set, and the texture entropy value of each edge point is calculated;

[0013] The texture entropy values ​​of all edge points in the global surface image are statistically analyzed to obtain the global average texture entropy value. A significant screening threshold is generated based on the global average texture entropy value and a preset scaling factor.

[0014] The texture entropy value of each edge point is compared with the significant screening threshold, and edge points with texture entropy values ​​less than or equal to the significant screening threshold are retained as candidate outliers.

[0015] Perform Hough transform on each candidate outlier point to construct a polar coordinate parameter space. Map the coordinates of each candidate outlier point to the polar coordinate parameter space and perform cumulative voting to filter out the line parameters in the polar coordinate parameter space that have more votes than the voting threshold.

[0016] The straight line parameter is back-projected onto the global image space of the surface, and candidate anomaly points whose deviation values ​​from the corresponding straight line parameter are within a preset deviation range are retained to form a linear candidate point set;

[0017] For the remaining candidate outliers that were not included in the linear candidate point set, perform eight-neighborhood connectivity analysis to divide the connected regions, extract candidate outliers within each connected region, and form a nonlinear candidate point set.

[0018] The coordinate information of candidate anomalies is merged from the linear candidate point set and the nonlinear candidate point set to form a set of suspected crack anomalies.

[0019] Specifically, the steps for adaptively segmenting potential anomalous regions in a global surface image include:

[0020] For a set of suspected crack anomalies, set a clustering neighborhood radius and a minimum number of points threshold;

[0021] Traverse each suspected crack anomaly point and perform density clustering; search for other suspected crack anomaly points within its clustering neighborhood radius, count the number of points within the clustering neighborhood radius, and when the number of points is greater than or equal to the minimum number of points threshold, divide the suspected crack anomaly point and the suspected crack anomaly points within the clustering neighborhood radius that meet the conditions into the same initial cluster, until all suspected crack anomaly points are clustered and the initial cluster set is obtained.

[0022] For each initial cluster in the initial cluster set, calculate the geometric parameters of the initial cluster, including elongation and curvature;

[0023] The initial clusters were classified into linear clusters, curved clusters, bifurcation clusters and irregular clusters based on geometric parameters.

[0024] An adaptive region partitioning operation is performed for initial clusters of different types to generate potential abnormal regions for each initial cluster.

[0025] Redundancy processing is performed on all the divided potential anomaly regions, the overlap rate of adjacent potential anomaly regions is calculated, and potential anomaly regions are merged; a minimum region area threshold is set to remove potential anomaly regions with an area smaller than the minimum region area threshold; thus forming a set of potential anomaly regions for the global surface image.

[0026] Specifically, the steps for generating potential anomalous regions for each initial cluster include:

[0027] For linear clusters, based on their bounding rectangle, a preset number of pixels are extended at both ends along the longest axis and a preset number of pixels are extended bidirectionally along the shortest axis to form a rectangular partition region, which serves as their potential abnormal region.

[0028] For curved clusters, the Douglas-Puk algorithm is used to fit polygons to the edge points of the curved clusters to obtain the minimum bounding polygon. Then, the polygon is expanded outward by a preset number of pixels to form a polygon division region, which serves as its potential anomaly region.

[0029] For bifurcated clusters, first perform minimum bounding polygon fitting on the main cluster and each sub-cluster of the bifurcated cluster, then calculate the intersection region of each polygon, and merge the polygon regions of the main cluster and the sub-cluster into a unified polygon partition region as its potential anomaly region.

[0030] For irregular clusters, calculate the minimum circumcircle of the initial cluster, expand the radius of the circumcircle by a preset number of pixels to form a circular or elliptical division region, which serves as its potential abnormal region.

[0031] Specifically, the steps to determine whether cracks exist in potential anomaly areas and to identify actual crack areas include:

[0032] The pre-trained YOLO algorithm model is invoked, and the potential anomaly region is taken as input. A single forward propagation calculation is performed, and the head detection output network of the YOLO algorithm model outputs the crack candidate box, confidence score and category probability corresponding to each suspected target in the potential anomaly region.

[0033] The confidence and class probability of the crack candidate boxes corresponding to each suspected target in the potential anomaly region output by the YOLO algorithm model are compared with the set confidence threshold and class probability threshold, respectively, and crack candidate boxes with confidence greater than or equal to the confidence threshold and class probability greater than or equal to the class probability threshold are initially screened.

[0034] For each crack candidate box after initial screening, the visual features within its coverage area are extracted, normalized to the [0,1] interval, and then fused into a feature vector with uniform dimension; the cosine similarity between the crack candidate box feature vector and the standard feature vectors of various interferences in the surface interference feature library is calculated; crack candidate boxes with a cosine similarity greater than or equal to the similarity threshold are removed.

[0035] The number of crack candidate boxes remaining in each potential anomaly region is counted. If the number of crack candidate boxes is greater than or equal to 1, it is determined that there is a crack in the potential anomaly region, and the crack candidate box is marked as a real crack region.

[0036] Specifically, the steps for calculating the crack quantization parameters of the actual crack region include:

[0037] The mapping ratio parameter between the pixel size of the global surface image and the actual physical size of the target object is retrieved, and the coordinate information of the crack candidate box corresponding to the real crack region is extracted; the crack length and width at the pixel scale are calculated for different morphological types.

[0038] Based on the mapping ratio parameter, the calculated crack length and width at the pixel scale are converted into actual physical crack quantization parameters to obtain the actual crack length and width; and the center pixel coordinates of the real crack region are calculated and converted into center physical coordinates through the mapping ratio parameter.

[0039] By integrating the actual length, width, and center physical coordinates of cracks in real crack regions, a structured crack quantification parameter dataset is formed.

[0040] Specifically, the steps for dynamically correcting the acquisition cycle include:

[0041] The structured crack quantization parameter dataset of each surface global image in the current target detection object surface image set is obtained, and the structured crack quantization parameter dataset is split according to the structural part identifier of the target detection object. The crack quantization parameters of the current real crack region are extracted for each structural part identifier; and the crack degree level of the real crack region is divided by multi-dimensional hierarchical classification.

[0042] Synchronously retrieve the structured crack quantification parameter data of the corresponding structural part identifier in the previous acquisition cycle and match it with the current acquisition cycle to construct a time-series comparison dataset by part;

[0043] Based on the time-series comparison dataset of different parts, the number of identified real crack areas in each structural part is increased between the current collection cycle and the previous collection cycle. At the same time, according to the crack severity level, the number of real crack areas under each crack severity level is calculated to determine whether the collection cycle update is triggered.

[0044] If so, the correction coefficient of the acquisition cycle is quantified by a multi-dimensional weighted normalization algorithm based on the increase in the total number of real crack areas identified in each structural part and the increase in the number of crack severity levels. This correction coefficient is then multiplied by the current acquisition cycle and rounded to obtain the updated acquisition cycle.

[0045] Specifically, the steps for performing time-series dynamic change analysis on the real crack regions of each surface global image in the surface image set and extracting the evolution characteristic indicators of each crack quantification parameter include:

[0046] Using the structural reference point of the target object as the spatial positioning origin, and combining the central physical coordinates of each real crack region, a unified spatial coordinate system is constructed. Spatial alignment is performed on the real crack regions of all time-series global surface images in the surface image set to generate a time-series subset of each reference crack region of the current acquisition period's global surface image.

[0047] Each temporal subset uses the corresponding baseline crack region as the core identifier, integrates the crack quantization parameters of the matching regions in all acquisition cycles, and the acquisition timestamps of each acquisition cycle; the crack quantization parameters in each temporal subset are subjected to acquisition cycle standardization processing, and a standardized crack quantization parameter sequence is generated through temporal smoothing;

[0048] Based on the standardized crack quantification parameter sequence, each crack quantification parameter is continuously fitted to generate the evolution trend curve of each crack quantification parameter over time, and the evolution characteristic index is extracted.

[0049] Specifically, the steps for generating temporal subsets of each reference crack region in the current acquisition cycle of the global surface image include:

[0050] Extract the center physical coordinates of each real crack region in all time-series global surface images in the surface image set, and map the center physical coordinates of each real crack region in time series to a unified spatial coordinate system.

[0051] Using the real crack region of the current acquisition cycle as the reference crack region, in a unified spatial coordinate system, a spatial distance matching threshold is set for each reference crack region. All real crack regions in each acquisition cycle are traversed, and the spatial Euclidean distance between their center physical coordinates and the center physical coordinates of each reference crack region is calculated. Coarse spatial matching of the reference crack regions is performed, and potential matching regions of the reference crack regions are marked.

[0052] Feature verification is performed on the potential matching regions obtained by coarse matching. Morphological feature parameters of the baseline crack region and each potential matching region are extracted. The similarity of morphological feature parameters between the baseline crack region and the potential matching regions is calculated. Spatial fine matching of the baseline crack region is performed, and the matching region of the baseline crack region is selected.

[0053] The matching regions confirmed in each acquisition cycle are associated and marked with the reference crack regions to generate a temporal subset of each real crack region in the current global surface image.

[0054] The concrete crack detection system based on the YOLO algorithm includes an image acquisition module, an abnormal region segmentation module, a crack detection module, a dynamic correction module, and a dynamic analysis module.

[0055] The image acquisition module is used to configure the acquisition cycle, acquire and generate a global image of the surface of the target object;

[0056] The abnormal region segmentation module is used to construct a set of surface images. It identifies a set of suspected crack anomalies in the global surface image through multi-feature fusion, and extracts clusters of suspected crack anomalies based on density clustering to identify clusters of suspected crack anomalies in order to adaptively segment potential abnormal regions in the global surface image.

[0057] The crack detection module performs initial crack detection on potential anomaly areas, outputs crack candidate boxes, corresponding confidence scores and category probabilities for each potential anomaly area, determines whether cracks exist in the potential anomaly area, identifies real crack areas and calculates crack quantification parameters.

[0058] The dynamic correction module is used to dynamically correct the acquisition cycle based on the crack quantization parameters of each real crack region in the global image of the target object surface.

[0059] The dynamic analysis module performs time-series dynamic change analysis on the real crack regions of each surface global image in the surface image set, extracts the evolution characteristic index of each crack quantification parameter, and provides anomaly warning for the target detected object.

[0060] The beneficial effects of this invention are:

[0061] This application identifies suspected crack anomaly sets through multi-feature fusion, extracts anomaly clusters using density clustering, and adaptively divides potential anomaly regions, effectively distinguishing between surface disturbances and crack features, reducing false positives and false negatives, and defining the focus area for subsequent detection. A pre-trained YOLO algorithm model is used for initial detection of potential anomaly regions, followed by double-threshold screening and cosine similarity-based secondary elimination to accurately identify real cracks. The actual quantitative parameters of different crack morphologies are calculated and integrated using mapping ratios to form a structured dataset, providing accurate data support for subsequent steps. Based on structured parameters, data is split and crack levels are classified. A time-series comparison set is constructed using data from the previous cycle, dynamically adjusting the collection cycle to achieve reasonable allocation of inspection resources. A unified coordinate system is constructed using structural reference points, spatial matching of real crack regions in the time series is performed, and time-series subsets are generated. After parameter standardization, evolution curves are fitted, features are extracted, and early warnings are issued, accurately capturing crack evolution characteristics and identifying potential safety hazards in advance. Overall, the accuracy, timeliness, and dynamic adaptability of the entire crack detection process are improved, providing comprehensive dynamic basis for the safe operation and maintenance of concrete structures. Attached Figure Description

[0062] Figure 1 This is a flowchart of the concrete crack detection method based on the YOLO algorithm of the present invention;

[0063] Figure 2 This is a flowchart illustrating the process of identifying a set of suspected crack anomalies in a global surface image according to the present invention.

[0064] Figure 3 A flowchart for generating potential anomalous regions for each initial cluster in this invention;

[0065] Figure 4 This is a flowchart for determining whether a potential abnormal area has cracks and identifying the actual crack area. Detailed Implementation

[0066] Please see Figures 1-4 This embodiment introduces a method for accurate detection of concrete cracks based on the YOLO algorithm, including:

[0067] Step S1: For the target object, configure the acquisition cycle, and use a drone equipped with a vision acquisition device to perform adaptive path planning. Following the predicted survey line coverage rules, acquire surface images of the target object, and use an image stitching algorithm to generate a global surface image of the target object. During the stitching process, feature point matching and distortion correction techniques are used to avoid crack breaks and pixel stretching distortion, ensuring the integrity and authenticity of cracks in the stitched global image, providing high-quality image data support for subsequent identification of potential crack areas.

[0068] In this embodiment, the target detection object refers to concrete engineering structures in coastal scenarios, including but not limited to bridge piers of cross-sea bridges, port and wharf foundations, undersea tunnel linings, concrete components of coastal transportation infrastructure, and concrete structures of coastal protection projects. The data collection period refers to the time interval during which the UAV performs concrete crack inspection and data collection operations on the target detection object, set based on the structural importance level of the target detection object, the severity of the coastal service environment, historical crack monitoring data, and crack risk assessment results. For example, the core load-bearing structure of a cross-sea bridge main pier or the main lining of an undersea tunnel has a data collection period of one month; the important structures of a port and wharf bearing foundation or the main structure of a coastal protection project have a data collection period of two months; and the core load-bearing structure of a coastal transportation infrastructure project has a data collection period of two months. The auxiliary structures of the construction of concrete components are collected over a period of 3 months. Due to the characteristics of coastal engineering structures, such as large size, wide distribution, and harsh service environment, they are in an extreme environment of high humidity, salt spray erosion, and alternating tides for a long time. In addition, some key structural parts are located at high altitudes, near water, or in concealed areas. Manual inspection has problems such as limited working space, high safety risks, incomplete coverage, and low efficiency. Therefore, UAVs equipped with high-definition visual acquisition equipment are used in combination with adaptive path planning to achieve no blind spots, long distance, safe and efficient acquisition of the surface of the target object. This ensures that the acquired single original image is clear and free from obvious environmental interference. Furthermore, the image stitching technology is used to form a complete global surface image, which fully preserves the spatial distribution and morphological characteristics of cracks. The survey line coverage rule is an execution criterion used by UAVs equipped with visual acquisition devices to plan inspection trajectories and set the spacing between adjacent acquisition trajectories when performing target detection surface image acquisition. Its core objective is to ensure that adjacent acquired images have an effective overlap rate that meets image stitching requirements, while ensuring that cracks corresponding to the predicted scale on the target object's surface can be completely and without omission, avoiding missed crack detections or redundant acquisition resources due to improper survey line spacing settings. This rule uses the minimum predicted crack width as the core input parameter to establish a negative correlation mapping relationship between crack width and survey line spacing. The mapping relationship is set based on the imaging resolution of the visual acquisition device, the effective field of view of a single acquired image, and adapts to the detection requirements of concrete cracks in coastal scenarios. It ensures that the effective field of view of a single image can completely cover the extension range of the corresponding crack width, and that the overlapping area of ​​adjacent survey lines can cover the potential extension direction of the crack, ensuring the integrity and continuity of crack features after subsequent image stitching. For example, for microcracks with an estimated width of less than 0.1 mm, the test line spacing can be set to 0.2 m; for conventional cracks with an estimated width of 0.1 mm to 5 mm, the test line spacing can be set to 1 m.

[0069] Step S2: Each global surface image is labeled according to its acquisition timestamp and the identifier ID of the target object to construct a set of surface images of the target object. The set of suspected crack anomalies in the global surface image is identified by multi-feature fusion. The cluster of suspected crack anomalies is extracted based on density clustering. The spatial correlation of suspected crack anomalies is explored. Clusters of suspected crack anomalies that conform to crack performance characteristics are extracted. The random shape of the crack is adapted. Based on the distribution pattern of the cluster of suspected crack anomalies, the potential abnormal regions of the global surface image are adaptively divided.

[0070] In this embodiment, the coastal environment with high humidity, salt spray, and alternating tides results in low contrast of the acquired surface images. Furthermore, the rough texture of the concrete surface, salt spots, water stains, and other interfering elements are easily confused with crack features, making single-feature recognition methods prone to false detections and missed detections. Moreover, crack morphology is highly random, encompassing various types such as linear, curved, bifurcated, short and thick, and mesh-like patterns. Traditional fixed windowing or single-morphology segmentation methods cannot completely cover cracks of different shapes, easily leading to crack fragmentation or excessively large segmentation ranges that introduce interference. Therefore, this step first constructs a surface image set by collecting timestamps, target detection object IDs, and environmental parameter annotations, providing data support for subsequent tracing, analysis, and error correction. Then, multi-feature fusion recognition is employed, integrating grayscale gradient, texture consistency, and linear continuity features to effectively distinguish cracks from interfering elements, improving the accuracy of identifying suspected crack anomalies. Subsequently, improved density clustering is used to mine the spatial correlation of anomalies, adapting to the linear extension and random shape features of cracks, and filtering effective crack clusters. Finally, potential anomaly regions are adaptively segmented according to cluster morphology, ensuring complete coverage of cracks of various shapes while narrowing the subsequent discrimination range to reduce unnecessary computational load. It achieves accurate identification and complete coverage of cracks of various random shapes in complex coastal environments, significantly reducing the false detection rate of interference objects. It lays a high-quality and focused data foundation for subsequent accurate crack identification based on the YOLO algorithm, while ensuring the traceability and repeatability of the detection process.

[0071] Preferably, the specific steps for identifying a set of suspected crack anomalies in a global surface image include:

[0072] An adaptive histogram equalization preprocessing is performed on the global surface image to enhance the grayscale difference between cracks and the concrete matrix in the coastal environment. Then, an edge detection algorithm is employed. Based on the global grayscale mean μ and standard deviation σ of the global surface image, a dynamic gradient threshold is set. The dynamic gradient threshold is set based on the global grayscale mean, adaptively adjusting the threshold boundary according to the dispersion of the overall grayscale distribution of the image. This ensures effective capture of weak grayscale abrupt edges between cracks and the concrete matrix while suppressing false edge interference from rough textures and salt spots on the concrete surface. The low threshold is used for initial edge screening, fully preserving all potential grayscale abrupt regions, while the high threshold is used to accurately identify strong edge features and filter out invalid interference areas. The two work together to achieve efficient and accurate edge point extraction. For example, the low threshold is set to μ-1.5σ, and the high threshold is set to μ+2.5σ. Based on the dynamic gradient threshold, edge points with pixel grayscale abrupt changes in the global surface image are extracted.

[0073] For each edge point in the aforementioned set of edge points, a neighborhood window of a fixed size is set, and the texture entropy value of each edge point is calculated. The actual physical coverage of the neighborhood window must completely cover the target area corresponding to the estimated minimum crack scale, while avoiding the introduction of too many irrelevant background pixels. The pixel size of the neighborhood window is determined by the resolution of the global surface image. For the same estimated minimum crack scale, the higher the image resolution, the larger the required neighborhood window pixel size. For the same image resolution, the larger the estimated minimum crack scale, the larger the required neighborhood window pixel size, thus ensuring that the texture entropy value calculation result can accurately and stably represent the texture features of the region where the edge point is located. For example, for the core detection requirement of estimating a minimum crack width of 0.1mm in the coastal scene of this solution, a 3×3 pixel neighborhood window is used. The texture entropy value is a quantitative indicator of the degree of disorder in the gray-level distribution within the neighborhood window of an edge point. By statistically analyzing the occurrence frequency of each gray level within the neighborhood window, the occurrence probability of each gray level is calculated. Then, the entropy value calculation formula is used to perform a weighted summation operation on the probabilities of each gray level to obtain the texture entropy value of each edge point.

[0074] The texture entropy values ​​of all edge points in the global surface image are statistically analyzed to calculate the global average texture entropy value. A significant screening threshold is generated based on this global average texture entropy value and a preset scaling factor. The texture entropy value of each edge point is compared with the significant screening threshold. Edge points with texture entropy values ​​less than or equal to the significant screening threshold are retained as candidate anomalies; edge points with texture entropy values ​​greater than the significant screening threshold are discarded. The scaling factor is a core hyperparameter used to adjust the sensitivity of the significant screening threshold. It is configured by combining coastal environmental parameters such as salt spray concentration and light intensity corresponding to the global surface image, as well as the statistical results of the texture entropy difference between interfering objects and cracks in historical detection data, so that the significant screening threshold can adapt to the interference suppression requirements of different coastal environments. For example, the scaling factor can be set to 0.6, which means multiplying the global average texture entropy value by this scaling factor to obtain the significant screening threshold used to distinguish cracks from interfering objects.

[0075] A Hough transform is performed on each candidate anomaly point to construct a polar coordinate parameter space. The coordinates of each candidate anomaly point are mapped to the polar coordinate parameter space and accumulated for voting. A voting threshold is set, and line parameters in the polar coordinate parameter space that exceed the voting threshold are selected. The line parameters refer to the core algorithm parameters describing the line in polar coordinates, specifically including the polar angle θ and the polar radius ρ, where the polar angle θ is the angle between the line and the positive direction of the horizontal coordinate axis of the global surface image, and the polar radius ρ is the vertical distance from the origin of the global surface image to the line. The line parameters are back-projected onto the global surface image space to determine the candidate anomalies in the global surface image that conform to the line parameters. The deviation values ​​between each candidate anomaly point and the corresponding line parameters are recorded, and candidate anomalies with deviation values ​​within the preset deviation range are retained to form a linear candidate point set. Cumulative voting refers to counting the number of candidate anomalies mapped to each grid cell in the polar coordinate parameter space. Each candidate anomaly corresponds to one grid cell, and this counting process is called cumulative voting. The voting threshold is a quantitative standard used to screen valid line parameters. For example, it is set by multiplying the median of the voting distribution of all grid cells in the polar coordinate parameter space by 1.5. The preset deviation range is used to determine whether a candidate anomaly meets the allowable deviation range of the corresponding line parameter. First, the global surface map is determined by the calibration plate and the parameters of the acquisition device. The actual physical size corresponding to a single pixel is used as the maximum allowable physical deviation, which is 1.2 times the estimated minimum crack width. This ratio is adapted to the slight offset characteristics of crack edges caused by salt spray erosion and water stain blurring in coastal scenes. Finally, this physical deviation is converted into pixel deviation in the image coordinate system according to the correspondence between pixels and physical size. This determines the upper and lower limits of the preset deviation range. For example, by calculating the mapping ratio between the pixel size of the global surface image and the actual physical size, and combining it with 1.2 times the estimated minimum crack width to convert into a pixel deviation value, the upper and lower limits of the preset deviation range are determined.

[0076] For the remaining candidate outliers that are not included in the linear candidate point set, perform eight-neighborhood connectivity analysis, initialize the connected region labeling matrix, traverse each unlabeled remaining candidate outlier, and search for all unlabeled candidate outliers in its eight neighborhoods starting from that remaining candidate outlier. Divide these remaining candidate outliers into the same connected region, repeat this process until all remaining candidate outliers are labeled, extract the candidate outliers in each connected region, and form a nonlinear candidate point set.

[0077] The coordinate information of candidate anomalies from the linear and nonlinear candidate point sets is merged, and isolated candidate anomalies are eliminated. Isolated candidate anomalies are defined as points that have no neighboring candidate anomalies and are not included in any straight line fitting or connected point clusters. Finally, a set of suspected crack anomalies that conforms to the crack distribution characteristics is formed.

[0078] Preferably, the specific steps for adaptively segmenting potential anomalous regions in a global surface image include:

[0079] For a set of suspected crack anomalies, a clustering neighborhood radius is set based on the mapping ratio between pixel size and actual physical size of the global surface image, combined with the estimated crack width range. The estimated crack width range is a crack width interval determined based on the service life of the coastal concrete structure, the degree of environmental erosion, and historical data. The lower limit is the estimated minimum crack width, and the upper limit is the maximum crack width allowed by structural safety. The clustering neighborhood radius is taken as three times the median of the estimated crack width range, and then converted to pixel value according to the correspondence between pixels and physical size. This ratio can completely cover the edge area of ​​the crack's lateral extension, avoiding the splitting of anomalies of the same crack into multiple discrete clusters. According to the estimated crack length range, the number of pixels corresponding to the estimated minimum crack length is used as the minimum number of pixels threshold. The estimated crack length range is a crack length interval determined based on structural characteristics and control requirements. The lower limit is the minimum identifiable effective crack length, and the upper limit is the maximum crack length detected in history. The number of pixels corresponding to the estimated minimum crack length is obtained by converting the minimum crack length according to the correspondence between pixels and physical size. For each suspected crack anomaly, perform density clustering. Search for other suspected crack anomalies within the radius of its clustering neighborhood, and count the number of points within the clustering neighborhood. When the number of points is greater than or equal to the minimum number of points threshold, classify the suspected crack anomaly and other suspected crack anomalies within the clustering neighborhood that meet the condition into the same initial cluster. The condition means that the suspected crack anomaly is located within the clustering neighborhood of the core point and has not been classified into other initial clusters. Repeat this process until all suspected crack anomalies have been clustered, and the initial cluster set is obtained.

[0080] For each initial cluster in the initial cluster set, calculate the geometric parameters of the initial cluster, including the length of the longest axis, the length of the shortest axis, the ratio of the longest axis to the shortest axis of the cluster's bounding rectangle (i.e., the extensibility), and the curvature of the fitted curve of the initial cluster's edge points (i.e., the average distance between the fitted curve and the edge points of the initial cluster).

[0081] The initial clusters are classified into linear clusters, curved clusters, bifurcation clusters, and irregular clusters based on geometric parameters. For example, a linear cluster is defined as having an extension of not less than 5 and a curvature of not more than a preset curvature threshold; a curved cluster is defined as having an extension of not less than 3 and a curvature of more than a preset curvature threshold; a bifurcation cluster is defined as having two or more sub-clusters and the sub-clusters sharing three or more common anomalies with the main cluster; and an irregular cluster is defined as having an extension of less than 3 and no fixed curvature pattern. The preset bending threshold is a quantitative critical value used to determine whether a cluster belongs to a bending morphology. First, the actual physical size corresponding to a single pixel in the global image of the surface is determined by calibration. Then, the bending critical value of the physical scale is calculated at 0.8 times the estimated minimum crack width. This ratio can distinguish the edge offset critical value between linear straight cracks and bending cracks, which is suitable for the morphological recognition requirements of micro-cracks in coastal scenarios. Finally, the physical bending critical value is converted into a pixel distance value in the image coordinate system according to the correspondence between pixels and physical size, thus completing the configuration of the preset bending threshold. Sub-clusters refer to independent clusters in a bifurcated cluster other than the main cluster, which share anomalies with the main cluster. The main cluster refers to the cluster with the most anomalies in the bifurcated cluster and serves as the core connecting other sub-clusters. The determination of sub-clusters and main clusters is determined by statistically analyzing the number of anomalies in each initial cluster and the number of shared anomalies between clusters. The cluster with the most anomalies is determined as the main cluster, and other initial clusters that share three or more anomalies with the main cluster are determined as sub-clusters.

[0082] An adaptive region partitioning operation is performed for initial clusters of different types to generate potential abnormal regions for each initial cluster.

[0083] Specifically, for linear clusters, based on their circumscribed rectangle, a preset number of pixels is added at both ends along the longest axis and a preset number of pixels is added bidirectionally along the shortest axis to form a rectangular partition region, which serves as its potential anomaly region. For curved clusters, the Douglas-Puk algorithm is used to fit polygons to the edge points of the curved clusters to obtain the minimum circumscribed polygon, which is then expanded outward by a preset number of pixels to form a polygon partition region, which serves as its potential anomaly region. For bifurcated clusters, the minimum circumscribed polygon is first fitted to the main cluster and each sub-cluster of the bifurcated cluster, and then the intersection region of each polygon is calculated. The polygon regions of the main cluster and the sub-clusters are merged into a unified polygon partition region, which serves as its potential anomaly region. For irregular clusters, the minimum circumscribed circle of the initial cluster is calculated, and the radius of the circumscribed circle is expanded by a preset number of pixels to form a circular or elliptical partition region, which serves as its potential anomaly region. Among them, the preset number of pixels refers to the quantized pixel value used to expand the boundary of various cluster division areas. Its setting needs to be adapted to the possible extension range of the crack and the blurring characteristics of the crack edge in the coastal environment. It is determined by calculating the mapping ratio between the pixel size of the global surface image and the actual physical size, combined with the estimated possible extension margin of the crack, that is, the statistical mean of the crack extension length in the historical detection data. It ensures that the extension range fully covers the crack body and the potential extension part, without introducing too much irrelevant background interference.

[0084] Redundancy processing is performed on all the divided potential anomaly regions. The overlap rate of adjacent potential anomaly regions is calculated. The overlap rate is the ratio of the area of ​​the overlapping part of adjacent potential anomaly regions to the area of ​​the smaller potential anomaly region among the two potential anomaly regions. When the overlap rate is not less than a preset overlap threshold, the two potential anomaly regions are merged into one potential anomaly region. A minimum region area threshold is set. This minimum region area threshold is determined by converting the estimated minimum crack area into pixel area. Potential anomaly regions with an area smaller than the minimum region area threshold are eliminated. Finally, all effective potential anomaly regions are integrated to form a set of potential anomaly regions for the global surface image. Among them, the preset overlap threshold is a quantitative critical value for determining whether adjacent potential abnormal regions need to be merged. It is configured by combining the distribution characteristics of cracks in coastal concrete engineering structures and the statistical results of the overlap rate of the same crack being divided into multiple adjacent potential abnormal regions in historical detection data. This ensures that adjacent potential abnormal regions corresponding to the same crack can be effectively merged, while avoiding the erroneous merging of potential abnormal regions of unrelated cracks. The minimum region area threshold is a quantitative critical value for determining whether a single potential abnormal region is a valid region. It is determined by first combining the relevant specifications for crack detection in coastal concrete engineering structures to determine the estimated minimum crack area, and then calculating the mapping ratio between the pixel size of the global surface image and the actual physical size. The estimated minimum crack area is then converted into the corresponding pixel area according to the mapping ratio, thus completing the configuration of the minimum region area threshold. This ensures that invalid small regions formed by noise points and tiny interference objects can be eliminated.

[0085] Step S3: Using a YOLO algorithm model specifically pre-trained with samples of coastal concrete cracks, the potential anomaly regions obtained in Step S2 are input into the YOLO algorithm model for initial crack detection. Through a single forward propagation calculation, the crack candidate boxes, corresponding confidence scores, and class probabilities within each potential anomaly region are output simultaneously. Based on the confidence score threshold, class probability threshold, and surface interference feature library, it is determined whether cracks exist in each potential anomaly region, and the real crack regions are identified. If they exist, based on the mapping ratio between the global surface image pixel size and the actual physical size recorded in Step S1, combined with the coordinate information of the crack candidate boxes, the crack quantization parameters of each real crack region are calculated, including crack length, width, and center position coordinates. At the same time, the confidence score and class probability of each crack region are recorded. Finally, the detection results containing the location of the real crack region, crack quantization parameters, and confidence information are output, providing data support for subsequent crack secondary verification and evolution analysis.

[0086] In this embodiment, the YOLO algorithm model refers to an end-to-end real-time target detection model built on a deep convolutional neural network. It features targeted structural and parameter optimizations for coastal concrete crack detection scenarios, enabling target localization, category determination, and confidence output through a single forward propagation. The model consists of a backbone feature extraction network, a neck feature fusion network, and a head detection output network. The backbone feature extraction network extracts multi-level deep features from images of potential anomaly regions. The neck feature fusion network performs multi-scale fusion enhancement on these multi-level deep features. The head detection output network performs regression and classification calculations on the multi-scale fused and enhanced multi-level deep features, outputting crack candidate boxes, confidence scores, and category probabilities. A labeled dataset is constructed by collecting crack samples and interference samples from the surface of coastal concrete structures. After performing scene-specific data enhancement processing such as salt spray, illumination, and blur on the labeled dataset, the enhanced labeled dataset is input into the YOLO algorithm model for iterative training. The network parameters are optimized by calculating the loss function and backpropagating until the YOLO algorithm model converges. Crack candidate boxes refer to the rectangular bounding boxes output by the YOLO algorithm model after detecting potential anomaly regions in an image, used to mark the location and extent of suspected cracks in the image. These boxes include the pixel coordinates of the top-left and bottom-right corners of the bounding box and the box's dimensions. Confidence score is the probability value that the YOLO algorithm model determines the presence of a target object within the output crack candidate box, comprehensively reflecting the accuracy of the crack candidate box's location and the reliability of the presence of a target within the box. Class probability is the probability value that the YOLO algorithm model determines the target object within the crack candidate box to be a concrete crack. It is a quantitative indicator used by the YOLO algorithm model to classify the target object within the box, distinguishing between concrete cracks and surface disturbances. The surface disturbance feature library is a structured database built specifically for the service environment of coastal concrete engineering structures, used to store the visual features of various disturbances on the concrete surface. It includes the grayscale features, texture features, morphological features, and pixel distribution features of common disturbances such as salt spots, water stains, concrete rough textures, and formwork joints. These features are stored and managed in the form of feature vectors.

[0087] Preferably, the specific steps for determining whether cracks exist in each potential abnormal area include:

[0088] The pre-trained YOLO algorithm model is invoked, and the single potential anomaly region image obtained in step S2 is input into the YOLO algorithm model. A single forward propagation calculation is performed, and the head detection output network of the YOLO algorithm model outputs the crack candidate box, confidence score and category probability corresponding to each suspected target in the potential anomaly region.

[0089] Set confidence thresholds and class probability thresholds. These thresholds are configured based on the detection accuracy and recall statistics of the YOLO algorithm model on the concrete crack verification dataset during pre-training. For example, the confidence threshold is set to 0.5 and the class probability threshold is set to 0.7. Compare the confidence and class probability of the crack candidate boxes corresponding to each suspected target in the potential anomaly region output by the YOLO algorithm model with the corresponding thresholds. Only retain crack candidate boxes with a confidence greater than or equal to the confidence threshold and a class probability greater than or equal to the class probability threshold to complete the initial screening of crack candidate boxes in each potential anomaly region.

[0090] For each crack candidate box after initial screening, visual features within the coverage area of ​​the crack candidate box are extracted, including gray-level distribution histogram features, 3×3 neighborhood texture entropy features, and circumscribed rectangle extension features. After normalizing various visual features to the [0,1] interval, they are fused into a feature vector with unified dimensions.

[0091] The pre-built surface interference feature library is invoked. This library contains standard feature vectors of interferences such as salt spots and water stains. The cosine similarity between the crack candidate box feature vector and the standard feature vectors of various interferences in the surface interference feature library is calculated.

[0092] Based on the statistical results of the feature differences between coastal concrete cracks and interference objects, a similarity threshold is configured. For example, the similarity threshold is set to 0.6. The calculated cosine similarity is compared with the similarity threshold, and crack candidate boxes with cosine similarity greater than or equal to the similarity threshold are removed to complete the secondary removal of interference objects.

[0093] The number of crack candidate boxes remaining after the above screening is counted in each potential anomaly area. If the number of crack candidate boxes is greater than or equal to 1, it is determined that there is a crack in the potential anomaly area, and the crack candidate box is marked as a real crack area; otherwise, it is determined that there is no crack in the potential anomaly area.

[0094] Preferably, the specific steps for calculating the crack quantification parameters of each real crack region include:

[0095] The mapping ratio parameter between the pixel size of the global surface image and the actual physical size of the target object is retrieved. At the same time, the coordinate information of the candidate crack boxes corresponding to the real crack areas that have been screened and calibrated is extracted, including the pixel coordinates of the upper left and lower right corners of the candidate boxes.

[0096] Based on the coordinate information of the candidate crack bounding boxes in the real crack region, the pixel dimension basic parameters of the candidate crack bounding boxes are calculated, including the pixel width and pixel height of the candidate crack bounding boxes. Furthermore, the pixel-scale crack pixel length and crack pixel width are calculated for different crack morphology types. For linear cracks and bifurcated cracks, the longest axis pixel length of the bounding rectangle of the candidate box is used as the crack pixel length, and the shortest axis pixel length is used as the crack pixel width. For curved cracks and irregular cracks, polygon fitting is first performed on the edge points of the real crack region, and the pixel arc length of the fitted contour is used as the crack pixel length, and the shortest axis pixel length of the smallest bounding rectangle of the fitted contour is used as the crack pixel width.

[0097] Based on the mapping ratio parameters, the calculated pixel length and width of the crack at the pixel scale are converted into actual physical crack quantization parameters, yielding the actual crack length and width. Simultaneously, the center pixel coordinates of the actual crack region are calculated and converted into center physical coordinates using the mapping ratio parameters. A physical coordinate system is established with the structural reference point of the target object as the origin, ensuring that the center physical coordinates are accurately correlated to the actual location of the concrete structure. This guarantees that the crack quantization parameters have clear engineering positioning significance and are suitable for the detection needs of coastal concrete engineering structures.

[0098] By integrating the actual length, width, and center physical coordinates of cracks after conversion from real crack regions, a structured crack quantification parameter dataset is formed. Simultaneously, corresponding confidence levels, category probabilities, and auxiliary information such as the target detection object identifier ID and acquisition timestamp are associated. When multiple real crack candidate boxes exist within the same potential anomaly region, the crack quantification parameters of adjacent cracks are merged based on the spatial correlation of the crack center physical coordinates, eliminating duplicate statistics and providing unified data support for subsequent crack risk assessment and evolution tracking.

[0099] Step S4: Based on the crack quantification parameters of each real crack region obtained in Step S3, dynamically adjust the acquisition cycle of the UAV inspection. Shorten the acquisition cycle for structural parts with larger crack scale, faster evolution rate, or in high-risk environments, and maintain or extend the acquisition cycle for parts with stable cracks without significant changes to ensure that inspection resources accurately match the structural risk level. At the same time, retrieve the real crack region data corresponding to the global surface images at different time sequences in the surface image set, partition the target object according to the structural parts, and perform time-series dynamic change analysis on the real crack regions of each global surface image in the surface image set. Compare the increase or decrease of crack length and width and the offset of the center position to identify abnormal evolution trends such as rapid crack expansion and concentrated appearance of new cracks. Combined with the safety control standards for cracks in coastal concrete engineering structures, generate abnormal warning signals for the target object and provide data support for the structural safety assessment and operation and maintenance decision-making of the target object.

[0100] In this embodiment, the traditional crack inspection of coastal concrete structures adopts a fixed acquisition cycle setting mode, without differentiated adjustment based on the actual quantitative parameters of cracks, evolution characteristics, and risk levels of structural locations. This easily leads to insufficient inspection frequency for high-risk crack areas and wasted inspection resources for low-risk, crack-free areas. Furthermore, the harsh service environment of coastal areas with high humidity, salt spray, and alternating tides accelerates the initiation and expansion of concrete cracks, and the traditional mode cannot capture the dynamic evolution characteristics of cracks in a timely manner. At the same time, the existing detection process only completes crack detection for a single batch of images, without systematic correlation analysis of detection data from different time sequences in the surface image set. It cannot achieve full-cycle tracking of cracks at the same location, making it difficult to identify abnormal trends such as rapid crack expansion and concentrated appearance of new cracks. This can easily lead to structural safety problems due to the failure to detect crack hazards in a timely manner. Based on this, this step is designed to detect cracks... By combining quantitative parameters with structural attributes to dynamically adjust the acquisition cycle, the system achieves precise allocation of inspection resources and differentiated inspections. Simultaneously, it leverages time-series data from surface image sets to analyze the dynamic changes in quantitative parameters of cracks at the same location. Combined with specific crack safety control standards for coastal concrete structures, it identifies abnormal evolution trends and generates early warning signals. This not only solves the resource mismatch problem associated with fixed acquisition cycles and improves the targeting and efficiency of crack inspections in coastal concrete structures, but also enables full-cycle dynamic tracking of crack evolution. It can promptly capture abnormal crack changes in harsh coastal environments and issue early safety warnings, fully leveraging the value of time-series data from surface image sets to achieve closed-loop utilization of inspection data. This provides accurate and dynamic data support for safety assessments and maintenance decisions in coastal concrete structures, effectively reducing structural safety risks caused by untimely crack treatment.

[0101] Preferably, the specific steps for dynamically correcting the data collection cycle of UAV inspections include:

[0102] This process acquires a structured crack quantization parameter dataset for each surface in the current target object's surface image set. The dataset is then split according to the structural component identifiers of the target object, and crack quantization parameters for each structural component identifier in the current real crack region are extracted. The structural component identifier is a unique code assigned to each component of the target object. The coding rules are formulated based on structural type, spatial location, and load-bearing capacity, enabling precise association with inspection data, structural attributes, and environmental monitoring information for corresponding components. This allows for refined data management and comparative analysis by component, avoiding data confusion between different components. Based on the crack quantization parameters of the real crack regions, the crack severity levels of the real crack regions are classified through multi-dimensional hierarchical methods. Specifically, taking into account the actual length and width of cracks, a crack classification threshold system adapted to the structural characteristics of coastal concrete engineering is configured. This crack classification threshold system is formulated in accordance with the industry standards for crack detection in coastal concrete engineering structures, and incorporates environmental impact correction factors for coastal salt spray erosion and tidal alternation, as well as structural load-bearing level adaptation coefficients. The environmental impact correction factor is matched with corresponding values ​​after classifying erosion levels based on monitoring data such as salt spray concentration and tidal immersion duration in the target detection area. The higher the erosion level, the smaller the factor. The structural load-bearing level adaptation coefficient is matched with corresponding values ​​after classifying the core load-bearing, important load-bearing, and auxiliary structure levels based on the structural design drawings. The higher the load-bearing level, the smaller the coefficient. By differentiating the crack size thresholds for different environments and different load-bearing parts through these two methods, each real crack area is divided into different crack severity levels, forming a graded list of real crack areas.

[0103] The structured crack quantification parameter data of the corresponding structural part identifier in the previous acquisition cycle is retrieved synchronously. Through time series data alignment technology, the current crack parameters of the same structural part identifier are accurately matched with those of the previous cycle to construct a time series comparison dataset for each part.

[0104] The data collection process involves statistically analyzing the increase in the number of identified actual crack areas at each structural location compared to the previous collection period. Simultaneously, based on the classified crack severity levels, the increase in the number of actual crack areas at each level is calculated, clarifying the level to which newly added crack areas belong and the increase or decrease of higher-level crack areas. A dynamic judgment threshold is set based on the variation range of the coastal environment. This threshold is initially set based on the normal fluctuation range of crack numbers at the same location and under the same environment in historical monitoring data. A correlation is established with salt spray concentration and tidal frequency; the threshold is lowered synchronously when salt spray concentration increases or tidal influence frequency increases, allowing for adaptive adjustment according to changes in salt spray concentration and tidal influence frequency. If the total increase in the number of identified actual crack areas at any structural location exceeds the corresponding dynamic judgment threshold, or if an increase occurs in the number of higher-level crack areas, a collection period update is triggered; otherwise, the current collection period remains unchanged, ensuring that the triggering conditions adapt to the dynamic changes in cracks under the coastal environment.

[0105] If a data collection cycle update is triggered, the correction coefficient for the data collection cycle is quantified using a multi-dimensional weighted normalization algorithm, based on the increase in the total number of actual crack areas and the increase in the number of crack severity levels for each structural component. This correction coefficient is then multiplied by the current data collection cycle and rounded to obtain the updated data collection cycle. The multi-dimensional weighted normalization algorithm first normalizes the increase in the total number of actual crack areas and the increase in the number of crack severity levels to eliminate dimensional differences between different dimensions. Then, it assigns differentiated weights to the increase in the number of crack severity levels based on the structural characteristics of coastal concrete engineering, assigning higher weights to higher-level crack areas. Simultaneously, it incorporates the load-bearing capacity of the structural component and the change in the severity of the coastal environment between the two data collection cycles as weight correction factors. The weight proportion is increased for core load-bearing structural components and areas where the coastal environment shows a deteriorating trend. The correction coefficient is obtained by multiplying the standardized incremental indicators by their corresponding weights and summing the results. The correction coefficient is negatively correlated with the overall risk level of crack evolution, ensuring that the higher the risk, the greater the reduction in the data collection cycle.

[0106] Preferably, the step of performing time-series dynamic change analysis on the real crack regions of each surface global image in the surface image set includes:

[0107] Using the structural reference point of the target object as the spatial positioning origin, the structural reference point refers to a fixed reference point with spatial uniqueness and physical stability pre-set for the target object. The selection principle is to choose non-core load-bearing parts that are not easily affected by environmental factors such as coastal salt spray erosion, tidal impact, and temperature and humidity changes. It can serve as a reference for unified spatial coordinate positioning, providing a consistent spatial reference for surface global images acquired at different time series and perspectives, ensuring the comparability and accuracy of spatial location data for various types of cracks. For example, for cross-sea bridge piers, pre-embedded metal positioning markers are selected at the junction of the pier body and foundation; for port and wharf piers, fixed markers are selected from the anti-corrosion treated concrete etching at the corners of the pier; and for subsea tunnel linings, standardized pre-embedded positioning markers are selected at the lining ring joints. A unified spatial coordinate system is constructed by combining the central physical coordinates of each real crack region. Spatial alignment is performed on the real crack regions of all time-series surface global images in the surface image set, generating time-series subsets of each reference crack region in the current acquisition period's surface global images, specifically including:

[0108] Extract the center physical coordinates of each real crack region in all time-series global surface images in the surface image set, and map the center physical coordinates of each real crack region in time series to a unified spatial coordinate system. This achieves a unified transformation of the spatial coordinates of all crack regions in time series, breaks down the barriers between independent coordinate systems of different time series images, and ensures that the location data of all crack regions are in the same reference dimension.

[0109] Using the real crack region of the current acquisition cycle as the benchmark crack region, a spatial distance matching threshold is set for each benchmark crack region in a unified spatial coordinate system. This spatial distance matching threshold is set by combining the structural scale of the target object, the crack detection accuracy, and the actual error range of the coastal patrol. Specifically, the maximum positional deviation allowed by the crack detection accuracy is taken as the basic component, and the maximum cumulative value of various errors such as UAV patrol positioning, image stitching distortion, and coordinate transformation is superimposed as the error correction component. Then, according to the structural scale, a corresponding proportion or fixed value is taken as the structural adaptation component. The three are added together to obtain the physical scale matching threshold, which can also be converted into a pixel scale value according to the spatial coordinate system mapping relationship. All real crack regions in each acquisition cycle are traversed, and the spatial Euclidean distance between their center physical coordinates and the center physical coordinates of each benchmark crack region is calculated. Real crack regions with spatial Euclidean distance values ​​less than or equal to the spatial distance matching threshold are marked as potential matching regions of the corresponding benchmark crack regions, completing the spatial coarse matching of the benchmark crack regions and quickly filtering out temporal crack regions with spatial correlation.

[0110] Feature verification is performed on the potential matching regions obtained by coarse matching. Morphological feature parameters of the baseline crack region and each potential matching region are extracted, including crack extension direction, circumscribed rectangle size ratio, crack morphology type, etc. The similarity of morphological feature parameters between the baseline crack region and the potential matching regions is calculated. The potential matching region with the largest similarity of morphological feature parameters is selected as the matching region of the baseline crack region. False matching regions that are spatially close but are actually different cracks are eliminated to achieve spatial fine matching of crack regions and ensure the uniqueness and accuracy of temporal matching. Specifically, the crack extension direction refers to the main extension direction of the actual crack area. Using the horizontal axis of the unified spatial coordinate system as a reference, the angle between the main axis obtained by fitting the crack contour pixels and the reference axis is quantitatively characterized, which can accurately reflect the core extension trend of the crack in space. The circumscribed rectangle size ratio refers to the ratio of the longest axis length to the shortest axis length of the smallest circumscribed rectangle of the actual crack area. It is a core quantitative indicator for characterizing the tensile extension characteristics of the crack and can intuitively reflect the elongation of the crack shape. The crack morphology type refers to the crack category based on the contour fitting characteristics, spatial distribution characteristics and structural correlation characteristics of the actual crack area. Specifically, it is divided into four categories: linear, curved, bifurcated and irregular. This is consistent with the crack morphology classification in the potential anomaly area division stage to ensure the uniformity of morphology representation throughout the entire process.

[0111] The matching regions confirmed in each acquisition cycle are associated with and marked with the baseline crack regions. For newly added real crack regions that appear in the acquisition cycle, temporal marking and spatial positioning are performed separately. Finally, the full-dimensional spatial alignment of real crack regions in all temporal global surface images is completed. Based on the spatial alignment result, the baseline crack region of each current acquisition cycle is bound to the matching region corresponding to it in all acquisition cycles, generating a temporal subset of each real crack region in the current global surface image.

[0112] Each time series subset uses the corresponding baseline crack region as its core identifier, integrates the crack quantification parameters of the matching regions in all acquisition cycles, and the acquisition timestamps of each acquisition cycle. For newly added real crack regions that are individually marked, an independent time series subset is constructed with the acquisition cycle in which it first appears as the time series starting point, ensuring the data integrity and time series correlation of each time series subset, and laying a standardized data foundation for subsequent parameter analysis.

[0113] For each time series subset, the crack quantification parameters are standardized using the acquisition cycle to eliminate the interference of differences in acquisition cycle intervals on crack evolution analysis. This involves extracting the inspection interval duration for all acquisition cycles within each time series subset and converting the absolute changes in crack quantification parameters into relative changes per unit time. Then, time series smoothing is applied to generate a standardized crack quantification parameter sequence, ensuring the accuracy of subsequent evolution trend analysis.

[0114] Based on the standardized crack quantification parameter sequence, the crack length, width, and center physical coordinate offset sequences are continuously fitted to generate evolution trend curves of each crack quantification parameter over time, intuitively presenting the crack's expansion, stabilization, and offset states. Evolutionary characteristic indicators are extracted from the evolution trend curves, including parameter evolution rate, rate change slope, evolution inflection point, and center coordinate offset direction and magnitude. Specifically, the parameter evolution rate refers to the change in the crack quantification parameters per unit time. It is the instantaneous slope of the evolution trend curve at the corresponding time node or the average slope within a specified time interval. It is the core indicator characterizing the speed of change of crack quantification parameters and can accurately reflect the evolution rhythm of cracks. The rate change slope refers to the change in the parameter evolution rate over time. It is the second derivative of the evolution trend curve and can characterize the increasing or decreasing trend of crack evolution rate. A positive slope represents an accelerating evolution rate, and a negative slope represents a slowing evolution rate. The evolution inflection point refers to the time node in the evolution trend curve where the parameter evolution rate changes significantly. It is the position where the curve curvature reaches its extreme value. Before and after the inflection point, the evolution rhythm or trend of cracks will change significantly. It is a key node for identifying abrupt changes in crack evolution state. The direction and magnitude of the center coordinate offset refer to the spatial movement characteristics of the center physical coordinates of the real crack area from the initial acquisition time sequence to the subsequent acquisition time sequence. The offset direction is represented by the angle based on the axis of the unified spatial coordinate system. The offset magnitude is the straight-line distance of the center physical coordinates from the initial position to the subsequent position under this direction, which can completely reflect the position movement characteristics of cracks in space.

[0115] Preferably, the specific steps for providing anomaly warnings for target detected objects include:

[0116] A threshold system for judging crack anomalies was constructed. This system is based on the safety control standards for cracks in coastal concrete engineering structures, and sets threshold indicators differently according to the load-bearing level of each structural part of the target object and the severity of the coastal service environment. The threshold indicators cover the threshold for single crack evolution characteristics and the threshold for regional crack distribution characteristics. The threshold for single crack evolution characteristics includes critical values ​​for crack length evolution rate, width evolution rate, center coordinate offset amplitude, and number of evolution inflection points. The threshold for regional crack distribution characteristics includes the critical values ​​for judging abnormal crack density per unit structural area, number of newly added high-level cracks, and crack cluster expansion. At the same time, for extreme coastal environments such as high salt spray concentration and high tide frequency, various thresholds are dynamically adjusted downward to ensure that the threshold system adapts to the abnormal evolution law of cracks in the coastal environment, which is different from the general crack judgment threshold mode.

[0117] The evolutionary characteristic indicators of single cracks obtained from the temporal dynamic change analysis of cracks are retrieved along with the crack distribution data of different structural regions. The evolutionary characteristic indicators of each crack temporal subset are precisely compared with the single crack evolutionary characteristic thresholds in the anomaly judgment threshold system to check for abnormal single crack evolution, such as evolution rate exceeding the critical value, center coordinate offset exceeding the threshold, or the number of evolution inflection points reaching the critical value. Simultaneously, the areas of high-level cracks appearing for the first time are marked. For each evolutionary characteristic indicator exceeding the threshold, anomaly type classification is performed to clarify the specific abnormal evolution type of the single crack, forming a single crack anomaly identification list. This ensures that all types of abnormal evolution signals at the single crack level are captured without omission, providing accurate single crack anomaly data for subsequent comprehensive early warning.

[0118] Based on the structural component identifiers of the target detected object, a cluster anomaly evolution analysis of cracks in structural component zones is conducted. This involves integrating the single-crack anomaly identification lists within each structural component zone, statistically analyzing the number, proportion, and anomaly type distribution characteristics of anomaly cracks within each zone, calculating the anomaly crack density per unit structural area, and determining whether it reaches the regional crack distribution characteristic threshold. Simultaneously, the cluster evolution trend of cracks within each zone is analyzed, identifying cluster anomaly characteristics such as the simultaneous rapid expansion of multiple cracks within the same zone, the concentrated emergence of new cracks, and the consistent spread of cracks in different evolution directions. Furthermore, the correlation of crack anomalies across zones needs to be verified to determine whether there is a trend of crack anomalies spreading from auxiliary structures to the core load-bearing structure. This results in a cluster anomaly analysis of structural component zones, compensating for the limitations of single-crack anomaly identification and capturing potential systemic crack risks in coastal concrete structures.

[0119] Anomaly risk classification is conducted by combining the anomaly level of single cracks, the degree of anomaly clustering in structural components, the importance level of structural components, and real-time coastal environmental monitoring data. Single crack anomalies are classified into three levels—mild, moderate, and severe—based on the magnitude of exceeding the threshold. Cluster anomalies are classified into three levels—local, regional, and systemic—based on density and spread trend. Differentiated weights are assigned to the importance level of structural components, the anomaly level of single cracks, the level of cluster anomalies, and the severity of the real-time coastal environment. Anomalies in core load-bearing structures and those under extreme coastal environments are given higher weights. The comprehensive risk value of each structural component is obtained through weighted calculation. Based on the comprehensive risk value, anomaly warnings are divided into three levels: Level 1 warning corresponds to systemic severe anomalies in core load-bearing structures; Level 2 warning corresponds to regional moderate anomalies in important structures; and Level 3 warning corresponds to localized mild anomalies in auxiliary structures, achieving precise classification of anomaly warnings.

[0120] Standardized, tiered anomaly early warning signals are generated. These signals include the warning level, anomaly location identifier, specific anomaly type, core data on single cracks and cluster anomalies, real-time coastal environmental causes, and corresponding handling recommendations. The signal identification rules are consistent with the identification standards of the inspection management system and the structural maintenance system. The anomaly early warning signals are simultaneously pushed to relevant maintenance management terminals. At the same time, the surface image set of the target object and the quantitative parameter dataset of structured cracks are updated in real time, supplementing information such as warning level, warning time, and anomaly area location. All data from this anomaly early warning is linked and archived with historical early warning data to form an early warning data traceability archive.

[0121] This embodiment introduces a concrete crack precision detection system based on the YOLO algorithm, including an image acquisition module, an abnormal region division module, a crack detection module, a dynamic correction module, and a dynamic analysis module;

[0122] The image acquisition module completes the initial configuration of the acquisition cycle for target detection objects in coastal scenes. Based on the preset prediction survey line coverage rules, it controls the UAV equipped with the vision acquisition device to perform adaptive path planning, realizing the acquisition of surface images of target detection objects without blind spots and at long distances. At the same time, it adopts an image stitching algorithm, and uses feature point matching and distortion correction technology during the stitching process to avoid problems such as crack breakage and pixel stretching distortion, accurately generating a global surface image of the target detection object. This ensures the integrity and authenticity of cracks in the stitched global image, providing high-quality image data support for the subsequent work of the modules.

[0123] The anomaly region segmentation module annotates each acquired global surface image with the acquisition timestamp, target object ID, and environmental parameters to construct a standardized set of surface images. Addressing the issues of low image contrast and numerous interfering objects in coastal environments, a multi-feature fusion technique is employed to identify suspected crack anomaly points in the global surface images. Then, density clustering is used to mine the spatial relationships of these suspected crack anomalies, extracting clusters of suspected crack anomalies that match crack characteristics. Finally, based on the distribution and morphological characteristics of these clusters, an adaptive segmentation operation is performed on the global surface images to generate a focused set of potential anomaly regions. This significantly reduces the false detection rate of interfering objects, narrowing the discrimination range and reducing the computational load for the precise crack detection module.

[0124] The crack detection module is equipped with a YOLO algorithm model specifically pre-trained on samples of coastal concrete cracks. Potential anomaly images are input into the model for a single forward propagation calculation, simultaneously outputting crack candidate boxes, confidence scores, and class probabilities. Initial screening of crack candidate boxes is performed by setting confidence and class probability thresholds. A pre-built surface interference feature library is then used to further eliminate these candidate boxes, accurately identifying real crack areas. Based on the mapping ratio between the pixel size of the global surface image and the actual physical size of the target object, combined with the coordinate information of the crack candidate boxes corresponding to the real crack areas, crack quantification parameters such as actual crack length, actual width, and center physical coordinates are calculated. All parameters, along with associated confidence scores, class probabilities, and acquisition timestamps, are integrated to generate a structured crack quantification parameter dataset, providing standardized data support for subsequent modules such as acquisition cycle correction and crack evolution analysis.

[0125] The dynamic correction module retrieves the structured crack quantification parameter dataset, splits the data according to the structural parts of the target object, extracts the quantification parameters of the actual crack areas corresponding to each structural part, and classifies the crack severity levels. Simultaneously, it retrieves the corresponding data from the previous acquisition cycle, constructs a segmented time-series comparison dataset using time-series data alignment technology, and statistically analyzes the increase in the total number of cracks under each structural part and the increase in the number of cracks at each level within adjacent acquisition cycles. Combined with a dynamic judgment threshold set by the change amplitude of the coastal environment, it determines whether to trigger the acquisition cycle update operation. If an update is triggered, it generates an acquisition cycle correction coefficient based on the increase in the number of cracks using a multi-dimensional weighted normalization algorithm. The correction coefficient is multiplied by the current acquisition cycle and rounded to obtain the updated acquisition cycle. This achieves dynamic adaptation between the inspection acquisition cycle and the crack risk level of each structural part of the target object, ensuring accurate allocation of inspection resources.

[0126] The dynamic analysis module uses the structural reference point of the target object as the spatial positioning origin and constructs a unified spatial coordinate system by combining the central physical coordinates of each real crack region. It performs coarse and fine spatial matching operations on the real crack regions of all time-series global surface images in the surface image set, completing full-dimensional spatial alignment and generating a unique time-series subset for each reference crack region. At the same time, it constructs an independent time-series subset for newly added real crack regions. It performs acquisition cycle standardization processing on the crack quantification parameters in each time-series subset to eliminate analysis interference caused by the difference in inspection intervals and generate a standardized crack quantification parameter sequence. By continuously fitting the parameter sequence, it generates evolution trend curves of parameters such as crack length, width, and central physical coordinate offset. It accurately extracts evolution feature indicators such as parameter evolution rate, rate change slope, and evolution inflection point from the curves, realizing a refined analysis of the evolution state of each crack of the target object and providing a quantitative analysis basis for the anomaly early warning module. Based on the safety control standards for cracks in coastal concrete engineering structures, and combined with the load-bearing capacity of each structural component of the target object and the severity of the coastal service environment, a crack anomaly judgment threshold system is constructed, including single crack evolution characteristic thresholds and regional crack distribution characteristic thresholds. The thresholds are dynamically adjusted downwards for extreme coastal environments. Evolutionary characteristic indicators output by the crack time-series dynamic change analysis module and regional crack distribution data for structural components are retrieved. Single crack anomalies are identified through precise comparison of indicators and thresholds, and a single crack anomaly identification list is generated. Simultaneously, cluster anomaly evolution analysis is conducted according to structural component identifiers to capture structural component distribution. The risk of clustered anomalies and systemic cracks across zones within the area is identified. Anomaly risk classification is conducted by combining single crack anomaly levels, cluster anomaly severity, structural component importance levels, and real-time coastal environmental monitoring data. Early warnings are divided into three levels, generating standardized graded anomaly early warning signals containing information such as warning level, anomaly type, and handling recommendations. These early warning signals are pushed to relevant operation and maintenance management terminals, while the surface image set and structured crack quantitative parameter dataset are updated in real time. This completes the association and archiving of the early warning data, providing accurate early warning support and data traceability for structural safety assessment and operation and maintenance decision-making of the target object.

[0127] Working principle and its effects:

[0128] This invention focuses on the precise detection of concrete cracks, dynamic adaptation of inspection cycles, and early warning of crack evolution anomalies. Relying on multi-stage technological collaboration, it constructs a complete technical system from target surface image acquisition to crack anomaly early warning. Through multi-feature fusion, density clustering, and pre-trained YOLO algorithms, it achieves precise crack identification, quantitative analysis, and dynamic early warning, fundamentally improving the accuracy, timeliness, and dynamic adaptability of detection.

[0129] This invention first configures the acquisition cycle for the target object, acquires surface images and generates a global image, then labels the global image and constructs a dataset. Suspected crack anomalies are identified through multi-feature fusion, and anomaly clusters are extracted using density clustering. Potential anomaly regions are adaptively divided according to morphology, effectively distinguishing interference from crack features, reducing false positives and false negatives, and improving detection targeting. Subsequently, the potential anomaly regions are input into a pre-trained YOLO algorithm model for initial screening. After initial screening using dual thresholds and secondary removal using cosine similarity, real cracks are accurately identified and quantified parameters are calculated, forming a structured dataset to support subsequent steps. Based on this dataset, data is split, crack levels are classified, and a time-series comparison set is constructed using data from the previous cycle. The acquisition cycle is dynamically adjusted to achieve reasonable allocation of inspection resources. Finally, a unified coordinate system is constructed using structural reference points, spatial matching of time-series crack regions is performed, time-series subsets are generated and standardized, evolution curves are fitted, features are extracted and warnings are issued, accurately capturing crack evolution trends and identifying potential safety hazards in advance.

[0130] This invention achieves a transformation in crack detection, from coarse identification to precise focusing; from fixed inspection cycles to dynamic adaptation; and from single-batch quantification to full-time evolution analysis, with each technical aspect progressing progressively. In practical applications, it effectively solves the problems of false detection and missed detection, poor cycle adaptability, and difficulty in capturing evolutionary anomalies in traditional detection methods. It achieves both accurate crack detection and standardized quantification, while optimizing the allocation of inspection resources. This provides comprehensive and dynamic data for the safe operation and maintenance and decision-making of various concrete engineering structures, significantly improving the overall level of crack detection and safety management.

[0131] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for accurate detection of concrete cracks based on the YOLO algorithm, characterized in that, include: Configure the acquisition cycle, acquire and generate a global image of the surface of the target object; Each global surface image is labeled with a collection timestamp and a target detection object ID to construct a set of surface images. Suspected crack anomaly points in the global surface images are identified by multi-feature fusion, and clusters of suspected crack anomalies are extracted based on density clustering to identify clusters of suspected crack anomalies in order to adaptively divide the potential abnormal regions of the global surface images. For potential anomaly regions, a pre-trained YOLO algorithm model is used to perform initial crack detection, outputting crack candidate boxes, corresponding confidence scores, and class probabilities within each potential anomaly region. This determines whether cracks exist in the potential anomaly region, identifies real crack regions, and calculates crack quantization parameters for real crack regions. Based on the crack quantization parameters of each real crack region in the global image of the surface of the target detected object, the acquisition period is dynamically corrected, and the time-series dynamic change analysis of the real crack region in each global image of the surface in the surface image set is performed to extract the evolution feature index of each crack quantization parameter and provide anomaly warning for the target detected object. The step of adaptively segmenting potential anomaly regions in the global surface image includes: For a set of suspected crack anomalies, set a clustering neighborhood radius and a minimum number of points threshold; Traverse each suspected crack anomaly point and perform density clustering; search for other suspected crack anomaly points within its clustering neighborhood radius, count the number of points within the clustering neighborhood radius, and when the number of points is greater than or equal to the minimum number of points threshold, divide the suspected crack anomaly point and the suspected crack anomaly points within the clustering neighborhood radius that meet the conditions into the same initial cluster, until all suspected crack anomaly points are clustered and the initial cluster set is obtained. For each initial cluster in the initial cluster set, calculate the geometric parameters of the initial cluster, including elongation and curvature; The initial clusters were classified into linear clusters, curved clusters, bifurcation clusters and irregular clusters based on geometric parameters. An adaptive region partitioning operation is performed for initial clusters of different types to generate potential abnormal regions for each initial cluster. Redundancy processing is performed on all the divided potential anomaly regions, the overlap rate of adjacent potential anomaly regions is calculated, and potential anomaly regions are merged; a minimum region area threshold is set to remove potential anomaly regions with an area smaller than the minimum region area threshold; thus forming a set of potential anomaly regions for the global surface image.

2. The method for accurate detection of concrete cracks based on the YOLO algorithm as described in claim 1, characterized in that, The specific steps for identifying the set of suspected crack anomalies in the global surface image include: Adaptive histogram equalization preprocessing is performed on the global surface image, and a dynamic gradient threshold is set based on the global gray mean and standard deviation of the global surface image to extract edge points where pixel gray abrupt changes occur in the global surface image. For each edge point, a neighborhood window is set, and the texture entropy value of each edge point is calculated; The texture entropy values ​​of all edge points in the global surface image are statistically analyzed to obtain the global average texture entropy value. A significant screening threshold is generated based on the global average texture entropy value and a preset scaling factor. The texture entropy value of each edge point is compared with the significant screening threshold, and edge points with texture entropy values ​​less than or equal to the significant screening threshold are retained as candidate outliers. Perform Hough transform on each candidate outlier point to construct a polar coordinate parameter space. Map the coordinates of each candidate outlier point to the polar coordinate parameter space and perform cumulative voting to filter out the line parameters in the polar coordinate parameter space that have more votes than the voting threshold. The straight line parameter is back-projected onto the global image space of the surface, and candidate anomaly points whose deviation values ​​from the corresponding straight line parameter are within a preset deviation range are retained to form a linear candidate point set; For the remaining candidate outliers that were not included in the linear candidate point set, perform eight-neighborhood connectivity analysis to divide the connected regions, extract candidate outliers within each connected region, and form a nonlinear candidate point set. The coordinate information of candidate anomalies is merged from the linear candidate point set and the nonlinear candidate point set to form a set of suspected crack anomalies.

3. The method for accurate detection of concrete cracks based on the YOLO algorithm as described in claim 1, characterized in that, The step of generating potential anomalous regions for each initial cluster includes: For linear clusters, based on their bounding rectangle, a preset number of pixels are extended at both ends along the longest axis and a preset number of pixels are extended bidirectionally along the shortest axis to form a rectangular partition region, which serves as their potential abnormal region. For curved clusters, the Douglas-Puk algorithm is used to fit polygons to the edge points of the curved clusters to obtain the minimum bounding polygon. Then, the polygon is expanded outward by a preset number of pixels to form a polygon division region, which serves as its potential anomaly region. For bifurcated clusters, first perform minimum bounding polygon fitting on the main cluster and each sub-cluster of the bifurcated cluster, then calculate the intersection region of each polygon, and merge the polygon regions of the main cluster and the sub-cluster into a unified polygon partition region as its potential anomaly region. For irregular clusters, calculate the minimum circumcircle of the initial cluster, expand the radius of the circumcircle by a preset number of pixels to form a circular or elliptical division region, which serves as its potential abnormal region.

4. The method for accurate detection of concrete cracks based on the YOLO algorithm as described in claim 1, characterized in that, The steps for determining whether a potential abnormal area contains cracks and identifying actual crack areas include: The pre-trained YOLO algorithm model is invoked, and the potential anomaly region is taken as input. A single forward propagation calculation is performed, and the head detection output network of the YOLO algorithm model outputs the crack candidate box, confidence score and category probability corresponding to each suspected target in the potential anomaly region. The confidence and class probability of the crack candidate boxes corresponding to each suspected target in the potential anomaly region output by the YOLO algorithm model are compared with the set confidence threshold and class probability threshold, respectively, and crack candidate boxes with confidence greater than or equal to the confidence threshold and class probability greater than or equal to the class probability threshold are initially screened. For each crack candidate box after initial screening, the visual features within its coverage area are extracted, normalized to the [0,1] interval, and then fused into a feature vector with uniform dimension; the cosine similarity between the crack candidate box feature vector and the standard feature vectors of various interferences in the surface interference feature library is calculated; crack candidate boxes with a cosine similarity greater than or equal to the similarity threshold are removed. The number of crack candidate boxes remaining in each potential anomaly region is counted. If the number of crack candidate boxes is greater than or equal to 1, it is determined that there is a crack in the potential anomaly region, and the crack candidate box is marked as a real crack region.

5. The method for accurate detection of concrete cracks based on the YOLO algorithm as described in claim 1, characterized in that, The steps for calculating the crack quantization parameters of the actual crack region include: The mapping ratio parameter between the pixel size of the global surface image and the actual physical size of the target object is retrieved, and the coordinate information of the crack candidate box corresponding to the real crack region is extracted; the crack length and width at the pixel scale are calculated for different morphological types. Based on the mapping ratio parameter, the calculated crack length and width at the pixel scale are converted into actual physical crack quantization parameters to obtain the actual crack length and width; and the center pixel coordinates of the real crack region are calculated and converted into center physical coordinates through the mapping ratio parameter. By integrating the actual length, width, and center physical coordinates of cracks in real crack regions, a structured crack quantification parameter dataset is formed.

6. The method for accurate detection of concrete cracks based on the YOLO algorithm as described in claim 5, characterized in that, The steps for dynamically correcting the acquisition period include: The structured crack quantization parameter dataset of each surface global image in the current target detection object's surface image set is obtained, and the structured crack quantization parameter dataset is split according to the structural part identifier of the target detection object. The crack quantization parameters of the current real crack region are extracted for each structural part identifier; and the crack degree level of the real crack region is divided into multiple dimensions and levels. Synchronously retrieve the structured crack quantification parameter data of the corresponding structural part identifier in the previous acquisition cycle and match it with the current acquisition cycle to construct a time-series comparison dataset by part; Based on the time-series comparison dataset of different parts, the number of identified real crack areas in each structural part is increased between the current collection cycle and the previous collection cycle. At the same time, according to the crack severity level, the number of real crack areas under each crack severity level is calculated to determine whether the collection cycle update is triggered. If so, the correction coefficient of the acquisition cycle is quantified by a multi-dimensional weighted normalization algorithm based on the increase in the total number of real crack areas identified in each structural part and the increase in the number of crack severity levels. This correction coefficient is then multiplied by the current acquisition cycle and rounded to obtain the updated acquisition cycle.

7. The method for accurate detection of concrete cracks based on the YOLO algorithm as described in claim 1, characterized in that, The step of performing time-series dynamic change analysis on the real crack regions of each surface global image in the surface image set and extracting the evolution characteristic index of each crack quantification parameter includes: Using the structural reference point of the target object as the spatial positioning origin, and combining the central physical coordinates of each real crack region, a unified spatial coordinate system is constructed. Spatial alignment is performed on the real crack regions of all time-series global surface images in the surface image set to generate a time-series subset of each reference crack region of the current acquisition period's global surface image. Each temporal subset uses the corresponding baseline crack region as the core identifier, integrates the crack quantization parameters of the matching regions in all acquisition cycles, and the acquisition timestamps of each acquisition cycle; the crack quantization parameters in each temporal subset are subjected to acquisition cycle standardization processing, and a standardized crack quantization parameter sequence is generated through temporal smoothing; Based on the standardized crack quantification parameter sequence, each crack quantification parameter is continuously fitted to generate the evolution trend curve of each crack quantification parameter over time, and the evolution characteristic index is extracted.

8. The method for accurate detection of concrete cracks based on the YOLO algorithm as described in claim 7, characterized in that, The step of generating a temporal subset of each reference crack region in the current acquisition cycle global surface image includes: Extract the center physical coordinates of each real crack region in all time-series global surface images in the surface image set, and map the center physical coordinates of each real crack region in time series to a unified spatial coordinate system. Using the real crack region of the current acquisition cycle as the reference crack region, in a unified spatial coordinate system, a spatial distance matching threshold is set for each reference crack region. All real crack regions in each acquisition cycle are traversed, and the spatial Euclidean distance between their center physical coordinates and the center physical coordinates of each reference crack region is calculated. Coarse spatial matching of the reference crack regions is performed, and potential matching regions of the reference crack regions are marked. Feature verification is performed on the potential matching regions obtained by coarse matching. Morphological feature parameters of the baseline crack region and each potential matching region are extracted. The similarity of morphological feature parameters between the baseline crack region and the potential matching regions is calculated. Spatial fine matching of the baseline crack region is performed, and the matching region of the baseline crack region is selected. The matching regions confirmed in each acquisition cycle are associated and marked with the reference crack regions to generate a temporal subset of each real crack region in the current global surface image.

9. A precise concrete crack detection system based on the YOLO algorithm, used to implement the precise concrete crack detection method based on the YOLO algorithm as described in any one of claims 1-8, characterized in that, It includes an image acquisition module, an anomaly region segmentation module, a crack detection module, a dynamic correction module, and a dynamic analysis module; The image acquisition module is used to configure the acquisition cycle, acquire and generate a global image of the surface of the target object; The abnormal region segmentation module is used to construct a set of surface images, identify a set of suspected crack anomalies in the global surface image through multi-feature fusion, and extract suspected crack anomaly clusters based on density clustering to identify suspected crack anomaly clusters in order to adaptively segment potential abnormal regions in the global surface image. The crack detection module performs initial crack detection on potential abnormal areas, outputs crack candidate boxes, corresponding confidence scores and category probabilities for each potential abnormal area, determines whether cracks exist in the potential abnormal area, identifies real crack areas and calculates crack quantification parameters. The dynamic correction module is used to dynamically correct the acquisition cycle based on the crack quantization parameters of each real crack region in the global image of the target object surface. The dynamic analysis module performs time-series dynamic change analysis on the real crack regions of each surface global image in the surface image set, extracts the evolution characteristic index of each crack quantification parameter, and provides anomaly warning for the target detected object.

Images