Bridge bearing crack detection system based on image processing

Abstract

The present application relates to the field of image processing and bridge structure nondestructive testing technology, and particularly relates to a bridge support crack detection system based on image processing, comprising: a data acquisition and calibration unit, used for acquiring an original gray image matrix of a target surface and generating a calibration data packet; an ideal photometric reconstruction unit, used for the calibration data packet and a preset photometric model based on Lambertian reflection, to generate an ideal photometric matrix of a nondestructive surface; a crack simulation unit, used for the ideal photometric matrix, preset crack morphology parameters and photometric attenuation rules, to generate a theoretical damaged simulation matrix; a double-track difference extraction unit, used for differentiating the theoretical damaged simulation matrix from the ideal photometric matrix to obtain a theoretical residual matrix; a coupling decision unit, used for calculating a coupling degree and outputting a crack determination result; a feedback updating unit, used for a simulation parameter library comprising at least a crack width, a crack depth and a fractal dimension; the present application effectively solves the problem of difficulty in extracting real physical dark valleys in a complex environment.

Description

Technical Field

[0001] This invention relates to the fields of image processing and non-destructive testing of bridge structures, specifically to a bridge bearing crack detection system based on image processing. Background Technology

[0002] Bridge bearing crack detection is usually based on image processing technology. The essence of this technology is to extract and calculate features from the original image of the target surface to determine whether there is structural damage. Image visual inspection is a commonly used non-destructive inspection method.

[0003] Existing image crack detection methods include directly extracting edges from the original image or using classification models lacking physical constraints for identification. These conventional image processing techniques can segment and judge diseased areas in simple backgrounds, and they solve the problem of locating two-dimensional gray-scale abrupt changes in a single image under ideal acquisition conditions.

[0004] Currently, for bridge inspection applications involving complex field conditions such as low light at night, damp reflections, and dust adhesion, conventional methods are unable to distinguish between the actual physical dark valleys of cracks and surface water stains, dust, or structural shadow edges. It is necessary to capture more physically interpretable depth photometric features to assist in the analysis. Therefore, how to establish a stable ideal photometric benchmark under multiple environmental noise interferences and accurately extract the actual crack features based on image difference has become a problem to be solved. Summary of the Invention

[0005] The purpose of this invention is to provide an image processing-based bridge bearing crack detection system to solve the following technical problems: The technical problem of directly extracting edges from the original image or relying solely on classification models lacking physical constraints can easily lead to misjudgments. The aim is to reduce the interference of water stains, dust, shadows, and aging textures on crack recognition results and improve the robustness of crack recognition under multiple environmental noise interferences.

[0006] The objective of this invention can be achieved through the following technical solutions: A bridge bearing crack detection system based on image processing includes: The data acquisition and calibration unit is used to acquire the original grayscale image matrix of the target surface, read the three-dimensional CAD geometric model, surface reflection parameters, camera spatial calibration parameters and ambient lighting parameters of the target surface, and map the three-dimensional CAD geometric model to the image pixel coordinate system based on the camera spatial calibration parameters to generate a calibration data package; An ideal photometric reconstruction unit is used to generate an ideal photometric matrix for a non-destructive surface based on the calibration data package and a preset photometric model based on Lambert reflection. The crack simulation unit is used to generate a theoretical damage simulation matrix based on the ideal photometric matrix, preset crack morphology parameters, and photometric attenuation rules. The dual-track differential extraction unit is used to differiate the original grayscale image matrix with the ideal photometric matrix to obtain the real residual matrix, and to differiate the theoretically damaged simulation matrix with the ideal photometric matrix to obtain the theoretical residual matrix; The coupling decision unit is used to calculate the coupling degree and output the crack determination result based on the structural similarity and / or gradient correlation between the actual residual matrix and the theoretical residual matrix; The feedback update unit is used to update the threshold strategy and simulation parameter library according to the crack determination result. The threshold strategy includes at least a high threshold, a low threshold, and a verification threshold. The simulation parameter library includes at least crack width, crack depth, and fractal dimension.

[0007] Preferably, the methods for generating the calibration data packet include: The original grayscale image matrix is ​​obtained by taking multiple images of the same target surface and then registering and fusing them. Read the 3D CAD geometric model and surface reflection parameters of the target surface; Obtain camera spatial calibration parameters and ambient lighting parameters; Based on the camera spatial calibration parameters, the 3D CAD geometric model is mapped to the image pixel coordinate system, and the surface reflection parameters and the ambient lighting parameters are associated with the corresponding pixel positions to generate the calibration data package.

[0008] Preferred methods for constructing the ideal photometric matrix include: Based on the calibration data packet, a preset photometric model based on Lambert reflection is invoked; The 3D CAD geometric model is projected onto the current viewpoint determined based on the camera spatial calibration parameters to generate a pixel-level surface normal vector field. By combining the surface reflection parameters and the ambient lighting parameters, the theoretical reference gray value of each pixel is calculated; The ideal photometric matrix is ​​generated based on the pixel-level surface normal field and the theoretical reference gray value.

[0009] Preferably, the methods for generating the theoretical damage simulation matrix include: Based on the ideal photometric matrix, the preset crack mechanical morphology modeling module is invoked; The crack width, crack depth, and fractal dimension are read from the simulation parameter library as the preset crack morphology parameters, wherein the fractal dimension is used to characterize the tortuosity of the crack path and the edge complexity. Construct a crack geometric skeleton based on the crack width, the crack depth, and the fractal dimension; Based on the crack geometry skeleton, perform crack interior shading calculation and photometric attenuation calculation based on the photometric attenuation rule; The photometric attenuation results corresponding to the crack geometric skeleton are superimposed on the ideal photometric matrix to generate the theoretical damage simulation matrix.

[0010] Preferably, the methods for generating the actual residual matrix and the theoretical residual matrix include: The original grayscale image matrix is ​​differiated from the ideal luminance matrix to generate the real residual matrix; The theoretical damage simulation matrix is ​​subtracted from the ideal photometric matrix to generate the theoretical residual matrix; High-frequency gradient extraction is performed on the real residual matrix to generate a real gradient field; High-frequency gradient extraction is performed on the theoretical residual matrix to generate a theoretical gradient field.

[0011] Preferably, the methods for calculating coupling degree include: Based on the actual gradient field and the theoretical gradient field, calculate the two-dimensional cross-correlation value; Based on the actual residual matrix and the theoretical residual matrix, the structural similarity value is calculated; The two-dimensional cross-correlation value and the structural similarity value are normalized respectively, and then weighted and fused according to preset weights to generate the coupling degree.

[0012] Preferably, the output method of the crack determination result includes: The high threshold is greater than the low threshold, and the review threshold is greater than the low threshold and less than the high threshold; When the coupling degree is higher than the high threshold, the true crack determination result is output; When the coupling degree is lower than the low threshold, output the noise determination result; When the coupling degree is higher than or equal to the low threshold and lower than or equal to the high threshold, the corresponding image region is extracted as a candidate region, the preset verification rules are called to perform a secondary simulation comparison on the candidate region, and the verification judgment result is output.

[0013] Preferably, the execution methods of the review rules include: Extract the local morphological skeleton of the candidate region and the local grayscale attenuation curve along the normal direction of the local morphological skeleton; The crack width, crack depth, and fractal dimension from the simulation parameter library are used as candidate crack parameter groups to perform local damage simulation on the candidate region; The simulation results of local damage are coupled and calculated again with the real residual submatrix corresponding to the candidate region extracted from the real residual matrix; When the result of the re-coupling calculation is higher than the verification threshold, the result of the true crack determination is output; When the result of the re-coupling calculation is lower than or equal to the verification threshold, the noise judgment result is output.

[0014] Preferably, the update methods for the feedback update unit include: A judgment log is generated based on the true crack determination result and the noise determination result; Based on the determination log, the high threshold, the low threshold, and the review threshold are corrected; Based on the judgment log, the parameter mapping relationship between the crack width, the crack depth, the fractal dimension, and the theoretical photometric attenuation model is corrected. Write the revised threshold strategy and simulation parameter library into the system configuration.

[0015] Preferably, the target surface is the bridge bearing surface, the original grayscale image matrix contains surface image data collected under conditions of low light, humidity, dust or shadow, and the noise determination result includes at least water stains, dust, light and shadow edges and non-target structural surface textures, the non-target structural surface textures including surface aging cracks.

[0016] The beneficial effects of this invention are: 1. This invention provides a complete bridge bearing crack detection system, which overcomes the shortcomings of traditional direct edge extraction which is easily interfered with. By constructing a physically interpretable ideal photometric matrix as a non-destructive benchmark, and combining theoretical damage simulation to perform dual-track differential and coupled decision, the interference of water stains, dust and shadows on the external field crack identification is reduced from the root, and the problem of difficulty in extracting real physical dark valleys in complex environments is effectively solved. 2. This invention uses a calibration method that involves multiple data acquisitions and pixel-level parameter binding. It combines the three-dimensional geometric model of the target surface, surface reflection, and ambient lighting parameters to generate a surface normal vector field, thereby constructing an ideal photometric matrix. This method can effectively suppress transient reflective noise and eliminate natural brightness fluctuations caused by normal curved surfaces and lighting changes in advance, providing a stable and physically constrained photometric benchmark for differential processing in complex field environments. 3. This invention utilizes preset morphological parameters to construct a crack geometric skeleton and performs internal shading and photometric attenuation calculations to generate a theoretical damage simulation matrix, which is then compared with the original image using differential and high-frequency gradient extraction. This active simulation and dual-track differential mechanism transforms planar two-dimensional mutations into three-dimensional physical template comparisons, accurately peels off the undamaged background, and effectively distinguishes between physical dark valleys caused by real cracks and ordinary two-dimensional stains on the surface. 4. The coupling decision mechanism of this invention normalizes and weights the two-dimensional cross-correlation value of the dual-track gradient field and the structural similarity value of the residual matrix. This mechanism comprehensively considers the consistency of the local light valley boundary features in the abnormal area and the conformity of the overall residual to the crack dark valley pattern. It uses multi-dimensional physical feature fusion evaluation to replace the single gray-scale threshold judgment, which greatly improves the robustness of suspected disease identification under low light and humid conditions. 5. This invention adopts a high and low dual threshold hierarchical judgment mechanism to extract the local morphological skeleton and normal grayscale decay curve of the candidate region in the critical interval, and uses adaptive local damage simulation for re-coupling and verification; this can quickly separate high-confidence true cracks from obvious noise, concentrate computing resources on easily confused and ambiguous regions and re-verify in a more refined dimension, effectively balancing the efficiency of large-scale detection and the false alarm control capability that meets the preset requirements. 6. This invention is based on the iterative correction of the judgment threshold and the mapping relationship of the photometric attenuation model parameters according to the judgment log, and outputs accurate classification results for non-target features such as water stains, dust, shadows and aging cracks. This not only enables the system to dynamically adapt to the characteristic changes caused by bridge aging or environmental alternation during long-term operation, but also makes the investigation results fully detailed, thereby directly serving and guiding subsequent differentiated and refined bridge maintenance auxiliary decisions. Attached Figure Description

[0017] The invention will now be further described with reference to the accompanying drawings.

[0018] Figure 1 This is a schematic diagram of the modules of the image processing-based bridge bearing crack detection system provided in the embodiments of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 A bridge bearing crack detection system based on image processing includes: a data acquisition and calibration unit, used to acquire the original grayscale image matrix of the target surface, read the three-dimensional CAD geometric model, surface reflection parameters, camera spatial calibration parameters and ambient lighting parameters of the target surface, and map the three-dimensional CAD geometric model to the image pixel coordinate system based on the camera spatial calibration parameters to generate a calibration data package; An ideal photometric reconstruction unit is used to generate an ideal photometric matrix for a non-destructive surface based on the calibration data package and a preset photometric model based on Lambert reflection. The crack simulation unit is used to generate a theoretical damage simulation matrix based on the ideal photometric matrix, preset crack morphology parameters, and photometric attenuation rules. The dual-track differential extraction unit is used to differiate the original grayscale image matrix with the ideal photometric matrix to obtain the real residual matrix, and to differiate the theoretically damaged simulation matrix with the ideal photometric matrix to obtain the theoretical residual matrix; The coupling decision unit is used to calculate the coupling degree and output the crack determination result based on the structural similarity and / or gradient correlation between the actual residual matrix and the theoretical residual matrix; The feedback update unit is used to update the threshold strategy and simulation parameter library according to the crack determination result. The threshold strategy includes at least a high threshold, a low threshold, and a verification threshold. The simulation parameter library includes at least crack width, crack depth, and fractal dimension.

[0021] This embodiment provides a bridge bearing crack detection mechanism based on image processing. Specifically, this mechanism is designed for the inspection of plate rubber bearings in highway viaducts. Under conditions of low light at night, dampness after rain, and dust on the surface, it can automatically determine whether there are structural cracks on the surface of a single bridge bearing.

[0022] The system executes sequentially in the order of acquisition and calibration, ideal reconstruction, crack simulation, dual-track differential, coupled decision, and feedback update, instead of directly extracting edges from the original image or relying solely on a classification model lacking physical constraints to provide results.

[0023] Specifically, the data acquisition and calibration unit first acquires the original grayscale image matrix captured by the industrial camera, and simultaneously retrieves the three-dimensional CAD geometric model of the support, surface reflection parameters, camera spatial calibration parameters, and ambient lighting parameters for the current inspection. For ease of explanation, an exemplary 3×3 pixel array can be used to represent the original grayscale image, for example, the original grayscale image matrix is ​​[98,101,97;103,62,100;99,102,98]. The grayscale value of the center pixel 62 is lower than the preset grayscale lower limit threshold, but it is still impossible to determine whether its source is a shadow, water stain or crack. After the system projects the three-dimensional CAD geometric model onto the image pixel coordinate system, it generates normal vectors, reflectivity and light incidence information corresponding to each pixel, forming a calibration data package. The ideal photometric reconstruction unit calculates the theoretical gray level of each pixel in a lossless state based on the Lambertian reflection model, thus obtaining the ideal photometric matrix. For example, the same 3×3 region may appear as [100,100,100;101,100,100;100,101,100] in a lossless state. This matrix does not depend on sample averaging, but comes from the physical deduction of support geometry and lighting conditions, and therefore can serve as a lossless reference benchmark under the current viewpoint. Based on this, the crack simulation unit reads the crack width, depth, and fractal dimension from the simulation parameter library, constructs a set of theoretical crack morphologies, and injects the crack effect into the ideal photometric matrix according to the photometric attenuation rule to generate the theoretical damage simulation matrix. Taking the above region as an example, if a narrow V-shaped crack is injected, the simulation matrix can be [100,100,100;101,58,100;100,101,100]. At this time, the difference between the theoretical damage simulation matrix and the ideal photometric matrix is ​​only caused by the crack and does not contain interference from water stains, dust, etc. The dual-track differential extraction unit establishes two parallel residual spaces; the real residual matrix is ​​obtained by subtracting the ideal photometric matrix from the original grayscale image matrix, for example [-2,1,-3;2,-38,0;-1,1,-2]; the theoretical residual matrix is ​​obtained by subtracting the ideal photometric matrix from the theoretical damage simulation matrix, for example [0,0,0;0,-42,0;0,0,0]; the real residual retains all anomalies in the actual shooting, while the theoretical residual only retains the ideal anomalies caused by the crack; The coupling decision unit calculates the structural similarity and / or gradient correlation between the two; for example, after normalizing the above small region, if the structural similarity is 0.81 and the gradient correlation value is 0.78, the coupling degree can be obtained as 0.80 after weighting; the system gives the judgment result of true crack, noise or pending review based on this; the feedback update unit writes the judgment result into the log for subsequent correction of threshold and simulation parameter mapping relationship.

[0024] As an anomaly handling strategy, if the corresponding 3D CAD geometric model cannot be read on site, the system can switch to manual verification and mark the set of images as calibration missing, and not enter the automatic judgment. If abnormal fluctuations in illumination parameters cause the ideal luminance matrix to be too bright or too dark overall, the system can first perform global brightness offset compensation on the entire image before entering the differential stage. If the coupling result is in the critical range, the system will not output the conclusion directly, but will instead enter the secondary verification process to avoid misjudging stains as cracks.

[0025] For example, during a routine night patrol of the same highway bridge, maintenance personnel continuously collected three grayscale images of the No. 2 bearing on the left side of the third span; after the system reads the three-dimensional CAD geometric model of the layered rubber-steel plate of the bearing, the camera spatial calibration parameters, and the illumination parameters under humid conditions, it constructs a non-destructive photometric reference for the current view of the bearing. The system actively simulates multiple sets of crack morphologies with different widths and depths, and couples them one by one with the residual space formed by the real image. If a certain set of simulation results is highly consistent with the measured residual in terms of pixel-level grayscale attenuation and skeleton morphology, then a crack judgment is output. Otherwise, if only diffuse scattering dark spots appear without V-shaped light valley features, then it is judged as noise.

[0026] The purpose of this step is to reduce the interference of water stains, dust, shadows and aging textures on crack identification results by first establishing a physically interpretable ideal benchmark and then comparing the actual deviation with the theoretical crack deviation.

[0027] In a preferred embodiment of the present invention, the calibration data packet is generated as follows: the original grayscale image matrix is ​​a matrix obtained by taking multiple pictures of the same target surface and registering and fusing them; the three-dimensional CAD geometric model and surface reflection parameters of the target surface are read; and camera spatial calibration parameters and ambient lighting parameters are obtained. Based on the camera spatial calibration parameters, the 3D CAD geometric model is mapped to the image pixel coordinate system, and the surface reflection parameters and the ambient lighting parameters are associated with the corresponding pixel positions to generate the calibration data package.

[0028] This embodiment provides a calibration data packet generation mechanism; specifically, in the aforementioned bridge bearing inspection process, relying solely on a single shot is easily affected by instantaneous water droplet reflections, vehicle headlights sweeping across, and camera shake within a preset error range, resulting in a lack of stable input for subsequent ideal photometric reconstruction; Therefore, this embodiment introduces a calibration method that binds multiple shots to pixel-level parameters, so that each pixel not only has a grayscale value, but also carries explanatory information from geometric models and environmental parameters.

[0029] Specifically, the system continuously acquires multiple frames of original grayscale image matrices from the same support surface, for example, obtaining matrix A, matrix B and matrix C in sequence; for ease of deduction, the grayscale values ​​of the same central pixel in the three frames can be set to 61, 63 and 62 respectively; the system can perform registration on these three frames, and if the registration offset is less than a preset threshold, for example, no more than 1 pixel, then they are regarded as the same observation area; Then, stable observations are obtained by median synthesis or weighted averaging, for example, the center pixel is synthesized to 62; this can suppress occasional bright spots or transient noise. The system reads the 3D CAD geometric model of the support, which includes the spatial coordinates of the upper steel plate, rubber layer, and edge chamfer, as well as surface reflection parameters; camera spatial calibration parameters include focal length, principal point position, distortion parameters, and camera pose parameters relative to the support; ambient lighting parameters include the principal light incident direction, ambient diffuse reflection intensity, and on-site auxiliary lighting status; using these parameters, the system maps the 3D points in the 3D CAD geometric model to image pixel coordinates; for example, if the coordinates of point P in the model are (10,5,0), after projection, it falls on the pixel coordinates (320,240), then the normal vector, reflection coefficient, and local illumination intensity corresponding to that point are also bound to that pixel;

[0030] The calibration data package can be understood as a set of data organized by pixels. In this application, specific exemplary data is used for illustration. If the observed gray level of pixel (2,2) is 62, the normal vector is oriented approximately positively, the reflection coefficient is 0.65, and the cosine value of the incident light angle is 0.92, then the subsequent photometric model can directly call these contents to calculate the theoretical lossless gray level of the pixel. As an anomaly handling strategy, if the registration offset between multiple frames is greater than the preset registration tolerance threshold, it indicates that the shooting posture is unstable, and the system can re-trigger the acquisition. If some pixels cannot be mapped to a 3D CAD geometric model due to severe occlusion, the corresponding pixels will be marked as invalid pixels and skipped in subsequent matrix calculations. If ambient lighting parameters are missing, the lighting parameters from the most recent valid inspection can be used as the initial values, and the confidence level of the relevant area will be lowered in this inspection. Specifically, during the nighttime re-inspection of the No. 2 support on the left side of the third span, the maintenance terminal continuously acquired three frames of images. Due to the brief illumination from maintenance vehicle lights under the bridge, the lower right corner of the second frame was locally overexposed. After registration, the system took the median of the three frames, preserving a stable grayscale distribution of the support body. It then used the existing design model of the support to map the edge of the rubber layer, the transition area of ​​the steel plate, and the exposed surface to the image coordinates. Subsequently, each pixel possessed the combined attributes of observed grayscale, geometric normal, material reflection, and ambient light, forming the calibration data package required for subsequent calculations.

[0031] The purpose of this step is to provide a stable, interpretable, and pixel-aligned input basis for ideal photometric reconstruction, so as to avoid subsequent analysis from deviating from reality due to sporadic noise or geometric mismatch in a single frame.

[0032] In a preferred embodiment of the present invention, the construction of the ideal photometric matrix includes: calling a preset photometric model based on Lambertian reflection based on the calibration data package; projecting the three-dimensional CAD geometric model onto the current viewing angle determined based on the camera spatial calibration parameters to generate a pixel-level surface normal vector field; By combining the surface reflection parameters and the ambient lighting parameters, the theoretical reference gray value of each pixel is calculated; based on the pixel-level surface normal vector field and the theoretical reference gray value, the ideal photometric matrix is ​​generated.

[0033] This embodiment provides an ideal photometric matrix construction mechanism. Specifically, after the calibration data package is generated, if only geometric projection is used without combining surface normal and illumination direction, the natural brightness and darkness changes of the same support surface will be mistakenly regarded as abnormal areas. Therefore, this embodiment further generates a pixel-level surface normal vector field and calculates the theoretical reference gray value that each pixel should present in the lossless state. After projecting the 3D CAD geometric model onto the current viewpoint, the system assigns a surface normal vector to each effective pixel. This can be illustrated using a simplified 2×2 region: the normal vector of the top-left pixel is approximately facing the camera, the normal vector of the top-right pixel is slightly deflected to the right, and the normal vector of the bottom-left pixel, located in the rounded corner transition area, is deflected downwards. Due to these different normal vectors, even if the surface is intact, the theoretical brightness of each pixel should not be exactly the same. The system calls a photometric model based on Lambertian reflection to calculate the theoretical baseline grayscale based on the pixel normal, surface reflectance, and ambient lighting parameters. If a pixel has a reflectance of 0.6, an ambient light intensity of 120, and a cosine value of 0.8 for the angle between its normal and the lighting direction, then the theoretical grayscale of that pixel can be approximately mapped to 58. If the cosine value of the angle between adjacent pixels is 0.95, then the theoretical grayscale can be 68. This forms the theoretical baseline brightness distribution of the entire image, rather than a uniform grayscale image.

[0034] Furthermore, the system integrates the normal vector field with the theoretical gray values ​​of each pixel into an ideal photometric matrix; this matrix not only retains the brightness gradient caused by the actual geometry of the support, but also eliminates local anomalies such as cracks, water stains and dust; for subsequent difference analysis, the closer it is to the actual undamaged state on site, the more concentrated the anomalies are retained in the residual. As an anomaly handling strategy, if multiple points compete for mapping after some edge pixels are projected, the surface points closest to the camera can be retained in a depth-first manner; if the local normal estimation is distorted due to the simplification of the 3D CAD geometric model, the system can perform neighborhood smoothing on the normal vector field, but without crossing the material boundary, to avoid accidentally smoothing the steel plate edge onto the rubber layer; if there is a strong directional change in ambient lighting, an image can be divided into multiple sub-regions to calculate the theoretical brightness separately, instead of using a single lighting parameter for the entire image. For example, in the aforementioned night patrol scenario, the side of the lower edge of the support closer to the streetlight is brighter than the side facing away from the light, and the corners of the support naturally have shadows due to the curved transition. The system uses the existing geometric model to restore these normal brightness changes and obtains an ideal luminosity matrix that conforms to the physical deduction law from the current perspective. After that, only the local dark valleys that deviate from this benchmark will be highlighted in the residual.

[0035] The purpose of this step is to pre-interpret normal geometric fluctuations and changes in illumination, so that subsequent difference results are more focused on the real disease signals.

[0036] In a preferred embodiment of the present invention, the generation method of the theoretical damage simulation matrix includes: based on the ideal photometric matrix, calling a preset crack mechanical morphology modeling module; reading the crack width, crack depth, and fractal dimension from the simulation parameter library as the preset crack morphology parameters, wherein the fractal dimension is used to characterize the tortuosity of the crack path and the edge complexity; and constructing a crack geometric skeleton according to the crack width, crack depth, and fractal dimension. Based on the crack geometry skeleton, perform crack interior shading calculation and photometric attenuation calculation based on the photometric attenuation rule; superimpose the photometric attenuation results corresponding to the crack geometry skeleton onto the ideal photometric matrix to generate the theoretical damage simulation matrix.

[0037] This embodiment provides a theoretical damage simulation matrix generation mechanism. Specifically, an ideal photometric matrix alone is insufficient to distinguish between dark valleys caused by real cracks and dark spots caused by two-dimensional stains. This is because both may appear as local darkening in real images. Therefore, this embodiment introduces a crack mechanical morphology modeling module, which does not directly draw low grayscale line segments on two-dimensional images, but actively generates damaged images that conform to fracture characteristics according to crack width, depth and fractal dimension. Specifically, the system reads a set of candidate parameters from the simulation parameter library; for example, the first set of parameters sets the crack width to 2 pixels, the crack depth to level 3, and the fractal dimension to 1.10, indicating that the path curvature is lower than the preset curvature threshold and the edge gradient variance is lower than the preset variance threshold; the second set of parameters sets the crack width to 3 pixels, the crack depth to level 5, and the fractal dimension to 1.35, indicating that the path curvature is higher than the preset curvature threshold and the edge gradient variance is higher than the preset variance threshold; the system generates a crack geometric skeleton on an ideal photometric matrix based on these parameters; the skeleton can be understood as the crack centerline and its local width distribution, and different fractal dimensions will lead to different skeleton turning frequencies; After the skeleton is generated, the system performs occlusion calculations inside the cracks; specifically, based on the geometric occlusion of the V-groove and the local energy loss caused by multiple reflections, the system introduces a photometric attenuation rule and performs calculations at the pixel level. Let the lateral and vertical relative displacement of the current pixel from the center of the nearest crack geometry be... Combined with the given crack width And the maximum center shielding attenuation coefficient determined by the crack depth. The luminance attenuation rate of this pixel It can be modeled as a decay function modulated by lateral distance, for example: when hour, If the distance exceeds this range, it is considered unaffected by the crack shielding. ;in, Pi It is a cosine function; The attenuation distribution function uses a cosine function to characterize the geometric occlusion and photometric attenuation effect that gradually weakens from the crack center to both sides; the system applies this attenuation rate to the ideal photometric value to generate the simulated grayscale result of the current pixel: ; in, This represents the simulated grayscale result of the current pixel. This represents the ideal gray value corresponding to the ideal photometric matrix; This indicates the luminance attenuation rate of the current pixel.

[0038] Taking a cross-section as an example, if the ideal grayscale cross-section is [100,100,100,100,100], and a V-shaped crack with an opening width less than the preset width threshold is injected and deepens downward, the theoretical cross-section can become [100,86,58,84,100]; the center position is the darkest, and the two sides gradually recover, which is different from the smooth dark area commonly seen in ordinary oil stains; If the crack depth is greater, the gray level in the middle can be further reduced to 48; if the crack width increases, the dark valley area expands laterally; for the bends in the path, the higher the fractal dimension, the more obvious the edge undulations, and the higher the skeleton complexity in the theoretical residual. The system superimposes the photometric attenuation results at each location back into the ideal photometric matrix to obtain the theoretical damage simulation matrix. To improve the matching capability, multiple sets of simulation matrices can be generated in parallel in practical applications, corresponding to different disease morphologies such as narrow cracks, deep cracks, and tortuous cracks, for subsequent coupled calculations to compare one by one. As a boundary condition processing rule, if the crack width generated by a certain set of parameters exceeds the size of the candidate region, it means that the parameters are not applicable to the current region, and the system can directly discard the set of simulations. If the depth parameter is too large and causes the theoretical gray level to drop to the invalid lower limit, then the gray level should be truncated according to the lowest resolvable gray level of the sensor. If a certain area is close to the edge of the support and there is natural shading, then the shading calculation needs to take into account the geometry of the support boundary at the same time to avoid mistaking the edge shadow as the crack depth effect. In the middle of the lower surface of the aforementioned support, the system retrieves three sets of crack morphologies from the parameter library for simulation. One set generates a slender dark valley along the rubber layer interface with a fluctuation amplitude within a preset range. Its lateral grayscale profile has a similarity to the suspected area in the field image that is higher than the preset matching threshold. Although another set also forms dark spots, the edges are too smooth and do not match the high-frequency skeleton in the field. The system retains the former as a better theoretical template and sends the corresponding theoretical damage simulation matrix into subsequent differential processing.

[0039] The purpose of this step is to transform the geometric features, grayscale attenuation, and path tortuosity of the crack into a calculable physical image template, so that subsequent judgments are no longer limited to simple grayscale abrupt changes.

[0040] In a preferred embodiment of the present invention, the generation of the real residual matrix and the theoretical residual matrix includes: differentiating the original grayscale image matrix from the ideal photometric matrix to generate the real residual matrix; and differentiating the theoretical damaged simulation matrix from the ideal photometric matrix to generate the theoretical residual matrix. High-frequency gradient extraction is performed on the real residual matrix to generate a real gradient field; high-frequency gradient extraction is also performed on the theoretical residual matrix to generate a theoretical gradient field.

[0041] This embodiment provides a dual-track residual and gradient field generation mechanism. Specifically, if the original image is directly compared with the theoretically damaged simulation image, the normal brightness background, overall exposure deviation and surface brightness gradient will interfere with the matching. Therefore, this embodiment first returns the real image and the theoretical crack image to the same lossless reference, and then extracts the high-frequency gradient to construct a comparable dual-track feature space. Specifically, the real residual matrix is ​​obtained by subtracting the ideal luminance matrix from the original grayscale image matrix; for example, if the original matrix is ​​[98,101,97;103,62,100;99,102,98] and the ideal matrix is ​​[100,100,100;101,100,100;100,101,100], then the real residual matrix is ​​[-2,1,-3;2,-38,0;-1,1,-2]. Similarly, if the theoretical damage simulation matrix is ​​[100,100,100;101,58,100;100,101,100], then the theoretical residual matrix is ​​[0,0,0;0,-42,0;0,0,0]. Through this processing, the normal background is effectively filtered out and the residual values ​​converge to the preset interval, highlighting the abnormal information. Based on the residual, the system further performs high-frequency gradient extraction; local difference, Sobel operator or other gradient extraction methods can be used; taking the central region of the real residual matrix as an example, the central pixel -38 is significantly different from the surrounding pixels, and its gradient magnitude will be significantly higher than the background; while the central position of the theoretical residual matrix also forms a steep gradient; the system thus generates the real gradient field and the theoretical gradient field respectively for subsequent related calculations; If the anomaly is caused by a gradual change in grayscale value due to water stains with an area greater than the preset area threshold, the extracted gradient amplitude is lower than the preset gradient threshold and cannot match the high-frequency gradient structure of the crack template. As an anomaly handling strategy, if the overall residual of the entire image after differencing is negative or positive, it indicates that there is a global brightness drift. The system can first perform residual mean zeroing processing. If noise points are densely distributed after gradient extraction, local smoothing or minimum connected area filtering can be added to avoid isolated noise points from increasing the relevance. If there are too few pixels in the candidate region to form a stable gradient structure, the region is marked as insufficient information and is transferred to reshoot or manual verification. In the suspected crack area of ​​the same support, the actual residual map shows a vertical dark valley, but it is surrounded by scattered dust points. After the system performs high-frequency gradient extraction on the actual residual, most of the dust points appear as isolated high-frequency points, while the dark valley forms a continuous gradient band. After the theoretical residual is processed in the same way, it also appears as a continuous narrow gradient band. The comparability of the two in high-frequency structure is significantly improved.

[0042] The purpose of this step is to first remove the non-destructive background and then extract the core grayscale transition boundary of the crack, so as to provide a more stable input for subsequent coupling calculation.

[0043] In a preferred embodiment of the present invention, the coupling degree is calculated by: calculating a two-dimensional cross-correlation value based on the actual gradient field and the theoretical gradient field; calculating a structural similarity value based on the actual residual matrix and the theoretical residual matrix; normalizing the two-dimensional cross-correlation value and the structural similarity value respectively, and weighting and fusing them according to a preset weight to generate the coupling degree.

[0044] This embodiment provides a coupling degree calculation mechanism; specifically, when only gradient correlation is considered, some edge shadows may also exhibit strong linear boundaries; When only structural similarity is considered, some large-area stains may be similar to the dark valley in low-frequency morphology; therefore, this embodiment incorporates the two-dimensional cross-correlation at the gradient level and the structural similarity at the residual level into the calculation, and then performs normalization and weighted fusion to improve the discrimination stability. Specifically, the system calculates a two-dimensional cross-correlation value for the real gradient field and the theoretical gradient field; this can be understood as aligning the two gradient maps in space and evaluating whether the peak and valley positions are synchronized; for example, if the cross-correlation value of a certain region is 0.84, it indicates that the real anomaly boundary and the theoretical crack boundary are relatively consistent in position and direction; the system then calculates the structural similarity value for the real residual matrix and the theoretical residual matrix, for example, obtaining 0.76, indicating that the two have a high structural similarity in brightness distribution and local structure; Since the two types of indicators may have different numerical ranges and sensitivities, the system needs to perform normalization processing first; specifically, for the calculated two-dimensional cross-correlation values... Its original evaluation range is usually between [-1, 1], and the system uses the mapping formula: ; in, This represents the normalized two-dimensional cross-correlation value; The original two-dimensional cross-correlation value is calculated and standardized to a scalar scale in the [0,1] interval. For the structural similarity value, a metric algorithm based on brightness distribution, contrast, and structural dimension can be used to obtain the result. For example, the standard structural similarity index algorithm in this field can be used for specific solutions; if the original structural similarity obtained by using the standard algorithm is in the range of [-1, 1], then the formula can be applied accordingly. Unify them to the standard range of [0,1] values Then, the final coupling degree is generated by using a weighted formula based on preset weights. : ; in, The coupling degree generated by weighted fusion, The normalized two-dimensional cross-correlation value. To unify the structural similarity values ​​to the standard interval [0,1]; and satisfy Preset weights and It can be an experience constant set by the system configuration based on the specific focus of the current inspection task, or a dynamic variable obtained by the feedback update unit through adaptive training optimization based on the contribution of cross-correlation values ​​and structural similarity values ​​in historical judgment logs to the actual judgment results; For example, when the cross-correlation weight is taken Structural similarity weighting In this case, the coupling degree can be calculated as 0.6 × 0.84 + 0.4 × 0.76 = 0.808. For inspection tasks that emphasize the characteristics of the Optics Valley boundary, the cross-correlation weight can be increased. For regions where overall texture consistency is emphasized, the structural similarity weight can be appropriately increased. .

[0045] Through this fusion method, the system considers both the correlation between local gradients and crack boundaries, and the structural similarity between the overall residual and the real dark valleys caused by cracks. If a place is just a water stain, it may have dark areas in its structure, but the gradient boundaries are blurred and the cross-correlation value is low. If a place is just a shadow edge, it may have clearer boundaries, but the overall residual shape does not conform to the characteristics of a V-shaped crack, and the structural similarity is low. It is difficult to obtain a high coupling degree after weighting both.

[0046] As an anomaly handling strategy, if the cross-correlation value of a certain region is higher than the preset first cross-correlation upper limit threshold but the structural similarity is lower than the preset second similarity lower limit threshold, or vice versa, the system can mark it as a conflict sample and send it to the review process. If the variance of a certain type of indicator is found to be too small across the entire region during normalization, it indicates that the indicator has insufficient discrimination. In this case, its weight can be temporarily reduced to avoid the amplification effect of a single outlier. If multiple theoretical templates achieve close coupling, the first few templates should be retained for the next stage of detailed review, rather than giving a single conclusion too early. For example, on the suspected crack in the middle of the aforementioned support, the system performs cross-correlation calculations on the actual gradient field and the three sets of theoretical gradient fields, and calculates the structural similarity between the actual residual and the three sets of theoretical residuals. The second set of templates has a cross-correlation of 0.87, a structural similarity of 0.79, and a fusion coupling degree of 0.838, which is the highest among the three sets. The first set only has a cross-correlation of 0.61, and although the third set has a structural similarity of 0.80, its cross-correlation is only 0.52. Based on this, the system prioritizes the second set of templates as the best match result.

[0047] The purpose of this step is to use multi-dimensional consistency rather than a single threshold to evaluate whether suspected anomalies conform to the physical characteristics of cracks, thereby improving the robustness of identification in complex environments.

[0048] In a preferred embodiment of the present invention, the output method of the crack determination result includes: wherein the high threshold is greater than the low threshold, and the verification threshold is greater than the low threshold and less than the high threshold; When the coupling degree is higher than the high threshold, the true crack determination result is output; when the coupling degree is lower than the low threshold, the noise determination result is output; when the coupling degree is higher than or equal to the low threshold and lower than or equal to the high threshold, the corresponding image region is extracted as a candidate region, the preset verification rules are called to perform a secondary simulation comparison on the candidate region, and the verification determination result is output.

[0049] This embodiment provides a hierarchical judgment output mechanism. Specifically, if the system only sets a single judgment threshold, critical samples are easily misjudged in complex scenarios such as low light and humidity. For example, when the coupling degree is close to the threshold, it may not be a typical crack or pure noise. Therefore, this embodiment adopts a hierarchical judgment method with high threshold, low threshold and intermediate candidate area to send uncertain samples into the review process. Specifically, the system pre-sets a high threshold and a low threshold, with the high threshold being greater than the low threshold. Pre-setting refers to the initial empirical judgment boundary determined in advance by statistically calculating using a historical set of known damaged and normal image samples of bridge bearings of the same type before the system is first deployed and run. In this embodiment, the high threshold can be set to 0.82 and the low threshold to 0.45. When the coupling degree of a certain region is greater than 0.82, it indicates that it is highly consistent with the theoretical crack template in terms of structure and gradient, and the system directly outputs the true crack judgment result. When the coupling degree is less than 0.45, it indicates that it is significantly different from the crack template, and the noise judgment result is directly output. When the coupling degree falls between 0.45 and 0.82, the system does not directly output the deterministic judgment result, but extracts the corresponding image region as a candidate region and enters the secondary simulation comparison. For example, a region with a coupling degree of 0.88 can be directly classified as a true crack; a region with a coupling degree of 0.31 can be directly classified as noise; and another region with a coupling degree of 0.67 is sent as a candidate region for verification. This is intended to perform preliminary screening and diversion of samples with an initial coupling degree higher than the high threshold or lower than the low threshold, and to trigger secondary verification processing for samples within the threshold range. In case of anomalies or edge situations, if the boundaries of multiple adjacent candidate regions in the same image overlap, the regions can be merged first to avoid repeated verification; if the coupling degree of a certain region is exactly equal to the high threshold or low threshold, it can be uniformly included in the verification according to the preset strategy to reduce the impact of boundary value jitter; if the number of candidate regions in an image is abnormally large, it indicates that the on-site noise is complex, and the system can automatically increase the priority of reshooting and classify the frame as a high interference image. For example, in the aforementioned support inspection, the system identified three suspected areas: area R1 with a coupling degree of 0.86, area R2 with a coupling degree of 0.39, and area R3 with a coupling degree of 0.68. The system directly generated a crack warning for R1, directly classified R2 as surface noise, and extracted a local image window for R3 and invoked more granular verification rules. In this way, on-site maintenance personnel only need to focus on a small number of truly uncertain areas, without having to check all dark spots one by one. The purpose of this step is to establish clear decision boundaries and intermediate buffer zones, balancing detection efficiency with false alarm control capabilities.

[0050] In a preferred embodiment of the present invention, the execution of the verification rule includes: extracting the local morphological skeleton of the candidate region and the local grayscale attenuation curve along the normal direction of the local morphological skeleton; calling the crack width, crack depth and fractal dimension in the simulation parameter library as a candidate crack parameter group, and performing local damage simulation on the candidate region; The simulation result of local damage is coupled again with the real residual submatrix corresponding to the candidate region extracted from the real residual matrix; when the result of the second coupling calculation is higher than the verification threshold, the real crack judgment result is output; when the result of the second coupling calculation is lower than or equal to the verification threshold, the noise judgment result is output.

[0051] This embodiment provides a candidate region verification mechanism; specifically, the reason why samples in the middle region are difficult to judge is often because their overall coupling degree is not low, but their local geometric features are not clear enough; for example, thin water stain edges, local shadows and early microcracks may all get similar scores; Therefore, in this embodiment, the local morphological skeleton and normal grayscale decay curve are further extracted within the candidate region, and a second judgment is completed using finer-grained local simulation. Specifically, the system first extracts a local morphological skeleton from the candidate region; this can be understood as extracting a center line from the candidate dark area to indicate the main extension direction; for example, in a 5×5 candidate window, the skeleton may extend from the upper left to the lower right. Along the normal direction of the skeleton, the system collects multiple local grayscale profiles; if the grayscale of one of the profiles changes from [98,90,61,88,97], a narrow and obvious dark valley can be seen forming in the center; if the profile of another candidate region is [96,88,82,86,94], it is closer to the gradual decay of stains. The system calls multiple sets of crack width, depth and fractal dimension from the parameter library to perform local damage simulation on the candidate region. At this time, the simulation range is reduced to a local window, and the large-scale template of the entire image is no longer used. Instead, it is adaptively generated based on the direction, length and grayscale profile of the candidate region. The generated local simulation results are coupled with the actual residual submatrix again for calculation; for example, the coupling degree of candidate region R3 is 0.68 in the initial judgment, and after entering the verification, the coupling degree of local simulation and actual residual submatrix reaches 0.74; if the verification threshold is set to 0.70, then true cracks are output; if it is only 0.55, then noise is output. This verification method emphasizes whether the skeleton is continuous, whether the normal grayscale forms a V-shaped decay, and whether the local simulation can reproduce the cross-section; therefore, many dark spots that resemble cracks but lack true depth decay characteristics will be eliminated at this stage.

[0052] As an anomaly handling strategy, if the candidate region skeleton extraction fails, it indicates that the region structure is too discrete or the area is too small. The system can directly judge it as noise or require a retake. If the region boundary is touched when sampling along the skeleton normal, the system can shorten the profile length and record the decrease in confidence. If multiple local simulations are all above the verification threshold, the output of the group with the highest coupling degree is selected, and its parameters are written to the log for subsequent updates; if all local simulations are below the verification threshold, it means that the candidate region does not meet the characteristics of the crack local light valley and should be classified as noise. For example, in region R3 of the second support on the left span of the third span, the system extracts a local skeleton about 11 pixels long and establishes normal profiles at its three key locations; Both of the cross-sections showed a sharp darkening at the center and a rapid rebound on both sides. The system performed local simulation with a parameter set of 2 pixels wide, 4 levels deep, and 1.18 fractal dimension. The resulting residual submatrix was coupled with the real residual submatrix again and exceeded the verification threshold. Therefore, R3 was upgraded from a pending sample to a real crack. The purpose of this step is to re-verify critical samples in a smaller, more refined space to reduce misjudgments caused by ambiguous areas.

[0053] In a preferred embodiment of the present invention, the update method of the feedback update unit includes: generating a judgment log based on the true crack judgment result and the noise judgment result; and correcting the high threshold, the low threshold and the verification threshold based on the judgment log. Based on the judgment log, the parameter mapping relationship between the crack width, crack depth, fractal dimension and theoretical photometric attenuation model is corrected; the corrected threshold strategy and simulation parameter library are written into the system configuration.

[0054] This embodiment provides a feedback update mechanism. Specifically, if the threshold and simulation parameters are fixed for a long time, the system may experience a decision shift under new operating conditions due to support aging, camera updates, or seasonal changes in lighting. Therefore, this embodiment generates a decision log based on the results of each inspection and iteratively corrects the threshold strategy and parameter mapping relationship accordingly. Specifically, the system generates one or more log records for each judgment completed. The log records include at least the detection time, support number, candidate area location, initial coupling degree, verification coupling degree, final result, and the parameter set used. A very simple log set can be used to illustrate this: the first record shows that the initial coupling degree of a certain area is 0.85, and it is finally judged to be a true crack. The second record shows that the initial coupling degree of a certain area is 0.47, which is 0.52 after verification and is finally determined to be noise; the third record shows that the initial coupling degree of a certain area is 0.79, which is 0.73 after verification and is finally determined to be a real crack; after accumulating a certain number of logs, the system can statistically analyze the coupling degree distribution between real cracks and noise samples. If a large number of genuine crack samples are found to be concentrated between 0.78 and 0.81, and the current high threshold is set to 0.82, it indicates that the high threshold is too strict, which may cause too many genuine cracks to enter the review or be delayed in confirmation. In this case, the high threshold can be appropriately lowered. If a large number of noisy samples are found to be hovering between 0.43 and 0.48, the low threshold and review threshold can be adjusted as appropriate. Similarly, if the logs show that a certain type of real crack always requires a larger depth parameter in the simulation to match the actual grayscale attenuation, the system can correct the mapping relationship between crack depth and photometric attenuation, making subsequent simulations more closely resemble the actual situation on site.

[0055] Furthermore, the updated threshold strategy and simulation parameter library will be written into the system configuration file or database configuration table so that they can be directly called in the next inspection; in this way, the system is not a static template, but can be continuously calibrated around the actual detection data of the bridge bearings. As an anomaly handling strategy, if the number of logs is insufficient at a certain time, the system will not automatically update to prevent small sample fluctuations from causing threshold distortion. If there are conflicts in the logs due to manual review and annotation, only high-confidence samples can be used for update calculation. If the difference between the old and new parameters is too large after an update, exceeding the preset safety limit, the system requires manual approval before writing it into the formal configuration to avoid abnormal data causing the overall strategy to drift. After multiple rounds of inspections of the bridge over six months, the system accumulated hundreds of logs for the same type of plate rubber bearing. Statistics showed that the coupling degree of real crack samples was slightly lower in the wet season than in the dry season, while some shallow water stain samples showed a slight increase in the initial coupling stage. Therefore, the system fine-tuned the low threshold from 0.45 to 0.48 and the review threshold from 0.70 to 0.72, while updating the mapping table between shallow crack width and grayscale attenuation. After that, the number of false alarms in the new round of inspections decreased significantly. The purpose of this step is to enable the system to gradually approximate the actual characteristics of the target bridge, target camera, and target environment during long-term operation, and to maintain the adaptability of the detection parameters.

[0056] In a preferred embodiment of the present invention, the target surface is the surface of a bridge bearing, the original grayscale image matrix contains surface image data collected under conditions of low light, humidity, dust or shadow, and the noise determination result includes at least water stains, dust, light and shadow edges and non-target structural surface textures, the non-target structural surface textures including surface aging cracks.

[0057] This embodiment provides a noise classification application mechanism for the complex field environment of bridge bearings. Specifically, although the aforementioned scheme can theoretically distinguish between cracked and non-cracked surfaces, the bridge bearing scenario has obvious industry-specific characteristics: first, weak light is common; second, rubber surfaces are easily affected by moisture, dust accumulation, and aging; and third, the obstruction of structures near the bearings often forms shadow edges. Therefore, this embodiment limits the target surface to the surface of the bridge bearing and clearly defines typical noise sources. Specifically, the original grayscale image matrix processed by the system can come from nighttime inspections, backlit areas under bridges, wet conditions after rain, or dusty environments. For these images, the noise judgment result output by the system is not simply marked as non-crack, but can be further subdivided into several categories. For example, water stains usually show unstable boundaries, relatively gentle internal attenuation, and obvious local reflections; dust often shows discrete spots or rough blocky areas with poor continuity; the edges of light and shadow generally have long straight boundaries, but the normal grayscale profile does not show a narrow and deep V-shaped structure; although surface aging cracks may form fine line mesh textures, they are shallow in depth, have random directions, and lack the continuous shading characteristics of a single crack. A simplified diagram can be used to distinguish them; if the local profile of a candidate area is [92,85,82,84,91], the darkening is relatively gradual, and the profiles in adjacent directions differ greatly, it is more likely to be water stains; if the profile is [96,70,54,72,95], with a clear sharp valley in the middle that extends continuously along the skeleton, it is more consistent with structural cracks; if the area presents many short, intersecting line segments and lacks a single main skeleton, it is more likely to be aging cracks rather than the target crack.

[0058] In the actual system output, the noise classification results can be compared with the support number, image coordinates, and cutoff. Figure 1 And save it; for example, output the second support of the third span left, area R2, noise type: water stains; or output area R5, noise type: surface aging cracks; this is helpful for distinguishing subsequent maintenance strategies: water stains can be cleaned or reshot, shadow edges can be adjusted for supplementary lighting, and aging cracks can be included in material aging observation rather than structural crack alarms. As an anomaly handling strategy, if a noise area has multiple characteristics, such as dust superimposed on a damp surface, the system can output a composite noise label; if the confidence of the noise category is insufficient, it can be uniformly labeled as non-crack noise to avoid forced subdivision; if an aging crack shows significant deepening and continuous expansion in subsequent inspections, the system can upgrade it to a key tracking object, but it is still necessary to determine whether it has evolved into a target crack through the aforementioned coupling and verification process. For example, after the night inspection of the bridge, the system output the results for multiple dark areas of the second support of the left span of the third span: the central area R1 was identified as a real crack; the edge area R2 was classified as water stains due to the flat normal profile and obvious reflection drift. The upper area R4 has a long straight boundary but does not have the characteristics of a deep crack valley, and is classified as a light and shadow edge; the bottom corner area R6 has a fine net-like texture and is classified as a surface aging crack; based on this, the maintenance personnel only carried out structural risk treatment on R1, and the other areas were treated by cleaning, re-photographing or routine observation. The purpose of this step is to stably implement the system on the specific industrial object of bridge bearings, and to enable the output results to serve subsequent differentiated maintenance decisions, rather than just remaining at the abstract binary classification level.

[0059] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A bridge bearing crack detection system based on image processing, characterized in that, include: The data acquisition and calibration unit is used to acquire the original grayscale image matrix of the target surface, read the three-dimensional CAD geometric model, surface reflection parameters, camera spatial calibration parameters and ambient lighting parameters of the target surface, and map the three-dimensional CAD geometric model to the image pixel coordinate system based on the camera spatial calibration parameters to generate a calibration data package; An ideal photometric reconstruction unit is used to generate an ideal photometric matrix for a non-destructive surface based on the calibration data package and a preset photometric model based on Lambert reflection. The crack simulation unit is used to generate a theoretical damage simulation matrix based on the ideal photometric matrix, preset crack morphology parameters, and photometric attenuation rules. The dual-track differential extraction unit is used to differiate the original grayscale image matrix with the ideal photometric matrix to obtain the real residual matrix, and to differiate the theoretically damaged simulation matrix with the ideal photometric matrix to obtain the theoretical residual matrix; The coupling decision unit is used to calculate the coupling degree and output the crack determination result based on the structural similarity and / or gradient correlation between the actual residual matrix and the theoretical residual matrix; The feedback update unit is used to update the threshold strategy and simulation parameter library according to the crack determination result. The threshold strategy includes at least a high threshold, a low threshold, and a verification threshold. The simulation parameter library includes at least crack width, crack depth, and fractal dimension.

2. The image processing-based bridge bearing crack detection system according to claim 1, characterized in that, The methods for generating the calibration data packet include: The original grayscale image matrix is ​​obtained by taking multiple images of the same target surface and then registering and fusing them. Read the 3D CAD geometric model and surface reflection parameters of the target surface; Obtain camera spatial calibration parameters and ambient lighting parameters; Based on the camera spatial calibration parameters, the 3D CAD geometric model is mapped to the image pixel coordinate system, and the surface reflection parameters and the ambient lighting parameters are associated with the corresponding pixel positions to generate the calibration data package.

3. The image processing-based bridge bearing crack detection system according to claim 2, characterized in that, The ideal photometric matrix is ​​constructed in the following ways: Based on the calibration data packet, a preset photometric model based on Lambert reflection is invoked; The 3D CAD geometric model is projected onto the current viewpoint determined based on the camera spatial calibration parameters to generate a pixel-level surface normal vector field. By combining the surface reflection parameters and the ambient lighting parameters, the theoretical reference gray value of each pixel is calculated; The ideal photometric matrix is ​​generated based on the pixel-level surface normal field and the theoretical reference gray value.

4. The image processing-based bridge bearing crack detection system according to claim 3, characterized in that, The methods for generating the theoretically damaged simulation matrix include: Based on the ideal photometric matrix, the preset crack mechanical morphology modeling module is invoked; The crack width, crack depth, and fractal dimension are read from the simulation parameter library as the preset crack morphology parameters, wherein the fractal dimension is used to characterize the tortuosity of the crack path and the edge complexity. Construct a crack geometric skeleton based on the crack width, the crack depth, and the fractal dimension; Based on the crack geometry skeleton, perform crack interior shading calculation and photometric attenuation calculation based on the photometric attenuation rule; The photometric attenuation results corresponding to the crack geometric skeleton are superimposed on the ideal photometric matrix to generate the theoretical damage simulation matrix.

5. The image processing-based bridge bearing crack detection system according to claim 4, characterized in that, The methods for generating the actual residual matrix and the theoretical residual matrix include: The original grayscale image matrix is ​​differiated from the ideal luminance matrix to generate the real residual matrix; The theoretical damage simulation matrix is ​​subtracted from the ideal photometric matrix to generate the theoretical residual matrix; High-frequency gradient extraction is performed on the real residual matrix to generate a real gradient field; High-frequency gradient extraction is performed on the theoretical residual matrix to generate a theoretical gradient field.

6. The image processing-based bridge bearing crack detection system according to claim 5, characterized in that, The calculation method for the coupling degree includes: Based on the actual gradient field and the theoretical gradient field, calculate the two-dimensional cross-correlation value; Based on the actual residual matrix and the theoretical residual matrix, the structural similarity value is calculated; The two-dimensional cross-correlation value and the structural similarity value are normalized respectively, and then weighted and fused according to preset weights to generate the coupling degree.

7. The image processing-based bridge bearing crack detection system according to claim 6, characterized in that, The output method for the crack determination result includes: The high threshold is greater than the low threshold, and the review threshold is greater than the low threshold and less than the high threshold; When the coupling degree is higher than the high threshold, the true crack determination result is output; When the coupling degree is lower than the low threshold, output the noise determination result; When the coupling degree is higher than or equal to the low threshold and lower than or equal to the high threshold, the corresponding image region is extracted as a candidate region, the preset verification rules are called to perform a secondary simulation comparison on the candidate region, and the verification judgment result is output.

8. The image processing-based bridge bearing crack detection system according to claim 7, characterized in that, The execution methods of the review rules include: Extract the local morphological skeleton of the candidate region and the local grayscale attenuation curve along the normal direction of the local morphological skeleton; The crack width, crack depth, and fractal dimension from the simulation parameter library are used as candidate crack parameter groups to perform local damage simulation on the candidate region; The simulation results of local damage are coupled and calculated again with the real residual submatrix corresponding to the candidate region extracted from the real residual matrix; When the result of the re-coupling calculation is higher than the verification threshold, the result of the true crack determination is output; When the result of the re-coupling calculation is lower than or equal to the verification threshold, the noise judgment result is output.

9. The image processing-based bridge bearing crack detection system according to claim 8, characterized in that, The update methods of the feedback update unit include: A judgment log is generated based on the true crack determination result and the noise determination result; Based on the determination log, the high threshold, the low threshold, and the review threshold are corrected; Based on the judgment log, the parameter mapping relationship between the crack width, the crack depth, the fractal dimension, and the theoretical photometric attenuation model is corrected. Write the revised threshold strategy and simulation parameter library into the system configuration.

10. The image processing-based bridge bearing crack detection system according to claim 9, characterized in that, The target surface is the bridge bearing surface. The original grayscale image matrix contains surface image data collected under conditions of low light, humidity, dust, or shadow. The noise determination result includes at least water stains, dust, light and shadow edges, and non-target structural surface textures. The non-target structural surface textures include surface aging cracks.

Images