Bridge structure damage intelligent identification method and system based on machine vision

Abstract

This invention discloses a machine vision-based intelligent identification method and system for bridge structural damage, specifically relating to the field of image recognition technology. It addresses the problem that existing methods struggle to handle the issues of biological occlusion and discontinuous distribution in mountainous bridge damage areas, leading to missed or false identifications. The method involves acquiring image sequences of the bridge surface, analyzing the correlation between the local density of epiphytic organisms and the surface's geometric curvature to identify occlusion areas, and then coupling the analysis of the growth direction of organisms within the occlusion area with the stress distribution direction of the bridge structure to determine suspected damage guidance areas. Furthermore, it adaptively determines the feature spatial scale by analyzing the inflection points of the multi-scale fractal dimension sequence of the guidance area image texture and removes biological occlusion to obtain candidate damage areas. Subsequently, it constructs a spatial relationship graph for the candidate areas to associate discontinuous features and form complete damage areas. Finally, it identifies the damage type based on the features of the complete areas.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and more specifically, to a method and system for intelligent identification of bridge structural damage based on machine vision. Background Technology

[0002] In the field of bridge structural safety operation and maintenance, machine vision technology has advantages such as non-contact operation, high efficiency, and the ability to collect and analyze data in batches. It is particularly suitable for applying machine vision technology to damage identification of bridge structures in mountainous areas. Mountainous bridges are affected by terrain, climate, and load conditions, resulting in complex types of structural damage. The humid and bacteria-prone environment easily fosters the growth of epiphytes such as moss and lichen in damaged areas. Furthermore, complex loads can easily lead to compound damage. These characteristics place higher demands on the applicability of machine vision recognition technology. Existing intelligent bridge structural damage identification methods based on machine vision typically involve acquiring images of the bridge structure, preprocessing the images, and extracting visual features of the damage to complete the damage identification. Their core relies on the capture and analysis of clear visual features of the damaged area, and they have already achieved relatively mature applications in bridge scenarios with relatively simple environments such as plains.

[0003] Existing bridge structural damage identification methods struggle to effectively address multi-level visual occlusion and discontinuous distribution of complex damage in mountainous bridge areas. They cannot accurately separate the visual features of occluders from the damage itself, nor can they effectively correlate discontinuous damage features. This results in missed and false positives in damage identification, failing to meet the accuracy requirements for bridge structural damage identification in mountainous areas. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, the present invention provides a machine vision-based intelligent identification method and system for bridge structure damage to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] Intelligent identification methods for bridge structural damage based on machine vision include:

[0007] S1. Obtain image sequences of the surface of bridge structures in mountainous areas;

[0008] S2. Calculate the local density of epiphytic organisms covering the surface of bridge structures in mountainous areas based on image sequences and determine the geometric curvature of the bridge structure surface in mountainous areas. By analyzing the correlation abrupt change points between local density and geometric curvature, determine whether there are epiphytic organism shading areas in the image sequence.

[0009] S3. When there is an area covered by epiphytes, the suspected damage-guiding area can be determined by analyzing the coupling relationship between the growth direction of the epiphytes in the area covered by epiphytes and the stress distribution direction of the bridge structure.

[0010] S4. The feature space scale is determined by analyzing the abrupt change in the multi-scale fractal dimension sequence of the texture of the image sequence corresponding to the suspected damage-guided region, and the epiphyte occlusion region is stripped under the feature space scale to obtain the damage ontology candidate region.

[0011] S5. Associate discontinuously distributed damage features with the candidate damage body region to form a complete damage region;

[0012] S6. Identify the type of bridge structural damage based on the complete damage area.

[0013] Furthermore, S1 includes:

[0014] A set of digital images with overlapping areas on the surface of a bridge structure in a mountainous area are collected along a preset inspection path using a stereo vision system mounted on a mobile platform.

[0015] The shooting position and attitude parameters of each frame of digital image are recorded synchronously to form an image sequence.

[0016] Furthermore, S2 includes:

[0017] Vegetation regions are obtained by segmenting single-frame digital images in an image sequence based on color space thresholds.

[0018] Within the vegetated area, the percentage of vegetation pixels is counted using a sliding window method to generate a local density map of epiphyte cover.

[0019] Meanwhile, based on the image sequence and shooting position and posture parameters, a triangular mesh model of the surface of the bridge structure in the mountainous area is obtained through stereo vision 3D reconstruction, and the Gaussian curvature at each vertex of the triangular mesh model is calculated to generate a geometric curvature distribution map.

[0020] On the local density map and geometric curvature distribution map of spatial registration, the Pearson correlation coefficient between the local density value and the geometric curvature value is calculated by sliding window, and the center point of the window where the absolute value of the Pearson correlation coefficient is lower than the preset correlation coefficient threshold is determined as the correlation mutation point;

[0021] If a relevant mutation point exists, it is determined that there is an area covered by epiphytic organisms in the image sequence.

[0022] Furthermore, S3 includes:

[0023] Within the area occluded by epiphytes, the main direction of image texture is extracted based on the gray-level co-occurrence matrix as the growth direction of the epiphytes;

[0024] Obtain the stress distribution direction of the bridge structure based on the bridge design model and standard load conditions obtained in advance through finite element analysis;

[0025] After spatial registration, the directional consistency metric between the growth direction and the stress distribution direction is calculated;

[0026] Within the area covered by epiphytic organisms, local areas where the directional consistency measurement exceeds a preset consistency threshold are identified as suspected damage-guided areas.

[0027] Furthermore, S4 includes:

[0028] Extract grayscale images from the image sequence corresponding to the suspected damage guidance area;

[0029] A multi-scale Gaussian pyramid decomposition is performed on the grayscale image, and the fractal dimension of the image is calculated based on the box counting method at each scale to obtain a multi-scale fractal dimension sequence.

[0030] A change point detection algorithm is applied to a multi-scale fractal dimension sequence to identify abrupt inflection points where the fractal dimension values ​​change significantly, and the scale parameter corresponding to the abrupt inflection point is determined as the feature space scale.

[0031] At the feature space scale, morphological opening operations are performed on the suspected damage-guided region using structuring elements constructed based on scale parameters to separate the epiphyte-shading region and obtain the damage ontology candidate region.

[0032] Furthermore, the grayscale image is decomposed into a multi-scale Gaussian pyramid. At each scale, the fractal dimension of the image is calculated based on box counting to obtain a multi-scale fractal dimension sequence. This includes: using the original grayscale image as the base layer, Gaussian blurring and fixed-rate downsampling are performed sequentially to construct an image Gaussian pyramid containing multiple scales; for each scale of the image in the pyramid, a series of grids with increasing side lengths are used to cover the image, and the number of grids containing the target texture under each grid side length is counted; by fitting the linear relationship between the logarithm of the grid side length and the logarithm of the number of grids, the slope is the fractal dimension at the corresponding scale; and the fractal dimension values ​​at all scales are collected to form a multi-scale fractal dimension sequence.

[0033] Furthermore, S5 includes:

[0034] Extract the centroid coordinates and principal axis directions of each candidate damage body region;

[0035] Calculate the Euclidean distance and relative azimuth between any two candidate regions of the damage body based on the centroid coordinates;

[0036] A spatial relationship graph constrained by distance threshold and azimuth consistency is constructed, and candidate regions of damage ontology that satisfy the constraints are determined to be related;

[0037] Regional clustering is performed based on spatial relationship graphs to merge interconnected candidate damage regions into complete damage regions.

[0038] Furthermore, a spatial relationship graph constrained by distance thresholds and azimuth consistency is constructed. Damage candidate regions that satisfy the constraints are determined to be associated. This includes: calculating the centroid coordinates of each damage candidate region and the principal axis direction of its smallest bounding rectangle; calculating the Euclidean distance between region pairs based on the centroid coordinates, and calculating the directional angle between region pairs based on the principal axis direction; when the Euclidean distance between region pairs is less than a preset distance threshold and their directional angle is less than a preset angle threshold, the corresponding region pair is determined to satisfy the azimuth consistency constraint; all region pairs that satisfy the constraints are added as edges to the spatial relationship graph, thereby completing the association determination.

[0039] Furthermore, S6 includes:

[0040] Extract the damage feature vector from the complete damaged region, which consists of the region area, perimeter, rectangularity, and texture contrast based on the gray-level co-occurrence matrix.

[0041] Input the damage feature vector into the support vector machine classification model that has been pre-trained with samples;

[0042] The matching degree between the damage feature vector and the preset typical damage type feature pattern is calculated using a support vector machine classification model.

[0043] Based on the characteristic pattern of the typical damage type with the highest matching degree, the bridge structure damage type corresponding to the complete damage area is output.

[0044] On the other hand, the present invention provides a machine vision-based intelligent identification system for bridge structural damage, comprising:

[0045] The image acquisition module is used to acquire image sequences of the surface of bridge structures in mountainous areas;

[0046] The occlusion determination module is used to calculate the local density of epiphytic organisms covering the surface of bridge structures in mountainous areas based on image sequences and determine the geometric curvature of the bridge structure surface in mountainous areas. By analyzing the correlation abrupt change points between local density and geometric curvature, it determines whether there are epiphytic occlusion areas in the image sequence.

[0047] The region determination module is used to identify suspected damage-guiding areas by analyzing the coupling relationship between the growth direction of epiphytes in the epiphyte-shaded area and the stress distribution direction of the bridge structure.

[0048] The candidate extraction module is used to determine the feature space scale by analyzing the mutation inflection point of the multi-scale fractal dimension sequence of the texture of the image sequence corresponding to the suspected damage-guided region, and to peel off the epiphyseal area under the feature space scale to obtain the damage body candidate region.

[0049] The region completion module is used to associate discontinuously distributed damage features with candidate regions of the damage ontology to form complete damage regions.

[0050] The type identification module is used to identify the type of damage to the bridge structure based on the complete damage area.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] 1. By constructing a progressive processing framework from occlusion detection, precise localization, feature stripping to damage correlation, this study effectively overcomes the core difficulties in damage identification of bridges in mountainous areas caused by biological cover and discontinuous distribution. It utilizes abrupt changes in the correlation between epiphytic cover density and bridge surface geometric curvature to indicate potential occlusion areas, providing accurate initial targets for subsequent analysis. Furthermore, by coupling biological growth direction with structural mechanical properties, it achieves deep reasoning and precise localization of damage locations under complex visual interference. Finally, through multi-scale texture analysis and spatial relationship modeling, it adaptively strips away occlusion and aggregates fragmented features to reconstruct the complete damage morphology, transforming environmental interference factors into analytical clues and achieving a more essential and robust identification of the damage ontology.

[0053] 2. By analyzing the broken correlation between density and curvature, abnormal areas covered by organisms can be more reliably identified, reducing missed detections caused by similar colors and textures. The introduction of coupled analysis between growth direction and structural stress direction provides a mechanical basis for damage localization beyond surface vision, improving the accuracy of the guidance area. An adaptive scale selection method based on multi-scale fractal dimension sequence mutation can find the optimal separation scale for different texture differences between damages and occluders, achieving more precise visual feature stripping. A spatial relationship graph based on distance and orientation constraints is constructed to associate discrete damage candidate regions, effectively restoring the overall outline of composite damage fractured due to occlusion. Finally, type identification is performed based on the multi-dimensional features of the complete damage region, improving the reliability of the classification results. Attached Figure Description

[0054] Figure 1 This is a flowchart of the intelligent bridge structure damage identification method based on machine vision according to the present invention;

[0055] Figure 2 This is a schematic diagram of the intelligent bridge structure damage recognition system based on machine vision according to the present invention. Detailed Implementation

[0056] 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.

[0057] Example 1: Figure 1 The present invention provides a machine vision-based intelligent identification method for bridge structural damage, comprising:

[0058] S1. Obtain image sequences of the surface of bridge structures in mountainous areas;

[0059] S2. Calculate the local density of epiphytic organisms covering the surface of bridge structures in mountainous areas based on image sequences and determine the geometric curvature of the bridge structure surface in mountainous areas. By analyzing the correlation abrupt change points between local density and geometric curvature, determine whether there are epiphytic organism shading areas in the image sequence.

[0060] S3. When there is an area covered by epiphytes, the suspected damage-guiding area can be determined by analyzing the coupling relationship between the growth direction of the epiphytes in the area covered by epiphytes and the stress distribution direction of the bridge structure.

[0061] S4. The feature space scale is determined by analyzing the abrupt change in the multi-scale fractal dimension sequence of the texture of the image sequence corresponding to the suspected damage-guided region, and the epiphyte occlusion region is stripped under the feature space scale to obtain the damage ontology candidate region.

[0062] S5. Associate discontinuously distributed damage features with the candidate damage body region to form a complete damage region;

[0063] S6. Identify the type of bridge structural damage based on the complete damage area.

[0064] In step S1, the specific implementation process for obtaining the image sequence of the surface of the bridge structure in the mountainous area is as follows:

[0065] A hardware system for image acquisition is deployed, comprising a stereo vision system stably mounted on a mobile platform, which is a bridge inspection vehicle. The stereo vision system consists of two industrial cameras of identical model and parameters, such as color cameras with a resolution of 4096×2160 pixels. The two cameras are held in a fixed relative position by a rigid bracket. The optical axes of the two cameras are parallel and on the same horizontal plane, and the horizontal distance between them, i.e., the baseline distance, is set to 0.5 meters. The focal length of the camera lens is selected based on the typical working distance between the bridge surface and the mobile platform; for example, when the working distance is approximately 5 meters, a fixed-focus lens with a focal length of 16 mm can be used. Before image acquisition begins, the stereo vision system is calibrated to obtain its internal and external parameters. The internal parameters include the focal length, principal point coordinates, and lens distortion coefficients of each camera, while the external parameters refer to the rotation matrix and translation vector between the two cameras. The calibration process adopts the Zhang Zhengyou calibration method, using a black and white checkerboard calibration board of known size. Multiple images of the calibration board are taken at different positions and orientations, and the internal and external parameters of the stereo vision system are calculated using the calibration algorithm.

[0066] A pre-defined inspection path is planned along the bridge structure, covering all key surface areas of the bridge to be inspected. During the movement of the mobile platform along this path, the stereo vision system is controlled to automatically trigger image capture at fixed time intervals, thereby acquiring a set of digital images of the bridge structure surface in the mountainous area. To ensure the integrity of subsequent 3D reconstruction, there is an overlap area on the bridge structure surface between two images acquired at adjacent trigger times. The area of ​​this overlap area is set to be greater than or equal to 80% of the area of ​​the bridge surface covered by a single frame. This overlap ratio requirement is met by controlling the moving speed of the mobile platform and the image acquisition frequency. For example, when the mobile platform moves at a constant speed of 0.2 meters per second and the image acquisition frequency is 2 frames per second, the overlap ratio requirement is guaranteed. Each acquired digital image frame is stored in a format containing red, green, and blue color channels.

[0067] While the stereo vision system acquires each frame of digital image, the system simultaneously records the corresponding shooting position and attitude parameters. The shooting position is obtained through a GPS receiver integrated into the mobile platform and rigidly connected to the stereo vision system. This GPS receiver provides real-time 3D coordinate data of the mobile platform in the global coordinate system. The attitude parameters are obtained through an inertial measurement unit (IMU) integrated into the mobile platform and rigidly connected to the stereo vision system. This IMU provides real-time three-axis attitude angle data of the mobile platform in the global coordinate system, namely pitch, roll, and yaw angles. To achieve strict synchronization, the image acquisition time of the stereo vision system, the position data recording time of the GPS receiver, and the attitude data recording time of the IMU are all triggered and timestamped by the same high-precision clock source. Each frame of digital image file is associated and bound to its corresponding timestamp, shooting position 3D coordinates, and attitude angle data through this timestamp. After completing the acquisition work along a preset inspection path, all digital image records arranged in chronological order with associated position and attitude parameters are summarized to form an image sequence for subsequent processing. The data structure of this image sequence is a list, where each record indexes a digital image file and its corresponding metadata, including timestamps, shooting locations, and pose parameters. Through the specific implementation described above, those skilled in the art can clearly and completely implement the image sequence acquisition process, and the acquired image sequence provides a raw data foundation with precise spatial correlation information for subsequent steps.

[0068] In step S2, the local density of epiphytic organisms covering the surface of the mountain bridge structure is calculated based on the image sequence, and the geometric curvature of the mountain bridge structure surface is determined. By analyzing the correlation abrupt change points between local density and geometric curvature, the specific implementation process for determining whether there is an epiphytic organism-covered area in the image sequence is as follows:

[0069] Single-frame digital images in an image sequence are processed for vegetation region segmentation. For each frame containing red, green, and blue color channels, the digital image is converted from the red-green-blue color space to a hue-saturation-lightness color space. Within this hue-saturation-lightness color space, threshold ranges are set to distinguish between vegetation and non-vegetation. These threshold ranges are based on color statistical analysis of epiphytic samples from typical bridge surfaces. For example, by analyzing the clustering intervals of vegetation pixels in the sample image along the hue component, the threshold range for the hue component is set to 80 to 160; by analyzing the clustering intervals of vegetation pixels in the sample image along the saturation component, the threshold range for the saturation component is set to 50 to 255; and by analyzing the clustering intervals of vegetation pixels in the sample image along the lightness component, the threshold range for the lightness component is set to 40 to 200. For each pixel in the image, if the hue component value, saturation component value, and lightness component value all fall within their respective threshold ranges, the pixel is classified as a vegetation pixel. The connected region consisting of all pixels classified as vegetation in a frame of a digital image is defined as the vegetation region of that frame of the digital image.

[0070] After obtaining the vegetation area, a local density map of epiphyte cover is generated. For the digital image containing the vegetation area, a fixed-size sliding window is defined. The size of the sliding window is set according to the size of typical epiphyte patches in the bridge surface image to ensure effective reflection of local density changes. For example, the sliding window is set to a square window with a side length of 41 pixels. The sliding window traverses the entire frame of the digital image from left to right and from top to bottom with a set step size. The step size needs to balance computational efficiency and spatial resolution of the density map. For example, the step size is set to 10 pixels. At the position of each sliding window, the number of vegetation pixels belonging to the vegetation area within the sliding window is counted, and the ratio of the number of vegetation pixels to the total number of pixels in the sliding window is calculated. This ratio is the proportion of vegetation pixels corresponding to the center position of the sliding window. After traversing the entire frame of the digital image, a two-dimensional matrix corresponding to the spatial resolution of the original image is obtained. The value of each element in the matrix is ​​the proportion of vegetation pixels at the image position corresponding to the element. This two-dimensional matrix is ​​the local density map of epiphyte cover.

[0071] Based on image sequences and shooting position and pose parameters, a triangular mesh model of the mountain bridge structure surface is obtained through stereo vision 3D reconstruction. Two adjacent digital images with sufficient overlap are selected from the image sequence as stereo pairs. For each stereo pair, the two digital images are stereo-corrected using the internal and external parameters of the stereo vision system obtained in step S1, ensuring row alignment. A semi-global stereo matching algorithm is used to calculate the disparity map on the stereo-corrected image. The semi-global stereo matching algorithm involves parameters including minimum disparity, maximum disparity, matching window size, and penalty coefficients. These parameters are set based on the baseline distance of the stereo vision system, the lens focal length, and the expected depth of field range of the bridge surface. For example, the minimum disparity is set to 0, the maximum disparity is set based on the disparity corresponding to the maximum measurable distance calculated from the baseline distance and focal length (e.g., 256), the matching window size is set to 9×9 pixels based on the image texture richness, the penalty coefficient P1 is set to 8, and the penalty coefficient P2 is set to 32. Based on the disparity map and the baseline distance and focal length parameters of the stereo vision system, a 3D point cloud of the scene is calculated and generated using the principle of triangulation. The 3D point clouds generated from all stereo pairs in the image sequence are combined with the shooting position and pose parameters of the 3D point clouds, transformed to a unified global coordinate system, and fused to form a complete 3D point cloud covering the bridge surface. Normal estimation and Poisson surface reconstruction are performed on the complete 3D point cloud to generate a continuous triangular mesh model of the mountain bridge structure surface.

[0072] The Gaussian curvature at each vertex of the triangular mesh model is calculated to generate a geometric curvature distribution map. For any vertex in the triangular mesh model, all adjacent triangular faces of the vertex are identified. Based on the 3D coordinates of the vertex and its adjacent vertices, the Gaussian curvature at the vertex is estimated using a local surface fitting method. Specifically, the local surface fitting calculation method involves projecting the vertex and its first-order neighboring vertices onto a tangent plane, fitting a quadratic surface, and obtaining the Gaussian curvature value by multiplying the two principal curvatures of the fitted surface. This calculation is performed on each vertex in the triangular mesh model to obtain the Gaussian curvature value for each vertex. The Gaussian curvature values ​​of the 3D vertices are then mapped onto a 2D image plane. Using the shooting position and pose parameters of the image sequence and the camera imaging model, each 3D vertex of the triangular mesh model is back-projected onto the pixel coordinates of each frame of digital image. For each frame of digital image, the Gaussian curvature values ​​of all 3D vertices corresponding to the pixel coordinate position are averaged to obtain the geometric curvature value of the pixel position. If there is no corresponding 3D vertex projection at a pixel position, the geometric curvature values ​​of surrounding pixels are interpolated and filled using bilinear interpolation. After traversing all pixels of the entire frame of digital image, a two-dimensional matrix with the same spatial resolution as the digital image is obtained. The value of each element in the matrix is ​​the geometric curvature value at the corresponding image position. This two-dimensional matrix is ​​the geometric curvature distribution map.

[0073] On the spatially registered local density map and geometric curvature distribution map, the Pearson correlation coefficient between local density values ​​and geometric curvature values ​​is calculated using a sliding window. Since both the local density map and geometric curvature distribution map correspond to the pixel coordinate system of the same frame of digital image, they are naturally registered spatially. A fixed-size sliding window is defined, large enough to contain the local statistical features used to calculate the correlation; for example, a square window with sides of 51 pixels. The sliding window traverses the entire frame of digital image from left to right and from top to bottom with a set step size, for example, 10 pixels. At each sliding window position, the percentage of all vegetation pixels in the local density map within the sliding window's coverage area is extracted, forming one set of values. Simultaneously, all geometric curvature values ​​at the same pixel position in the geometric curvature distribution map within the sliding window's coverage area are extracted, forming another set of values. The Pearson correlation coefficient between these two sets of values ​​is calculated. The Pearson correlation coefficient is calculated as follows: first, the average of each of the two sets is calculated; then, the deviation of each value in a set from the average of its respective set is calculated; next, the sum of the products of the deviations of the two sets is calculated; finally, the sum of these products is divided by the square root of the product of the sum of the squares of the deviations of the two sets. After traversing the entire frame of the digital image, a new two-dimensional matrix is ​​obtained, and the value of each element in the matrix is ​​the Pearson correlation coefficient at the center position of the corresponding sliding window.

[0074] The center point of the window where the absolute value of the Pearson correlation coefficient is lower than a preset correlation coefficient threshold is identified as a correlation abrupt change point. The preset correlation coefficient threshold is set based on the statistical characteristics of a large number of normal bridge surface samples. Specifically, multiple images of undamaged bridge surfaces with uniform biological coverage are acquired, and the Pearson correlation coefficient distribution between the local density and geometric curvature values ​​of the images is calculated. The 5th percentile of the Pearson correlation coefficient distribution is taken as the preset correlation coefficient threshold. For each value in the Pearson correlation coefficient matrix, the absolute value is taken. If the absolute value is less than the preset correlation coefficient threshold, the position of the center point of the corresponding sliding window in the image is identified as a correlation abrupt change point. If at least one correlation abrupt change point is detected in a frame of digital image, it is determined that there is an epiphyte-covered area in the corresponding region of the image sequence. The entire process from vegetation region segmentation to correlation abrupt change point identification is repeated for each frame of digital image in the image sequence, thereby completing the initial identification of epiphyte-covered areas on the surface of the entire mountain bridge structure.

[0075] In step S3, when there is an area covered by epiphytes, the specific implementation process for determining the suspected damage guiding area is as follows: By analyzing the coupling relationship between the growth direction of the epiphytes in the area covered by epiphytes and the stress distribution direction of the bridge structure, the suspected damage guiding area is determined.

[0076] The image regions identified as having epiphyte occlusion areas in step S2 are processed. For these epiphyte occlusion areas, grayscale images are extracted from the corresponding original digital images. Based on the grayscale images, a grayscale co-occurrence matrix (GCMM) is calculated within the epiphyte occlusion areas to extract the main direction of the image texture, which is then used as the growth direction of the epiphytes. When calculating the GCMM, spatial distance parameters and directional angle parameters for calculating pixel pairs need to be set. The spatial distance parameter is set according to the typical spatial periodicity of the epiphyte texture in the image; for example, it is set to 1 pixel. The directional angle parameters are selected from a set of uniformly distributed directions, such as 0 degrees, 45 degrees, 90 degrees, and 135 degrees. For each pixel position within the epiphyte occlusion area, within a local neighborhood window centered on the pixel position, the frequency of different grayscale pixel pairs is statistically analyzed according to the set spatial distance and directional angle parameters, thereby generating a GCMM corresponding to each directional angle parameter. For each generated GCMM with each directional angle parameter, the contrast characteristics of the GCMM are calculated. By comparing the contrast feature values ​​across all directional angle parameters, the directional angle parameter corresponding to the maximum contrast feature value is defined as the main direction of the image texture at the pixel location, i.e., the growth direction of the epiphyte. All pixel locations within the epiphyte's occlusion area are traversed, and a growth direction angle value is assigned to each pixel location, thereby generating a growth direction distribution map.

[0077] This process involves obtaining the stress distribution directions of a bridge structure based on a bridge design model and standard load conditions obtained through finite element analysis. The bridge design model is a precise three-dimensional geometric model created from the bridge's design drawings. Standard load conditions refer to various load combinations used in engineering practice to verify structural strength, such as structural self-weight dead load and vehicle lane loads. The bridge design model is imported into finite element analysis software, and loads and constraints are applied according to the standard load conditions to perform static calculations. After the calculations are completed, the stress state of each element or node of the bridge structure is extracted from the finite element analysis results. The direction of the maximum principal tensile stress at each location is calculated and defined as the stress distribution direction at that location. The stress distribution directions at all locations of the entire bridge structure are correlated with its three-dimensional coordinates to form a stress distribution direction dataset for the bridge structure. This dataset is a vector field closely associated with the three-dimensional bridge model.

[0078] After spatial registration, the directional consistency metric between the growth direction and the stress distribution direction is calculated. The purpose of spatial registration is to map the growth direction distribution map in the image pixel coordinate system and the stress distribution direction dataset of the bridge structure in the bridge's three-dimensional coordinate system to the same reference system. Using the shooting position and attitude parameters of each frame of digital image recorded in step S1, as well as the camera imaging model, a mapping relationship between the image pixel coordinates and the bridge's three-dimensional coordinates is established. For each pixel in the growth direction distribution map, the direction of the light ray corresponding to the pixel in the bridge's three-dimensional space is calculated using the camera's internal parameters, external parameters, and position and attitude. Combined with the triangular mesh model generated in step S2, the three-dimensional spatial position coordinates of the bridge surface observed by the pixel are determined. Based on the three-dimensional spatial position coordinates, the stress distribution direction at the three-dimensional spatial position coordinates is obtained through spatial interpolation in the pre-obtained stress distribution direction dataset of the bridge structure. After spatial registration, each pixel within the epiphyte's occlusion area simultaneously possesses a growth direction angle value and a stress distribution direction angle value. The directional consistency metric is defined by calculating the angle between the growth direction angle value and the stress distribution direction angle value. The included angle is calculated using a vector dot product method. The growth direction angle value is converted into a unit direction vector, and the stress distribution direction angle value is also converted into a unit direction vector. The dot product of these two unit direction vectors is then calculated. The dot product value ranges from -1 to +1. The closer the dot product value is to +1, the more consistent the growth direction and stress distribution direction are. The dot product value is directly used as a measure of directional consistency.

[0079] Within the epiphyte-covered area, local regions where the directional consistency metric exceeds a preset consistency threshold are identified as suspected damage-guided regions. The preset consistency threshold is a threshold value used to determine whether the growth direction and stress distribution direction are significantly consistent. The preset consistency threshold is set based on prior knowledge or statistical analysis of undamaged areas. Specifically, certain known intact areas without stress anomalies on the bridge structure surface are selected, and the statistical distribution of the consistency metric values ​​between the growth direction and the theoretical stress distribution direction under epiphyte coverage is calculated. The metric value corresponding to the 95th percentile of the statistical distribution is taken as the preset consistency threshold. After obtaining the directional consistency metric value for each pixel, a directional consistency metric map with the same spatial resolution as the growth direction distribution map is generated. Pixels with directional consistency metric values ​​greater than the preset consistency threshold in the directional consistency metric map are marked. Connectivity analysis is performed on the marked pixels, and spatially adjacent pixels with directional consistency metric values ​​all exceeding the preset consistency threshold are grouped together to form one or more connected local regions. These connected local regions are the suspected damage-guided regions.

[0080] In step S4, the feature space scale is determined by analyzing the abrupt change in inflection points of the multi-scale fractal dimension sequence of the texture of the image sequence corresponding to the suspected damage-guided region. The specific implementation process of peeling off the epiphyseal area under the feature space scale to obtain the candidate region of the damage body is as follows:

[0081] Process the suspected damage guidance regions identified in step S3. Extract grayscale images from the image sequence corresponding to the suspected damage guidance regions. For each suspected damage guidance region, find one or more frames of digital images covering the suspected damage guidance region in the original digital image sequence. Crop the image block containing the suspected damage guidance region from the digital image. Convert the color image block from the red-green-blue color space to a grayscale image. The grayscale conversion uses a weighted average method. The weight coefficients in the weighted average method are set based on the coefficients of the standard luminance formula in colorimetry. For example, the weight coefficient for the red channel is set to 0.299, the weight coefficient for the green channel is set to 0.587, and the weight coefficient for the blue channel is set to 0.114. Calculate the grayscale value of each pixel in the image block according to the formula: grayscale value equals 0.299 multiplied by the red channel value plus 0.587 multiplied by the green channel value plus 0.114 multiplied by the blue channel value, thus obtaining the grayscale image corresponding to the suspected damage guidance region.

[0082] A multi-scale Gaussian pyramid decomposition is performed on a grayscale image. At each scale, the fractal dimension of the image is calculated using box counting, resulting in a multi-scale fractal dimension sequence. The specific implementation process of multi-scale Gaussian pyramid decomposition is as follows: The original grayscale image is used as the base layer of the Gaussian pyramid, i.e., layer 0. When constructing the Gaussian pyramid, the size and variance parameter of the Gaussian blur convolution kernel, as well as the fixed downsampling factor, need to be set. The size of the Gaussian blur convolution kernel is set to an odd number based on the image size and the desired smoothness, for example, 5 pixels by 5 pixels. The variance parameter of the Gaussian blur convolution kernel is calculated from the kernel size, for example, set to 1.0. The fixed downsampling factor is usually set to 2. The construction process involves performing a Gaussian blur on the base layer grayscale image with specified parameters to obtain a smoothed image; then downsampling is performed on the smoothed image, i.e., taking one pixel every other pixel, resulting in an image with half the size, which serves as layer 1 of the Gaussian pyramid. The first layer image is repeatedly subjected to Gaussian blur and fixed-rate downsampling to obtain the second layer image. This process is repeated to construct an image Gaussian pyramid containing multiple scales. The number of pyramid layers is determined by the size of the original image and the minimum analyzable scale; for example, a Gaussian pyramid with four scales can be constructed.

[0083] For each scale of the Gaussian pyramid, box counting is used to calculate the fractal dimension of that scale. The specific implementation process of box counting is as follows: First, prepare a series of square grids with increasing side lengths. The grid side length sequence is a geometric sequence with a common ratio greater than 1. The common ratio is set to analyze texture scales across multiple orders of magnitude; for example, the common ratio is set to 2, and the grid side length sequence is set to 1 pixel, 2 pixels, 4 pixels, 8 pixels, 16 pixels, and 32 pixels. For each side length value in the grid side length sequence, cover the scale image with a square grid of that length. Count the minimum number of grids required to cover the entire scale image. The rule for counting the grids is that if a square grid contains at least one pixel with a gray value greater than the global gray-level mean threshold of the image, then that grid is counted. The global gray-level mean threshold of the image is obtained by calculating the arithmetic mean of the gray values ​​of all pixels in the image at that scale. Traverse all side length values ​​in the grid side length sequence to obtain a set of corresponding grid count values.

[0084] By fitting a linear relationship between the logarithm of the grid side length and the logarithm of the grid number, the slope of this relationship represents the fractal dimension at the corresponding scale. Specifically, the calculation method involves taking the logarithm to base 10 for each side length value in the grid side length sequence, obtaining a logarithmic sequence of grid side length values; and taking the logarithm to base 10 for the corresponding grid number values, obtaining a logarithmic sequence of grid number values. The least squares method is used to linearly fit the logarithmic sequence of grid side length values ​​and the logarithmic sequence of grid number values. The least squares method solves for the line parameters by minimizing the sum of squared residuals. The fitted line represents the fractal dimension of the image at that scale. The fractal dimension values ​​at all scales in the Gaussian pyramid of the image are collected and arranged in order from finest to coarsest scale, forming a numerical sequence, which is the multi-scale fractal dimension sequence.

[0085] A change point detection algorithm is applied to multi-scale fractal dimension sequences to identify abrupt inflection points where the fractal dimension values ​​change significantly. The Bayesian change point detection algorithm can be used. The specific implementation process of the Bayesian change point detection algorithm is as follows: The multi-scale fractal dimension sequence is modeled as a Gaussian distribution with different means before and after the unknown point. Key parameters that need to be set for the algorithm include prior distribution parameters and the significance level of the hypothesis test. The prior distribution parameters can assume that the change point positions follow a uniform distribution on the sequence index, and the mean of the sequence segments follows a broad Gaussian prior distribution. For example, the mean of the Gaussian prior distribution can be set to the global mean of the multi-scale fractal dimension sequence, and the variance can be set to 10 times the global variance of the multi-scale fractal dimension sequence. The significance level of the hypothesis test is usually set according to conventional choices in statistics, for example, a significance level of 0.05. The algorithm calculates the posterior probability of each position in the sequence as a change point and identifies points with a posterior probability exceeding the preset significance level of 0.05 as potential abrupt inflection points. If multiple potential mutation inflection points are detected, the point with the highest posterior probability is selected as the final mutation inflection point. The mutation inflection point corresponds to its index number in the multi-scale fractal dimension sequence, which is directly related to the scale level of the image Gaussian pyramid. For example, if the multi-scale fractal dimension sequence contains 4 values, corresponding to levels 0 to 3 of the pyramid, and the detected mutation inflection point index is 2, then the pyramid scale corresponding to the mutation inflection point is level 2. The pyramid scale level corresponding to the mutation inflection point is determined as the feature space scale used for subsequent texture stripping.

[0086] At the feature space scale, morphological opening operations are performed on suspected damage-guided regions using structuring elements constructed based on scale parameters to separate epiphyte-occluded regions and obtain candidate damage bodies. Morphological opening is an operation of erosion followed by dilation. The shape and size of the structuring element need to be constructed according to the feature space scale. The shape of the structuring element is usually chosen to be circular or square to accommodate isotropic textures. The size of the structuring element is proportional to the feature space scale. For example, if the feature space scale corresponds to the k-th layer of the image's Gaussian pyramid, then the size S of the structuring element is calculated using the formula: S=2 k×S0. S0 is the base size, which is set based on the expected minimum size of the damage features in the original image. The expected minimum size can be obtained by analyzing the pixel statistics of the damage area in a historical damage image library; for example, S0 is set to 3 pixels. The erosion operation scans each pixel of the grayscale image with a structuring element, replacing the grayscale value of the center pixel with the minimum grayscale value of the pixels within the coverage area of ​​the structuring element. The dilation operation scans the eroded image with a structuring element, replacing the grayscale value of the center pixel with the maximum grayscale value of the pixels within the coverage area of ​​the structuring element. After the opening operation, fine textures smaller than the structuring element in the grayscale image are filtered out, while areas larger than the structuring element are retained. The image after the opening operation is binarized. The binarization method can use a global thresholding method, with the threshold set to the valley value of the grayscale histogram of the image after the opening operation. The binarized image is then logically ANDed with the mask image of the original suspected damage guidance area. The logical AND operation requires that the output is true only when the corresponding pixel values ​​of the two input images are both true. The resulting connected regions are the candidate regions of the separated damage body.

[0087] In step S5, the specific implementation process of associating discontinuously distributed damage features with the candidate damage body region to form a complete damage region is as follows:

[0088] The damage candidate regions obtained in step S4 are processed. The damage candidate regions are connected regions in a binary image. The centroid coordinates and principal axis direction of each damage candidate region are extracted. For each damage candidate region, the centroid coordinates are calculated by obtaining the coordinate set of all pixels constituting the region, calculating the arithmetic mean of the x-coordinates of all pixels in the coordinate set as the centroid x-coordinate, and calculating the arithmetic mean of the y-coordinates of all pixels in the coordinate set as the centroid y-coordinate. The principal axis direction is calculated by calculating the second-order central moment of the image for the damage candidate region. Specifically, the elements of the covariance matrix are calculated based on the pixel coordinates and centroid coordinates. The covariance matrix is ​​a 2x2 matrix. Eigenvalue decomposition is performed on the covariance matrix, and the direction of the eigenvector corresponding to the largest eigenvalue is the principal axis direction of the damage candidate region. The principal axis direction is represented by the angle between the principal axis and the positive x-axis of the image, with the angle ranging from 0 to 180 degrees.

[0089] The Euclidean distance and relative azimuth between any two candidate damage bodies are calculated based on centroid coordinates. For any two distinct candidate damage bodies, let the centroid coordinates of region A be Xa and Ya, and the centroid coordinates of region B be Xb and Yb. The Euclidean distance Dab between region A and region B is calculated using the formula: The calculation of relative azimuth angles requires considering the principal axis directions of both regions. Let θa be the principal axis direction angle of region A and θb be the principal axis direction angle of region B. The relative azimuth angle αab from region A to region B is defined by calculating the angle between the vector direction from the centroid of region A to the centroid of region B and the principal axis direction of region A itself. The direction angle φab of the vector from the centroid of region A to the centroid of region B is calculated using the formula: φab = arctan((Yb-Ya) / (Xb-Xa)). The result is converted to the range of 0 to 360 degrees according to the coordinate quadrant. The relative azimuth angle αab = |φab - θa|. If |φab - θa| is greater than 180 degrees, then |φab - θa| is subtracted from 360 degrees. The final value of αab is between 0 and 180 degrees.

[0090] A spatial relationship graph constrained by distance thresholds and azimuth consistency is constructed, and candidate damage bodies that satisfy the constraints are considered correlated. The distance threshold is a threshold value used to determine whether two regions are spatially close enough. The distance threshold is set based on prior knowledge of the dimensions of bridge structural components and the typical crack extension length. Specifically, the pixel distances between different segments belonging to the same crack in a historical bridge damage image database are analyzed, the statistical average of these pixel distances is calculated, and the distance threshold is set to 1.5 times this statistical average. Azimuth consistency is constrained by a preset angle threshold. The preset angle threshold is set based on the characteristic that cracks usually have consistency along their extension direction. By analyzing the distribution of the angle between the principal axes of adjacent segments in historical crack images, the 90th percentile is taken as the preset angle threshold. The spatial relationship graph is an undirected graph, where the set of vertices represents all candidate damage bodies, and the edges of the graph indicate the correlation between two regions. The process of determining association involves traversing all possible candidate pairs of damage entities. For a pair of regions A and B, if both of the following conditions are met, then regions A and B are considered to be associated: Condition 1, the Euclidean distance Dab between regions A and B is less than a preset distance threshold; Condition 2, the relative azimuth angle αab between regions A and B is less than a preset angle threshold. Each pair of regions that meets the conditions is added as an edge to the spatial relationship graph.

[0091] Region clustering is performed based on the spatial relationship graph to merge interconnected candidate damage bodies into complete damage regions. The region clustering uses a depth-first search algorithm to traverse the spatial relationship graph. The specific implementation process is as follows: An empty set is initialized to store the found complete damage regions. A marker array is initialized, with a length equal to the number of candidate damage bodies, to record whether each candidate vertex has been visited. For each unvisited vertex in the spatial relationship graph, a depth-first search is performed starting from that vertex. The depth-first search recursively visits all vertices directly or indirectly connected by edges and adds these vertices to the same temporary set. When a connected component has been searched, all candidate damage bodies in the temporary set are considered to belong to multiple fragments of the same complete damage. The bounding polygons of all regions in the temporary set are merged. The merging operation calculates the minimum convex hull that covers all pixels in the set; specifically, the Graham scan algorithm can be used to calculate the convex hull vertices. The region within the convex hull is defined as a complete damage region. The formed complete damage region is added to the set storing complete damage regions, and all vertices in the connected component are marked as visited. Continue traversing the spatial relationship graph to the next unvisited vertex, repeating the above process until all vertices have been visited. Ultimately, the set storing complete damaged regions contains one or more complete damaged regions.

[0092] In step S6, the specific implementation process for identifying the bridge structure damage type based on the intact damage area is as follows:

[0093] The complete damaged region formed in step S5 is processed. A damage feature vector is extracted from the complete damaged region. This vector consists of four feature values ​​in sequence: region area, perimeter, rectangularity, and texture contrast based on the gray-level co-occurrence matrix. The region area is calculated by counting the total number of pixels constituting the complete damaged region. The perimeter is calculated by first obtaining the boundary pixel sequence of the complete damaged region using a boundary tracking algorithm, then calculating and summing the straight-line distances between adjacent pixels in the sequence. The distance between adjacent pixels is 1 pixel, or approximately 1.414 pixels, and the sum is the perimeter. The rectangularity is calculated by first determining the smallest horizontal rectangle completely enclosing the region, and then calculating the ratio of the area of ​​the complete damaged region to the area of ​​the smallest horizontal rectangle.

[0094] Calculating texture contrast based on the gray-level co-occurrence matrix (GLCM) requires extracting gray-level image patches from corresponding regions of the original digital image. When calculating the GLCM of a gray-level image patch, a pixel distance parameter and an orientation angle parameter are set; for example, the pixel distance parameter is set to 1 and the orientation angle parameter to 0 degrees. Under the set pixel distance and orientation angle, all pixel pairs in the gray-level image patch that satisfy the positional relationship are counted. Each pair of pixels consists of two specific gray-level values. The frequency of each combination of gray-level values ​​is counted, thus forming the GLCM. Texture contrast is calculated based on the obtained GLCM. The calculation process is as follows: traverse each cell of the GLCM, multiply the frequency value recorded in the cell by the square of the difference between the row number and column number of the cell, and then sum the results of these products for all cells. The final sum is the texture contrast.

[0095] The extracted damage feature vectors are input into a pre-trained Support Vector Machine (SVM) classification model. The training process of the SVM classification model precedes the identification process and is as follows: A sample library of bridge images containing various known damage types and corresponding regions is collected. Four feature values, identical to those described above, are extracted from each sample region to construct a set of training feature vectors. The values ​​of each feature dimension in the set are standardized. Standardization involves calculating the mean and dispersion of all values ​​in that dimension, then subtracting the mean from each value and dividing by the dispersion. During training, a specific mathematical function is selected to map the feature vectors to a high-dimensional space to find the classification boundary. This function includes a width parameter controlling the mapping complexity, and the model itself has a penalty parameter controlling the tolerance for misclassification. By repeatedly dividing the training data into training and validation parts, the effects of different combinations of width and penalty parameters are tested, and the parameter combination that achieves the highest classification accuracy in the validation part is selected. Using the finally selected parameter combination, a classification model is built on all standardized training feature vectors, and all necessary parameters of the model are saved, including the mean and dispersion of each dimension used for standardization, the type and parameters of the mapping function, and the key vectors and their weights for determining the classification boundary.

[0096] The width parameter refers to the built-in parameter of the selected radial basis kernel function. This parameter directly affects the density of feature vector distribution in the mapping space and the shape of the classification boundary by adjusting the decay rate of the exponential function. Its specific value is determined by forming a parameter grid together with the penalty parameter within a preset range, and evaluating the classification accuracy under different parameter combinations on the validation set based on cross-validation. Finally, the parameter value that makes the validation set classification accuracy the highest is selected.

[0097] The pre-trained support vector machine classification model is obtained through the following specific process: First, a sample library of bridge surface images containing multiple accurately labeled damage types and corresponding regions is constructed. Damage types include at least cracks, spalling, and corrosion. Damage feature vectors, which are exactly the same as those defined in step S6, are extracted from each labeled region, namely, region area, perimeter, rectangularity, and gray-level co-occurrence matrix texture contrast calculated based on specific parameters, to form the original training feature set. Next, Z-score standardization is performed on each feature dimension of the training feature set, that is, the arithmetic mean and standard deviation of all feature values ​​in each dimension are calculated respectively, and then each feature value is subtracted from the mean of its respective dimension and divided by the standard deviation of that dimension. Then, a radial basis function is selected as the kernel function, and the penalty parameter and kernel width parameter of the model are optimized using the cross-validation grid search method. Specifically, the standardized feature set is randomly divided into multiple training-validation subset combinations, and training and validation are performed by traversing the preset parameter value range. The parameter combination with the highest average classification accuracy on the validation set is selected. Finally, the final support vector machine model is trained on the entire standardized training feature set using this optimal parameter combination, and all necessary parameters of the model are saved, including the standardized mean and standard deviation of each feature dimension, all support vectors, corresponding coefficients, selected kernel function type and parameters, and bias term. This yields a classification model that can be directly used for subsequent recognition.

[0098] During identification, the lesion feature vector to be identified is first standardized according to the average value and dispersion of each dimension saved during training. The standardized vector is then input into the loaded support vector machine classification model. The model calculates the propensity score for each preset lesion type corresponding to the input vector based on its internally stored decision function; this score is the matching degree. The specific formula for calculating the decision function is: f(x)=Σ(αi*yi*K(xi,x))+Bb, where f(x) is the decision function, x is the input feature vector, xi is the support vector saved by the model, αi is the corresponding coefficient, yi is the class label of the support vector, K is the kernel function, Bb is the bias term, and i is the index variable used to traverse all support vectors.

[0099] The matching scores of the feature vector of the complete damaged area with all preset damage types are compared, and the damage type with the highest score is determined as the damage type of the area. For example, the preset types include cracks, spalling, and corrosion. If the corresponding crack has the highest score, the damage is determined to be a crack. This determination result is output as the final identification conclusion of the damage.

[0100] Example 2: Figure 2 A schematic diagram of the intelligent bridge structural damage identification system based on machine vision of the present invention is provided. The intelligent bridge structural damage identification system based on machine vision includes:

[0101] The image acquisition module is used to acquire image sequences of the surface of bridge structures in mountainous areas;

[0102] The occlusion determination module is used to calculate the local density of epiphytic organisms covering the surface of bridge structures in mountainous areas based on image sequences and determine the geometric curvature of the bridge structure surface in mountainous areas. By analyzing the correlation abrupt change points between local density and geometric curvature, it determines whether there are epiphytic occlusion areas in the image sequence.

[0103] The region determination module is used to identify suspected damage-guiding areas by analyzing the coupling relationship between the growth direction of epiphytes in the epiphyte-shaded area and the stress distribution direction of the bridge structure.

[0104] The candidate extraction module is used to determine the feature space scale by analyzing the mutation inflection point of the multi-scale fractal dimension sequence of the texture of the image sequence corresponding to the suspected damage-guided region, and to peel off the epiphyseal area under the feature space scale to obtain the damage body candidate region.

[0105] The region completion module is used to associate discontinuously distributed damage features with candidate regions of the damage ontology to form complete damage regions.

[0106] The type identification module is used to identify the type of damage to the bridge structure based on the complete damage area.

[0107] The calculations involved in the embodiments are all dimensionless numerical calculations, and the preset parameters and thresholds in the calculations are set by those skilled in the art according to the actual situation.

[0108] It should be noted that this invention can be deployed on the device itself to realize embedded applications, or it can run on a PC or other terminal with a user interface, thereby meeting various hardware environments and usage requirements.

[0109] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions according to the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wireless or wired transmission; wired transmission methods include optical fiber, twisted pair, coaxial cable, etc.; wireless transmission includes infrared, microwave, etc. Computer-readable storage media can be any available medium that a computer can access or a data storage device such as a server or data center that contains one or more sets of available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0110] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0111] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0112] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0113] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0114] If a function is implemented as a software module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0115] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0116] In conclusion, the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent identification of bridge structure damage based on machine vision, characterized in that, include: S1. Obtain image sequences of the surface of bridge structures in mountainous areas; S2. Calculate the local density of epiphytic organisms covering the surface of bridge structures in mountainous areas based on image sequences and determine the geometric curvature of the bridge structure surface. By analyzing the abrupt changes in the correlation between local density and geometric curvature, determine whether there are epiphytic occlusion areas in the image sequences, including: Vegetation regions are obtained by segmenting single-frame digital images in an image sequence based on color space thresholds. Within the vegetated area, the percentage of vegetation pixels is counted using a sliding window method to generate a local density map of epiphyte cover. Meanwhile, based on the image sequence and shooting position and pose parameters, a triangular mesh model of the surface of the bridge structure in the mountainous area is obtained through stereo vision 3D reconstruction, and the Gaussian curvature at each vertex of the triangular mesh model is calculated to generate a geometric curvature distribution map. On the local density map and geometric curvature distribution map of spatial registration, the Pearson correlation coefficient between the local density value and the geometric curvature value is calculated by sliding window, and the center point of the window where the absolute value of the Pearson correlation coefficient is lower than the preset correlation coefficient threshold is determined as the correlation mutation point; If a relevant mutation point exists, it is determined that there is an area covered by epiphytic organisms in the image sequence; S3. When there is an area covered by epiphytes, the suspected damage-guiding area can be determined by analyzing the coupling relationship between the growth direction of the epiphytes in the area covered by epiphytes and the stress distribution direction of the bridge structure. S4. The feature space scale is determined by analyzing the abrupt change in inflection point of the multi-scale fractal dimension sequence of the texture of the image sequence corresponding to the suspected damage-guided region, and the epiphyte occlusion region is stripped under the feature space scale to obtain the damage ontology candidate region. S5. Associate discontinuously distributed damage features with the candidate damage body region to form a complete damage region; S6. Identify the type of bridge structural damage based on the complete damage area. 2.The method of claim 1, wherein, S1 includes: A set of digital images with overlapping areas on the surface of a bridge structure in a mountainous area are collected along a preset inspection path using a stereo vision system mounted on a mobile platform. The shooting position and attitude parameters of each frame of digital image are recorded synchronously to form an image sequence. 3.The method of claim 1, wherein, S3 includes: Within the area occluded by epiphytes, the main direction of image texture is extracted based on the gray-level co-occurrence matrix as the growth direction of the epiphytes; Obtain the stress distribution direction of the bridge structure based on the bridge design model and standard load conditions obtained in advance through finite element analysis; After spatial registration, the directional consistency metric between the growth direction and the stress distribution direction is calculated; Within the area covered by epiphytic organisms, local areas where the directional consistency measurement exceeds a preset consistency threshold are identified as suspected damage-guided areas. 4.The method of claim 1, wherein, S4 includes: Extract grayscale images from the image sequence corresponding to the suspected damage guidance area; A multi-scale Gaussian pyramid decomposition is performed on the grayscale image, and the fractal dimension of the image is calculated based on the box counting method at each scale to obtain a multi-scale fractal dimension sequence. A change point detection algorithm is applied to a multi-scale fractal dimension sequence to identify abrupt inflection points where the fractal dimension values ​​change significantly, and the scale parameter corresponding to the abrupt inflection point is determined as the feature space scale. At the feature space scale, morphological opening operations are performed on the suspected damage-guided region using structuring elements constructed based on scale parameters to separate the epiphyte-shading region and obtain the damage ontology candidate region. 5.The method of claim 4, wherein, A multi-scale Gaussian pyramid decomposition is performed on the grayscale image. The fractal dimension of the image is calculated based on box counting at each scale to obtain a multi-scale fractal dimension sequence. This process includes: using the original grayscale image as the base layer, Gaussian blurring and fixed-rate downsampling are performed sequentially to construct an image Gaussian pyramid containing multiple scales; for each scale of the image in the pyramid, a series of grids with increasing side lengths are used to cover the image, and the number of grids containing the target texture under each grid side length is counted; by fitting the linear relationship between the logarithm of the grid side length and the logarithm of the number of grids, the slope is the fractal dimension at the corresponding scale; and the fractal dimension values ​​at all scales are collected to form a multi-scale fractal dimension sequence. 6.The method of claim 1, wherein S5 include: Extract the centroid coordinates and principal axis directions of each candidate damage body region; Calculate the Euclidean distance and relative azimuth between any two candidate regions of the damage body based on the centroid coordinates; A spatial relationship graph constrained by distance threshold and azimuth consistency is constructed, and candidate regions of damage ontology that satisfy the constraints are determined to be related; Regional clustering is performed based on spatial relationship graphs to merge interconnected candidate damage regions into complete damage regions. 7.The method of claim 6, wherein, A spatial relationship graph constrained by distance thresholds and azimuth consistency is constructed. Damage candidate regions that satisfy the constraints are determined to be associated. This includes: calculating the centroid coordinates of each damage candidate region and the principal axis direction of its minimum bounding rectangle; calculating the Euclidean distance between region pairs based on the centroid coordinates and the directional angle between region pairs based on the principal axis direction; determining that the corresponding region pair satisfies the azimuth consistency constraint when the Euclidean distance between region pairs is less than a preset distance threshold and the directional angle is less than a preset angle threshold; and adding all region pairs that satisfy the constraints as edges to the spatial relationship graph to complete the association determination. 8.The method of claim 1, wherein S6 include: Extract the damage feature vector from the complete damaged region, which consists of the region area, perimeter, rectangularity, and texture contrast based on the gray-level co-occurrence matrix. Input the damage feature vector into the support vector machine classification model that has been pre-trained with samples; The matching degree between the damage feature vector and the preset typical damage type feature pattern is calculated using a support vector machine classification model. Based on the characteristic pattern of the typical damage type with the highest matching degree, the bridge structure damage type corresponding to the complete damage area is output.

9. A machine vision-based intelligent bridge structure damage identification system for implementing the machine vision-based intelligent bridge structure damage identification method according to any one of claims 1-8, characterized in that, include: The image acquisition module is used to acquire image sequences of the surface of bridge structures in mountainous areas; The occlusion determination module is used to calculate the local density of epiphytic organisms covering the surface of bridge structures in mountainous areas based on image sequences and determine the geometric curvature of the bridge structure surface in mountainous areas. By analyzing the correlation abrupt change points between local density and geometric curvature, it determines whether there are epiphytic occlusion areas in the image sequence. The region determination module is used to identify suspected damage-guiding areas by analyzing the coupling relationship between the growth direction of epiphytes in the epiphyte-shaded area and the stress distribution direction of the bridge structure when there is an epiphyte-shaded area. The candidate extraction module is used to determine the feature space scale by analyzing the mutation inflection point of the multi-scale fractal dimension sequence of the texture of the image sequence corresponding to the suspected damage-guided region, and to peel off the epiphyseal area under the feature space scale to obtain the damage body candidate region. The region completion module is used to associate discontinuously distributed damage features with candidate regions of the damage ontology to form complete damage regions. The type identification module is used to identify the type of damage to the bridge structure based on the complete damage area.

Images