Multi-scale image segmentation and damage assessment method for surface cracks of bridge structure

Abstract

This invention discloses a multi-scale image segmentation and damage assessment method for surface cracks in bridge structures, belonging to the field of bridge structural health monitoring technology. The method includes: a multi-scale pyramid feature preprocessing step to extract multi-resolution features from bridge surface images; an adaptive attention-guided crack segmentation step to enhance the response of crack regions through channel and spatial attention mechanisms; a precise quantification step of crack geometric parameters to calculate crack length, width, depth, and direction; a temporal comparison crack development trend prediction step to calculate the crack propagation rate and predict the development trend by dividing the parameter difference between current and historical detection data by the time interval; and a multi-dimensional damage comprehensive assessment step to calculate a damage score, determine the damage level, and provide feedback to adjust segmentation parameters, providing scientific data support for bridge safety assessment and maintenance decisions.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bridge structure health monitoring, and in particular to a multi-scale image segmentation and damage assessment method for bridge structure surface cracks. BACKGROUND

[0002] As an important part of transportation infrastructure, the structural safety of bridges is directly related to the safety of public life and property and the stability of social and economic operation. During long-term service, concrete bridges are inevitably affected by environmental erosion, load action, material aging and other factors, resulting in various types of cracks on the structure surface. The existence of these cracks not only affects the durability and bearing capacity of the bridge structure, but also may cause more serious structural damage or even collapse accidents. Therefore, timely and accurate detection of bridge surface cracks and assessment of their damage degree are of great significance for the safe operation and scientific maintenance of bridges.

[0003] From the formation mechanism of cracks, bridge concrete surface cracks can be mainly divided into two categories: load cracks and non-load cracks. Load cracks are structural cracks caused by vehicle load, wind load and other loads exceeding the designed bearing capacity, which usually occur at key positions with high stress concentration and pose a greater threat to structural safety. Non-load cracks mainly include shrinkage cracks, temperature cracks and chemical corrosion cracks. Shrinkage cracks are surface cracks caused by volume shrinkage during the hardening process of concrete due to water evaporation. Temperature cracks are cracks caused by excessive temperature difference between the inside and outside of concrete due to environmental temperature changes. Chemical corrosion cracks are cracks caused by chemical reactions such as steel corrosion expansion or alkali-aggregate reaction. Different types of cracks have significant differences in morphological characteristics, development rules and damage degree, which requires the crack detection system to have the ability to identify and distinguish multiple types of cracks.

[0004] From the scale characteristics of cracks, the width span of bridge surface cracks usually ranges from sub-millimeter level fine hairline cracks to centimeter level large structural cracks, and the length ranges from a few centimeters to several meters. Although fine hairline cracks currently have little impact on structural safety, they are often early signs of structural damage, and timely detection and monitoring of such cracks are of great value for preventive maintenance. Large structural cracks usually indicate that the bridge has serious damage and immediate reinforcement or repair measures are needed. The diversity of crack scales puts high requirements on the multi-scale feature capturing ability of the detection system.

[0005] In the prior art, Chinese invention patent CN118505701A discloses a concrete crack identification method based on computer image processing. The method uses a high-resolution camera to obtain a concrete surface image, enhances the image contrast through histogram equalization, eliminates noise using morphological opening and closing operations, performs edge detection using the Canny algorithm and the Sobel algorithm, measures the geometric parameters of the cracks through connected domain labeling and skeleton extraction, and performs semantic segmentation using the U-Net model. Finally, the results are analyzed comprehensively through logical judgment and voting mechanism.

[0006] However, the above prior art solution has the following technical defects. In terms of feature extraction, the existing solution uses a single scale image processing method, which cannot effectively capture the full spectrum of feature information from fine hairline cracks to large structural cracks, resulting in significant differences in detection ability for cracks of different scales. In terms of segmentation accuracy, the standard U-Net network lacks adaptive attention ability for crack regions, and is prone to missed detection and false detection under complex background interference and uneven lighting conditions. In terms of parameter measurement, the geometric parameter measurement accuracy of the existing solution is limited, and it is difficult to achieve sub-millimeter level measurement requirements, and lacks the ability to estimate the important parameter of crack depth. In terms of damage assessment, the existing solution can only provide static results for single detection, and cannot track the development and evolution trend of cracks, making it difficult to predict the development speed and safety risk of structural damage.

[0007] To solve the above technical problems, the present application provides a multi-scale image segmentation and damage assessment method for bridge structure surface cracks, which realizes accurate detection of bridge surface cracks and scientific assessment of damage degree by constructing a multi-scale pyramid feature extraction framework, an adaptive attention guided segmentation mechanism, a high-precision geometric parameter quantization algorithm, and a time series comparison trend prediction model. SUMMARY

[0008] The purpose of the present application is to provide a multi-scale image segmentation and damage assessment method for bridge structure surface cracks to solve the technical problems of insufficient multi-scale crack detection ability, limited segmentation accuracy, limited geometric parameter measurement accuracy, and lack of development trend prediction ability in the prior art.

[0009] To achieve the above object, the present application provides a bridge structure surface crack multi-scale image segmentation and damage evaluation method, comprising: a multi-scale pyramid feature preprocessing step, obtaining a to-be-detected image of a bridge concrete structure surface, performing Gaussian pyramid downsampling processing on the to-be-detected image to generate a plurality of feature maps of different resolution levels, performing Laplacian pyramid edge enhancement processing on each level feature map respectively, and generating a multi-scale feature pyramid through cross-level feature fusion; an adaptive attention guided crack segmentation step, inputting the multi-scale feature pyramid into an encoder-decoder structure segmentation network, performing importance weighting on each level feature channel through an adaptive channel attention mechanism in the encoding stage, enhancing the feature response of the crack region through a spatial attention mechanism in the decoding stage, and outputting a crack probability distribution map and a crack segmentation mask; a crack geometric parameter accurate quantization step, performing morphological thinning processing on the crack segmentation mask to generate a crack skeleton map, calculating the neighborhood shortest distance along each skeleton point of the crack skeleton map to obtain a crack width distribution, performing path tracking calculation along the crack skeleton map to obtain a crack length, estimating a crack depth based on the gray gradient feature of the crack region, and calculating the main direction of the crack skeleton to obtain a crack strike angle; a time series comparison crack development trend prediction step, obtaining historical detection data of the same detection position, comparing the crack geometric parameter set obtained by the current detection with the historical detection data in time series, calculating a crack length expansion rate according to the difference between the crack length obtained by the current detection and the corresponding crack length in the historical detection data divided by the detection time interval, calculating a crack width growth rate according to the difference between the crack width obtained by the current detection and the corresponding crack width in the historical detection data divided by the detection time interval, calculating a crack depth development rate according to the difference between the crack depth obtained by the current detection and the corresponding crack depth in the historical detection data divided by the detection time interval, predicting the development trend of the crack based on the crack length expansion rate, the crack width growth rate and the crack depth development rate, and calculating the time to reach the preset dangerous threshold according to the difference between the current crack geometric parameter value and the preset dangerous threshold divided by the corresponding expansion rate; a multi-dimensional damage comprehensive evaluation step, based on the length, width and depth parameters in the crack geometric parameter set and the expansion rate output by the development trend prediction step, calculating a multi-dimensional damage comprehensive score through weighted summation, determining the damage grade according to the comparison result of the comprehensive score and the preset damage grade threshold, and feeding back the damage grade and the development trend information to the adaptive attention guided crack segmentation step to dynamically adjust the segmentation sensitivity parameter.

[0010] Preferably, the number of levels of the Gaussian pyramid is 4, and the downsampling ratios of each level are 1 / 4, 1 / 8, 1 / 16 and 1 / 32 of the original resolution respectively.

[0011] Preferably, the encoder of the segmentation network adopts a residual network structure, and the depth of the residual network structure is 18 to 50 layers.

[0012] Preferably, the measurement accuracy of the crack width reaches 0.1-0.5 mm, and the measurement accuracy of the crack length reaches 1-5 mm.

[0013] Preferably, the damage grade is divided into four grades, the first grade is slight damage, the second grade is moderate damage, the third grade is serious damage, and the fourth grade is dangerous damage.

[0014] The present application has the advantages that: through the multi-scale pyramid feature extraction framework, full-spectrum feature information from fine hair cracks to large structural cracks is effectively captured, the crack detection accuracy reaches more than 98%; through the adaptive attention-guided segmentation mechanism, the segmentation accuracy and robustness under complex background conditions are significantly improved; through the neighborhood shortest distance measurement algorithm and the gray gradient depth estimation algorithm, the geometric parameter measurement accuracy reaches the sub-millimeter level; through the time series comparison trend prediction model, the predictive evaluation of crack development and evolution is realized, and scientific data support is provided for bridge structure safety evaluation and maintenance decision-making. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is the overall flowchart of the method of the present application.

[0016] Figure 2 is the detailed flowchart of the multi-scale pyramid feature preprocessing step.

[0017] Figure 3 is the detailed flowchart of the crack geometric parameter accurate quantization step.

[0018] Figure 4 is the detailed flowchart of the time series comparison crack development trend prediction step.

[0019] Figure 5 is the detailed flowchart of the multi-dimensional damage comprehensive evaluation step. DETAILED DESCRIPTION

[0020] Please refer to the attached Figures 1-5 In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the protection scope of the present application.

[0021] The present application provides a multi-scale image segmentation and damage evaluation method for bridge structure surface cracks, as shown in Figure 1As shown, this method comprises five deeply coupled core steps: a multi-scale pyramid feature preprocessing step, an adaptive attention-guided crack segmentation step, a precise quantification step of crack geometric parameters, a temporal comparison step for crack development trend prediction, and a multi-dimensional damage comprehensive assessment step. These five steps form a closed-loop collaborative architecture of forward data transfer and backward feedback adjustment. The output of each step serves as the key input for the next, while the evaluation results of the subsequent steps can inversely influence the processing parameters of the preceding steps, thereby achieving continuous optimization of the overall detection and evaluation performance.

[0022] I. Multi-scale pyramid feature preprocessing steps: The design of step 1 in the multi-scale pyramid feature preprocessing step fully considers the diverse scale characteristics of bridge surface cracks, such as... Figure 2 As shown. In one embodiment of the present invention, this step mainly completes four sub-tasks: image acquisition, Gaussian pyramid construction, Laplacian pyramid edge enhancement, and cross-level feature fusion, providing rich multi-scale feature representations for subsequent crack segmentation.

[0023] In the image acquisition subtask, images of the bridge's concrete structure surface to be inspected are acquired using an industrial-grade high-resolution camera or aerial photography equipment mounted on a drone. Preferably, the resolution of the acquisition device is no less than 4000×3000 pixels to ensure clear capture of the texture features of minute cracks. Illumination conditions need to be calibrated during the acquisition process; good image quality can be obtained within an illumination intensity range of 500 to 2000 lux. The acquired raw images undergo color space conversion processing, changing from the RGB color space to the grayscale space to reduce the computational complexity of subsequent processing.

[0024] In the Gaussian pyramid construction subtask, the grayscale image to be detected undergoes multi-level downsampling. The Gaussian pyramid construction process follows the standard image pyramid algorithm, with each level downsampled by a factor of 2 relative to the previous level. In this embodiment, the number of Gaussian pyramid levels is set to 4, with resolutions of 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the original image resolution, respectively. Before each downsampling, a Gaussian filter is first used to smooth the image to avoid aliasing. The kernel size of the Gaussian filter is 5×5, and the standard deviation is 1.4. Through this multi-level downsampling process, crack features can be captured at different resolution scales. Low-resolution levels are beneficial for capturing the overall morphology of large-scale structural cracks, while high-resolution levels are beneficial for preserving the detailed information of fine cracks.

[0025] In the Laplacian pyramid edge enhancement subtask, edge enhancement processing is performed on the feature maps of each level of the Gaussian pyramid. The calculation process of the Laplacian pyramid involves upsampling the lower-level images of the Gaussian pyramid and subtracting them from the higher-level images to obtain a difference image containing edge detail information. The edge enhancement strategy proposed in this invention is implemented through the following formula: .in, For the first Hierarchical images of the Laplace pyramid For the first A hierarchical image of a Gaussian pyramid. For the first A hierarchical image of a Gaussian pyramid. This indicates a bilinear interpolation upsampling operation. Laplacian pyramid images contain rich edge and texture information, significantly enhancing linear structural features such as cracks.

[0026] In the cross-level feature fusion subtask, Gaussian pyramid features and Laplacian pyramid features are fused to generate the final multi-scale feature pyramid. The cross-level fusion strategy employed in this invention first upsamples the Laplacian pyramid images from each level to the original image resolution at a uniform size, and then fuses them through weighted superposition. The fusion weights are set considering the differences in the contribution of features from different levels to crack detection; lower-level features have higher weights to preserve detail information, while higher-level features have lower weights to provide global context.

[0027] The specific calculation process for cross-level feature fusion is as follows. Let the first level be... The Laplacian pyramid image after hierarchical upsampling is The fusion weight is The fused multi-scale feature map The calculation is as follows: .

[0028] in, In this embodiment, the total number of pyramid levels is [number]. Fusion weights The settings follow a principle of decreasing from high resolution to low resolution. In this embodiment, Set to 0.4, Set to 0.3, Set to 0.2, It is set to 0.1. This weight configuration ensures that the detailed features of the high-resolution layer dominate the fusion result, while preserving the global semantic information provided by the low-resolution layer.

[0029] Furthermore, this invention introduces an adaptive weight adjustment mechanism during feature fusion. This mechanism automatically adjusts the fusion weights based on the local variance of the feature maps at each level, assigning higher weights to levels with larger local variance in the crack region. (Adaptive weight adjustment coefficient) The calculation formula is: .

[0030] in, This represents the calculation of the local variance of the image. The final adaptive fusion weights. Then, normalization is performed to ensure that the weight sum is 1. The fused multi-scale feature pyramid serves as the input data for the adaptive attention-guided crack segmentation step.

[0031] II. Adaptive Attention-Guided Crack Segmentation Step 2: Adaptive attention-guided crack segmentation step 2 employs a deep learning segmentation network with an encoder-decoder structure. Unlike the existing technology that directly uses the standard U-Net network, this invention introduces an adaptive channel attention mechanism and a spatial attention mechanism in the encoding and decoding stages, respectively, which significantly improves the network's ability to focus on crack regions and the segmentation accuracy.

[0032] In the encoding stage, a residual network is used as the backbone network for feature extraction. Preferably, the depth of the residual network is chosen to be 34 or 50 layers, which controls the number of model parameters and computational complexity while ensuring feature extraction capability. The encoder receives a multi-scale feature pyramid as input and extracts high-level semantic features step by step through successive convolutional layers, batch normalization layers, and activation function layers. At each feature level in the encoding stage, this invention introduces an adaptive channel attention mechanism to weight the importance of feature channels.

[0033] The calculation process of the adaptive channel attention mechanism is as follows: First, the input feature map... Perform global average pooling to obtain the channel description vector. Then, the channel description vectors are input into a two-layer fully connected network to model the relationships between channels, generating channel attention weight vectors. The formula for calculating channel attention weights is: ,in, This is the weight matrix of the first fully connected layer. This is the weight matrix of the second fully connected layer. The channel compression ratio is set to 16 in this embodiment. To correct the linear activation function, The Sigmoid activation function is used. Finally, the channel attention weight vector is multiplied channel-by-channel with the input feature map to obtain the channel-weighted feature map. ,in This represents an element-wise multiplication operation. Through the channel attention mechanism, the network can adaptively enhance the response of feature channels related to crack detection while suppressing the response of background interference channels.

[0034] In the decoding stage, a progressive upsampling and skip connections are used to restore the spatial resolution of the feature maps. Each level of the decoder receives information from the feature maps of the corresponding levels of the encoder through skip connections, and then concatenates and fuses them with the upsampled higher-level features. At each feature level in the decoding stage, this invention introduces a spatial attention mechanism to enhance the feature response of the crack region.

[0035] The computation process of the spatial attention mechanism is as follows. First, max pooling and average pooling operations are performed on the input feature map along the channel dimension to obtain two spatial description maps. and Then, the two spatial description maps are concatenated along the channel dimension, and a spatial attention weight map is generated using a 7×7 convolutional layer and a sigmoid activation function. The formula for calculating spatial attention weights is: .in, This represents a convolution operation with a kernel size of 7×7. This represents the concatenation operation along the channel dimension. Finally, the spatial attention weight map is multiplied pixel-by-pixel with the input feature map to obtain the spatially weighted feature map. Through the spatial attention mechanism, the network can focus on the spatial region where the crack is located in the image, effectively suppressing interference from the background region.

[0036] The final layer of the decoder outputs two results: a crack probability distribution map and a crack segmentation mask. Each pixel value in the crack probability distribution map represents the confidence level that the location belongs to a crack, ranging from 0 to 1. The crack segmentation mask is obtained by thresholding the probability distribution map. The default threshold is set to 0.5, and this threshold can be dynamically adjusted based on feedback information from the multi-dimensional damage comprehensive assessment steps.

[0037] During model training, a combination of cross-entropy loss and Dice loss is used as the optimization objective. This combined loss function simultaneously optimizes pixel-level classification accuracy and region-level overlap, showing better performance for elongated targets such as cracks. The calculation formula is: ,in, For cross-entropy loss, For Dice's loss, and This is the balance coefficient between the two loss terms. In this embodiment, Set to 0.5. Set it to 0.5.

[0038] Cross-entropy loss The calculation formula is: .in, Total number of pixels For the first The real label of each pixel For the first The predicted probability of each pixel.

[0039] Dice loss The calculation formula is: .in, This is a smoothing coefficient used to avoid division by zero errors; in this embodiment, it is set to... .

[0040] The optimizer chosen is the Adam algorithm, whose adaptive learning rate characteristic is beneficial for handling sparse gradient problems. The Adam algorithm adaptively adjusts the learning rate through first-order moment estimation and second-order moment estimation, where the decay coefficient of the first-order moment estimation... Set to 0.9, the attenuation coefficient for second-order moment estimation Set to 0.999, numerical stability parameter Set as The initial learning rate was set to 0.001, and a learning rate decay strategy was adopted, reducing the learning rate to 0.5 times its original value every 20 training epochs. The batch size was set to 16, which allowed for stable operation on a GPU with 12GB of video memory. The number of iteration epochs was set to 100. Training could be terminated earlier depending on the convergence of the validation set loss function, specifically using an early stopping strategy: training would terminate when the validation set loss function showed no improvement for 10 consecutive epochs.

[0041] To prevent overfitting and improve the model's generalization ability, multiple data augmentation strategies are employed during training. The execution probability of random horizontal flipping is set to 0.5 to simulate scenes from different shooting directions. The random rotation angle range is set to -15 degrees to +15 degrees to enhance the model's ability to identify different crack orientations. The scale factor of random scaling is set to 0.8 to 1.2 to enable the model to adapt to crack scale changes at different shooting distances. The adjustment range of random brightness is set to 0.7 to 1.3 times the original brightness to improve the model's robustness to different lighting conditions. Furthermore, this invention also employs two augmentation strategies: random contrast adjustment (range 0.8 to 1.2) and random Gaussian noise addition (standard deviation range 0 to 10). These data augmentation operations are executed with independent random probabilities in each training epoch, effectively expanding the diversity of training samples and improving the model's robust performance in real-world application scenarios.

[0042] In terms of training data preparation, this invention employs professional annotation tools to perform pixel-level annotation of bridge crack images. The annotation process follows a strict quality control procedure, with each image independently annotated by two annotators and then cross-validated. Regions with inconsistent annotation results are arbitrated by experts. The annotation dataset contains 3000 bridge crack images, divided into training, validation, and test sets in a 7:2:1 ratio. The training set is used for model parameter optimization, the validation set for hyperparameter tuning and early stopping detection, and the test set for final performance evaluation. There is no image overlap between the three datasets, ensuring the fairness of the evaluation results.

[0043] III. Precise Quantification Steps for Crack Geometric Parameters: Step 3 involves precisely measuring the geometric parameters of the segmented crack regions, such as... Figure 3 As shown in the diagram. The input to this step is a crack segmentation mask and the original grayscale image, and the output is a set of crack geometric parameters including length, width, depth, and orientation. The geometric parameter quantization algorithm proposed in this invention can achieve sub-millimeter level measurement accuracy, providing precise quantitative data support for damage assessment.

[0044] In the morphological thinning subtask, the crack segmentation mask is used to extract the skeleton to obtain the crack centerline representation. This invention employs the Zhang-Suen thinning algorithm for iterative thinning, which gradually peels away crack boundary pixels to obtain a crack skeleton map with a single pixel width. The iteration termination condition for thinning is that the image pixels no longer change after two consecutive iterations. The thinned crack skeleton map preserves the crack's topological structure and orientation information, providing a basis for subsequent geometric parameter calculations.

[0045] In the crack width measurement subtask, this invention proposes a neighborhood shortest distance measurement algorithm, which outperforms existing average width methods and orthogonal projection methods in terms of measurement accuracy. The core idea of ​​the algorithm is to search for crack boundary point pairs within a neighborhood perpendicular to the local tangent direction along each skeleton point of the crack skeleton, and calculate the Euclidean distance between the boundary point pairs as the local crack width at that point. The specific calculation process is as follows:

[0046] For any skeleton point on the crack skeleton First, calculate the local tangent direction at that point. The tangent direction is determined by the forward and backward neighbors of the skeleton point:

[0047] ,

[0048] in, and These are the skeleton points. Forward and backward Neighboring points, The value of is between 3 and 5. Then, the normal direction perpendicular to the tangent direction is calculated. ,in and These are the tangent direction vectors. and Components. Search ranges are set on both the positive and negative sides of the normal direction. Search within this range for two boundary points that intersect with the crack boundary. and The local crack width at this skeleton point is calculated as follows:

[0049] ,

[0050] in, Indicates Euclidean distance. Search range. The setting needs to be determined based on the expected maximum width of the crack, and in this embodiment it is set to 50 pixels.

[0051] By statistically analyzing the widths of all skeleton points on the crack skeleton, the distribution characteristics of the crack width can be obtained, including the maximum width, minimum width, average width, and standard deviation of the width. When converting pixel width to physical width, the pixel equivalent obtained from camera calibration needs to be used for conversion. Preferably, the measurement system of the present invention achieves a width measurement accuracy of 0.1 mm to 0.5 mm through high-precision calibration.

[0052] In the crack length measurement subtask, this invention employs a skeleton path tracing algorithm to calculate the actual crack length. The algorithm first identifies all endpoints and branch points in the crack skeleton diagram, dividing the skeleton into several individual branch segments. For each branch segment, it traces the skeleton path pixel-by-pixel from the endpoint, accumulating the Euclidean distance between adjacent pixels to obtain the length of that branch segment. For adjacent pixels in the horizontal or vertical direction, the distance is 1 pixel unit; for adjacent pixels in the diagonal direction, the distance is... The unit is pixels. The formula for calculating the branch segment length is:

[0053] ,

[0054] in, This represents the number of skeleton points contained in this branch segment. For the first The coordinates of each skeleton point. The total length of the crack is the sum of the lengths of all branch segments. For complex crack networks with intersections and branches, this invention decomposes them into multiple independent cracks through connected component analysis and calculates their lengths separately. Preferably, the measurement system of this invention achieves a length measurement accuracy of 1 mm to 5 mm.

[0055] In the crack depth estimation subtask, this invention proposes a depth estimation algorithm based on gray-level gradient features. Crack depth cannot be accurately measured directly from a single 2D image, but there is a correlation between the gray-level features of the crack region and its depth. The deeper the crack, the worse its internal lighting conditions, which is reflected in the image as lower gray-level values ​​and greater gray-level contrast with the surrounding background. This invention utilizes this characteristic to establish an estimation model for crack gray-level features and depth.

[0056] The depth estimation algorithm first extracts the grayscale values ​​of each pixel within the crack region. and the average background gray value of the neighborhood outside the crack boundary Then calculate the relative grayscale contrast of the crack region: ,in, The grayscale mean of the crack region is given. Finally, the crack depth is estimated using a pre-calibrated grayscale-depth mapping function. .

[0057] in, , and The parameters of the mapping function are obtained by calibration using standard crack samples of known depths. In this embodiment, Set to 15. Set to 0.8, Set to 0.5, with depth units in millimeters. It should be noted that depth estimation based on grayscale features has certain limitations; its accuracy is affected by factors such as lighting conditions and contaminants inside the crack. The estimated results should be used as reference values ​​rather than precise measurements.

[0058] In the crack direction calculation subtask, the principal direction of the crack skeleton is calculated using principal component analysis. The coordinates of all points on the crack skeleton are used to form a point set matrix, and the covariance matrix of this point set is calculated. The direction of the first principal component of the covariance matrix is ​​the principal direction of the crack. The crack direction is represented by the angle with the horizontal direction, ranging from 0 degrees to 180 degrees.

[0059] The specific process of calculating the principal directions of cracks using principal component analysis is described below. Assume the crack skeleton contains... There are 1 skeleton point, and the coordinates of each skeleton point are: ,in First, calculate the mean of the skeleton point coordinates: Then calculate the covariance matrix. The elements: .in, , , .

[0060] Eigenvalue decomposition of the covariance matrix yields two eigenvalues. and and the corresponding feature vectors and Among them, the larger eigenvalues corresponding feature vector This is the principal direction vector of the crack. Crack direction angle. The calculation is as follows: ,in, and These are the principal direction vectors. of and Portion. angle The value range is adjusted to between 0 and 180 degrees, where 0 degrees represents the horizontal direction and 90 degrees represents the vertical direction. Furthermore, this invention also calculates the crack direction dispersion index, defined as the ratio of two characteristic values. This index reflects the consistency of crack orientation; the smaller the ratio, the more consistent the crack orientation.

[0061] IV. Time-series comparison crack development trend prediction steps: Step 4 of the time-series comparison crack development trend prediction step involves analyzing the changes in crack geometric parameters at the same detection location at different time points to predict the crack development trend and safety risks, such as... Figure 4 As shown. The inputs to this step include the set of geometric parameters of the crack obtained from the current detection and historical detection data of the same location obtained from the historical database. The output is a set of development trend prediction indicators.

[0062] In the historical data acquisition subtask, historical inspection records corresponding to the current inspection location are retrieved from the bridge health monitoring database. Location matching is achieved through image registration technology, which matches feature points and performs geometric transformations between the current inspection image and historical inspection images to determine the spatial correspondence. The bridge health monitoring database uses a relational database management system for data storage and indexing. The information stored in the database includes fields such as image files, inspection time, inspection location coordinates, crack geometric parameters, and damage assessment results. The database design follows normalization principles, improving data retrieval efficiency by establishing spatial and temporal indexes.

[0063] The image registration process employs a feature-point-based registration method, with the specific steps described below. First, the SIFT or ORB algorithm is used to extract feature points and their descriptors from the current and historical detection images. The SIFT algorithm detects keypoints by constructing a scale-space pyramid and calculates the gradient orientation histogram of the keypoint neighborhood as a descriptor, exhibiting strong scale and rotation invariance. The ORB algorithm combines FAST corner detection with BRIEF descriptors, offering high computational efficiency and suitability for real-time applications. In this embodiment, the SIFT algorithm is preferred for achieving higher registration accuracy; however, the ORB algorithm can be used for scenarios with high real-time requirements. Then, a nearest neighbor distance ratio matching strategy is employed for feature point matching, filtering out reliable corresponding point pairs. The nearest neighbor distance ratio threshold is set to 0.75; a reliable match is considered achieved when the ratio of the nearest neighbor distance to the second nearest neighbor distance is less than this threshold. Based on the matched feature point pairs, the RANSAC algorithm is used to estimate the homography transformation matrix between the two images. The RANSAC algorithm eliminates erroneous matches through random sampling consistency checks, with 1000 iterations and an inlier distance threshold of 3 pixels. Finally, a transformation matrix is ​​used to map the crack locations in historical detection images to the current image coordinate system, establishing the spatial correspondence.

[0064] To ensure registration accuracy, this invention establishes a registration quality evaluation mechanism. The registration quality evaluation indicators include the number of matched feature points, the proportion of inliers, and the reprojection error. When the number of matched feature points is less than 20 or the proportion of inliers is less than 60%, the system indicates insufficient registration quality, requiring manual verification or re-acquiring of the image. When the reprojection error exceeds 5 pixels, the system issues a registration warning, indicating that the registration result may be unreliable.

[0065] Historical inspection data should include data from at least two inspection cycles. The inspection cycle is determined based on the bridge's importance level and usage condition, preferably ranging from 1 to 12 months. For bridges of higher importance or areas with obvious defects, the inspection cycle can be shortened to 1 month; for ordinary bridges in good condition, the inspection cycle can be extended to 6 to 12 months.

[0066] In the time-series parameter comparison subtask, the geometric parameters of the currently detected cracks are compared with the corresponding parameters in historical detection data. For existing cracks with established correspondences, the changes in each geometric parameter over time are calculated; for newly discovered cracks, they are marked as new cracks and the time of their first discovery is recorded. The main parameters for time-series comparison analysis include the change in crack length. Crack width variation and crack depth variation .

[0067] In the crack propagation rate calculation subtask, the crack propagation rate is calculated based on the changes in time-series parameters and the detection time interval. Crack length propagation rate. Crack width growth rate and crack depth development rate The calculation formulas are as follows: , , ,

[0068] in, , and These represent the length, width, and depth of the crack currently detected. , and These are the corresponding parameters obtained from the previous detection; and These represent the current detection time and the last detection time, respectively, with the time unit being days.

[0069] When historical monitoring data includes three or more monitoring periods, the trend of crack propagation rate can be further analyzed to determine whether crack development is in an accelerating, uniform, or decelerating phase. (Propagation rate acceleration) The calculation formula is: ,in, The length expansion rate of the most recent cycle, This represents the length propagation rate of the previous cycle. A positive propagation rate acceleration indicates that crack propagation is accelerating, a negative value indicates that crack propagation is decelerating, and a value close to zero indicates that the crack is propagating at a constant speed.

[0070] In the development trend prediction subtask, based on the current geometric parameter values ​​and propagation rate, the time it takes for the crack to reach a preset danger threshold is predicted. The danger thresholds set by this invention include a length danger threshold. Width danger threshold and depth hazard threshold The specific values ​​are determined based on bridge structural design specifications and maintenance standards. The estimated time to reach the danger threshold. Calculated based on the time when any threshold is first reached: When the estimated time to reach the danger threshold is less than the preset warning period, the system outputs a warning message about the development trend.

[0071] V. Multi-dimensional Damage Comprehensive Assessment Steps: Step 5 of the multi-dimensional damage comprehensive assessment calculates the multi-dimensional damage comprehensive score and determines the damage level based on the crack geometric parameters and development trend prediction results, such as... Figure 5 As shown in the diagram. This step also feeds the evaluation results back to the adaptive attention-guided crack segmentation step, achieving closed-loop optimization.

[0072] In the multi-dimensional scoring calculation subtask, this invention proposes a multi-dimensional damage scoring model that comprehensively considers both the static geometric features and dynamic development characteristics of cracks. The input dimension of the scoring model includes crack length. Maximum width of crack Average crack depth Length expansion rate and width growth rate The parameters of each dimension are first normalized, mapped to the range of 0 to 1:

[0073] , , , , .

[0074] in, , , , and These are the normalized reference upper limits for each parameter, set according to bridge structural specifications. In this embodiment, Set to 1000 mm. Set to 5 mm, Set to 20 mm. Set to 10 mm / day Set to 0.1 mm / day.

[0075] The normalized parameters of each dimension are used to calculate a multidimensional damage comprehensive score through weighted summation. :

[0076] .in, , , , and The weight coefficients for each dimension satisfy... The weighting coefficients reflect the importance of each dimension parameter to structural safety. In this embodiment, crack width is considered the most important damage indicator, and its weight... Set to 0.3; the development rate parameter reflects the dynamic trend of damage, and its weight... and The weights are set to 0.15 and 0.15 respectively; the weights for crack length and depth. and Set them to 0.2 and 0.2 respectively.

[0077] In the damage level determination subtask, damage is assessed based on a multi-dimensional comprehensive damage score. The damage level is determined by comparing the result with a preset damage level threshold. This invention classifies the damage level into four levels:

[0078] The first level is minor injury, corresponding to the comprehensive score range. This indicates that the crack is in its initial development stage and has no significant impact on structural safety. Routine observation and monitoring are recommended.

[0079] The second level is moderate injury, corresponding to the comprehensive score range. This indicates that the cracks have developed to a certain extent, requiring increased monitoring frequency, and preventative maintenance can be considered.

[0080] The third level is severe injury, corresponding to the comprehensive score range. This indicates that the cracks have developed to a relatively serious extent, and it is recommended to arrange repairs as soon as possible and limit the bridge load.

[0081] Level 4 is considered dangerous injury, corresponding to the comprehensive score range. This indicates that the cracks have reached a dangerous level and emergency measures should be taken immediately, including closing traffic if necessary.

[0082] In the feedback adjustment subtask, damage level and development trend information are fed back to the adaptive attention-guided crack segmentation step to dynamically adjust the segmentation sensitivity parameters. The feedback adjustment strategy is as follows: When the damage level is level three or four, it indicates that there is relatively severe crack damage in the current detection area. To avoid missing the propagation of minor cracks, the system automatically lowers the confidence threshold for crack segmentation. Preferably, the confidence threshold is lowered from the default value of 0.5 to 0.3 to improve the detection sensitivity for minor cracks. When the damage level is level one or two, it indicates that the crack damage in the current detection area is relatively minor. To reduce false detections due to background noise, the system maintains or increases the confidence threshold for crack segmentation. Preferably, the confidence threshold is maintained at the default value of 0.5 or increased to 0.6. Through this closed-loop feedback mechanism, the system can adaptively optimize the detection strategy based on historical evaluation results, achieving a better balance between sensitivity and accuracy.

[0083] The core value of the feedback adjustment mechanism lies in enabling the detection system to learn and adapt. In practical engineering applications, different bridges exhibit significant differences in structural characteristics, service environments, and damage modes, making it difficult for fixed detection parameters to achieve optimal performance in all scenarios. Through a closed-loop feedback mechanism, the system can automatically adjust detection parameters based on historical inspection experience of specific bridges, achieving personalized customization of the detection strategy. For example, for a particular bridge, if historical inspections have revealed a rapid crack development rate, the system will automatically increase its detection sensitivity to ensure timely detection of new cracks; conversely, if the crack condition of another bridge remains stable over a long period, the system will appropriately decrease its sensitivity to reduce false alarms caused by misjudgments due to background texture. This adaptive adjustment mechanism significantly enhances the system's practicality and reliability in complex and ever-changing real-world engineering environments.

[0084] Furthermore, the feedback adjustment mechanism also supports manual intervention and parameter overriding functions. When technicians believe, based on on-site observations, that the parameters automatically adjusted by the system are unsuitable for the current testing scenario, they can manually set the confidence threshold and other testing parameters. The system will then prioritize the manually set parameter values. This human-machine collaborative design concept leverages the efficiency advantages of automated systems while preserving the judgment and control rights of professionals, ensuring the reliability and traceability of the test results.

[0085] To verify the effectiveness of the method of this invention, experimental evaluation was conducted on a test dataset containing 1000 bridge crack images. The test dataset covers bridge crack samples of different types, scales, and degrees of damage; the images were collected from multiple bridges of different structural types and service ages. The dataset includes crack samples of various types, such as fine hairline cracks, ordinary network cracks, and large through cracks, covering various crack morphologies that may be encountered in actual engineering projects. The image resolution ranges from 2000×1500 pixels to 4000×3000 pixels to accommodate the output specifications of different detection equipment.

[0086] The test dataset was subdivided and statistically analyzed according to crack type, with load cracks accounting for approximately 35%, shrinkage cracks approximately 30%, temperature cracks approximately 20%, and chemical corrosion cracks approximately 15%. Statistically, cracks less than 0.2 mm accounted for approximately 25%, medium cracks between 0.2 mm and 1 mm accounted for approximately 45%, and large cracks greater than 1 mm accounted for approximately 30%. This diverse data distribution ensures that the evaluation results comprehensively reflect the performance of the method of this invention under different application scenarios.

[0087] The experimental environment was configured as follows: Intel Core i9-12900K processor (3.2GHz); NVIDIA RTX 3090 graphics card (24GB VRAM); 64GB RAM; Ubuntu 20.04 operating system; and PyTorch version 1.12 deep learning framework. With this hardware configuration, the complete processing time for a single 4000×3000 resolution image was approximately 1.2 seconds, meeting the real-time requirements of engineering applications.

[0088] To comprehensively evaluate the performance advantages of the method of this invention, multiple sets of comparative experiments were set up. The comparative methods included traditional edge detection methods, the standard U-Net semantic segmentation method, and the method disclosed in the comparative document CN118505701A. The traditional edge detection method employs a combination of the Canny algorithm and morphological post-processing, while the standard U-Net method uses the original U-Net network structure without adding an attention mechanism. All comparative methods were evaluated on the same test dataset and using the same evaluation metrics to ensure the fairness of the comparison results.

[0089] Experimental results show that the crack detection accuracy of the method of this invention reaches 98.3%, the recall rate reaches 96.7%, and the F1 score reaches 97.5%. Compared with the method disclosed in prior art CN118505701A, the method of this invention improves the accuracy by 3.2 percentage points and the recall rate by 4.5 percentage points. Compared with the traditional Canny edge detection method, the method of this invention improves the accuracy by 15.6 percentage points and the recall rate by 18.3 percentage points. Compared with the standard U-Net method, the method of this invention improves the accuracy by 2.8 percentage points and the recall rate by 3.6 percentage points. This improvement is mainly attributed to the accurate focusing of the crack region by the adaptive attention mechanism.

[0090] Regarding the accuracy of geometric parameter measurements, the average absolute error for crack width measurement is 0.15 mm, and the relative error is controlled within 10%; the average absolute error for crack length measurement is 2.3 mm, and the relative error is controlled within 5%. Compared with existing technologies, the method of this invention improves detection accuracy by 3.2 percentage points, width measurement accuracy by 40%, and length measurement accuracy by 35%. The average absolute error for depth estimation is 1.2 mm, and the relative error is controlled within 15%. Although the accuracy of depth estimation is relatively low, it is sufficient as an auxiliary reference indicator to meet the needs of engineering applications.

[0091] This invention also provides a detailed performance evaluation for cracks of different scales. For fine cracks less than 0.2 mm in width, the detection recall rate of the method in this invention reaches 93.5%, significantly better than the 78.2% of the comparative method. This is attributed to the effective capture of fine features by multi-scale pyramid feature extraction. For medium-sized cracks with a width between 0.2 mm and 1 mm, the detection accuracy of the method in this invention reaches 99.1%, and the F1 score reaches 98.6%. For large cracks with a width greater than 1 mm, both the detection accuracy and recall rate of the method in this invention exceed 99.5%.

[0092] Furthermore, this invention underwent robustness testing under varying lighting conditions and background complexity. Even under conditions of strong illumination, weak illumination, shadow occlusion, and complex background textures, the detection accuracy of the method remains above 95%, significantly outperforming traditional edge detection and standard semantic segmentation methods. This is attributed to the effective suppression of environmental interference by multi-scale pyramid feature extraction and adaptive attention mechanisms.

[0093] Robustness testing was conducted using various extreme scenarios. In strong lighting conditions, the overall image brightness exceeds 220 gray levels, with significant highlight reflections in some areas. Traditional edge detection methods often misidentify highlight edges as cracks under these conditions. Our method, through the combined action of multi-scale feature fusion and an attention mechanism, effectively distinguishes highlight edges from real cracks, keeping the false detection rate below 3% in strong lighting scenarios. In low lighting conditions, the overall image brightness is below 50 gray levels, significantly reducing the contrast between cracks and the background, leading to a substantial increase in the false detection rate of traditional methods. Our method, through Laplacian pyramid edge enhancement, effectively improves the feature response of crack areas, maintaining a recall rate above 92% in low lighting scenarios. In shadowed scenarios, images contain shadow areas cast by buildings, equipment, or clouds, and shadow edges are easily misidentified as cracks. Our method, through a spatial attention mechanism, learns the differences in spatial distribution features between cracks and shadows, effectively suppressing false detections in shadow areas. In complex background texture scenarios, concrete surfaces are subject to interference factors such as aggregate textures, formwork marks, and stains. The edge features of these background textures are similar to those of cracks. The method of this invention learns the differences in channel response patterns between cracks and background textures through a channel attention mechanism, maintaining an accuracy of over 94% even in complex background scenarios.

[0094] This invention has also undergone field verification tests on multiple bridges in actual service, including highway bridges crossing rivers, urban viaducts, and railway overpasses. Field test results show that the method of this invention can adapt to different bridge structure types and testing environments. The test results show a high degree of consistency with the manual interpretation results of professional inspectors, with a consistency rate exceeding 96%. Field tests also verified the effectiveness of the time-series comparison trend prediction function. For bridges that have undergone multiple phases of inspection, the system's predicted crack development trends basically match the actual observation results, providing reliable data support for bridge maintenance decisions.

[0095] In summary, the multi-scale image segmentation and damage assessment method for bridge surface cracks provided by this invention achieves accurate detection and scientific assessment of bridge surface cracks through a five-step closed-loop collaborative architecture consisting of multi-scale pyramid feature extraction, adaptive attention-guided segmentation, precise geometric parameter quantification, temporal trend prediction, and multi-dimensional comprehensive assessment. This provides reliable technical support for bridge structural safety monitoring and maintenance decisions.

Claims

1. A multi-scale image segmentation and damage assessment method for surface cracks in bridge structures, characterized in that, include: The multi-scale pyramid feature preprocessing step involves acquiring the image to be detected on the surface of the bridge concrete structure, performing Gaussian pyramid downsampling on the image to be detected to generate feature maps at multiple resolution levels, performing Laplacian pyramid edge enhancement on each feature map level, and generating a multi-scale feature pyramid through cross-level feature fusion. The adaptive attention-guided crack segmentation step involves inputting the multi-scale feature pyramid into the segmentation network of the encoder-decoder structure. In the encoding stage, the feature channels at each level are weighted by an adaptive channel attention mechanism. In the decoding stage, the feature response of the crack region is enhanced by a spatial attention mechanism, and the crack probability distribution map and crack segmentation mask are output. The adaptive channel attention mechanism includes: performing global average pooling on the input feature map to obtain a channel description vector; inputting the channel description vector into a two-layer fully connected network to generate a channel attention weight vector; and multiplying the channel attention weight vector with the input feature map channel by channel to obtain a channel-weighted feature map. The spatial attention mechanism includes: performing max pooling and average pooling on the input feature map along the channel dimension to obtain two spatial description maps; concatenating the two spatial description maps and passing them through a convolutional layer to generate a spatial attention weight map; and multiplying the spatial attention weight map with the input feature map pixel by pixel to obtain a spatially weighted feature map. The precise quantification steps for crack geometric parameters involve morphological refinement of the crack segmentation mask to generate a crack skeleton map, calculating the shortest neighborhood distance along each skeleton point of the crack skeleton map to obtain the crack width distribution, performing path tracing calculation along the crack skeleton map to obtain the crack length, estimating the crack depth based on the gray-level gradient features of the crack region, and calculating the main direction of the crack skeleton to obtain the crack direction angle. The calculation of the shortest neighborhood distance to obtain the crack width distribution includes: calculating the tangent direction at each skeleton point along the crack skeleton, searching for crack boundary point pairs within the neighborhood perpendicular to the tangent direction, and calculating the Euclidean distance between the boundary point pairs as the local crack width at that skeleton point. The time-series comparison crack development trend prediction steps involve acquiring historical detection data at the same detection location, comparing the currently detected crack geometric parameter set with the historical detection data in a time series, calculating the crack length expansion rate by dividing the difference between the currently detected crack length and the corresponding crack length in the historical detection data by the detection time interval, calculating the crack width growth rate by dividing the difference between the currently detected crack width and the corresponding crack width in the historical detection data by the detection time interval, and calculating the crack depth development rate by dividing the difference between the currently detected crack depth and the corresponding crack depth in the historical detection data by the detection time interval. Based on the crack length expansion rate, the crack width growth rate, and the crack depth development rate, the crack development trend is predicted, and the estimated time to reach the preset danger threshold is calculated by dividing the difference between the current crack geometric parameter value and the preset danger threshold by the corresponding expansion rate. The multi-dimensional damage comprehensive assessment step, based on the length, width, and depth parameters in the crack geometric parameter set and the crack length propagation rate output by the development trend prediction step, calculates a multi-dimensional damage comprehensive score by weighted summation, determines the damage level based on the comparison result of the comprehensive score and the preset damage level threshold, and feeds the damage level and development trend information back to the adaptive attention-guided crack segmentation step to dynamically adjust the segmentation sensitivity parameters.

2. The multi-scale image segmentation and damage assessment method for surface cracks in bridge structures according to claim 1, characterized in that, In the multi-scale pyramid feature preprocessing step, the Gaussian pyramid has 4 levels, and the downsampling ratios of each level are 1 / 4, 1 / 8, 1 / 16 and 1 / 32 of the original resolution, respectively.

3. The multi-scale image segmentation and damage assessment method for surface cracks in bridge structures according to claim 1, characterized in that, In the adaptive attention-guided crack segmentation step, the encoder of the segmentation network adopts a residual network structure with a depth of 18 to 50 layers.

4. The multi-scale image segmentation and damage assessment method for surface cracks in bridge structures according to claim 1, characterized in that, In the precise quantification step of the crack geometry parameters, the measurement accuracy of the crack width reaches 0.1 mm to 0.5 mm, and the measurement accuracy of the crack length reaches 1 mm to 5 mm.

5. The multi-scale image segmentation and damage assessment method for surface cracks in bridge structures according to claim 1, characterized in that, In the time-series comparison crack development trend prediction step, the time span of the historical detection data is no less than two detection cycles, and the detection cycle is 1 month to 12 months.

6. The multi-scale image segmentation and damage assessment method for surface cracks in bridge structures according to claim 1, characterized in that, In the multi-dimensional damage assessment process, the damage level is divided into four levels: the first level is minor damage, the second level is moderate damage, the third level is severe damage, and the fourth level is dangerous damage.

7. The multi-scale image segmentation and damage assessment method for surface cracks in bridge structures according to claim 1, characterized in that, Feeding the damage level and development trend information back to the adaptive attention-guided crack segmentation step to dynamically adjust the segmentation sensitivity parameters includes: when the damage level is level three or four, lowering the confidence threshold for crack segmentation to improve the detection sensitivity for fine cracks; when the damage level is level one or two, maintaining or increasing the confidence threshold for crack segmentation to reduce false detections.

Images