Deep learning-based bridge crack feature refined quantification method and system

Abstract

The invention discloses a bridge crack feature fine quantification method and system based on deep learning, and relates to the technical field of bridge detection, and the method comprises the steps: carrying out the preprocessing of bridge surface image data; inputting the preprocessed image data into a U-Net network fused with a multi-scale channel space attention module for semantic segmentation and outputting a crack segmentation mask graph; performing morphological refinement operation to obtain fracture skeleton line data, and calculating a trend angle and curvature distribution; measuring the crack width along the normal direction of the skeleton point to generate a width distribution thermodynamic diagram; accumulating skeleton point intervals to calculate the total length and identifying branch features; according to the method, the segmentation IoU reaches 85% or above, the width precision is superior to 0.05 mm, and the method has the crack development trend prediction capability.

Description

Technical Field

[0001] This invention relates to the field of bridge inspection technology, specifically to a method and system for refined quantification of bridge crack features based on deep learning. Background Technology

[0002] As a crucial component of transportation infrastructure, the structural safety of bridges directly impacts public safety and socio-economic operations. With increasing service life and the effects of vehicle loads and environmental factors, bridge structures are prone to developing cracks. These cracks not only affect the bridge's appearance but can also trigger a chain reaction, leading to steel corrosion and concrete deterioration, severely threatening the bridge's load-bearing capacity and durability. Therefore, timely and accurate detection and quantification of bridge crack characteristics are of paramount importance for bridge maintenance decisions and safety assessments.

[0003] Chinese patent CN119477886A discloses a method and system for monitoring building cracks based on machine learning. This technical solution employs multi-band image data enhancement and preprocessing, utilizing images from different bands of ultraviolet, infrared, and visible light to enrich crack feature information; it uses a deep learning-based adaptive segmentation model to segment crack regions from multi-layer image data, with the adaptive segmentation model being a fully convolutional network (FCN) or U-Net; it extracts geometric features of crack length, width, depth, direction, and shape complexity for feature vectorization; it learns crack features based on a self-supervised learning model; it uses a generative adversarial network (GAN) to synthesize crack images for comparative learning; it models crack morphology and dynamics using physics and mechanics formulas; and it performs crack correlation analysis by constructing a crack topology graph.

[0004] However, the existing technologies have the following technical problems. First, the existing methods do not fully consider the slender morphological characteristics of bridge cracks in the design of the crack segmentation network architecture, and lack targeted attention mechanism optimization, resulting in insufficient segmentation accuracy of fine cracks, especially prone to missed detections and false detections under complex backgrounds and lighting conditions. Second, existing methods mostly use simple distance transformations or boundary point spacing statistics to measure crack width, failing to perform precise point-by-point measurements along the normal direction of the crack skeleton, resulting in measurement direction deviations, making it difficult to meet the accuracy requirements of better than 0.05mm in engineering inspection standards. Third, existing methods lack systematic analysis of crack orientation angles and curvature distribution, failing to comprehensively characterize the geometric morphological characteristics of cracks, affecting subsequent crack damage assessment and maintenance decisions. Fourth, existing methods lack the ability to conduct time-series comparative analysis of multiple inspection results for the same inspection area, failing to quantitatively assess the crack propagation rate and direction, and making it difficult to predict and warn of crack development trends. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a deep learning-based method for refined quantification of bridge crack features, thereby solving the technical problems mentioned in the background section.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for refined quantization of bridge crack features based on deep learning, comprising the following steps: S1, image preprocessing step: acquiring bridge surface image data, performing grayscale normalization and contrast enhancement processing on the bridge surface image data to generate preprocessed image data; S2, crack semantic segmentation step: inputting the preprocessed image data into a deep learning semantic segmentation network, the deep learning semantic segmentation network adopting a U-Net encoder-decoder architecture, and fusing a multi-scale channel spatial attention module in the skip connection layer, extracting multi-level feature maps through the encoder and restoring spatial resolution through layer-by-layer upsampling by the decoder, outputting a crack segmentation mask map; S3, skeleton extraction and morphological analysis step: performing morphological thinning operations on the crack segmentation mask map to obtain crack skeleton line data, calculating the tangential direction of each skeleton point along the crack skeleton line data and determining the crack direction angle value, and calculating based on the tangential change of adjacent skeleton points. S4, Normal Width Measurement Step: Calculate the normal direction vector of each skeleton point in the crack skeleton line data, search the boundary pixels of the crack segmentation mask bidirectionally along the normal direction vector of the skeleton points, calculate the crack width value at each skeleton point based on the boundary pixel spacing and pixel scale factor, and generate the maximum width, average width, and width distribution heatmap by statistically analyzing the crack width values; S5, Length and Topology Analysis Step: Accumulate the spacing between adjacent skeleton points in the crack skeleton line data and multiply by the pixel scale factor to obtain the total crack length value, identify the bifurcation nodes and intersection nodes in the crack skeleton line data to extract crack branch feature data; S6, Temporal Development Analysis Step: Obtain historical crack segmentation masks and crack skeleton line data at different times in the same detection area, perform spatial registration and temporal comparison between the current detection results and historical detection results, calculate the crack propagation rate value and crack propagation direction data, and generate crack development trend curve data.

[0007] This invention also provides a deep learning-based system for refined quantification of bridge crack features. Constructed based on the aforementioned method, it includes an image preprocessing module, a crack semantic segmentation module, a skeleton extraction and morphological analysis module, a normal width measurement module, a length and topology analysis module, and a temporal evolution analysis module. The image preprocessing module receives bridge surface image data and performs grayscale normalization and contrast enhancement to generate preprocessed image data. The crack semantic segmentation module receives the preprocessed image data output from the image preprocessing module and outputs a crack segmentation mask map using a deep learning semantic segmentation network employing a U-Net architecture and incorporating a multi-scale channel spatial attention module. The skeleton extraction and morphological analysis module receives the crack segmentation mask map output from the crack semantic segmentation module, performs morphological thinning operations to obtain crack skeleton line data, and calculates the crack direction angle and crack curvature distribution data. The normal width measurement module receives the crack skeleton line data output from the skeleton extraction and morphological analysis module and the crack segmentation mask map output from the crack semantic segmentation module, calculates the normal direction vector of each skeleton point, measures the crack width value along the normal direction, and generates a width distribution heatmap. The length and topology analysis module receives crack skeleton line data output from the skeleton extraction and morphology analysis module, calculates the total crack length by accumulating the distances between adjacent skeleton points, and identifies bifurcation and intersection nodes to extract crack branch feature data. The temporal development analysis module receives the current output and historical detection results from the crack semantic segmentation module and the skeleton extraction and morphology analysis module, performs spatial registration and temporal comparison to calculate crack propagation rate and direction data, and generates crack development trend curve data.

[0008] Compared with existing technologies, this invention has the following advantages: By integrating a multi-scale channel spatial attention module into the U-Net encoder-decoder architecture, this invention effectively enhances the network's feature extraction capability for slender crack morphology, achieving a crack segmentation cross-union ratio of over 85%, significantly improving crack detection accuracy in complex backgrounds. This invention eliminates errors caused by measurement direction deviations in traditional methods by performing point-by-point width measurement along the normal direction of the skeleton points, achieving a width measurement accuracy better than 0.05mm. This invention achieves a complete characterization of crack geometry by calculating the crack direction angle and curvature distribution. This invention achieves quantitative analysis of crack propagation rate and direction through spatial registration and temporal comparison of multi-temporal detection results, providing technical support for predicting the development trend of bridge crack defects. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating the method for refining and quantifying bridge crack features based on deep learning proposed in this invention.

[0010] Figure 2This is a schematic diagram of the structure of the bridge crack feature refinement and quantification system based on deep learning proposed in this invention. Detailed Implementation

[0011] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0012] Please see Figure 1 A method for refined quantification of bridge crack features based on deep learning includes the following steps.

[0013] S1. Image preprocessing steps.

[0014] The technical purpose of this step is to standardize the original images of the bridge surface collected, eliminate the differences in image quality caused by different acquisition devices, lighting conditions and shooting distances, and provide standardized input data for subsequent deep learning segmentation networks.

[0015] Specifically, the first step is to acquire image data of the bridge surface. Preferably, the acquisition device can be an industrial camera or a high-resolution digital camera, with an image resolution of not less than 2048×2048 pixels, to ensure that minute cracks with a width of 0.1mm can be captured. In one embodiment of the present invention, the acquired bridge surface image data is an RGB three-channel color image with a pixel bit depth of 24 bits.

[0016] Subsequently, the bridge surface image data is subjected to grayscale normalization. Considering that crack detection mainly relies on grayscale information rather than color information, this invention first converts the RGB image to a grayscale image. The grayscale conversion uses a weighted average method, with weight coefficients of 0.299 for the R channel, 0.587 for the G channel, and 0.114 for the B channel. In one embodiment of this invention, the grayscale normalization process uses the minimum-maximum normalization method to map the original grayscale values ​​to the range of 0 to 1. The normalization formula is expressed as follows: , in, Normalized coordinates The pixel value at that location ranges from [0,1] and is dimensionless. The original grayscale image in coordinates The pixel value at that location ranges from [0, 255] and is dimensionless. The minimum grayscale value in the current image; This represents the maximum grayscale value in the current image. This normalization process eliminates grayscale distribution shifts caused by differences in illumination intensity between different images, ensuring that the network input data has a uniform numerical range.

[0017] Furthermore, the normalized image data undergoes contrast enhancement processing. In one embodiment of the present invention, the contrast enhancement processing employs an adaptive histogram equalization method, namely the CLAHE method. Unlike global histogram equalization, the CLAHE method divides the image into multiple small regions and performs histogram equalization on each region separately, effectively avoiding the problem of local over-enhancement that may occur with global equalization. In this invention, the block size for adaptive histogram equalization is set to 8×8 pixels, a size that has been experimentally verified to effectively enhance local details while maintaining overall contrast. The contrast limiting threshold clipLimit is set to a value within the range of 2.0 to 4.0, preferably 3.0. This threshold is used to limit the contrast amplification factor of the local histogram, preventing excessive noise amplification. When clipLimit is set to 3.0, it can effectively suppress concrete surface texture noise while enhancing the contrast between cracks and the background.

[0018] After the above grayscale normalization and contrast enhancement processes, preprocessed image data is generated. In one embodiment of the present invention, the preprocessed image data is a single-channel 32-bit floating-point image with the same size as the original image. In the preprocessed image data, the crack area appears as a linear region that is darker than the surrounding concrete background, and the grayscale contrast is significantly enhanced, providing high-quality input data for the subsequent segmentation network.

[0019] S2, the semantic segmentation step of the crack.

[0020] The technical objective of this step is to use a deep learning semantic segmentation network to perform pixel-level classification on preprocessed image data, accurately separate the crack region in the image from the background, and generate a binarized crack segmentation mask.

[0021] Specifically, preprocessed image data is input into a deep learning semantic segmentation network. In this invention, the deep learning semantic segmentation network adopts the U-Net encoder-decoder architecture as its basic network structure. The U-Net architecture has symmetrical encoder and decoder paths. The encoder extracts image features and reduces spatial resolution layer by layer through continuous convolution and pooling operations. The decoder restores spatial resolution layer by layer through upsampling and convolution operations. Skip connections directly pass the feature maps of each layer of the encoder to the corresponding layer of the decoder, effectively preserving detailed information.

[0022] In one embodiment of the present invention, the encoder of the deep learning semantic segmentation network comprises four downsampling stages. The first stage inputs a single-channel preprocessed image, passes it through two 3×3 convolutional layers, outputs a 64-channel feature map, and then performs 2×2 max pooling to reduce the spatial size to half its original value. The second stage inputs the 64-channel feature map, passes it through two 3×3 convolutional layers, outputs a 128-channel feature map, and then performs 2×2 max pooling to further reduce the spatial size. The third stage inputs the 128-channel feature map, passes it through two 3×3 convolutional layers, outputs a 256-channel feature map, and then performs 2×2 max pooling. The fourth stage inputs the 256-channel feature map, passes it through two 3×3 convolutional layers, outputs a 512-channel feature map, and then performs 2×2 max pooling. The bottom of the encoder is a bridging layer containing two 3×3 convolutional layers with a total of 1024 channels. Batch normalization and ReLU activation functions are applied after each convolutional layer.

[0023] The decoder comprises four upsampling stages, each employing bilinear interpolation to restore the feature map size. Preferably, bilinear interpolation, compared to transposed convolution, avoids checkerboard artifacts and generates a smoother upsampling result. In the first upsampling stage, the 1024-channel feature map output from the bridging layer is upsampled and concatenated with the 512-channel feature map from the fourth stage of the encoder along the channel dimension. After passing through two 3×3 convolutional layers, a 512-channel feature map is output. Similar operations are performed sequentially in the second to fourth upsampling stages, restoring the number of channels to 256, 128, and 64, respectively. The final output layer is a 1×1 convolution, mapping the 64-channel feature map to a single-channel segmentation probability map. After Sigmoid activation, the output value range is [0,1] for pixel-level crack probabilities.

[0024] The innovation of this invention lies in fusing a multi-scale channel spatial attention module into the skip connection layer. The technical function of this module is to adaptively enhance the response of crack-related feature channels and suppress background noise channels, while simultaneously focusing on the crack region in space. The multi-scale channel spatial attention module comprises two cascaded sub-modules: a channel attention branch and a spatial attention branch.

[0025] The specific implementation of the channel attention branch is as follows: Let the input feature map be... ,in For the number of channels, and These represent the feature map height and width, respectively. First, global average pooling and global max pooling are performed on the input feature map to obtain two... Channel description vector and The two vectors are then fed into a multilayer perceptron (MLP) with shared weights for feature transformation. The MLP consists of two fully connected layers, with the intermediate hidden layer having [number of channels]. , For compression ratio, in this invention, we take... ReLU activation is used between the two fully connected layers. Finally, the two outputs of the MLP are summed and activated by Sigmoid to obtain the channel weight vector. The channel attention output is the element-wise product of the input feature map and the channel weight vector: , in, This represents broadcast multiplication. Each element in the array takes values ​​in the range [0,1], representing the importance weight of the corresponding channel.

[0026] The specific implementation of the spatial attention branch is as follows: using the output of the channel attention branch As input, the average and maximum values ​​at each spatial location are first calculated along the channel dimension, resulting in two... Spatial description diagram and The two descriptive graphs are then stitched together along the channel dimension to form... The feature map is then processed by a 7×7 convolutional layer for spatial feature fusion. The 7×7 kernel size is chosen to obtain a larger receptive field to capture the linear distribution features of the cracks. The convolutional output is then activated by a sigmoid function to obtain the spatial weight map. The spatial attention output is: , In this invention, a multi-scale channel spatial attention module is embedded into four skip connection positions of the U-Net to perform attention weighting on the feature maps passed from the encoder to the decoder. Through this module, the network can adaptively focus on crack-related feature channels and spatial regions, effectively improving the detection capability of slender cracks.

[0027] Finally, the probability map output by the Sigmoid algorithm is binarized with a threshold of 0.5 to obtain the crack segmentation mask. In one embodiment of the invention, the crack segmentation mask is a binary image of the same size as the input image, where a pixel value of 1 represents a crack region and a pixel value of 0 represents a background region. The network training uses a weighted combination of the binary cross-entropy loss function and the Dice loss function, with a weight ratio of 0.5:0.5. The training dataset contains 2000 labeled bridge crack images, which are expanded to 10000 images after data augmentation. Data augmentation methods include random rotation, random flipping, random cropping, random brightness adjustment, and elastic deformation. The network is trained for 100 epochs on a workstation equipped with an NVIDIA RTX 3090 graphics card, with an initial learning rate of 0.001, using the Adam optimizer and cosine annealing as the learning rate decay strategy. After training, the crack segmentation mask achieves a segmentation intersection-union ratio (IoU) of 86.2% on the test set, meeting the requirement of at least 85%.

[0028] S3. Skeleton extraction and morphological analysis steps.

[0029] The technical objective of this step is to extract the central skeleton line of the crack from the crack segmentation mask image, and to calculate the crack's orientation angle and curvature distribution based on the skeleton line, providing geometric basis data for subsequent width measurement and topology analysis.

[0030] Specifically, morphological thinning is first performed on the crack segmentation mask to obtain crack skeleton line data. In one embodiment of the present invention, the morphological thinning operation employs the Zhang-Suen thinning algorithm. This algorithm is a classic iterative thinning algorithm that gradually removes boundary pixels of the target region by alternating between two sub-iterations, while maintaining the connectivity and endpoints of the target, ultimately thinning the crack region into skeleton lines with a single pixel width. The specific steps of the Zhang-Suen algorithm are as follows: In the first sub-iteration, boundary pixels that meet specific conditions are marked; in the second sub-iteration, boundary pixels that meet another set of conditions are marked; after each iteration, all marked pixels are deleted; the above process is repeated until no new pixels are deleted. Preferably, before performing the thinning operation, a morphological closing operation is first performed on the crack segmentation mask to fill small holes, with the closing operation structuring element being a 5×5 ellipse. The crack skeleton line data obtained after the thinning operation is a binary image, where points with a pixel value of 1 constitute the skeleton line.

[0031] Subsequently, the tangential direction of each skeleton point is calculated along the crack skeleton line data, and the crack orientation angle is determined. In this invention, the skeleton line is first traversed in an ordered manner to obtain the coordinate sequence of the skeleton points. ,in The total number of skeleton points. The crack orientation angle is calculated as follows: a predetermined number of neighboring skeleton points before and after the current skeleton point are used for least-squares line fitting, and the arctangent of the slope of the fitted line is taken as the tangential angle of the current skeleton point. In one embodiment of the present invention, the number of neighboring skeleton points is set to 5 points before and after, that is, a total of 11 points participate in the fitting. Let the equation of the fitted line be... Then the tangential angle The calculation formula is: , in, For the first The tangential angle of each skeleton point, with a value range of [value missing]. The unit is radians (rad). Let be the slope of the line obtained by least-squares fitting, which is dimensionless. The normal equation for least-squares fitting is: , in, This represents the number of neighborhood points involved in the fitting process, and its value is 11. Traverse the neighborhood of the current point ; and The first The x and y coordinates of each neighboring skeleton point are used. For skeleton points near the endpoints of the skeleton line, if the number of neighboring points is insufficient, a one-sided neighborhood is used for fitting.

[0032] Furthermore, crack curvature distribution data is calculated based on the tangential variation of adjacent skeleton points. Curvature reflects the degree of bending of the crack orientation and is an important indicator characterizing crack morphology. In this invention, the crack curvature distribution data is calculated as follows: the coordinate sequence of skeleton points in the crack skeleton line data is parameterized, the first and second derivatives at each skeleton point are calculated using the central difference method, and the curvature value of each skeleton point is calculated based on the plane curve curvature formula. Specifically, let the skeleton line be parameterized as follows: ,in The parameter is used. The first derivative is calculated using the central difference method: , The second derivative is calculated using the central difference method: , in, The parameter interval between adjacent skeleton points is taken in this invention. Based on the curvature formula of a plane curve, the first... Curvature of a skeleton point The calculation is as follows: , in, For the first The curvature of each skeleton point is expressed in units of 1 / pixel, ranging from [0,1]. A larger curvature value indicates a greater degree of crack bending at that point. The curvature values ​​of each point are arranged sequentially along the skeleton line to form crack curvature distribution data. This data can intuitively reflect the bending variation pattern of the crack along its path, providing a basis for crack type identification and disease assessment.

[0033] S4. Normal width measurement steps.

[0034] The technical objective of this step is to accurately measure the crack width point-by-point along the normal direction of the crack skeleton, generating statistical parameters and a visual distribution map of the crack width. Compared with traditional distance transformation methods, normal width measurement can eliminate errors caused by measurement direction deviations, significantly improving measurement accuracy.

[0035] Specifically, the normal direction vector of each skeleton point in the crack skeleton line data is first calculated. In this invention, the normal direction vector of a skeleton point is determined by rotating the tangential angle of the current skeleton point by 90 degrees to obtain the normal angle, and then constructing a unit normal vector based on the normal angle. Let the first... The tangential angle of each skeleton point is Then the normal angle for: , Unit normal vector for: , in, It is a unit vector with a magnitude of 1 and is dimensionless. The normal direction vector is perpendicular to the tangential direction of the crack skeleton line and points in the direction of crack width.

[0036] Then, a bidirectional search is performed along the normal direction vector of the skeleton point to segment the boundary pixels of the mask image for the crack. The specific implementation is as follows: starting from the current skeleton point... Starting from the beginning, perform pixel searches along both the positive and negative normal directions. During the positive search, visit the positions sequentially. ,in Starting from 1, the value increments gradually by 1 pixel. When the pixel value of the crack segmentation mask at the access location changes from 1 to 0, that boundary location is recorded as the positive boundary point. The negative search employs a similar approach, searching along the negative normal and recording the negative boundary points. .

[0037] The crack width at each skeleton point is calculated based on the boundary pixel spacing and pixel scale factor. In one embodiment of the present invention, the pixel scale factor... The value is determined by the camera parameters and shooting distance during image acquisition, and the unit is mm / pixel. The typical value range is 0.05 mm / pixel to 0.2 mm / pixel; in this embodiment, it is taken as... mm / pixel. No. The crack width at each skeleton point The calculation formula is: , in, For the first The crack width at each skeleton point is expressed in mm and the value range is a positive real number. is the pixel scale factor, in mm / pixel; and These are the coordinates of the positive and negative boundary points, respectively, in pixels. This calculation method measures the width along the true normal direction of the crack, eliminating the systematic error caused by the deviation of the measurement direction from the normal direction in traditional methods.

[0038] Furthermore, statistical analysis of crack width values ​​generates a heatmap showing the maximum width, average width, and width distribution. Maximum width The maximum width value at all skeleton points: , Average width The arithmetic mean of the width values ​​at all skeleton points: , in, This represents the total number of skeleton points. The width distribution heatmap is generated by using the crack segmentation mask as a base image and mapping the color of each skeleton point and its neighboring pixels to a pseudo-color corresponding to the width value. This invention uses a JET color map for color mapping; the larger the width, the more reddish the color, and the smaller the width, the more blue the color. The width distribution heatmap can intuitively display the width variation pattern along the crack, facilitating engineers to quickly locate severely affected sections of the crack.

[0039] In one embodiment of the present invention, the accuracy of width measurement was verified on a test dataset containing 200 bridge cracks. Using manual measurements with vernier calipers as the reference true value, the average absolute error of width measurement using the method of the present invention was 0.038 mm, and the maximum absolute error was 0.045 mm, both better than the accuracy requirement of 0.05 mm. Compared with the traditional method based on distance transformation, the measurement error was reduced by approximately 40%.

[0040] S5, Length and Topology Analysis Steps.

[0041] The technical objective of this step is to calculate the total length of the crack based on the crack skeleton line data, identify the bifurcation nodes and intersection nodes in the skeleton line, and extract the topological features of the crack.

[0042] Specifically, the total crack length is obtained by first summing the distances between adjacent skeleton points in the crack skeleton line data and multiplying by a pixel scale factor. Let the skeleton point coordinate sequence be... Spacing between adjacent skeleton points for: , in, For the first The first skeleton point and the second The Euclidean distance between each skeleton point, in pixels. Total crack length numerical value. for: , in, This represents the total length of the crack, in mm. , where is the pixel scale factor, and the unit is mm / pixel. It should be noted that for crack skeleton lines containing branches, the length of each branch needs to be calculated separately before summing. In one embodiment of the present invention, the average relative error of the length measurement accuracy on the test dataset is 2.1%.

[0043] Subsequently, bifurcation and intersection nodes in the crack skeleton line data were identified to extract crack branch feature data. The bifurcation node identification method was as follows: the number of skeleton points within the eight-neighborhood of each skeleton point in the crack skeleton line data was counted; skeleton points with more than two eight-neighborhood skeleton points were marked as bifurcation nodes. The eight-neighborhood was defined as the eight locations within a 3×3 area centered on the current point, excluding the center point. Let the coordinates of the current skeleton point be... Number of eight neighboring skeleton points The calculation is as follows: , in, For crack skeleton line data in coordinates The pixel value at that location can be either 0 or 1. When At that time, skeleton points It is marked as a fork node.

[0044] The method for identifying intersection nodes is as follows: a node is marked as an intersection node when the number of skeleton line branches connected to it is greater than or equal to 4. The number of branches is determined by analyzing the connectivity patterns of skeleton points within the eight neighborhoods of the intersection node. In one embodiment of the present invention, the number of branches is determined by rotating through the eight neighborhoods and counting the number of transitions from 0 to 1. Specifically, the eight neighborhoods are arranged into a sequence in clockwise or counterclockwise order. Count the number of transitions from 0 to 1 in the sequence. : , in, This is an indicator function that takes the value 1 if the condition is met, and 0 otherwise. Number of jumps. That is, the number of branches. When At that time, the fork node is further marked as a cross node.

[0045] Crack branching characteristics include the location coordinates of bifurcation and intersection nodes, the length of each branch, and its directional angle. These topological features can reflect the development pattern of cracks, and network cracks usually indicate more severe structural damage.

[0046] S6. Steps for time series development analysis.

[0047] The technical objective of this step is to conduct time-series comparative analysis of crack data from multiple tests in the same testing area, quantitatively assess the crack propagation rate and direction, generate crack development trend curves, and provide predictive basis for bridge maintenance decisions.

[0048] Specifically, the process begins by acquiring historical crack segmentation mask images and crack skeleton line data at different times within the same detection area. In one embodiment of the invention, historical detection results are stored in a structural health monitoring database, including detection timestamps, crack segmentation mask images, crack skeleton line data, and quantification parameters. A typical detection cycle is once a month or once a quarter.

[0049] Subsequently, spatial registration is performed between the current and historical detection results. Since the image acquisition positions and angles may differ at different times, spatial registration is necessary to ensure pixel-level correspondence of the cracks. In this invention, spatial registration employs an affine transformation method based on feature point matching. First, ORB feature points are detected in the current and historical images. Then, feature point correspondences are established through brute-force matching or FLANN matching. Finally, the affine transformation matrix is ​​estimated based on the RANSAC method, and a geometric transformation is performed on the historical images. Affine transformation can correct for translation, rotation, and scaling differences, ensuring that the registered historical crack segmentation mask has the same coordinate system as the current detection result.

[0050] Furthermore, the registered current detection results are compared with historical detection results over time to calculate the crack propagation rate. The crack propagation rate is calculated as the difference between the total crack length at the current moment and the total crack length at a historical moment, divided by the time interval. Let the current moment be... Historical moment The corresponding total crack lengths are respectively and Then the crack propagation rate for: , in, The crack propagation rate is expressed in mm / d (millimeters per day). and These are the total crack lengths at the current and historical moments, respectively, in mm; The time interval is expressed in days (d). A positive propagation rate indicates that the crack is extending, while a zero propagation rate indicates that the crack is in a stable state.

[0051] The crack propagation direction data is determined by comparing the change in the endpoint positions of the crack skeleton line data at the current moment with those at historical moments. Specifically, this involves identifying the set of endpoints of the current crack skeleton line. and the set of endpoints of the historical crack skeleton line Calculate the displacement vectors of the corresponding endpoints: , Displacement vector The direction is the direction of crack propagation at that endpoint. The angle representing the propagation direction output by this invention is defined with due east as 0 degrees and counterclockwise as the positive direction.

[0052] Finally, crack development trend curve data is generated. In one embodiment of the present invention, the crack development trend curve data includes the following: a length growth curve with time as the horizontal axis and total crack length as the vertical axis; a width growth curve with time as the horizontal axis and maximum width as the vertical axis; and a rate change curve with time as the horizontal axis and propagation rate as the vertical axis. These curve data can intuitively reflect the historical development pattern of cracks, and can be extrapolated through curve fitting to predict the future development trend of cracks. The present invention uses a linear regression or exponential regression model for trend prediction, and triggers an early warning when the prediction result exceeds a preset safety threshold.

[0053] Please see Figure 2 A deep learning-based system for fine-grained quantification of bridge crack features is constructed based on the deep learning-based method for fine-grained quantification of bridge crack features described in the above-mentioned method embodiments. The system includes an image preprocessing module, a crack semantic segmentation module, a skeleton extraction and morphological analysis module, a normal width measurement module, a length and topology analysis module, and a temporal evolution analysis module.

[0054] The image preprocessing module receives bridge surface image data and performs grayscale normalization and contrast enhancement processing to generate preprocessed image data. In one embodiment of the present invention, the image preprocessing module receives a 2048×2048 resolution RGB image from an industrial camera, first performs grayscale conversion, then uses the minimum-maximum normalization method to map the grayscale values ​​to the [0,1] range, and finally uses the CLAHE method for contrast enhancement. The block size is 8×8 pixels, and the clipLimit is 3.0. The output of the image preprocessing module is single-channel 32-bit floating-point preprocessed image data.

[0055] The crack semantic segmentation module receives preprocessed image data from the image preprocessing module and outputs a crack segmentation mask map through a deep learning semantic segmentation network employing a U-Net architecture and incorporating a multi-scale channel spatial attention module. Preferably, the crack semantic segmentation module is deployed on a GPU-accelerated computing device, with an inference time of approximately 50ms per image. The architecture parameters and training method of the deep learning semantic segmentation network are consistent with those described in step S2 of the method embodiment, achieving a segmentation intersection-union ratio (IoU) of 86.2%. The output of the crack semantic segmentation module is a binary crack segmentation mask map of the same size as the input image.

[0056] The skeleton extraction and morphological analysis module receives the crack segmentation mask image output by the crack semantic segmentation module, performs morphological thinning operations to obtain crack skeleton line data, and calculates crack orientation angle values ​​and crack curvature distribution data. In one embodiment of the present invention, the skeleton extraction and morphological analysis module uses the Zhang-Suen thinning algorithm to extract single-pixel width skeleton lines, uses 11-point neighborhood least squares fitting to calculate the tangential angle of each skeleton point, and uses the central difference and planar curve curvature formula to calculate the curvature distribution. The output of the skeleton extraction and morphological analysis module includes crack skeleton line data, a sequence of orientation angle values ​​for each skeleton point, and a sequence of curvature distribution data.

[0057] The normal width measurement module receives crack skeleton line data output by the skeleton extraction and morphological analysis module and crack segmentation mask map output by the crack semantic segmentation module. It calculates the normal direction vector of each skeleton point and measures the crack width along the normal direction, generating a width distribution heatmap. In one embodiment of the invention, the normal width measurement module determines the normal direction based on a 90-degree rotation of the tangential angle, searches for boundary pixels bidirectionally along the positive and negative normal directions, and calculates the physical width using a pixel scale factor. The output of the normal width measurement module includes the crack width value, maximum width, average width, and width distribution heatmap at each skeleton point. The width measurement accuracy is better than 0.05 mm.

[0058] The length and topology analysis module receives crack skeleton line data output by the skeleton extraction and morphology analysis module, calculates the total crack length by accumulating the distances between adjacent skeleton points, and identifies bifurcation nodes and intersection nodes to extract crack branch feature data. In one embodiment of the invention, the length and topology analysis module calculates the total length by accumulating the Euclidean distance between adjacent skeleton points and multiplying it by a pixel scale factor, identifies bifurcation nodes by counting the number of eight-neighbor skeleton points, and identifies intersection nodes by analyzing the number of jumps in the eight-neighbor area. The output of the length and topology analysis module includes the total crack length, a list of bifurcation node positions, a list of intersection node positions, and the length and orientation angle of each branch.

[0059] The temporal development analysis module receives the current output and historical detection results from the crack semantic segmentation module and the skeleton extraction and morphological analysis module. It performs spatial registration and temporal comparison to calculate the crack propagation rate and crack propagation direction data, generating crack development trend curve data. In one embodiment of the invention, the temporal development analysis module reads historical detection data from the structural health monitoring database, performs spatial registration using ORB feature point matching and affine transformation, calculates the propagation rate based on the total length difference and time interval, and determines the propagation direction based on the endpoint displacement vector. The output of the temporal development analysis module includes the crack propagation rate value, crack propagation direction data, and crack development trend curve data presented in time series form.

[0060] In this system embodiment, the modules form a serial coupling relationship in the data flow: the output of the image preprocessing module serves as the input of the crack semantic segmentation module; the output of the crack semantic segmentation module is simultaneously transmitted to the skeleton extraction and morphological analysis module and the normal width measurement module; the output of the skeleton extraction and morphological analysis module is transmitted to the normal width measurement module, the length and topology analysis module, and the temporal development analysis module. This deeply coupled modular architecture ensures the integrity of the data flow and processing efficiency. Simultaneously, the temporal development analysis module achieves closed-loop feedback by reading historical detection results, enabling the system to dynamically track crack development trends.

[0061] Compared with existing technologies, the system of this invention has the following technical advantages: high-precision crack segmentation is achieved through the U-Net network enhanced by the multi-scale channel spatial attention module, with an IoU of 86.2%; a width measurement accuracy better than 0.05 mm is achieved through the normal width measurement algorithm; a complete characterization of crack geometry is achieved through curvature distribution and orientation angle analysis; and quantitative prediction of crack propagation trend is achieved through the temporal development analysis module, providing a scientific basis for bridge maintenance decisions.

[0062] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. A method for refined quantification of bridge crack features based on deep learning, characterized in that, Includes the following steps: S1. Image preprocessing step: Collect bridge surface image data, perform grayscale normalization and contrast enhancement processing on the bridge surface image data, and generate preprocessed image data. S2. Crack semantic segmentation step: Input the preprocessed image data into the deep learning semantic segmentation network. The deep learning semantic segmentation network adopts the U-Net encoder-decoder architecture and integrates a multi-scale channel spatial attention module in the skip connection layer. The encoder extracts multi-level feature maps and the decoder upsamples layer by layer to restore the spatial resolution and outputs a crack segmentation mask map. S3. Skeleton extraction and morphological analysis steps: Perform morphological refinement operation on the crack segmentation mask to obtain crack skeleton line data, calculate the tangential direction of each skeleton point along the crack skeleton line data and determine the crack direction angle value, and calculate crack curvature distribution data based on the tangential change of adjacent skeleton points. S4. Normal width measurement step: Calculate the normal direction vector of each skeleton point in the crack skeleton line data, search the boundary pixels of the crack segmentation mask map bidirectionally along the normal direction vector of the skeleton point, calculate the crack width value at each skeleton point according to the boundary pixel spacing and pixel scale factor, and generate the maximum width, average width and width distribution heat map by statistically analyzing the crack width values. S5. Length and topology analysis steps: Accumulate the spacing between adjacent skeleton points in the crack skeleton line data and multiply by the pixel scale factor to obtain the total crack length value; identify the bifurcation nodes and intersection nodes in the crack skeleton line data to extract crack branch feature data. S6. Temporal development analysis step: Obtain historical crack segmentation mask images and crack skeleton line data at different times in the same detection area; perform spatial registration and temporal comparison between the current detection results and historical detection results; calculate crack propagation rate values ​​and crack propagation direction data; and generate crack development trend curve data.

2. The method for refined quantification of bridge crack features based on deep learning according to claim 1, characterized in that, In step S1, the grayscale normalization process uses the minimum-maximum normalization method to map pixel values ​​to the range of 0 to 1, and the contrast enhancement process uses the adaptive histogram equalization method. The block size of the adaptive histogram equalization is set to 8×8 pixels, and the contrast limit threshold is set to a value in the range of 2.0 to 4.

0.

3. The method for refined quantification of bridge crack features based on deep learning according to claim 1, characterized in that, In step S2, the encoder of the deep learning semantic segmentation network includes four downsampling stages, with the number of channels in each stage being 64, 128, 256 and 512 respectively. The decoder includes four upsampling stages, and each stage uses bilinear interpolation to restore the feature map size. The segmentation crossover ratio of the crack segmentation mask map is not less than 85%.

4. The method for refined quantification of bridge crack features based on deep learning according to claim 1, characterized in that, In step S2, the multi-scale channel spatial attention module includes a channel attention branch and a spatial attention branch. The channel attention branch extracts channel statistical features through global average pooling and global max pooling and generates channel weight vectors through a multilayer perceptron. The spatial attention branch generates a spatial description map by operating on the average and maximum values ​​along the channel dimension and generates a spatial weight map through convolution. The channel weight vector and the spatial weight map are applied to the input feature map in sequence.

5. The method for refined quantification of bridge crack features based on deep learning according to claim 1, characterized in that, In step S3, the morphological thinning operation uses the Zhang-Suen thinning algorithm to iteratively delete boundary pixels until the crack skeleton line data with a single pixel width is obtained. The crack direction angle value is calculated by taking a preset number of neighboring skeleton points before and after the current skeleton point and performing least squares line fitting. The arctangent value of the slope of the fitted line is used as the tangential angle of the current skeleton point.

6. The method for refined quantification of bridge crack features based on deep learning according to claim 1, characterized in that, In step S3, the calculation method for crack curvature distribution data is as follows: the coordinate sequence of skeleton points of crack skeleton line data is parameterized, the first and second derivatives at each skeleton point are calculated using the central difference method, the curvature values ​​of each skeleton point are calculated based on the plane curve curvature formula, and curvature distribution data is formed along the skeleton line.

7. The method for refined quantification of bridge crack features based on deep learning according to claim 1, characterized in that, In step S4, the normal direction vector of the skeleton point is determined by rotating the tangential angle of the current skeleton point by 90 degrees to obtain the normal angle, and constructing a unit normal vector based on the normal angle. The measurement accuracy of the crack width value is better than 0.05mm.

8. The method for refined quantification of bridge crack features based on deep learning according to claim 1, characterized in that, In step S5, the method for identifying bifurcation nodes is as follows: count the number of skeleton points in the eight neighborhoods of each skeleton point in the crack skeleton line data, and mark the skeleton points with more than 2 skeleton points in the eight neighborhoods as bifurcation nodes. The method for identifying intersection nodes is as follows: mark the intersection nodes when the number of skeleton line branches connected at the bifurcation node is greater than or equal to 4.

9. The method for refined quantification of bridge crack features based on deep learning according to claim 1, characterized in that, In step S6, spatial registration uses an affine transformation method based on feature point matching to geometrically align the crack segmentation mask images at different times. The crack propagation rate is calculated by dividing the difference between the total crack length at the current time and the total crack length at a historical time by the time interval. The crack propagation direction is determined by comparing the change direction of the endpoint positions of the crack skeleton line data at the current time and the historical time.

10. A deep learning-based system for refined quantification of bridge crack features, characterized in that, The method for refined quantification of bridge crack features based on deep learning, as described in any one of claims 1 to 9, is constructed, comprising: The image preprocessing module is used to receive bridge surface image data and perform grayscale normalization and contrast enhancement processing to generate preprocessed image data. The crack semantic segmentation module is used to receive the preprocessed image data output by the image preprocessing module, and output a crack segmentation mask map through a deep learning semantic segmentation network that adopts the U-Net architecture and integrates a multi-scale channel spatial attention module. The skeleton extraction and morphological analysis module is used to receive the crack segmentation mask map output by the crack semantic segmentation module, perform morphological thinning operation to obtain crack skeleton line data, and calculate crack orientation angle values ​​and crack curvature distribution data. The normal width measurement module is used to receive the crack skeleton line data output by the skeleton extraction and morphological analysis module and the crack segmentation mask map output by the crack semantic segmentation module, calculate the normal direction vector of each skeleton point and measure the crack width value along the normal direction, and generate a width distribution heat map. The length and topology analysis module is used to receive the crack skeleton line data output by the skeleton extraction and morphology analysis module, accumulate the spacing between adjacent skeleton points to calculate the total crack length, and identify bifurcation nodes and intersection nodes to extract crack branch feature data. The temporal development analysis module is used to receive the current output and historical detection results of the crack semantic segmentation module and the skeleton extraction and morphological analysis module, perform spatial registration and temporal comparison to calculate the crack propagation rate and crack propagation direction data, and generate crack development trend curve data.

Images