Intelligent crack detection and early warning system for bridge safety operation and maintenance

Abstract

The invention relates to the technical field of image recognition, and discloses an intelligent crack detection and early warning system for bridge safety operation and maintenance, and the system comprises a crack signal-to-noise index field calculation module which obtains a crack signal-to-noise index field through calculation; the enhanced graph calculation module is used for calculating to obtain an enhanced graph; the segmentation mask calculation module is used for calculating and obtaining segmentation masks; the feature extraction module is used for extracting a skeleton and Euclidean distance transformation; the risk prediction module obtains a node risk and a global maximum risk through a risk prediction model; and the risk grade dividing module is used for dividing risk grades. According to the method, the crack visibility is quantified by constructing a crack signal-to-noise index field, the crack definition of a low-visibility area is improved based on self-adaptive enhancement of the index, and image distortion caused by global enhancement is avoided; the self-adaptive threshold segmentation reduces the boundary shrinkage of the segmentation mask, compensates the crack width by combining the width correction driven by the signal-to-noise index, captures the inter-segment space correlation through the risk prediction model, and guarantees the final risk and grade judgment to be accurate.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and more specifically, to an intelligent crack detection and early warning system for bridge safety operation and maintenance. Background Technology

[0002] In bridge safety maintenance, nighttime and early morning inspections often rely on fixed cameras or drones to collect visible light images. In this scenario, strong streetlights, vehicle lights, and structural reflections can easily create a high dynamic range environment. Stray light generated by strong light sources through the optical system and the structural surface will raise the black level of dark areas. At the same time, under low illumination conditions, the photon shot noise and readout noise of the photosensitive sensor are superimposed on each other. The dual effects compress the brightness difference between cracks and the background, resulting in the crack edges being blurred in the original image.

[0003] Existing crack detection processes typically involve performing simple enhancements on the original image before inputting it into a segmentation network. The output probability map is then binarized using a fixed or empirical threshold to obtain a segmentation mask. Since the probability distribution in low-visibility areas tends to be flat, the mask is prone to shrinkage at the boundaries after binarization. Subsequently, when estimating the crack width through skeleton extraction and Euclidean distance transformation, the width value is twice the distance from the skeleton to the boundary. The shrinkage of the mask boundary directly translates into a reduction in the distance measurement value, resulting in a systematic underestimation of the crack width.

[0004] In bridge operation and maintenance, threshold rules based on the maximum crack width are widely used for early warning classification. When the width quantized on the image side is underestimated due to boundary contraction, the same physical crack will be classified into a lower risk level, which directly affects risk summary and subsequent handling decisions. Even after pixel-to-millimeter calibration conversion, this kind of deviation cannot be eliminated. Its root cause is the misjudgment of the position and shape of the crack boundary in the image rather than geometric scale error. Existing methods lack a systematic bias suppression and calibration mechanism for visibility problems, making it difficult to stably output reliable crack width quantization results under low illumination and backlight conditions, which in turn leads to deviations in early warning classification. Summary of the Invention

[0005] This invention provides an intelligent crack detection and early warning system for bridge safety operation and maintenance, solving the technical problems mentioned in the background.

[0006] This invention provides an intelligent crack detection and early warning system for bridge safety operation and maintenance, comprising: The crack signal-to-noise index field calculation module calculates the local mean field of the input image based on the mean convolution kernel, calculates the equivalent noise by combining the noise gain constant and the readout noise constant, takes the gray-level gradient direction as the normal and samples at intervals to obtain the brightness difference, and calculates the crack signal-to-noise index field accordingly. The enhancement map calculation module calculates adaptive gain based on the local mean field and the crack signal-to-noise ratio field, and then combines it with high-pass residuals and clips it to obtain the enhancement map; The segmentation mask calculation module inputs the enhancement map into the segmentation network to obtain network predictions, calculates an adaptive threshold based on the crack signal-to-noise ratio field, and binarizes the network predictions using the adaptive threshold to obtain the segmentation mask. The feature extraction module extracts the skeleton and Euclidean distance transformation based on the segmentation mask, generates the correction factor and visibility confidence based on the crack signal-to-noise index field, and calculates the correction width. It performs segment-level aggregation calculation on the skeleton connected segments to obtain the total segment length, maximum correction width, average correction width, average visibility confidence, and minimum crack signal-to-noise index within the segment. The risk prediction module uses the skeleton connected segments as nodes and the total segment length, maximum correction width, average correction width, average visibility confidence and minimum crack signal-to-noise ratio within the segment as node features. It constructs a crack segment map according to the endpoint spacing and obtains the node risk and the global maximum risk through the risk prediction model. The risk level classification module calculates segment-level visibility penalty and global visibility penalty based on the crack signal-to-noise index field. It performs log-probability shift on node risk and global maximum risk respectively to obtain node-level calibration risk and global prior calibration results. Then, it performs linear weighting in log-probability space to obtain the final risk and classifies the risk level according to fixed boundaries.

[0007] Furthermore, the input image is subjected to discrete convolution with the mean convolution kernel to obtain the local mean, wherein the mean convolution kernel has a radius parameter and the sum of all elements of the convolution kernel is 1; Equivalent noise at pixel x The calculation formula is as follows: Where k represents the noise gain constant, Indicates the readout noise constant. This represents the local mean at pixel x. By summing the local means of all pixels, we obtain the local mean field. Calculate the gray-level gradient and the norm of the gray-level gradient at pixel x in the input image. Divide the gray-level gradient by the sum of the norm of the gray-level gradient and the numerical stability term to obtain the normal at pixel x. Samples are taken on both sides of the normal direction at a normal sampling interval greater than 0, and the absolute value of the image brightness difference between the two sampling points is calculated to obtain the brightness difference. Divide the brightness difference at pixel x by the sum of the equivalent noise and the numerical stability term at pixel x to obtain the crack signal-to-noise ratio at pixel x. Summarize the crack signal-to-noise ratios of all pixels to obtain the crack signal-to-noise ratio field.

[0008] Furthermore, the difference between the input image and the local mean field is calculated to obtain the high-pass residual; Adaptive gain at pixel x The calculation formula is as follows: ,in and This represents the local mean and crack signal-to-noise ratio at pixel x. Indicates increased strength. Represents the numerically stable term; The adaptive gain is multiplied by the high-pass residual and then superimposed on the input image. An enhancement image is obtained by cropping the image. The value of each pixel in the enhancement image ranges from 0 to 1. Enhance the pixel value at pixel x in the image. The calculation formula is as follows: ,in represents the high-pass residual at pixel x, and clip represents the clipping function.

[0009] Furthermore, the enhanced image is input into the segmentation network to obtain network predictions, with each pixel corresponding to a network prediction value ranging from 0 to 1; Adaptive threshold at pixel x The calculation formula is as follows: ,in Indicates the baseline threshold. This indicates the threshold relaxation amount. This indicates the first steepness adjustment parameter; The network prediction value is compared with the adaptive threshold. When the network prediction value is greater than or equal to the adaptive threshold, the corresponding position is output as 1, otherwise it is output as 0, thus obtaining the segmentation mask.

[0010] Furthermore, the segmentation mask is processed by skeleton extraction to extract a set of center lines one pixel wide as the skeleton; Calculate the Euclidean distance from each pixel in the segmentation mask to the nearest background pixel to obtain the Euclidean distance transformation result, and obtain the Euclidean distance transformation value corresponding to the skeleton position through bilinear interpolation; Obtain the crack signal-to-noise ratio corresponding to the skeleton position, and calculate the correction factor by combining the correction amplitude and the second kurtosis adjustment parameter. The correction factor corresponding to the skeleton position x is... The calculation formula is as follows: ,in This represents the signal-to-noise ratio of the crack at location x in the skeleton. Indicates the correction range. denoted by , where tanh represents the hyperbolic tangent activation function; and both the correction magnitude and the second kurtosis adjustment parameter are constants greater than 0. Visibility confidence at skeleton position x The calculation formula is as follows: ,in This represents the third kurtosis adjustment parameter, where the third kurtosis adjustment parameter is a constant greater than 0; The original width is obtained by multiplying the Euclidean distance transformation value corresponding to the skeleton position by 2, and the corrected width is obtained by multiplying the correction factor by the original width.

[0011] Furthermore, the skeleton is divided into multiple connected segments, and a discrete arc length increment greater than 0 is set. An aggregation operation is then performed on each connected segment, including: The total length of the segment is obtained by summing all discrete arc length increments within the connected segment. The maximum value among all correction widths within the connected segment is selected as the maximum correction width. Calculate the sum of the products of each correction width and the corresponding discrete arc length increment within the connected segment, and then divide by the total length of the segment to obtain the average correction width; Calculate the sum of the products of each visibility confidence level within the connected segment and the corresponding discrete arc length increment, and then divide by the total segment length to obtain the average visibility confidence level. The minimum value among the crack signal-to-noise index field values ​​at all skeleton locations within the connected segment is selected as the minimum crack signal-to-noise index within the segment.

[0012] Furthermore, each skeleton connected segment is treated as a graph node, and the total segment length, maximum correction width, average correction width, average visibility confidence, and minimum crack signal-to-noise ratio within the segment corresponding to each skeleton connected segment are used as the node features of that node, forming a node feature vector. Calculate the coordinates of the two endpoints of each skeleton connected segment. For any two different skeleton connected segments, calculate the Euclidean distance between the four endpoints of the two segments. Take the minimum value of the four Euclidean distances as the minimum distance between the endpoints of the two skeleton connected segments. Set an adjacency distance greater than 0. When the minimum distance between the endpoints of two skeleton connected segments is less than or equal to the adjacency distance, it is determined that there is an edge between the two corresponding nodes. No edge is set for the same node itself. Based on this, construct the adjacency matrix.

[0013] Furthermore, a node feature matrix is ​​constructed, where each row of the matrix corresponds to the node feature vector of a node; the degree matrix is ​​calculated, which is a diagonal matrix, where each element on the diagonal is the sum of all elements in the corresponding row of the adjacency matrix; The normalized adjacency matrix is ​​obtained by multiplying the inverse of the degree matrix with the adjacency matrix. The number of updates is set, and the node feature matrix is ​​used as the initial node representation. The graph is updated through the risk prediction model according to the set number of updates. During each update, the normalized adjacency matrix, the current layer node representation, and the model parameters of the corresponding layer are calculated. The calculation result is superimposed with the bias parameters of the corresponding layer and then processed by a nonlinear function to obtain the node representation of the next layer. After the update is completed, the node representation of the final layer is calculated with the readout weight parameters. The calculation results are then superimposed with the readout bias parameters and processed by a nonlinear function to obtain the node risk corresponding to each node. The maximum value among all node risks is selected as the global maximum risk.

[0014] Furthermore, skeleton connected segments Corresponding segment-level visibility penalty The calculation formula is as follows: ,in This represents the total length of segments corresponding to the skeleton connected segments. This represents the signal-to-noise ratio of the crack at location x in the skeleton. tanh represents the discrete arc length increment, and tanh represents the hyperbolic tangent activation function. This represents the fourth kurtosis adjustment parameter, where the fourth kurtosis adjustment parameter is a constant greater than 0; Calibrated node risk The calculation formula is as follows: ,in This indicates the node risk corresponding to the skeleton connectivity segment. The first penalty weight parameter is represented by Sigmoid, and the Sigmoid activation function is represented by Sigmoid. The maximum value among all node calibration risks is selected as the node-level calibration risk; Multiply the total length of each skeleton connected segment by the corresponding segment-level visibility penalty, sum all the multiplication results, and then divide by the sum of the total lengths of all skeleton connected segments to obtain the global visibility penalty. Global prior calibration results The calculation formula is as follows: Where R represents the global maximum risk, This represents a global visibility penalty. This represents the second penalty weight parameter.

[0015] Furthermore, log-odds transformation is performed on the node-level calibration risk and the global prior calibration result respectively; a fusion coefficient between 0 and 1 is set, and the result of subtracting the fusion coefficient from 1 is multiplied by the log-odds transformation result of the node-level calibration risk, and then the result of multiplying the fusion coefficient by the log-odds transformation result of the global prior calibration result is added to obtain the final log-odds value. The final logarithmic odds value is then processed by the Sigmoid function to obtain the final risk. Log-odds transformation of node-level calibration risk The calculation formula is as follows: ,in This indicates the risk of calibration at the node level; Log-probability transformation of global prior calibration results The calculation formula is as follows: ,in This represents the global prior calibration result; Define two fixed boundaries, where the first fixed boundary is greater than 0, and the second fixed boundary is greater than the first fixed boundary but less than 1; When the final risk is greater than or equal to 0 and less than the first fixed boundary, it is classified as Level 1; when the final risk is greater than or equal to the first fixed boundary and less than the second fixed boundary, it is classified as Level 2; when the final risk is greater than or equal to the second fixed boundary, it is classified as Level 3.

[0016] The beneficial effects of this invention are as follows: This invention quantifies crack visibility by constructing a crack signal-to-noise index field, providing a unified judgment basis for the entire process; and based on this index, adaptive enhancement can specifically improve the clarity of cracks in low-visibility areas, avoiding image distortion caused by global enhancement; adaptive threshold segmentation reduces the shrinkage of segmentation mask boundaries, and combined with width correction driven by the signal-to-noise index, accurately compensates for the systematic underestimation of crack width; a graph structure is constructed with crack segments as nodes, and the spatial correlation between segments is captured through a risk prediction model, followed by segment-level and global visibility penalty calibration and dual-source fusion to ensure accurate final risk and level determination. Overall, this invention significantly improves the accuracy of crack detection and the reliability of early warning in low-light backlight scenarios, providing accurate and reliable decision support for bridge safety operation and maintenance. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the intelligent crack detection and early warning system for bridge safety operation and maintenance according to the present invention. Figure 2 This is a flowchart of the intelligent crack detection and early warning method for bridge safety operation and maintenance according to the present invention. Figure 3 This is the segment-level feature visualization radar chart of the present invention; Figure 4 This is a scatter plot of the total segment length error of the present invention; Figure 5 This is a scatter plot of the maximum correction width error of the present invention.

[0018] The diagram includes: crack signal-to-noise ratio field calculation module 101, enhancement map calculation module 102, segmentation mask calculation module 103, feature extraction module 104, risk prediction module 105, and risk level classification module 106. Detailed Implementation

[0019] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0020] It should be noted that, unless otherwise defined, the technical or scientific terms used in one or more embodiments of the present invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in one or more embodiments of the present invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" indicate that the element or object preceding the term encompasses the elements or objects listed following the term and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0021] like Figures 1-2 As shown, the intelligent crack detection and early warning system for bridge safety operation and maintenance includes: The crack signal-to-noise index field calculation module 101 is used to calculate the local mean field of the input image based on the mean convolution kernel, calculate the equivalent noise by combining the noise gain constant and the readout noise constant, take the gray-level gradient direction as the normal and sample at intervals to obtain the brightness difference, and calculate the crack signal-to-noise index field accordingly. Enhancement map calculation module 102 is used to calculate adaptive gain based on local mean field and crack signal-to-noise index field, combine high-pass residual and crop to obtain enhancement map; The segmentation mask calculation module 103 is used to input the enhancement map into the segmentation network to obtain network prediction, calculate the adaptive threshold based on the crack signal-to-noise index field, and binarize the network prediction through the adaptive threshold to obtain the segmentation mask. The feature extraction module 104 is used to extract the skeleton and Euclidean distance transformation according to the segmentation mask, generate the correction factor and visibility confidence according to the crack signal-to-noise index field, calculate the correction width, perform segment-level aggregation calculation on the skeleton connected segments, and obtain the total segment length, maximum correction width, average correction width, average visibility confidence and minimum crack signal-to-noise index within the segment. The risk prediction module 105 is used to construct a crack segment map based on the endpoint spacing, using the skeleton connected segments as nodes and the total segment length, maximum correction width, average correction width, average visibility confidence and minimum crack signal-to-noise index within the segment as node features. It also obtains the node risk and the global maximum risk through the risk prediction model. The risk level classification module 106 is used to calculate the segment-level visibility penalty and the global visibility penalty based on the crack signal-to-noise index field, perform log-probability shift on the node risk and the global maximum risk respectively to obtain the node-level calibration risk and the global prior calibration result, and perform linear weighting in the log-probability space to obtain the final risk, and classify the risk level according to the fixed boundary.

[0022] It should be noted that drone inspections can be used to collect image data. Following a pre-planned inspection route, the drone can comprehensively cover the bridge deck, piers, towers, and other structures. During flight, the drone's high-definition camera can be set to an appropriate resolution (e.g., no less than 20 megapixels) and frame rate (e.g., more than 20 frames per second) to capture images of the bridge surface at different heights and angles. This allows for the rapid acquisition of images of large areas of the bridge structure, is not limited by terrain, and can approach high-risk areas of the bridge (e.g., high piers, cable stays). Compared to manual inspections, this method is more efficient and safer, and can provide rich and clear image data for subsequent inspections. Further details will not be elaborated here.

[0023] In one embodiment of the present invention, the input image is subjected to discrete convolution operation with a mean convolution kernel to obtain a local mean, wherein the mean convolution kernel has a radius parameter and the sum of all elements of the convolution kernel is 1; Equivalent noise at pixel x The calculation formula is as follows: Where k represents the noise gain constant, Indicates the readout noise constant. This represents the local mean at pixel x. By summing the local means of all pixels, we obtain the local mean field. Calculate the gray-level gradient and the norm of the gray-level gradient at pixel x in the input image. Divide the gray-level gradient by the sum of the norm of the gray-level gradient and the numerical stability term to obtain the normal at pixel x. Samples are taken on both sides of the normal direction at a normal sampling interval greater than 0, and the absolute value of the image brightness difference between the two sampling points is calculated to obtain the brightness difference. Divide the brightness difference at pixel x by the sum of the equivalent noise and the numerical stability term at pixel x to obtain the crack signal-to-noise ratio at pixel x. Summarize the crack signal-to-noise ratios of all pixels to obtain the crack signal-to-noise ratio field.

[0024] It should be noted that the noise gain constant is used to map image brightness to photon noise intensity, quantifying the relationship between brightness and noise. It is a custom parameter, preferably set to 0.1. The specific value can be determined through device calibration or data learning, for example, by fitting multiple sets of image data with known noise characteristics. The readout noise constant represents the noise constant generated when the imaging device reads the signal. It is determined by the device characteristics and needs to be provided by the device manufacturer or determined through noise testing, for example, by testing under dark conditions. The numerical stability term is set to 10 to the power of -6. The normal direction represents the direction of the fastest change in image intensity, i.e., the unit vector of the grayscale gradient. The normal sampling interval represents the half-interval of sampling along the normal direction, in pixels. It is a custom parameter, preferably set to 1 pixel, determined based on empirical values ​​of image resolution and crack width. For example, when the crack width is typically 2 pixels, the normal sampling interval is set to 1 pixel.

[0025] It should be noted that the mean convolution kernel uses a square uniform kernel. For example, when the radius parameter is set to 3, the convolution kernel is a 3×3 matrix, with each element being 1 / 9, ensuring that the sum of all elements is 1. The discrete convolution operation slides the mean convolution kernel across the image. For each pixel position, the convolution kernel is multiplied by the image pixel values ​​of the corresponding area around that pixel, and then all products are summed to obtain the local mean result of that pixel. Through this operation, the averaging of local areas of the image can be achieved, which is used to extract the local mean features of the image. The gray-level gradient is calculated using the Sobel operator, calculating the gradient components in the x and y directions respectively. The magnitude of the gray-level gradient of each pixel can be obtained by taking the square root of the sum of the squares of the gradient components in the x and y directions, and the gradient direction can be determined by the ratio of the x and y direction components using the arctangent function. In addition, the crack signal-to-noise ratio field quantization is used to approximate the degree of collapse of crack visibility in near-high-brightness and low-light scenes, providing unified basic data for subsequent image enhancement, crack segmentation, geometric measurement, and other processes.

[0026] In one embodiment of the present invention, the difference between the input image and the local mean field is calculated to obtain the high-pass residual; Adaptive gain at pixel x The calculation formula is as follows: ,in and This represents the local mean and crack signal-to-noise ratio at pixel x. Indicates increased strength. Represents the numerically stable term; The adaptive gain is multiplied by the high-pass residual and then superimposed on the input image. An enhancement image is obtained by cropping the image. The value of each pixel in the enhancement image ranges from 0 to 1. Enhance the pixel value at pixel x in the image. The calculation formula is as follows: ,in represents the high-pass residual at pixel x, and clip represents the clipping function.

[0027] It should be noted that the high-pass residual represents the difference between the input image and the local mean field, extracting image edge and detail information; the adaptive gain represents the gain value for specifically enhancing the visibility of crack edges; the enhancement intensity represents a calibrated value that controls the overall enhancement amplitude to be greater than 0. Preferably, the enhancement intensity is set between 0.5 and 2. Setting it too small will result in weak enhancement and cracks will still be difficult to distinguish, while setting it too large will easily distort the image. For example, it can be set to 1.5 for low-light images and 0.8 for highlight images; the cropping function is used to control the pixel value between 0 and 1. By calculating the high-pass residual to extract crack details, and combining it with the adaptive gain constructed from the crack signal-to-noise ratio and the local mean, targeted enhancement is achieved only for areas where cracks are difficult to see. This avoids the distortion of global enhancement and improves the distinguishability of low-visibility cracks, providing a clearer input for subsequent crack segmentation, and ultimately ensuring the accuracy and reliability of bridge crack detection.

[0028] In one embodiment of the present invention, the enhanced image is input into the segmentation network to obtain network predictions, and each pixel corresponds to a network prediction value with a value range between 0 and 1. Adaptive threshold at pixel x The calculation formula is as follows: ,in Indicates the baseline threshold. This indicates the threshold relaxation amount. This indicates the first steepness adjustment parameter; The network prediction value is compared with the adaptive threshold. When the network prediction value is greater than or equal to the adaptive threshold, the corresponding position is output as 1, otherwise it is output as 0, thus obtaining the segmentation mask.

[0029] It should be noted that the adaptive threshold is a binary judgment standard that changes with pixel position. The baseline threshold, threshold relaxation amount, and first steepness adjustment parameter are all user-defined parameters. Preferably, the baseline threshold is set to 0.6, because the probability output by the segmentation network is around 0.5 to distinguish between positive and negative classes. Combined with the adaptive adjustment requirements, 0.6 can ensure basic accuracy in areas with obvious cracks, while leaving room for threshold relaxation in low-visibility areas. Preferably, the threshold relaxation amount is set to 0.2. If it is too small, the threshold adjustment in low-visibility areas will be insufficient, making it difficult to detect blurry cracks. If it is too large, it will easily introduce a large number of background false detections. Preferably, the first steepness adjustment parameter is set to 3. If it is too small, the threshold adjustment in low-visibility areas will be insignificant. If it is too large, it may cause drastic threshold fluctuations. The segmentation network can use a U-Net structure, which consists of an encoding path (multiple downsampling to extract features), a decoding path (multiple upsampling to restore resolution), and skip connections (fusion of features at different scales). During training, a weighted binary cross-entropy loss is used to improve the segmentation weights in low-visibility areas, which will not be elaborated here.

[0030] In one embodiment of the present invention, the segmentation mask is processed by a skeleton extraction operation to extract a set of center lines with a width of one pixel as the skeleton; Calculate the Euclidean distance from each pixel in the segmentation mask to the nearest background pixel to obtain the Euclidean distance transformation result, and obtain the Euclidean distance transformation value corresponding to the skeleton position through bilinear interpolation; Obtain the crack signal-to-noise ratio corresponding to the skeleton position, and calculate the correction factor by combining the correction amplitude and the second kurtosis adjustment parameter. The correction factor corresponding to the skeleton position x is... The calculation formula is as follows: ,in This represents the signal-to-noise ratio of the crack at location x in the skeleton. Indicates the correction range. denoted by , where tanh represents the hyperbolic tangent activation function; and both the correction magnitude and the second kurtosis adjustment parameter are constants greater than 0. Visibility confidence at skeleton position x The calculation formula is as follows: ,in This represents the third kurtosis adjustment parameter, where the third kurtosis adjustment parameter is a constant greater than 0; The original width is obtained by multiplying the Euclidean distance transformation value corresponding to the skeleton position by 2, and the corrected width is obtained by multiplying the correction factor by the original width.

[0031] It should be noted that the skeleton extraction employs morphological thinning algorithms, such as the Zhang-Suen thinning algorithm, which iteratively removes edge pixels that meet the criteria for deletability, ultimately obtaining a one-pixel-wide skeleton while preserving the topological structure of the cracks. The Euclidean distance transformation can utilize a fast, scan-based algorithm, such as a two-pass algorithm. The first pass calculates a temporary distance along a certain direction, and the second pass optimizes along the opposite direction to obtain the accurate Euclidean distance, improving computational efficiency. Bilinear interpolation first determines the coordinates of four pixels surrounding the skeleton position, then calculates the distance weight between each point and the skeleton position (e.g., the reciprocals of the horizontal and vertical distances), and finally sums the weighted Euclidean distance values ​​of the four pixels to obtain the interpolated Euclidean distance value. The correction amplitude is a custom parameter; preferably, it is set to 0.5. Setting the kurtosis parameter too small results in insufficient correction and fails to effectively compensate for the underestimation of width in low-visibility areas; setting it too large leads to overestimation of width. The second kurtosis adjustment parameter is a custom parameter, preferably set to 3. Setting it too small results in an overly gentle effect on the correction factor, leading to insufficient correction strength in low-visibility areas; setting it too large causes drastic fluctuations in the correction factor. The third kurtosis adjustment parameter is also a custom parameter, preferably set to 2. Through skeleton extraction, Euclidean distance transformation, and correction based on the crack signal-to-noise ratio, accurate measurement of crack width and quantification of visibility reliability are achieved. This not only compensates for the underestimation of width in low-visibility areas but also provides segment-level geometric and reliability indicators for subsequent risk assessment, ensuring accurate determination of the true size and visibility of cracks even under complex lighting conditions.

[0032] In one embodiment of the present invention, the skeleton is divided into multiple connected segments, a discrete arc length increment greater than 0 is set, and an aggregation operation is performed on each connected segment, including: The total length of the segment is obtained by summing all discrete arc length increments within the connected segment. The maximum value among all correction widths within the connected segment is selected as the maximum correction width. Calculate the sum of the products of each correction width and the corresponding discrete arc length increment within the connected segment, and then divide by the total length of the segment to obtain the average correction width; Calculate the sum of the products of each visibility confidence level within the connected segment and the corresponding discrete arc length increment, and then divide by the total segment length to obtain the average visibility confidence level. The minimum value among the crack signal-to-noise index field values ​​at all skeleton locations within the connected segment is selected as the minimum crack signal-to-noise index within the segment.

[0033] It should be noted that a connected segment represents a connected sub-path in the skeleton, with its two ends being endpoints, and is the basic unit of segment-level aggregation. Skeleton connected segment partitioning can employ connected region labeling algorithms, such as the two-pass scanning method, based on the pixel adjacency relationships of 4-neighborhoods or 8-neighborhoods, dividing the skeleton into multiple connected segments, with the endpoints as the segment boundaries. Discrete arc length increments represent the approximate arc length between adjacent pixels on the skeleton, which can be obtained using Euclidean distance calculations. For example, in a 512×512 resolution bridge crack image, the arc length increment between adjacent pixels on the skeleton is precisely calculated based on their coordinate differences (such as (1,0), (0,1), (1,1), etc.). The exact calculation of Euclidean distance is not elaborated here; the total segment length represents the sum of all discrete arc length increments within the connected segment, used to quantify the length of the crack segment; the maximum corrected width represents the maximum corrected width within the connected segment, reflecting the widest part of the crack within the segment; the average corrected width represents the length-weighted average of the corrected widths within the connected segment, reflecting the overall level of width within the segment; the average visibility confidence represents the length-weighted average of the visibility confidence within the connected segment, reflecting the overall reliability of visibility within the segment; the minimum crack signal-to-noise ratio within the segment represents the minimum crack signal-to-noise ratio within the connected segment, reflecting the visibility of the most blurred area within the segment.

[0034] It should be noted that by performing segment-level aggregation on the skeleton connected segments, the geometric and credibility information of cracks is elevated from the pixel level to the segment level, providing structured feature input for subsequent graph structure construction and risk warning. This segment-level aggregation not only retains the key geometric attributes of cracks, but also integrates the information of visibility credibility and the most ambiguous areas, upgrading crack risk assessment from single-pixel judgment to segment-based structured analysis, which is more in line with the risk assessment of crack clusters in bridge engineering.

[0035] In one embodiment of the present invention, each skeleton connected segment is treated as a graph node, and the total segment length, maximum correction width, average correction width, average visibility confidence, and minimum crack signal-to-noise ratio within the segment corresponding to each skeleton connected segment are used as the node features of that node to form a node feature vector. Calculate the coordinates of the two endpoints of each skeleton connected segment. For any two different skeleton connected segments, calculate the Euclidean distance between the four endpoints of the two segments. Take the minimum value of the four Euclidean distances as the minimum distance between the endpoints of the two skeleton connected segments. Set an adjacency distance greater than 0. When the minimum distance between the endpoints of two skeleton connected segments is less than or equal to the adjacency distance, it is determined that there is an edge between the two corresponding nodes. No edge is set for the same node itself. Based on this, construct the adjacency matrix.

[0036] It should be noted that the adjacency distance is a custom parameter, typically set between 5 and 20 pixels. For example, for a bridge detection image with a pixel resolution of 0.1 mm / pixel, 10 pixels correspond to 1 mm, which can effectively capture the spatial relationship between adjacent crack segments on the same bridge component. The endpoint coordinates of the skeleton connected segment can be obtained by traversing all pixels of the skeleton connected segment, counting the number of adjacent pixels in the 4-neighbor or 8-neighbor domains of each pixel, and the pixel with 1 adjacent pixel is the endpoint, and its row and column coordinates are read as the endpoint coordinates. The risk prediction model uses a graph convolutional network (GCN). The minimum distance between endpoints is the minimum Euclidean distance of the four endpoints of two connected segments. The adjacency matrix determines whether nodes are connected based on the minimum distance between endpoints and the adjacency distance. The node risk is the risk value of a single crack segment, and the global maximum risk is the maximum value of the risks of all nodes, reflecting the overall highest risk of bridge cracks.

[0037] In one embodiment of the present invention, a node feature matrix is ​​constructed, wherein each row of the matrix corresponds to the node feature vector of a node; a degree matrix is ​​calculated, wherein the degree matrix is ​​a diagonal matrix, and each element on the diagonal is the sum of all elements in the corresponding row of the adjacency matrix; The normalized adjacency matrix is ​​obtained by multiplying the inverse of the degree matrix with the adjacency matrix. The number of updates is set, and the node feature matrix is ​​used as the initial node representation. The graph is updated through the risk prediction model according to the set number of updates. During each update, the normalized adjacency matrix, the current layer node representation, and the model parameters of the corresponding layer are calculated. The calculation result is superimposed with the bias parameters of the corresponding layer and then processed by a nonlinear function to obtain the node representation of the next layer. After the update is completed, the node representation of the final layer is calculated with the readout weight parameters. The calculation results are then superimposed with the readout bias parameters and processed by a nonlinear function to obtain the node risk corresponding to each node. The maximum value among all node risks is selected as the global maximum risk.

[0038] It should be noted that the update count is a custom parameter, typically set between 3 and 5. Too few updates will prevent nodes from fully aggregating risk information from adjacent segments, leading to insufficient prediction accuracy; too many updates will increase model complexity and make it prone to overfitting. The degree matrix is ​​a diagonal matrix, with diagonal elements being the sum of elements in the corresponding rows of the adjacency matrix. If a diagonal element is 0, it is set to 1 to avoid the absence of an inverse matrix, ensuring that the sum of elements in each row is 1. The calculation formula for the graph propagation update of the risk prediction model is: Node representation at layer t+1 = Nonlinear function (Normalized adjacency matrix × Node representation at layer t × Model parameters + Bias parameters), where the model parameters are the weight matrix, the bias parameters are the bias vector, and the nonlinear function is such as ReLU. The calculation formula for node risk is a nonlinear function (Final layer node representation × Readout weight parameters + Readout bias parameters), where the nonlinear function is Sigmoid, outputting a risk probability of 0 to 1. The node feature matrix integrates the geometric and confidence information of the crack segment, and the normalized adjacency matrix is ​​used for information weighting in graph propagation. The update count determines the depth of information aggregation, and the model parameters and bias parameters are used for the linear transformation and offset of the node representation.

[0039] It should be noted that the risk prediction model based on graph structure achieves structured aggregation of crack segment risks. It not only preserves the individual attributes of crack segments by utilizing the node feature matrix, but also captures the spatial correlation between segments through the normalized adjacency matrix and multiple rounds of updates. This upgrades risk assessment from a single crack segment to a global judgment based on the collaboration of multiple segments, thus aligning with the safety impact of crack clusters in bridge engineering and improving the accuracy of early warning.

[0040] In one embodiment of the present invention, the skeleton connection segment Corresponding segment-level visibility penalty The calculation formula is as follows: ,in This represents the total length of segments corresponding to the skeleton connected segments. This represents the signal-to-noise ratio of the crack at location x in the skeleton. tanh represents the discrete arc length increment, and tanh represents the hyperbolic tangent activation function. This represents the fourth kurtosis adjustment parameter, where the fourth kurtosis adjustment parameter is a constant greater than 0; Calibrated node risk The calculation formula is as follows: ,in This indicates the node risk corresponding to the skeleton connectivity segment. The first penalty weight parameter is represented by Sigmoid, and the Sigmoid activation function is represented by Sigmoid. The maximum value among all node calibration risks is selected as the node-level calibration risk; Multiply the total length of each skeleton connected segment by the corresponding segment-level visibility penalty, sum all the multiplication results, and then divide by the sum of the total lengths of all skeleton connected segments to obtain the global visibility penalty. Global prior calibration results The calculation formula is as follows: Where R represents the global maximum risk, This represents a global visibility penalty. This represents the second penalty weight parameter.

[0041] It should be noted that the fourth steepness adjustment parameter is a custom parameter that controls the steepness of the impact of the crack signal-to-noise ratio on the segment-level visibility penalty; preferably, the fourth steepness adjustment parameter is set to 3. The first penalty weight parameter is a custom parameter that controls the magnitude of the impact of the segment-level visibility penalty on the node risk calibration; preferably, the first penalty weight parameter is set to 0.5. The second penalty weight parameter is a custom parameter that controls the magnitude of the impact of the global visibility penalty on the global prior calibration; preferably, the second penalty weight parameter is set to 0.3. The segment-level visibility penalty reflects the risk penalty caused by insufficient visibility in the skeleton connectivity segment; the larger the value, the greater the risk to the node. The larger the calibration range, the better; the calibrated node risk represents the node risk value after segment-level visibility penalty calibration, which is closer to the actual risk brought by visibility; the node-level calibration risk represents the maximum value among all calibrated node risks, reflecting the highest risk after segment-level calibration; the global visibility penalty represents the result of the visibility penalty of all skeleton connected segments weighted by the total segment length, reflecting the impact of global visibility on risk; the global prior calibration result represents the global risk after global visibility penalty calibration, which is used for final risk fusion; that is, through segment-level and global two-layer visibility penalty calibration, the fine adjustment of risk from single segment to global is achieved.

[0042] In one embodiment of the present invention, log-odds transformation is performed on the node-level calibration risk and the global prior calibration result respectively; a fusion coefficient between 0 and 1 is set, and the result of subtracting the fusion coefficient from 1 is multiplied by the log-odds transformation result of the node-level calibration risk, and then the result of multiplying the fusion coefficient by the log-odds transformation result of the global prior calibration result is added to obtain the final log-odds value. The final logarithmic odds value is then processed by the Sigmoid function to obtain the final risk. Log-odds transformation of node-level calibration risk The calculation formula is as follows: ,in This indicates the risk of calibration at the node level; Log-probability transformation of global prior calibration results The calculation formula is as follows: ,in This represents the global prior calibration result; Define two fixed boundaries, where the first fixed boundary is greater than 0, and the second fixed boundary is greater than the first fixed boundary but less than 1; When the final risk is greater than or equal to 0 and less than the first fixed boundary, it is classified as Level 1; when the final risk is greater than or equal to the first fixed boundary and less than the second fixed boundary, it is classified as Level 2; when the final risk is greater than or equal to the second fixed boundary, it is classified as Level 3.

[0043] It should be noted that the fusion coefficient represents the weight used to balance the calibration risk at the node level and the global prior calibration result, with a value ranging from 0 to 1, determining the contribution ratio of the two in the final risk. Preferably, the fusion coefficient is set to 0.5. The final log-odds value represents the weighted sum of the log-odds at the node level and the log-odds at the global prior, which is the intermediate value of the fused information. The final risk represents the risk value after Sigmoid processing, which is the final result after the fusion of node and global information. The first fixed boundary and the second fixed boundary are both user-defined parameters. Preferably, the first fixed boundary is set to 1 / 3 and the second fixed boundary is set to 2 / 3. Level 1 represents the level when the final risk is lower than the first fixed boundary, indicating a low risk that requires attention. Level 2 represents the level when the final risk is between the first and second fixed boundaries, indicating a medium risk that requires warning. Level 3 represents the level when the final risk is higher than the second fixed boundary, indicating a high risk that requires alarm. Through linear fusion of the log-odds space and clear risk level classification, the transformation from multi-source risk information to a unified quantitative level is realized, namely, a three-level classification of attention, warning, and alarm, enabling bridge maintenance personnel to efficiently judge crack risks and formulate handling strategies.

[0044] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0045] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A smart crack detection and early warning system for bridge safety operation and maintenance, characterized in that, include: The crack signal-to-noise index field calculation module calculates the local mean field of the input image based on the mean convolution kernel, calculates the equivalent noise by combining the noise gain constant and the readout noise constant, takes the gray-level gradient direction as the normal and samples at intervals to obtain the brightness difference, and calculates the crack signal-to-noise index field accordingly. The enhancement map calculation module calculates adaptive gain based on the local mean field and the crack signal-to-noise ratio field, and then combines it with high-pass residuals and clips it to obtain the enhancement map; The segmentation mask calculation module inputs the enhancement map into the segmentation network to obtain network predictions, calculates an adaptive threshold based on the crack signal-to-noise ratio field, and binarizes the network predictions using the adaptive threshold to obtain the segmentation mask. The feature extraction module extracts the skeleton and Euclidean distance transformation based on the segmentation mask, generates the correction factor and visibility confidence based on the crack signal-to-noise index field, and calculates the correction width. It performs segment-level aggregation calculation on the skeleton connected segments to obtain the total segment length, maximum correction width, average correction width, average visibility confidence, and minimum crack signal-to-noise index within the segment. The risk prediction module uses the skeleton connected segments as nodes and the total segment length, maximum correction width, average correction width, average visibility confidence and minimum crack signal-to-noise ratio within the segment as node features. It constructs a crack segment map according to the endpoint spacing and obtains the node risk and the global maximum risk through the risk prediction model. The risk level classification module calculates segment-level visibility penalty and global visibility penalty based on the crack signal-to-noise index field. It performs log-probability shift on node risk and global maximum risk respectively to obtain node-level calibration risk and global prior calibration results. Then, it performs linear weighting in log-probability space to obtain the final risk and classifies the risk level according to fixed boundaries.

2. The intelligent crack detection and early warning system for bridge safety operation and maintenance according to claim 1, characterized in that, The input image is discretely convolved with a mean convolution kernel to obtain a local mean, where the mean convolution kernel has a radius parameter and the sum of all elements of the kernel is 1; Equivalent noise at pixel x The calculation formula is as follows: Where k represents the noise gain constant, Indicates the readout noise constant. This represents the local mean at pixel x. By summing the local means of all pixels, we obtain the local mean field. Calculate the gray-level gradient and the norm of the gray-level gradient at pixel x in the input image. Divide the gray-level gradient by the sum of the norm of the gray-level gradient and the numerical stability term to obtain the normal at pixel x. Samples are taken on both sides of the normal direction at a normal sampling interval greater than 0, and the absolute value of the image brightness difference between the two sampling points is calculated to obtain the brightness difference. Divide the brightness difference at pixel x by the sum of the equivalent noise and the numerical stability term at pixel x to obtain the crack signal-to-noise ratio at pixel x. Summarize the crack signal-to-noise ratios of all pixels to obtain the crack signal-to-noise ratio field.

3. The intelligent crack detection and early warning system for bridge safety operation and maintenance according to claim 1, characterized in that, Calculate the difference between the input image and the local mean field to obtain the high-pass residual; Adaptive gain at pixel x The calculation formula is as follows: ,in and This represents the local mean and crack signal-to-noise ratio at pixel x. Indicates increased strength. Represents the numerically stable term; The adaptive gain is multiplied by the high-pass residual and then superimposed on the input image. An enhancement image is obtained by cropping the image. The value of each pixel in the enhancement image ranges from 0 to 1. Enhance the pixel value at pixel x in the image. The calculation formula is as follows: ,in represents the high-pass residual at pixel x, and clip represents the clipping function.

4. The intelligent crack detection and early warning system for bridge safety operation and maintenance according to claim 1, characterized in that, The augmented image is input into the segmentation network to obtain the network prediction. Each pixel corresponds to a network prediction value that ranges from 0 to 1. Adaptive threshold at pixel x The calculation formula is as follows: ,in Indicates the baseline threshold. This indicates the threshold relaxation amount. This indicates the first steepness adjustment parameter; The network prediction value is compared with the adaptive threshold. When the network prediction value is greater than or equal to the adaptive threshold, the corresponding position is output as 1, otherwise it is output as 0, thus obtaining the segmentation mask.

5. The intelligent crack detection and early warning system for bridge safety operation and maintenance according to claim 1, characterized in that, The segmentation mask is processed by skeleton extraction to extract a set of center lines with a width of one pixel as the skeleton. Calculate the Euclidean distance from each pixel in the segmentation mask to the nearest background pixel to obtain the Euclidean distance transformation result, and obtain the Euclidean distance transformation value corresponding to the skeleton position through bilinear interpolation; Obtain the crack signal-to-noise ratio corresponding to the skeleton position, and calculate the correction factor by combining the correction amplitude and the second kurtosis adjustment parameter. The correction factor corresponding to the skeleton position x is... The calculation formula is as follows: ,in This represents the signal-to-noise ratio of the crack at location x in the skeleton. Indicates the correction range. denoted by , where tanh represents the hyperbolic tangent activation function; and both the correction magnitude and the second kurtosis adjustment parameter are constants greater than 0. Visibility confidence at skeleton position x The calculation formula is as follows: ,in This represents the third kurtosis adjustment parameter, where the third kurtosis adjustment parameter is a constant greater than 0; The original width is obtained by multiplying the Euclidean distance transformation value corresponding to the skeleton position by 2, and the corrected width is obtained by multiplying the correction factor by the original width.

6. The intelligent crack detection and early warning system for bridge safety operation and maintenance according to claim 1, characterized in that, The skeleton is divided into multiple connected segments, and a discrete arc length increment greater than 0 is set. An aggregation operation is performed on each connected segment, including: The total length of the segment is obtained by summing all discrete arc length increments within the connected segment. The maximum value among all correction widths within the connected segment is selected as the maximum correction width. Calculate the sum of the products of each correction width and the corresponding discrete arc length increment within the connected segment, and then divide by the total length of the segment to obtain the average correction width; Calculate the sum of the products of each visibility confidence level within the connected segment and the corresponding discrete arc length increment, and then divide by the total segment length to obtain the average visibility confidence level. The minimum value among the crack signal-to-noise index field values ​​at all skeleton locations within the connected segment is selected as the minimum crack signal-to-noise index within the segment.

7. The intelligent crack detection and early warning system for bridge safety operation and maintenance according to claim 1, characterized in that, Each skeleton connected segment is treated as a graph node, and the total segment length, maximum correction width, average correction width, average visibility confidence, and minimum crack signal-to-noise ratio within the segment corresponding to each skeleton connected segment are used as the node features to form a node feature vector. Calculate the coordinates of the two endpoints of each skeleton connected segment. For any two different skeleton connected segments, calculate the Euclidean distance between the four endpoints of the two segments. Take the minimum value of the four Euclidean distances as the minimum distance between the endpoints of the two skeleton connected segments. Set an adjacency distance greater than 0. When the minimum distance between the endpoints of two skeleton connected segments is less than or equal to the adjacency distance, it is determined that there is an edge between the two corresponding nodes. No edge is set for the same node itself. Based on this, construct the adjacency matrix.

8. The intelligent crack detection and early warning system for bridge safety operation and maintenance according to claim 1, characterized in that, Construct a node feature matrix, where each row of the matrix corresponds to the node feature vector of a node; Calculate the degree matrix, which is a diagonal matrix, where each element on the diagonal is the sum of all elements in the corresponding row of the adjacency matrix; The normalized adjacency matrix is ​​obtained by multiplying the inverse of the degree matrix with the adjacency matrix. The number of updates is set, and the node feature matrix is ​​used as the initial node representation. The graph is updated through the risk prediction model according to the set number of updates. During each update, the normalized adjacency matrix, the current layer node representation, and the model parameters of the corresponding layer are calculated. The calculation result is superimposed with the bias parameters of the corresponding layer and then processed by a nonlinear function to obtain the node representation of the next layer. After the update is completed, the node representation of the final layer is calculated with the readout weight parameters. The calculation results are then superimposed with the readout bias parameters and processed by a nonlinear function to obtain the node risk corresponding to each node. The maximum value among all node risks is selected as the global maximum risk.

9. The intelligent crack detection and early warning system for bridge safety operation and maintenance according to claim 1, characterized in that, skeleton connected segments Corresponding segment-level visibility penalty The calculation formula is as follows: ,in This represents the total length of segments corresponding to the skeleton connected segments. This represents the signal-to-noise ratio of the crack at location x in the skeleton. tanh represents the discrete arc length increment, and tanh represents the hyperbolic tangent activation function. This represents the fourth kurtosis adjustment parameter, where the fourth kurtosis adjustment parameter is a constant greater than 0; Calibrated node risk The calculation formula is as follows: ,in This indicates the node risk corresponding to the skeleton connectivity segment. The first penalty weight parameter is represented by Sigmoid, and the Sigmoid activation function is represented by Sigmoid. The maximum value among all node calibration risks is selected as the node-level calibration risk; Multiply the total length of each skeleton connected segment by the corresponding segment-level visibility penalty, sum all the multiplication results, and then divide by the sum of the total lengths of all skeleton connected segments to obtain the global visibility penalty. Global prior calibration results The calculation formula is as follows: Where R represents the global maximum risk, This represents a global visibility penalty. This represents the second penalty weight parameter.

10. The intelligent crack detection and early warning system for bridge safety operation and maintenance according to claim 1, characterized in that, Log-odds transformation is performed on the node-level calibration risk and the global prior calibration result respectively; a fusion coefficient between 0 and 1 is set, and the result of subtracting the fusion coefficient from 1 is multiplied by the log-odds transformation result of the node-level calibration risk, and then added to the fusion coefficient multiplied by the log-odds transformation result of the global prior calibration result to obtain the final log-odds value. The final logarithmic odds value is then processed by the Sigmoid function to obtain the final risk. Log-odds transformation of node-level calibration risk The calculation formula is as follows: ,in This indicates the risk of calibration at the node level; Log-probability transformation of global prior calibration results The calculation formula is as follows: ,in This represents the global prior calibration result; Define two fixed boundaries, where the first fixed boundary is greater than 0, and the second fixed boundary is greater than the first fixed boundary but less than 1; When the final risk is greater than or equal to 0 and less than the first fixed boundary, it is classified as Level 1; when the final risk is greater than or equal to the first fixed boundary and less than the second fixed boundary, it is classified as Level 2; when the final risk is greater than or equal to the second fixed boundary, it is classified as Level 3.

Images