Bridge apparent disease early warning method and system based on image analysis and storage medium

Abstract

The application provides a bridge apparent disease early warning method based on image analysis. The method comprises the following steps: acquiring an original apparent image of a bridge monitoring, separating a low-frequency background component and a high-frequency disease component from the original apparent image, performing background adaptive enhancement on the high-frequency disease component, extracting deep features, and simultaneously magnifying the extracted deep features and the low-frequency background by 4 times resolution through hierarchical high-frequency compensation up-sampling; outputting a disease probability soft segmentation map based on the enhanced feature map, extracting a boundary transition area from the disease probability soft segmentation map, determining a sub-pixel signed distance field of the boundary transition area, optimizing the smoothness of the sub-pixel signed distance field, taking the zero contour line of the signed distance field as a sub-pixel level disease boundary, performing spatial continuity smoothing according to the pixel level width, finally outputting disease geometric parameters, and triggering a corresponding grade warning. The technical problems of low recall rate and large width measurement error of small-scale diseases can be solved.

Description

Technical Field

[0001] This invention relates to the field of image analysis technology, and specifically to a method, system, and storage medium for early warning of bridge surface defects based on image analysis. Background Technology

[0002] Identifying surface defects in bridges is a core part of routine bridge maintenance and safety assessment. Early detection of defects such as microcracks smaller than 0.2mm allows for timely maintenance measures, significantly reducing the risk of structural deterioration and operation and maintenance costs.

[0003] While deep learning-based image recognition solutions are widely used in bridge defect detection, they still have significant limitations in small-scale, early-stage defect identification scenarios. Conventional super-resolution reconstruction algorithms use an end-to-end full-image reconstruction mode. The smoothing process of the concrete background can easily cover up the edges of weak and subtle defects, resulting in the dilution of the features of cracks smaller than 0.2mm and a recall rate of less than 40%, which is difficult to meet the needs of early warning. Existing disease segmentation schemes mostly use pixel-level hard segmentation, combined with discrete offset correction of boundaries, which has problems such as rough edges and poor continuity. The average error of crack width measurement is about ±0.8mm, which is far from meeting the warning accuracy requirement of 0.1mm. Furthermore, traditional methods have low feature extraction efficiency for slender linear targets such as cracks, and do not incorporate prior knowledge from the field of bridge engineering. Background interference such as stains, water stains, and construction joints can easily cause a large number of false alarms, and the recognition robustness is insufficient in complex scenarios.

[0004] Therefore, there is an urgent need for image analysis-based methods, systems, and storage media for early warning of bridge surface defects, which are tailored to the morphology and scene characteristics of early-stage minor defects in bridges. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes a bridge appearance defect early warning method, system, and storage medium based on image analysis, in order to solve the technical problems of low recall rate for small-scale defects and large width measurement error in existing technologies.

[0006] The technical solution adopted in this invention includes: The original appearance image of the bridge monitoring is obtained, and the low-frequency background component and high-frequency defect component are separated from the original appearance image. The high-frequency defect component is subjected to background adaptive enhancement, and then deep features are extracted. At the same time, the extracted deep features and low-frequency background are upsampled by 4 times resolution through hierarchical high-frequency compensation. The upsampled low-frequency background and deep features are fused to obtain a high-resolution image with edge fidelity. Multi-scale linear features are extracted from high-resolution images. These multi-scale linear features are then double-calibrated using channel domain disease statistics and spatial domain linear shape prior attention to obtain enhanced feature maps that suppress background interference and strengthen disease features. Based on the enhanced feature map, the disease probability soft segmentation map is output. The boundary transition region is extracted from the disease probability soft segmentation map, and the sub-pixel symbolic distance field of the boundary transition region is determined. Then, the smoothness of the sub-pixel symbolic distance field is constrained and optimized, and the zero contour line of the symbolic distance field is used as the sub-pixel level disease boundary. Based on the subpixel disease boundary and crack skeleton, the pixel-level width is calculated by projection along the crack skeleton normal, converted into physical width, and then spatially smoothed to finally output the disease geometric parameters and trigger the corresponding level of early warning.

[0007] Furthermore, the original appearance image is separated into low-frequency background components and high-frequency defect components. The low-frequency background component is obtained by convolving the original appearance image with a large variance Gaussian kernel. The high-frequency defect component is obtained by subtracting the low-frequency background component from the original appearance image. Then, the high-frequency defect component is adaptively enhanced by using the crack high-frequency feature enhancement coefficient and the concrete background suppression coefficient to obtain the enhanced high-frequency component. The enhanced high-frequency component is input into multiple cascaded crack-sensing residual units. The gradient magnitude of the enhanced high-frequency component is calculated through a single crack-sensing residual unit, and a gradient adaptive weighting factor is generated. The convolutional features are weighted by channel attention weights and then connected to the enhanced high-frequency component to output deep high-frequency features. Then, the deep high-frequency features are upsampled twice by 2x pixel rearrangement. After each upsampling, a shallow high-frequency compensation term of the corresponding scale is injected to restore the edge details of the disease step by step, and finally a high-frequency feature map magnified by 4x is obtained. The low-frequency background component is magnified four times and mapped to the high-frequency feature map to obtain a reconstructed high-frequency detail map. This map is then superimposed element-wise to obtain a high-resolution image. The joint loss includes L1 loss for full-image reconstruction, edge fidelity loss for diseased areas, and background pseudo-texture suppression loss. The adaptive background enhancement of the high-frequency disease component is expressed as follows: Ihigh=ωc·Ihigh,raw⊙1-ωb·Ilow255 Where Ihigh,raw represents the original high-frequency component, Ilow represents the low-frequency background component, ωc represents the high-frequency feature enhancement coefficient of cracks, ωb represents the background suppression coefficient of concrete, and ⊙ represents the element-wise Hadamard product operation.

[0008] The output expression of the single crack sensing residual unit is: Fi=Fi-1+Fconv,i⊙Wgrad,i⊙CAi Where Fi-1 is the input feature of the i-th residual unit, Fconv,i is the output feature of the main path convolution, Wgrad,i is the gradient adaptive weighting factor, and CAi is the channel attention weight; the gradient adaptive weighting factor Wgrad,i is generated by normalizing the gradient magnitude of the input feature, and the region with higher edge intensity has a greater feature gain.

[0009] Furthermore, the extraction of multi-scale linear features from high-resolution images includes setting directional dilated convolution branches in four directions: 0°, 45°, 90°, and 135°. The convolution kernels of each branch expand only along the corresponding direction, while maintaining the original spacing in the vertical direction. Linear features in different directions are extracted respectively, and channels are stitched together to obtain multi-directional fused features. Global average pooling is performed on the multi-directional fusion features, and channel weights are generated by combining the predefined disease channel prior vectors. The multi-directional fusion features are then weighted and calibrated channel by channel. This includes calculating the structural tensor and linearity index of each pixel of the multi-directional fusion features, generating the final spatial attention weights by combining the initial spatial weights, and weighting the multi-directional fusion features pixel by pixel to enhance the response of continuous linear regions and suppress scattered noise.

[0010] Furthermore, the aforementioned soft segmentation map based on enhanced feature map outputting disease probability first outputs a pixel-level disease probability map by convolution and sigmoid activation of the enhanced feature map, preserving the probability gradient information of the boundary region; then, pixels with disease probabilities in the range of 0.1 to 0.9 are selected to form a boundary transition region mask; the transition region mask is then used to calculate the sub-pixel symbolic distance field based on the enhanced feature map regression; the positive values ​​of the distance field are internal values ​​of the disease, the negative values ​​are background values, and the zero value is the precise boundary value; and the gradient direction of the sub-pixel symbolic distance field is calculated, and the gradient direction is used as the boundary normal to constrain the continuous angle between the normals of adjacent pixels, ensuring that the boundary is smooth and conforms to physical laws.

[0011] Furthermore, the pixel-level width is calculated by projecting along the normal direction of the crack skeleton based on the sub-pixel disease boundary and crack skeleton, and then converted into physical width and spatial continuity smoothed to finally output the disease geometric parameters. First, the physical size of the single pixel corresponding to the high-resolution image is calculated based on the physical resolution of the original apparent image and the super-resolution magnification factor, and the physical resolution is calibrated. Then, the contour line extraction algorithm is used to extract the zero contour lines of the symbol distance field. Based on the zero contour lines, the sub-pixel boundary point sets on the left and right sides are obtained. Probability of disease Figure 2Value-enhanced, single-pixel crack skeletons are extracted, and the boundary normal angle of each skeleton point is calculated based on the gradient direction of the signed distance field. A straight line is drawn along the normal of each skeleton point, and its intersection with the sub-pixel boundaries on both sides is used to obtain the width endpoints. The pixel width is then calculated and converted into the physical width. Finally, the width sequence is smoothed by weighted average of adjacent points, as expressed by: Wi,smooth=0.2·Wi-1+0.6·Wi+0.2·Wi+1 Where Wi is the original physical width of the i-th skeleton point, Wi-1 and Wi+1 are the physical widths of adjacent skeleton points, and Wi,smooth is the smoothed width value. The crack length, average width, maximum width, and 95th percentile width are statistically obtained. Based on the crack length, average width, maximum width, and 95th percentile width, a preset threshold is matched, and a graded warning is triggered.

[0012] The aforementioned graded early warning rules include Level 1, Level 2, and Level 3 early warnings. The Level 3 early warning is triggered when any of the following conditions are met, corresponding to the early development of minor diseases, and prompts an increase in the frequency of daily inspections: the average width of the crack is ≥0.1mm and the total length of the crack is ≥50mm. The width of the crack at the 95th percentile is ≥0.12mm; The aforementioned Level II early warning is triggered when any of the following conditions are met, indicating that the corresponding disease is showing an expanding trend, prompting an on-site special review and testing: The average width of the crack is ≥0.15mm; The total length of the crack is ≥100mm; The width of the crack at the 95th percentile is ≥0.15mm; The Level 1 warning is triggered when any of the following conditions are met: the corresponding disease has reached the early stage limit of the standard or has the risk of large-scale expansion, prompting the initiation of an evaluation of the maintenance and treatment plan: The maximum width of the crack is ≥0.2mm, reaching the standard threshold for early-stage disease; The average crack width is ≥0.18mm and the total crack length is ≥200mm; The width of the crack at the 95th percentile is ≥0.18mm and the total length of the crack is ≥150mm.

[0013] Furthermore, the present invention provides a bridge surface defect early warning system based on image analysis, including a high-low frequency separation super-resolution reconstruction module, which performs high-low frequency separation, high-frequency enhancement, hierarchical upsampling and fusion reconstruction on the original bridge surface image to output a high-resolution image; The multi-directional dual-domain attention enhancement module extracts multi-directional linear disease features, enhances disease features and suppresses background interference through dual prior attention in the channel domain and spatial domain, and outputs an enhanced feature map. The distance field regression module outputs a soft segmentation map of disease probability, regresses the subpixel continuous symbolic distance field and applies normal consistency constraints, and extracts zero contour lines as subpixel disease boundaries. The quantitative early warning module is used to calculate the physical width based on the sub-pixel boundary and skeleton normal, and after spatial smoothing, statistically analyze the disease parameters and trigger graded early warnings.

[0014] Furthermore, the present invention provides a computer storage medium storing a computer program that can run on a processor, wherein the processor executes the computer program to implement the image analysis-based bridge appearance defect early warning method described above.

[0015] As can be seen from the above technical solution, the beneficial technical effects of the present invention are as follows: This invention employs a high-low frequency separation progressive super-resolution method, which enhances and amplifies only the high-frequency components of the disease to avoid smoothing the background and obscuring fine edges. Combined with gradient-weighted residual units to strengthen edge features, the pixel ratio of 0.1~0.2mm fine cracks is increased by 4 times after 4x magnification, greatly improving the recall rate and effectively capturing early-stage diseases.

[0016] By replacing discrete offset regression with a continuous symbolic distance field and combining it with normal consistency physical constraints, the boundary positioning accuracy can reach the 0.01 pixel level. Combined with contour normal projection and spatial smoothing calculations, the average error of crack width measurement is reduced to ±0.08 mm, meeting the 0.1 mm early warning accuracy requirement, and significantly improving width quantification accuracy. Multi-directional oriented hollow convolution is used to adapt to the linear morphology of cracks, combined with a dual attention mechanism of channel statistical prior and spatial shape prior, embedding knowledge from the bridge engineering domain. This effectively suppresses background interference such as stains, water stains, and rough surfaces, greatly reducing the false alarm rate and significantly enhancing the anti-interference capability in complex scenarios. Attached Figure Description

[0017] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0018] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a system structure diagram of the present invention; Figure 3 This is a structural diagram of the storage medium of the present invention. Detailed Implementation

[0019] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are merely illustrative of the technical solution of the present invention and are therefore intended to limit the scope of protection of the present invention.

[0020] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0021] Example 1: like Figure 1 As shown, this embodiment provides a method for early warning of bridge surface defects based on image analysis; High- and low-frequency separation progressive super-resolution reconstruction: The input for this step is the original acquired RGB image of the bridge appearance. , dimension ,in For image height, Image width; the physical resolution of the original image is obtained by pre-calibration using a calibration plate. The unit is mm / pixel. This step outputs a 4x super-resolution image. , dimension .

[0022] The specific implementation process is as follows: 1. High and low frequency component separation and adaptive enhancement are performed using a Gaussian kernel with a standard deviation of 3. For the original image Convolution is performed to obtain the low-frequency background component of the concrete:

[0023] in For two-dimensional convolution operations, the low-frequency background component is subtracted from the original image to obtain the original high-frequency component: Introducing high-frequency characteristic enhancement coefficient of crack Concrete background inhibition coefficient Background adaptive enhancement is performed on the original high-frequency components:

[0024] in This is for element-wise Hadamard product calculation. In the above formula, the brightness of the smooth concrete area is high, with an inhibition factor of 1. Smaller values ​​can suppress background surface noise; higher contrast at the disease edge can enhance the disease edge features.

[0025] Specifically, it will enhance high-frequency components. Input 16 cascaded crack sensing residual units and extract deep high-frequency features. For the i-th residual unit (i=1,2,...,16), the input features are... .

[0026] Specifically, the processing procedure is as follows: Main path convolution extracts features:

[0027] Calculate the gradient magnitude of the input features:

[0028] in , These are the Sobel gradient operators in the x and y directions, respectively.

[0029] Specifically, the gradient adaptive weighting factor is then generated:

[0030] Regions with higher edge intensity have larger weighting factors and stronger feature gains.

[0031] The channel attention weights are calculated using the following expression:

[0032] in, It is the Sigmoid activation function. For global average pooling, This is for global max pooling.

[0033] The residual connection is then recalculated to output the final features:

[0034] After cascading 16 residual units, the deep high-frequency characteristics are output. , dimension .

[0035] In this embodiment, a graded high-frequency compensation upsampling is performed, with 2x pixel rearrangement upsampling performed twice. After each upsampling, a shallow high-frequency compensation is injected to avoid edge blurring.

[0036] First 2x upsampling: ;

[0037]

[0038]

[0039] in, This is a primary high-frequency compensation term, directly anchoring the edge of the defect. A second 2x upsampling (cumulative 4x magnification):

[0040]

[0041]

[0042] Output high-frequency feature map magnified 4 times , dimension .

[0043] Specifically, in this embodiment, the high-low frequency fusion reconstruction includes amplifying the low-frequency background component by a factor of 4: Mapping high-frequency feature maps to RGB detail maps: ; in, As the convolution bias term, high and low frequencies are element-wise superimposed to obtain the final high-resolution image: During the training phase, the joint loss function uses a three-part joint loss to optimize network parameters: in, For high-resolution ground truth images, weighting coefficients .

[0044] The fidelity loss at the edge of the diseased area (calculated only in the true value diseased area) is expressed as follows: in, This is a binary mask for the true value of the disease. It is an L1 norm.

[0045] Background pseudo-texture suppression loss (calculated only in ground truth background regions):

[0046] Minimize high-frequency components in the background region to avoid generating false textures.

[0047] In this embodiment, the input for the multi-directional dual-domain prior attention feature enhancement step is a high-resolution image. The output is an enhanced feature map. , dimension .

[0048] The specific implementation process is as follows: First, multi-directional dilated convolution feature extraction is performed, with four directional dilated convolution branches set at 0°, 45°, 90°, and 135°. The convolution kernel size is 3×3, and the dilation rate is 2. The convolution kernel of each branch dilates only along the corresponding direction, maintaining a 1-pixel spacing in the vertical direction, so that the receptive field extends along the direction of the gap, improving the efficiency of linear feature extraction. For each direction... Output direction features:

[0049] The four branch features are concatenated along the channel dimension to obtain multi-directional fused features: , dimension .

[0050] In this embodiment, the channel domain defect statistical prior attention calibration is based on the channel statistical prior of bridge defects (cracks have a strong response in the blue channel, and corrosion has a strong response in the red channel), and a predefined defect channel prior vector is defined. The dimension and the number of feature channels are consistent, and the channels with strong disease response have higher weights.

[0051] Global average pooling is applied to the fused features:

[0052] Generate channel weights by combining prior vectors:

[0053] Channel calibration characteristics are obtained by weighting each channel individually. In this embodiment, the spatial domain linear shape prior attention enhancement first calculates the initial spatial weights:

[0054] in, , These represent average pooling and max pooling along the channel dimension, respectively. The structure tensor and linearity index for each pixel are calculated: the gradient covariance matrix is ​​calculated for each pixel. Solving for eigenvalues Define linearity:

[0055] The closer the linearity is to 1, the closer the region is to an ideal straight line, and the better it matches the crack morphology. A crack linearity enhancement factor is introduced. Adjust spatial weights:

[0056] The final enhanced feature map is obtained by weighting pixels one by one:

[0057] In this embodiment, the sub-pixel symbolic distance field regression under normal constraints uses the enhanced feature map as the specific input. The output is a disease probability map. Distance field with subpixel symbol The specific implementation process is as follows: First, soft segmentation based on disease probability is performed, followed by dimensionality reduction through 1×1 convolution, and then 3×3 convolution with Sigmoid activation. Output pixel-level disease probability map:

[0058]

[0059] ] The probability that pixel (x,y) belongs to a disease is preserved, and the gradient information of the boundary region is retained.

[0060] In this embodiment, boundary transition region extraction is performed, including filtering pixels with a disease probability in the range of 0.1 to 0.9 to form a boundary transition region mask:

[0061] High-precision distance field regression is performed only on this region to reduce computational redundancy.

[0062] Then, subpixel symbolic distance field regression is performed, specifically including obtaining the symbolic distance field based on the enhanced feature map through 3×3 convolution regression:

[0063] This represents the signed distance from pixel (x, y) to the actual boundary of the lesion, in pixels; positive values ​​are used inside the lesion, and negative values ​​are used in the background. The contour lines are the precise subpixel disease boundaries.

[0064] In this embodiment, the gradient direction of the normal consistency constraint distance field is the boundary normal, and the normal component of each pixel is calculated: To ensure the continuity of normals between adjacent pixels, define the normal consistency loss: The smaller the angle between the normals of adjacent pixels, the closer the cosine value is to 1, the smaller the loss, ensuring that the boundary is smooth and conforms to the physical characteristics of the crack.

[0065] In this embodiment, the joint loss function during the training phase

[0066] in, For binary cross-entropy splitting loss, the weight coefficients are... , The distance field regression loss is calculated only in the boundary transition region.

[0067] The true subpixel distance field is generated by calculating the precise boundaries marked manually.

[0068] In this embodiment, the disease width is quantified and warned using contour integral quantification, and the input for this step is a disease probability map. Symbolic distance field Compared to the original physical resolution The output includes the geometric parameters of the disease and the warning level.

[0069] Specifically, after performing a 4x super-resolution calibration of the physical resolution, the physical resolution of the high-resolution image is: The unit is mm / pixel, meaning that a single pixel corresponds to a physical size reduced by a factor of 4.

[0070] Sub-pixel boundary extraction employs the MarchingSquares contour extraction algorithm to extract the zero contour lines of the symbolic distance field, thereby obtaining continuous sub-pixel defect boundaries, which are then discretized into sets of boundary points on the left and right sides. , The boundary positioning accuracy can reach 0.01 pixels.

[0071] Crack skeleton and normal calculation first determine the probability of disease. Figure 2 Value-based, the expression is as follows:

[0072] The ZhangSuen thinning algorithm was used to extract the single-pixel crack skeleton: The skeleton point set is denoted as N is the total number of skeleton points. For each skeleton point... The boundary normal angle is calculated based on the gradient of the signed distance field: This refers to the direction in which the crack width is measured, which is strictly perpendicular to the crack direction.

[0073] Normal projection width calculation for each skeleton point Along the normal direction Draw a straight line It intersects with the sub-pixel boundaries on the left and right sides respectively. , Two points, with a pixel-level width equal to the Euclidean distance between them:

[0074] Convert to physical width: .

[0075] In this embodiment, spatial smoothing and parameter statistics are performed. Specifically, a three-frame weighted smoothing method is used to filter the width sequence, which conforms to the physical law of gradual crack width change. Statistical analysis of core geometric parameters of the disease: Total crack length: Average width: Maximum width: 95th percentile width: In this embodiment, the three-level early warning determination is based on the four dimensions of crack total length, average width, maximum width, and 95th percentile width, constructing a three-level early warning mechanism. The triggering rules and maintenance response levels for each level are as follows: Level 3 early warning, nascent stage warning: Triggered when any of the following conditions are met, corresponding to the initial development of early, minor diseases, prompting an increase in the frequency of daily inspections and recording of disease development trends: Average crack width And the total length of the crack ; 95th percentile width of crack ; Level 2 warning, development stage warning: Triggered when any of the following conditions are met, indicating a clear trend of disease expansion, prompting an on-site special review and investigation of the disease's causes: Average crack width And the total length of the crack ; 95th percentile width of crack ; Level 1 Warning, Severe Warning: Triggered when any of the following conditions are met, indicating that the corresponding disease has reached the early stage limit of the standard or has the risk of large-scale expansion, prompting the initiation of maintenance and treatment plan assessment and structural safety review: Maximum crack width The bridge maintenance standards have been met for early-stage defects. Average crack width And the total length of the crack ; 95th percentile width of crack And the total length of the crack .

[0076] Example 2 This invention provides an image analysis-based early warning system for bridge surface defects, such as... Figure 2It includes a high- and low-frequency separation super-resolution reconstruction module, which performs high- and low-frequency separation, high-frequency enhancement, hierarchical upsampling and fusion reconstruction on the original bridge appearance image, and outputs a high-resolution image. The multi-directional dual-domain attention enhancement module extracts multi-directional linear disease features, enhances disease features and suppresses background interference through dual prior attention in the channel domain and spatial domain, and outputs an enhanced feature map. The distance field regression module outputs a soft segmentation map of disease probability, regresses the subpixel continuous symbolic distance field and applies normal consistency constraints, and extracts zero contour lines as subpixel disease boundaries. The quantitative early warning module is used to calculate the physical width based on the sub-pixel boundary and skeleton normal, and after spatial smoothing, statistically analyze the disease parameters and trigger graded early warnings.

[0077] Example 3 In this embodiment, as Figure 3 The present invention provides a computer storage medium storing a computer program that can run on a processor, wherein the processor executes the computer program to implement the image analysis-based bridge appearance defect early warning method described above.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A method for early warning of bridge apparent defects based on image analysis, characterized in that, include: The original appearance image of the bridge monitoring is obtained. The low-frequency background component and the high-frequency defect component are separated from the original appearance image. The high-frequency defect component is subjected to background adaptive enhancement. Then, deep features are extracted. At the same time, the extracted deep features and low-frequency background are upsampled by 4 times resolution through hierarchical high-frequency compensation. The upsampled low-frequency background and deep features are fused to obtain a high-resolution image. Multi-scale linear features are extracted from high-resolution images, and these features are then double-calibrated using channel domain disease statistics and spatial domain linear shape prior attention to obtain an enhanced feature map. Based on the enhanced feature map, the disease probability soft segmentation map is output. The boundary transition region is extracted from the disease probability soft segmentation map, and the sub-pixel symbolic distance field of the boundary transition region is determined. Then, the smoothness of the sub-pixel symbolic distance field is constrained and optimized, and the zero contour line of the symbolic distance field is used as the sub-pixel level disease boundary. Based on the subpixel disease boundary and crack skeleton, the pixel-level width is calculated by projection along the crack skeleton normal. Spatial continuity smoothing is performed based on the pixel-level width, and finally the disease geometric parameters are output, triggering the corresponding level of early warning.

2. The method for early warning of bridge apparent defects based on image analysis according to claim 1, characterized in that, The process involves separating the low-frequency background component and the high-frequency disease component from the original appearance image, convolving the original appearance image with a large variance Gaussian kernel to obtain the low-frequency background component, and then subtracting the low-frequency background component from the original appearance image to obtain the high-frequency disease component. Then, by using the crack high-frequency feature enhancement coefficient and the concrete background inhibition coefficient, the high-frequency disease component is adaptively enhanced to obtain the enhanced high-frequency component. The enhanced high-frequency component is input into multiple cascaded crack-sensing residual units. The gradient magnitude of the enhanced high-frequency component is calculated through a single crack-sensing residual unit, and a gradient adaptive weighting factor is generated. The convolutional features are weighted by channel attention weights and then connected to the enhanced high-frequency component to output deep high-frequency features. Then, the deep high-frequency features are upsampled twice by 2x pixel rearrangement. After each upsampling, a shallow high-frequency compensation term of the corresponding scale is injected to restore the edge details of the disease step by step, and finally a high-frequency feature map magnified by 4x is obtained. The low-frequency background component is magnified four times and mapped to the high-frequency feature map to obtain a reconstructed high-frequency detail map. The high-resolution image is then obtained by superimposing the low-frequency background component element by element.

3. The method for early warning of bridge apparent defects based on image analysis according to claim 2, characterized in that, The aforementioned background adaptive enhancement of high-frequency defect components is expressed as follows: in, For the original high-frequency components, For low-frequency background components, This represents the high-frequency characteristic enhancement coefficient of the crack. The background inhibition coefficient of concrete. This is for element-wise Hadamard product operation.

4. The method for early warning of bridge apparent defects based on image analysis according to claim 2, characterized in that, The output expression of the single crack sensing residual unit is: in, For the first Input features of each residual unit The main path convolution outputs features. This is the gradient adaptive weighting factor. Channel attention weights; gradient adaptive weighting factor It is generated by normalizing the gradient magnitude of the input features.

5. The method for early warning of bridge apparent defects based on image analysis according to claim 1, characterized in that, The method for extracting multi-scale linear features from high-resolution images includes setting directional dilated convolution branches in four directions: 0°, 45°, 90°, and 135°. The convolution kernels of each branch expand only along the corresponding direction, while maintaining the original spacing in the vertical direction. Linear features in different directions are extracted separately, and channels are stitched together to obtain multi-directional fused features. Global average pooling is performed on the multi-directional fusion features, and channel weights are generated by combining the predefined disease channel prior vectors. The multi-directional fusion features are then weighted and calibrated channel by channel. This includes calculating the structural tensor and linearity index of each pixel in the multi-directional fusion features, and generating the final spatial attention weights by combining the initial spatial weights.

6. The method for early warning of bridge apparent defects based on image analysis according to claim 1, characterized in that, The aforementioned soft segmentation map based on enhanced feature map outputting disease probability first outputs a pixel-level disease probability map by convolution and sigmoid activation of the enhanced feature map, preserving the probability gradient information of the boundary region; then, pixels with disease probabilities in the range of 0.1 to 0.9 are selected to form a boundary transition region mask; the transition region mask is then used to calculate the sub-pixel symbolic distance field based on the enhanced feature map regression; the positive values ​​of the distance field are internal values ​​of the disease, the negative values ​​are background values, and the zero value is the precise boundary value; and the gradient direction of the sub-pixel symbolic distance field is calculated and used as the boundary normal.

7. The method for early warning of bridge apparent defects based on image analysis according to claim 1, characterized in that, The method described above calculates pixel-level width by projecting along the normal direction of the crack skeleton based on sub-pixel disease boundary and crack skeleton, converts it into physical width, performs spatial continuity smoothing, and finally outputs disease geometric parameters. First, based on the physical resolution of the original apparent image and the super-resolution magnification factor, the physical size of the single pixel corresponding to the high-resolution image is calculated, and the physical resolution is calibrated. Then, the contour line extraction algorithm is used to extract the zero contour lines of the symbol distance field. Based on the zero contour lines, the sub-pixel boundary point sets on the left and right sides are obtained. The disease probability map is binarized, and single-pixel crack skeletons are extracted. The boundary normal angle of each skeleton point is calculated based on the gradient direction of the signed distance field. A straight line is drawn along the normal of each skeleton point, and its intersection with the sub-pixel boundaries on both sides is used to obtain the width endpoints. The pixel width is then calculated and converted into the physical width. Finally, the width sequence is smoothed by weighted average of adjacent points, as shown in the expression: in, For the first The original physical width of each skeleton point , The physical width of adjacent skeleton points. This is the smoothed width value; The crack length, average width, maximum width, and 95th percentile width are statistically obtained. Based on the crack length, average width, maximum width, and 95th percentile width, a preset threshold is matched, and a graded warning is triggered.

8. The method for early warning of bridge apparent defects based on image analysis according to claim 7, characterized in that, The aforementioned graded early warning rules include Level 1, Level 2, and Level 3 early warnings. The Level 3 early warning is triggered when any of the following conditions are met, corresponding to the early development of minor diseases, and prompts an increase in the frequency of daily inspections: the average width of the crack is ≥0.1mm and the total length of the crack is ≥50mm. The width of the crack at the 95th percentile is ≥0.12mm; The aforementioned Level II early warning is triggered when any of the following conditions are met, indicating that the corresponding disease is showing an expanding trend, and prompting an on-site special review and testing to be arranged: The average width of the crack is ≥0.15mm; Total crack length ≥ 100mm; The width of the crack at the 95th percentile is ≥0.15mm; The Level 1 warning is triggered when any of the following conditions are met: the corresponding disease has reached the early stage limit of the standard or has the risk of large-scale expansion, prompting the initiation of an evaluation of the maintenance and treatment plan: The maximum width of the crack is ≥0.2mm, reaching the standard threshold for early-stage disease; The average crack width is ≥0.18mm and the total crack length is ≥200mm; The width of the crack at the 95th percentile is ≥0.18mm and the total length of the crack is ≥150mm.

9. A bridge apparent defect early warning system based on image analysis, characterized in that, It includes a high- and low-frequency separation super-resolution reconstruction module, which performs high- and low-frequency separation, high-frequency enhancement, hierarchical upsampling and fusion reconstruction on the original bridge appearance image, and outputs a high-resolution image; The multi-directional dual-domain attention enhancement module extracts multi-directional linear disease features, enhances disease features and suppresses background interference through dual prior attention in the channel domain and spatial domain, and outputs an enhanced feature map. The distance field regression module outputs a soft segmentation map of disease probability, regresses the subpixel continuous symbolic distance field and applies normal consistency constraints, and extracts zero contour lines as subpixel disease boundaries. The quantitative early warning module is used to calculate the physical width based on the sub-pixel boundary and skeleton normal, and after spatial smoothing, statistically analyze the disease parameters and trigger graded early warnings.

10. A computer storage medium storing a computer program executable on a processor, characterized in that, When the processor executes the computer program, it implements the bridge appearance defect early warning method based on image analysis as described in any one of claims 1-8.

Images