Concrete structure long crack assessment method based on deep learning

Abstract

The invention discloses a concrete structure long crack evaluation method based on deep learning, and the method comprises the steps: carrying out the correction of a concrete crack image, obtaining a corrected concrete crack image, carrying out the recognition of an attention crack, obtaining a concrete crack feature image, extracting the feature point description of the concrete crack feature image, and carrying out the feature point matching, performing image fusion to obtain a concrete long crack image, performing fuzzy segmentation on the concrete long crack image to obtain a long crack pixel image, performing quantitative evaluation on the long crack pixel image to obtain a crack quantitative pixel result, and calibrating a camera to obtain camera model parameters; and converting the crack quantification pixel result according to the camera model parameters to obtain a crack quantification real result, and taking the crack quantification real result as a concrete structure long crack evaluation result. The method can improve the accuracy and efficiency of concrete long crack evaluation, and is of great significance for promoting intelligent development in the field of constructional engineering.

Description

Technical Field

[0001] The present invention relates to the field of non-destructive measurement, and in particular to a method for evaluating long cracks in concrete structures based on deep learning. Background Art

[0002] Concrete is highly susceptible to cracks, particularly long cracks, in complex natural environments and during long-term use. These cracks can severely shorten the structure's service life. Therefore, accurately assessing long cracks in concrete structures has become crucial for ensuring building safety and making timely repair decisions. In recent years, deep learning technology has rapidly advanced in areas such as image recognition and feature extraction. Deep learning-based identification of long cracks in concrete structures holds great promise, opening up new avenues for efficient and accurate assessment of long cracks in concrete structures.

[0003] Current concrete crack detection technology has a series of limitations. First, traditional image processing methods have difficulty accurately extracting crack features when faced with long cracks due to image acquisition limitations and complex environmental interference. Furthermore, problems such as misalignment and information loss are prone to occur during the image splicing process, making it impossible to achieve accurate assessment of long cracks. Furthermore, although some existing technologies have introduced preliminary deep learning models, the depth and accuracy of crack feature extraction still need to be improved. In particular, for complex situations such as long cracks, which require the integration of multiple local image information for overall assessment, there is a lack of effective overall solutions. Therefore, the present invention proposes a method for assessing long cracks in concrete structures based on deep learning. By constructing an attention crack recognition model and a fuzzy segmentation model, combined with camera calibration technology, it achieves accurate correction of the texture base of long concrete cracks, efficient fusion of feature images, and high-precision quantitative assessment of length, width, and angle. This technology breaks through the limitations of traditional designs and significantly improves the accuracy and efficiency of long concrete crack assessment. It is of great significance for ensuring the safety of concrete structures, reducing maintenance costs, and promoting the intelligent development of the construction engineering field. Summary of the Invention

[0004] The purpose of this invention is to provide a method for evaluating long cracks in concrete structures based on deep learning.

[0005] To achieve the above object, the present invention is implemented according to the following technical solutions:

[0006] The present invention comprises the following steps:

[0007] Extracting a texture base of a concrete crack image, correcting the concrete crack image according to the texture base to obtain a corrected concrete crack image, and processing the corrected concrete crack image using an attention crack recognition model to obtain a concrete crack feature image;

[0008] Extracting feature points of the concrete crack feature image to obtain feature point descriptions, matching feature points of different concrete crack feature images according to the feature point descriptions, and performing image fusion to obtain a concrete long crack image;

[0009] The long crack image of the concrete is processed using a fuzzy segmentation model to obtain a long crack pixel image, and the long crack pixel image is quantitatively evaluated to obtain a crack quantization pixel result; the quantitative evaluation includes length quantization, width quantization and angle quantization;

[0010] The camera is calibrated to obtain camera model parameters, the crack quantization pixel results are converted according to the camera model parameters to obtain crack quantization true results, and the crack quantization true results are used as the concrete structure long crack assessment results.

[0011] Furthermore, the method for obtaining a corrected concrete crack image includes:

[0012] The concrete crack image is subjected to fast Fourier transform to obtain frequency domain representation, and the periodic texture frequency with concentrated energy is detected. The image is filtered using a multi-scale and multi-directional Gabor filter bank to obtain texture features. The maximum value of the multi-scale response is taken to generate the texture base map, which is expressed as:

[0013] T base (x,y)=max θ,σ (I*G θ,σ )

[0014] Where T base is the texture base image, I is the concrete crack image, G θ,σ is the Gabor filter, θ is the filtering direction, and σ is the filtering scale;

[0015] Guided filtering is used to process the texture base image to obtain the illumination component, and the reflection component is calculated based on the concrete crack image and the illumination component.

[0016] A crack probability mask is generated according to the reflection component, and a denoised image is obtained by performing bilateral filtering on the concrete crack image according to the crack probability mask value.

[0017] The modified concrete crack image is obtained by performing three-level feature fusion on the denoised image according to the texture base map, reflection component and crack probability mask. The expression is:

[0018] I corrected =I denoise +α1·(Ref-0.5)·Mask+α2·[CLAHE(T base )-I denoise ]·(1-Mask)

[0019] Among them I correctedTo correct the concrete crack image, I denoise is the denoised image, α1 and α2 are gain coefficients, Ref is the reflection component, Mask is the crack probability mask, CLAHE(T base ) represents the texture base map T base Perform adaptive histogram equalization.

[0020] Furthermore, the method for obtaining a characteristic image of a concrete crack comprises the following steps:

[0021] Construct an attention crack recognition model, specifically including an input layer, a Steam convolution layer, an improved Fused-MBConv module, an improved MBConv module, an MBConv stacking module, a global pooling layer, a fully connected layer, and an output layer; the Steam convolution layer is used to extract primary features to achieve downsampling; the improved Fused-MBConv module is used to extract mid-order features for detail preservation, including an expansion layer, a triple attention enhancement unit, and a residual connection layer; the improved MBConv module is used to extract high-order features, including an expansion layer, a deep convolution layer, a triple attention enhancement unit, a compression layer, and a residual connection layer; the triple attention enhancement unit includes channel attention, spatial attention, and crack-guided adaptive attention;

[0022] The channel attention is used to enhance the weight of crack-related channels, improve crack channel response, reduce concrete texture interference, and output channel weighted features;

[0023] The spatial attention is used to focus on the spatial position of the crack, reduce the weight of the background area, and output spatial weighted features;

[0024] The crack-guided adaptive attention is used to adapt to the crack direction and output morphological adaptive features. The specific steps include:

[0025] The spatial weighted features are used to estimate the crack direction field, and the expression is:

[0026]

[0027] Where θ(i,j) is the azimuth angle at the image position (i,j), G xx (i, j) is the second-order partial derivative of the Gaussian filtered image in the x direction at position (i, j), G yy (i, j) is the second-order partial derivative of the Gaussian filtered image in the y direction at position (i, j), G xy (i, j) is the mixed second-order partial derivative of the Gaussian filtered image in the x and y directions at position (i, j);

[0028] A deformable convolution kernel is generated according to the direction field, and crack edge enhancement convolution is performed based on the deformable convolution kernel to obtain morphological adaptive features. The expression is:

[0029]

[0030]

[0031] Where y(i,j) is the morphological adaptive feature of the output feature map at coordinate (i,j), w k is the sampling point weight factor, The sampled image is at coordinate The pixel value at They are the convolution kernel coordinate increments Δp k The components in x and y, is the gradient operator of the pixel value of the sampled image at coordinate (i, j), γ is the learnable sharpening coefficient, λ1 is the learnable scaling parameter, R(θ)) is the rotation matrix generated according to the direction field θ, is the standard convolution coordinate;

[0032] The corrected concrete crack image is input into the attention crack recognition model to obtain the concrete crack feature image.

[0033] Furthermore, the method for obtaining feature point descriptions includes:

[0034] The Sift algorithm is used to process the concrete crack feature image to obtain feature point description. The specific steps are as follows:

[0035] The original concrete crack feature image is used as the first image group, the first image group is downsampled to obtain the second image group, the second image group is downsampled to obtain the third image group, and the three image groups are Gaussian filtered using different standard deviations;

[0036] Perform differential processing on each set of images one by one to obtain the DOG of each set of images, screen the local mutation part of each layer in each set of DOG, compare the neighborhood of the local mutation part to obtain the undetermined feature points, and use the Hessian matrix to eliminate the edge response points in the undetermined feature points to obtain the image feature points;

[0037] Take the nearest Gaussian layer neighborhood of the image feature point, calculate the gradient modulus and gradient angle of all points in the neighborhood, accumulate the gradient modulus and sum it according to the gradient angle and perform smoothing filtering, determine the direction group corresponding to the maximum value position as the direction of the image feature point, use the image feature point direction, group information, layer information and XY coordinate information as feature point information, and normalize the feature point information to obtain the feature point description.

[0038] Furthermore, the method for obtaining an image of a long concrete crack comprises:

[0039] According to the description of the feature points, the distance ratio between the feature points and the surrounding feature points in the feature image of the same layer of concrete crack is calculated, and the registration information is composed of the distance ratio between the feature points and the surrounding feature points and the direction of the feature points;

[0040] Calculate the cosine similarity of the registration information of the feature points of the same layer of concrete crack feature images in image groups of different sizes, and define three image feature points with a cosine similarity greater than 0.95 as the same feature point;

[0041] Calculate the comprehensive similarity between the feature points of the reference image and the feature points of the image to be registered in different size groups, calculate the average comprehensive similarity of the three groups at different sizes, and take the group of feature points with an average comprehensive similarity greater than 0.95 as a group of registration points;

[0042] The pixel distance between the reference image and the registration point of the image to be registered is calculated as the translation parameter, the directional angle difference between the reference image and the registration point of the image to be registered is calculated as the direction parameter, and the ratio of the pixel distance between the reference image and the registration point to the surrounding registration points is calculated as the scaling parameter. The translation parameter, the direction parameter, and the scaling parameter form the affine parameter. A global affine transformation is performed according to the radial parameters to convert the image to be registered into the reference image coordinate system, and the overlapping area between the two images is determined to generate an overlapping image of concrete crack characteristics.

[0043] The radial basis function is used to detect the cracks in the overlapping area and determine the crack area. The local deformation of the crack area and the non-crack area is then partitioned and elastically optimized. The expression is:

[0044]

[0045] in is the partition elasticity optimization target, b=[b2,b2,…,b k ,…,b m ] is the coefficient vector, m is the number of crack key points, b k is the radial basis function φ k The coefficient of (x,y) at (x,y) coordinates, d i is the actual observation value, λ2 is the regularization parameter, (x k ,y k ) is the coordinate of the kth crack key point, τ k is the adaptive parameter at k crack key points, Wide k is the actual width of the crack at k key points of the crack, is the average value of all crack widths, μ base is the influence range parameter;

[0046] The binary weights of the crack areas of the overlapping images of concrete crack features after partition elastic optimization are calculated, and a multi-scale fusion is performed using a multi-scale pyramid to obtain a concrete crack fusion image. The concrete crack fusion image is then subjected to edge protection processing to obtain a concrete long crack image. The edge protection processing specifically includes forcibly using the main image data when the crack edge pixels do not exceed the corresponding threshold, and performing gradual fusion when the transition zone pixels are within the corresponding threshold range.

[0047] Furthermore, the method for obtaining a long crack pixel image includes:

[0048] The fuzzy segmentation model specifically includes an input layer, an encoder, a decoder, an edge refinement module and an output layer; the fuzzy segmentation model is a weighted hybrid loss function of binary cross entropy, Dice loss and edge enhancement loss; the fuzzy segmentation model uses AdamW optimizer to optimize model parameters;

[0049] The encoder is embedded with a crack-guided adaptive fuzzy kernel algorithm, which gradually extracts deep semantic features of the image through hierarchical downsampling operations, captures the global morphological features of cracks, and outputs crack pixel optimized features. The crack-guided adaptive fuzzy kernel algorithm optimizes the crack features extracted by the encoder based on the crack probability distribution and local texture complexity. The specific steps are as follows:

[0050] Calculate the crack probability prediction and local variance respectively, and determine the adaptive fuzzy kernel according to the crack probability prediction and local variance to perform fuzzy convolution operation. The expression is:

[0051]

[0052] Among them F blur (x,y) is the eigenvalue of the (x,y) position after the fuzzy convolution operation, is the adaptive fuzzy kernel with the internal offset (i, j) at the (x, y) position, β1 and β2 are the fuzzy weights, is the fuzzy constant, P crack (x,y) is the crack probability prediction at the (x,y) position, σ 2 (x,y) is the local variance at the (x,y) position, F(x+i,y+j) is the eigenvalue at the (x+i,y+j) position, is the radius coefficient, k is the neighborhood radius calculated based on the variance, a and b are the radius of the convolution kernel in the row and column dimensions respectively, and W p (m,n) is the weight value of the convolution kernel at the position (m,n), and F(x+m,y+n) is the eigenvalue at the position (x+m,y+n);

[0053] The decoder gradually restores the spatial resolution of the crack pixel optimization features through upsampling and feature fusion operations, accurately locates the crack boundary and outputs the fuzzy long crack pixel features;

[0054] The edge refinement module performs crack edge enhancement on the blurred long crack pixel features through deformable convolution to obtain a long crack pixel image;

[0055] The concrete long crack image is input into the fuzzy segmentation model to obtain the long crack pixel image.

[0056] Furthermore, the method for performing quantitative evaluation to obtain crack quantification pixel results includes:

[0057] Length quantization is performed to obtain the crack pixel length. The specific steps include: binarizing the long crack pixel image, extracting the crack skeleton line using the cv2.distanceTransform function, and calculating the crack pixel length;

[0058] The crack pixel width is obtained by performing width quantization, and the specific steps include: selecting a local crack pixel image, traversing all points to sum the crack pixel points to obtain the local crack pixel area, extracting the skeleton line of the local crack pixel image to calculate the local crack pixel length, and calculating the ratio of the local crack pixel area to the local crack pixel length to obtain the local crack pixel width;

[0059] Angle quantification is performed to obtain the crack angle. The specific steps include: traversing all pixel points of the long crack pixel image, storing all crack coordinates in a crack point set, taking the first coordinate in the crack point set as the starting coordinate of the crack and the last coordinate as the ending coordinate of the crack, and calculating the angle of the crack using the inverse tangent function based on the starting coordinate point and the ending coordinate point to obtain the crack angle;

[0060] The crack pixel length, crack pixel width and crack angle are combined to form the crack quantization pixel results.

[0061] Furthermore, the method for obtaining the true quantitative result of cracks includes:

[0062] The camera's internal and external parameters are calibrated respectively to obtain the camera model parameters. The camera coordinate system is transformed into the pixel coordinate system according to the camera model parameters, and the conversion coefficient between the camera distance and the pixel distance is determined. The crack quantification pixel results are converted into the real crack quantification results according to the conversion coefficient. The real crack quantification results are used as the long crack assessment results of the concrete structure.

[0063] The beneficial effects of the present invention are:

[0064] The present invention is a method for assessing long cracks in concrete structures based on deep learning. Compared with the existing technology, the present invention has the following technical effects:

[0065] The present invention can improve the image preprocessing capability in the assessment of long cracks in concrete structures through the steps of image correction, feature point extraction and matching, image registration and fusion, image segmentation, parameter quantization and parameter conversion, and can increase the speed of the assessment of long cracks in concrete structures, thereby improving the efficiency and accuracy of the assessment of long cracks in concrete structures. The optimization of the assessment technology of long cracks in concrete structures can greatly save resources, improve work efficiency, and realize the intelligent assessment of long cracks in concrete structures, providing strong technical support and guarantee for the assessment of long cracks in concrete structures, and is of great significance for ensuring the safety of concrete structures, reducing maintenance costs, and promoting the intelligent development of the field of construction engineering. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 This is a flowchart of the steps of a method for assessing long cracks in concrete structures based on deep learning in the present invention. DETAILED DESCRIPTION

[0067] The present invention will be further described below through specific examples. The illustrative examples and descriptions of the present invention are used to explain the present invention but are not intended to limit the present invention.

[0068] The present invention provides a method for assessing long cracks in concrete structures based on deep learning, comprising the following steps:

[0069] like Figure 1 As shown, in this embodiment, the following steps are included:

[0070] Extracting a texture base of a concrete crack image, correcting the concrete crack image according to the texture base to obtain a corrected concrete crack image, and processing the corrected concrete crack image using an attention crack recognition model to obtain a concrete crack feature image;

[0071] Extracting feature points of the concrete crack feature image to obtain feature point descriptions, matching feature points of different concrete crack feature images according to the feature point descriptions, and performing image fusion to obtain a concrete long crack image;

[0072] The long crack image of the concrete is processed using a fuzzy segmentation model to obtain a long crack pixel image, and the long crack pixel image is quantitatively evaluated to obtain a crack quantization pixel result; the quantitative evaluation includes length quantization, width quantization and angle quantization;

[0073] The camera is calibrated to obtain camera model parameters, the crack quantization pixel results are converted according to the camera model parameters to obtain crack quantization true results, and the crack quantization true results are used as the concrete structure long crack assessment results.

[0074] In this embodiment, the method for obtaining a corrected concrete crack image includes:

[0075] The concrete crack image is subjected to fast Fourier transform to obtain frequency domain representation, and the periodic texture frequency with concentrated energy is detected. The image is filtered using a multi-scale and multi-directional Gabor filter bank to obtain texture features. The maximum value of the multi-scale response is taken to generate the texture base map, which is expressed as:

[0076] T base (x,y)=max θ,σ (I*G θ,σ )

[0077] Where T base is the texture base image, I is the concrete crack image, G θ,σ is the Gabor filter, θ is the filtering direction, and σ is the filtering scale;

[0078] Guided filtering is used to process the texture base image to obtain the illumination component, and the reflection component is calculated based on the concrete crack image and the illumination component.

[0079] A crack probability mask is generated according to the reflection component, and a denoised image is obtained by performing bilateral filtering on the concrete crack image according to the crack probability mask value.

[0080] The modified concrete crack image is obtained by performing three-level feature fusion on the denoised image according to the texture base map, reflection component and crack probability mask. The expression is:

[0081] I corrected =I denoise +α1·(Ref-0.5)·Mask+α2·[CLAHE(T base )-I denoise ]·(1-Mask)

[0082] Among them I corrected To correct the concrete crack image, I denoise is the denoised image, α1 and α2 are gain coefficients, Ref is the reflection component, Mask is the crack probability mask, CLAHE(T base ) represents the texture base map T base Perform adaptive histogram equalization;

[0083] In the actual evaluation, taking the identification of long cracks in a building beam structure as an example, the concrete crack image to be identified is subjected to fast Fourier transform, filtered using a Gabor filter bank to obtain texture features, and the maximum value of the multi-scale response is taken to generate the texture base map T base , use guided filtering to process the texture base map to obtain the light component Light, and calculate the reflection component based on the concrete crack image and the light component ∈ is a small constant, and the crack probability mask Mask is generated according to the reflection component. The expression is:

[0084]

[0085] Where τ is the crack probability mask threshold rate;

[0086] The concrete crack image area (non-crack area) with a mask of 0 is subjected to unilateral filtering, and the concrete crack image area with a mask of 1 is subjected to bilateral filtering to obtain the denoised image I denoise ;

[0087] Take the gain coefficients α1 = 0.5 and α2 = 0.3, and perform three-level feature fusion to obtain the corrected concrete crack image I corrected .

[0088] In this embodiment, the method for obtaining a concrete crack characteristic image includes the following steps:

[0089] Construct an attention crack recognition model, specifically including an input layer, a Steam convolution layer, an improved Fused-MBConv module, an improved MBConv module, an MBConv stacking module, a global pooling layer, a fully connected layer, and an output layer; the Steam convolution layer is used to extract primary features to achieve downsampling; the improved Fused-MBConv module is used to extract mid-order features for detail preservation, including an expansion layer, a triple attention enhancement unit, and a residual connection layer; the improved MBConv module is used to extract high-order features, including an expansion layer, a deep convolution layer, a triple attention enhancement unit, a compression layer, and a residual connection layer; the triple attention enhancement unit includes channel attention, spatial attention, and crack-guided adaptive attention;

[0090] The channel attention is used to enhance the weight of crack-related channels, improve crack channel response, reduce concrete texture interference, and output channel weighted features;

[0091] The spatial attention is used to focus on the spatial position of the crack, reduce the weight of the background area, and output spatial weighted features;

[0092] The crack-guided adaptive attention is used to adapt to the crack direction and output morphological adaptive features. The specific steps include:

[0093] The spatial weighted features are used to estimate the crack direction field, and the expression is:

[0094]

[0095] Where θ(i,j) is the azimuth angle at the image position (i,j), G xx(i, j) is the second-order partial derivative of the Gaussian filtered image in the x direction at position (i, j), G yy (i, j) is the second-order partial derivative of the image after Gaussian filtering in the y direction at position (i, j), G xy (i, j) is the mixed second-order partial derivative of the Gaussian filtered image in the x and y directions at position (i, j);

[0096] A deformable convolution kernel is generated according to the direction field, and crack edge enhancement convolution is performed based on the deformable convolution kernel to obtain morphological adaptive features. The expression is:

[0097]

[0098] Where y(i,j) is the morphological adaptive feature of the output feature map at coordinate (i,j), w k is the sampling point weight factor, The sampled image is at coordinate The pixel value at They are the convolution kernel coordinate increments Δp k The components in x and y, is the gradient operator of the pixel value of the sampled image at coordinate (i, j), γ is the learnable sharpening coefficient, λ1 is the learnable scaling parameter, R(θ) is the rotation matrix generated according to the direction field θ, is the standard convolution coordinate;

[0099] The corrected concrete crack image is input into the attention crack recognition model to obtain the concrete crack feature image;

[0100] In the actual evaluation, an improved attention crack recognition model based on EfficientNetV2-S is used. The convolution kernel / step size of the Steam convolution layer is 3×3 / 2, the convolution kernel / step size of the extension layer of the improved Fused-MBConv module is 3×3 / 1, and the convolution kernel of the residual connection layer is 3×3. The convolution kernel / step size of the extension layer of the improved MBConv module is 1×1 / 1, the convolution kernel of the depthwise convolution layer is 3×3, the convolution kernel of the compression layer is 1×1, and the convolution kernel of the residual connection layer is 1×1.

[0101] The channel attention compression ratio is 0.2, the Sigmoid output weight range is used, the spatial attention convolution kernel is 7×7, the Sigmoid output weight range is used, the crack-guided adaptive attention basic convolution kernel size is 3×3, the learnable sharpening coefficient γ=0.6, the learnable scaling parameter λ1=0.15, and the direction field calculation window is 5×5;

[0102] The corrected concrete crack image (320×320×3) is input into the attention crack recognition model to obtain the concrete crack feature image (12×12×256).

[0103] In this embodiment, the method for obtaining feature point descriptions includes:

[0104] The Sift algorithm is used to process the concrete crack feature image to obtain feature point description. The specific steps are as follows:

[0105] The original concrete crack feature image is used as the first image group, the first image group is downsampled to obtain the second image group, the second image group is downsampled to obtain the third image group, and the three image groups are Gaussian filtered using different standard deviations;

[0106] Perform differential processing on each set of images one by one to obtain the DOG of each set of images, screen the local mutation part of each layer in each set of DOG, compare the neighborhood of the local mutation part to obtain the undetermined feature points, and use the Hessian matrix to eliminate the edge response points in the undetermined feature points to obtain the image feature points;

[0107] Take the nearest Gaussian layer neighborhood of the image feature point, calculate the gradient modulus and gradient angle of all points in the neighborhood, accumulate the gradient modulus according to the gradient angle and perform smoothing filtering, determine the direction group corresponding to the maximum value position as the image feature point direction, use the image feature point direction, group information, layer information and XY coordinate information as feature point information, and normalize the feature point information to obtain the feature point description;

[0108] In the actual evaluation, the original concrete crack feature images (12×12×256) were used as the first image group. The second image group (6×6×128) was obtained by downsampling to 1 / 2 the size of the first image group. The third image group (3×3×64) was obtained by downsampling to 1 / 2 the size of the second image group. The three groups of images were Gaussian filtered with standard deviations of 1.6, 3.2, and 6.4, respectively.

[0109] The number of DOG layers obtained by differential processing is one layer less than the number of images of the corresponding size. The neighborhood of the local mutation part specifically includes the 3×3 neighborhood feature points of the same layer of the local mutation point, the 3×3 neighborhood of the upper layer, and the 3×3 neighborhood of the lower layer (a total of 26 pixels). When the local mutation point is an extreme point, it is determined to be a pending feature point.

[0110] The Hessian matrix is ​​used to eliminate the edge response points in the undetermined feature points to obtain the image feature points. The gradient modulus and gradient angle accumulation values ​​of all points in the neighborhood of the image feature points (3×3 neighborhood feature points in the same layer, 3×3 neighborhood in the upper layer, and 3×3 neighborhood in the lower layer) are calculated. The accumulated values ​​are smoothed by the Gaussian kernel function, and the direction group corresponding to the maximum value position is taken as the direction of the image feature point. The feature point information is normalized to obtain the feature point description.

[0111] In this embodiment, the method for obtaining an image of a long concrete crack includes:

[0112] According to the description of the feature points, the distance ratio between the feature points and the surrounding feature points in the feature image of the same layer of concrete crack is calculated, and the registration information is composed of the distance ratio between the feature points and the surrounding feature points and the direction of the feature points;

[0113] Calculate the cosine similarity of the registration information of the feature points of the same layer of concrete crack feature images in image groups of different sizes, and define three image feature points with a cosine similarity greater than 0.95 as the same feature point;

[0114] Calculate the comprehensive similarity between the feature points of the reference image and the feature points of the image to be registered in different size groups, calculate the average comprehensive similarity of the three groups at different sizes, and take the group of feature points with an average comprehensive similarity greater than 0.95 as a group of registration points;

[0115] The pixel distance between the reference image and the registration point of the image to be registered is calculated as the translation parameter, the directional angle difference between the reference image and the registration point of the image to be registered is calculated as the direction parameter, and the ratio of the pixel distance between the reference image and the registration point to the surrounding registration points is calculated as the scaling parameter. The translation parameter, the direction parameter, and the scaling parameter form the affine parameter. A global affine transformation is performed according to the radial parameters to convert the image to be registered into the reference image coordinate system, and the overlapping area between the two images is determined to generate an overlapping image of concrete crack characteristics.

[0116] The radial basis function is used to detect the cracks in the overlapping area and determine the crack area. The local deformation of the crack area and the non-crack area is then partitioned and elastically optimized. The expression is:

[0117]

[0118] in is the partition elasticity optimization target, b=[b2,b2,…,b k ,…,b m ] is the coefficient vector, m is the number of crack key points, b k is the radial basis function φ k The coefficient of (x,y) at (x,y) coordinates, d i is the actual observation value, λ2 is the regularization parameter, (x k ,y k ) is the coordinate of the kth crack key point, τ k is the adaptive parameter at k crack key points, Wide k is the actual width of the crack at k key points of the crack, is the average value of all crack widths, μ base is the influence range parameter;

[0119] The binary weights of the crack areas of the overlapping images of concrete crack features after partition elastic optimization are calculated, and a multi-scale fusion is performed using a multi-scale pyramid to obtain a concrete crack fusion image. The concrete crack fusion image is then subjected to edge protection processing to obtain a concrete long crack image. The edge protection processing specifically includes forcibly using the main image data when the crack edge pixels do not exceed the corresponding threshold, and performing gradual fusion when the transition zone pixels are within the corresponding threshold range.

[0120] In the actual evaluation, taking the feature point A1 (pixel coordinates (120, 85), direction 60°) of the first image group as an example, the distances of the five nearest feature points in the surrounding neighborhood are calculated to be 5px, 7px, 8.5px, 9.2px, and 10px, respectively. The corresponding distance ratio vectors are [1, 1.4, 1.7, 1.84, 2]. The distance ratio vectors and feature points constitute the registration information.

[0121] The cosine similarities of the corresponding registration information of the first-layer feature point A1 of the first image group (12×12×256) and the first-layer feature point A2 of the second image group (6×6×128), the first-layer feature point A2 of the second image group and the first-layer feature point A3 of the third image group (3×3×64), and the first-layer feature point A1 of the first image group and the first-layer feature point A3 of the third image group are calculated to be 0.99, 0.98, 0.985, and 0.988, respectively. The first-layer feature point A1 of the first image group, the first-layer feature point A2 of the second image group, and the first-layer feature point A3 of the third image group are defined as the same feature point A.

[0122] Taking the first layer image as the reference image and the second layer image as the image to be registered, taking the registration of the reference image feature point A and the image feature point B to be registered as an example, the comprehensive similarity (0.7*distance ratio similarity+0.3*direction similarity) of the reference image feature point A1 and the image feature point B1 to be registered in the first image group is calculated to be 0.968, the comprehensive similarity of the reference image feature point A2 and the image feature point B2 to be registered in the second image group is 0.972, and the comprehensive similarity of the reference image feature point A3 and the image feature point B3 to be registered in the third image group is 0.975. The average comprehensive similarity is 0.972>0.95, and the reference image feature point A and the image feature point B to be registered are determined to be a set of registration points. The feature points of all layers are traversed to complete the image registration;

[0123] The pixel distance, direction angle difference, and pixel distance ratio between the reference image registration point A and the registration point B in the image to be registered are calculated to obtain affine parameters (7.67, 5), 2.5°, and 1.008. A global affine transformation is performed based on the affine parameters, and the overlapping area is determined to generate an overlapping image of concrete crack characteristics.

[0124] The radial basis function is used to detect the cracks in the overlapping area to determine the crack area, and the regularization parameter λ2 = 0.2 in the crack area, λ2 = 0.8 in the non-crack area, and the average value of all crack widths is taken. Influence range parameter μ base =5px to perform partition elastic optimization and obtain the overlapping image of concrete crack characteristics after partition elastic optimization;

[0125] The binary weight of the crack area in the image is calculated, and multi-scale fusion is performed using a multi-scale pyramid. Edge protection processing is performed to obtain an image of long cracks in concrete. When the crack edge pixels do not exceed 3px, the main image data is forced to be used, and when the transition zone pixels are between 15px and 30px, gradient fusion is performed.

[0126] In this embodiment, the method for obtaining a long crack pixel image includes:

[0127] The fuzzy segmentation model specifically includes an input layer, an encoder, a decoder, an edge refinement module and an output layer; the fuzzy segmentation model is a weighted hybrid loss function of binary cross entropy, Dice loss and edge enhancement loss; the fuzzy segmentation model uses AdamW optimizer to optimize model parameters;

[0128] The encoder is embedded with a crack-guided adaptive fuzzy kernel algorithm, which gradually extracts deep semantic features of the image through hierarchical downsampling operations, captures the global morphological features of cracks, and outputs crack pixel optimized features. The crack-guided adaptive fuzzy kernel algorithm optimizes the crack features extracted by the encoder based on the crack probability distribution and local texture complexity. The specific steps are as follows:

[0129] Calculate the crack probability prediction and local variance respectively, and determine the adaptive fuzzy kernel according to the crack probability prediction and local variance to perform fuzzy convolution operation. The expression is:

[0130]

[0131] Among them F blur (x,y) is the eigenvalue of the (x,y) position after the fuzzy convolution operation, is the adaptive fuzzy kernel with the internal offset (i, j) at the position (x, y), β1 and β2 are the fuzzy weights, is the fuzzy constant, P crack (x,y) is the crack probability prediction at the (x,y) position, σ 2 (x,y) is the local variance at the (x,y) position, F(x+i,y+j) is the eigenvalue at the (x+i,y+j) position, is the radius coefficient, k is the neighborhood radius calculated based on the variance, a and b are the radius of the convolution kernel in the row and column dimensions respectively, and Wp (m,n) is the weight value of the convolution kernel at the position (m,n), and F(x+m,y+n) is the eigenvalue at the position (x+m,y+n);

[0132] The decoder gradually restores the spatial resolution of the crack pixel optimization features through upsampling and feature fusion operations, accurately locates the crack boundary and outputs the fuzzy long crack pixel features;

[0133] The edge refinement module performs crack edge enhancement on the blurred long crack pixel features through deformable convolution to obtain a long crack pixel image;

[0134] Input the concrete long crack image into the fuzzy segmentation model to obtain the long crack pixel image;

[0135] In the actual evaluation, the loss weights of binary cross entropy, Dice loss and edge reinforcement loss in the weighted hybrid loss function are 0.4, 0.4 and 0.2 respectively, and the optimizer learning rate is e -4 , weight decay takes e -5 ;

[0136] The encoder sequentially includes a first convolutional layer (including two 3×3 convolution kernels and a ReLU function), a first fuzzy module, a first pooling layer (2×2), a second convolutional layer (including two 3×3 convolution kernels and a ReLU function), a second fuzzy module, a second pooling layer (2×2), a third convolutional layer (including two 3×3 convolution kernels and a ReLU function), and a third pooling layer (2×2). The first fuzzy module uses a fuzzy kernel algorithm to perform fuzzy preprocessing on the input features. The second fuzzy module uses a crack-guided adaptive fuzzy kernel algorithm to perform intermediate fuzzy processing on the input features.

[0137] Take the fuzzy weights β1 = 0.3 and β2 = 0.5, and the fuzzy constants Based on the neighborhood radius k=3 calculated by variance, and the radii a=3 and b=3 of the convolution kernel in the row and column dimensions, a crack-guided adaptive fuzzy convolution operation is performed to optimize the crack features extracted by the encoder;

[0138] The decoder includes, in sequence, a first transposed convolutional layer (2×2 transposed convolution kernel), a first connection layer, a first depthwise convolutional layer (including two 3×3 convolution kernels and a ReLU function), a second transposed convolutional layer (2×2 transposed convolution kernel), a second connection layer, and a second depthwise convolutional layer (including two 3×3 convolution kernels and a ReLU function); the first connection layer is connected to the second convolutional layer; the second connection layer is connected to the first convolutional layer;

[0139] The edge refinement module includes an offset prediction layer (3×3 convolution kernel), a deformable convolution layer (3×3 deformable convolution kernel), and an edge enhancement layer (3×3 convolution integration channel);

[0140] The concrete long crack image is input into the fuzzy segmentation model to obtain the long crack pixel image.

[0141] In this embodiment, the method for performing quantitative evaluation to obtain crack quantification pixel results includes:

[0142] Length quantization is performed to obtain the crack pixel length. The specific steps include: binarizing the long crack pixel image, extracting the crack skeleton line using the cv2.distanceTransform function, and calculating the crack pixel length;

[0143] The crack pixel width is obtained by performing width quantization, and the specific steps include: selecting a local crack pixel image, traversing all points to sum the crack pixel points to obtain the local crack pixel area, extracting the skeleton line of the local crack pixel image to calculate the local crack pixel length, and calculating the ratio of the local crack pixel area to the local crack pixel length to obtain the local crack pixel width;

[0144] Angle quantification is performed to obtain the crack angle. The specific steps include: traversing all pixel points of the long crack pixel image, storing all crack coordinates in a crack point set, taking the first coordinate in the crack point set as the starting coordinate of the crack and the last coordinate as the ending coordinate of the crack, and calculating the angle of the crack using the inverse tangent function based on the starting coordinate point and the ending coordinate point to obtain the crack angle;

[0145] The crack pixel length, crack pixel width and crack angle are combined into crack quantization pixel results;

[0146] In the actual evaluation, the length quantization steps include: transforming the long crack pixel image into a binary image, with the crack as black with a pixel value of 0 and the background as a pixel value of 255. Using the cv2.distanceTransform function to calculate the distance from each pixel in the image to the nearest 0 pixel point, the point with the maximum pixel value in each row is taken as the point of the skeleton line (the pixel value of the point in the non-crack area is 0, and the pixel value of the center point of each row of cracks is the largest. For example, if the pixel value of a row of crack points is 1234321, the pixel point with a pixel value of 4 is the point of the skeleton line. Connect all the points of the skeleton line to form a skeleton line). The pixel length of the skeleton line extracted from the long crack pixel image is 1486px.

[0147] The crack pixel image skeleton line is divided into 5 segments to obtain a local crack pixel length of 297px. All points of the 5 segments of the local crack pixel image are traversed and the crack pixel points are summed to obtain a local crack pixel area of ​​892px. 2 、1206px 2 、1480px 2 、1332px 2 、921px2 , the ratio of local crack pixel area to local crack pixel length was calculated to obtain the local crack pixel widths of 3px, 4.06px, 4.98px, 4.48px, and 3.10px;

[0148] Traverse all the pixels of the long crack pixel image to obtain the crack point set, take the starting coordinates (35120) and the ending point (290, 275), and use the inverse tangent function to calculate the crack angle to be 31.2°.

[0149] In actual evaluation, the method for obtaining the true quantitative results of cracks includes:

[0150] The camera's internal and external parameters are calibrated to obtain the camera model parameters. The camera coordinate system is transformed into the pixel coordinate system based on the camera model parameters, and the conversion coefficient between the camera distance and the pixel distance is determined. The crack quantification pixel results are converted into the real crack quantification results based on the conversion coefficient. The real crack quantification results are used as the long crack assessment results of the concrete structure.

[0151] In actual evaluation, the camera is calibrated to obtain the camera intrinsic parameter matrix, which is expressed as:

[0152]

[0153] where f x 、f y are the focal lengths of the x and y axes, respectively, c x 、c y are the x and y coordinates of the principal point respectively;

[0154] Calibrate the camera to obtain the camera extrinsic parameter matrix, which is expressed as:

[0155]

[0156] Where R is the rotation matrix, t is the distance from the camera to the object (mm);

[0157] The conversion from the world coordinate system to the camera coordinate system only involves rotation and translation, and does not affect the size of the distance value. Therefore, the conversion from the world coordinate system to the camera coordinate system is considered in the calculation of the present invention. The conversion from the camera coordinate system to the pixel coordinate system is determined according to the camera model parameters to determine the conversion coefficient as 0.522 mm / px. The crack quantization pixel result is converted into the crack quantization true result according to the conversion coefficient: long crack length = crack pixel length * conversion coefficient = 1486 * 0.522 = 775.7 mm, local crack width = local crack pixel width * conversion coefficient = [3, 4.06, 4.98, 4.48, 3.10] * 0.522 = [1.57, 2.12, 2.60, 2.34, 1.62] (mm), the crack angle does not need to be converted to 31.2°, and the crack quantization true result is used as the long crack assessment result of the concrete structure.

[0158] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for assessing long cracks in concrete structures based on deep learning, characterized in that: The following steps are involved: S1. Extracting a texture base of a concrete crack image, correcting the concrete crack image according to the texture base to obtain a corrected concrete crack image, and processing the corrected concrete crack image using an attention crack recognition model to obtain a concrete crack feature image; S2. Extracting feature points of the concrete crack feature image to obtain feature point descriptions, matching feature points of different concrete crack feature images according to the feature point descriptions, and performing image fusion to obtain a concrete long crack image; S3. Processing the concrete long crack image using a fuzzy segmentation model to obtain a long crack pixel image, and performing quantitative evaluation on the long crack pixel image to obtain a crack quantization pixel result; the quantitative evaluation includes length quantization, width quantization, and angle quantization; S4. Calibrate the camera to obtain camera model parameters, convert the crack quantization pixel results according to the camera model parameters to obtain crack quantization true results, and use the crack quantization true results as the concrete structure long crack assessment results.

2. The method for evaluating long cracks in concrete structures based on deep learning according to claim 1, characterized in that: The method for obtaining a corrected concrete crack image comprises: The concrete crack image is subjected to fast Fourier transform to obtain frequency domain representation, and the periodic texture frequency with concentrated energy is detected. The image is filtered using a multi-scale and multi-directional Gabor filter bank to obtain texture features. The maximum value of the multi-scale response is taken to generate the texture base map, which is expressed as: T base (x,y)=max θ,σ (I-G θ,σ ) Where T base is the texture base image, I is the concrete crack image, G θ,σ is the Gabor filter, θ is the filtering direction, and σ is the filtering scale; Guided filtering is used to process the texture base image to obtain the illumination component, and the reflection component is calculated based on the concrete crack image and the illumination component. A crack probability mask is generated according to the reflection component, and a denoised image is obtained by performing bilateral filtering on the concrete crack image according to the crack probability mask value. The modified concrete crack image is obtained by performing three-level feature fusion on the denoised image according to the texture base map, reflection component and crack probability mask. The expression is: I corrected =I denoise +α1·(Ref-0.5)·Mask+α2·[CLAHE(T base )-I denoise ]·(1-Mask) Among them I corrected To correct the concrete crack image, I denoise is the denoised image, α1 and α2 are gain coefficients, Ref is the reflection component, Mask is the crack probability mask, CLAHE(T base ) represents the texture base map T base Perform adaptive histogram equalization.

3. A method for assessing long cracks in concrete structures based on deep learning according to claim 2, characterized in that: The method for obtaining a concrete crack characteristic image comprises the following steps: Construct an attention crack recognition model, specifically including an input layer, a Steam convolution layer, an improved Fused-MBConv module, an improved MBConv module, an MBConv stacking module, a global pooling layer, a fully connected layer, and an output layer; the Steam convolution layer is used to extract primary features to achieve downsampling; the improved Fused-MBConv module is used to extract mid-order features for detail preservation, including an expansion layer, a triple attention enhancement unit, and a residual connection layer; the improved MBConv module is used to extract high-order features, including an expansion layer, a deep convolution layer, a triple attention enhancement unit, a compression layer, and a residual connection layer; the triple attention enhancement unit includes channel attention, spatial attention, and crack-guided adaptive attention; The channel attention is used to enhance the weight of crack-related channels, improve crack channel response, reduce concrete texture interference, and output channel weighted features; The spatial attention is used to focus on the spatial position of the crack, reduce the weight of the background area, and output spatial weighted features; The crack-guided adaptive attention is used to adapt to the crack direction and output morphological adaptive features. The specific steps include: The spatial weighted features are used to estimate the crack direction field, and the expression is: Where θ(i,j) is the azimuth angle at the image position (i,j), G xx (i, j) is the second-order partial derivative of the Gaussian filtered image in the x direction at position (i, j), G yy (i, j) is the second-order partial derivative of the image after Gaussian filtering in the y direction at position (i, j), G xy (i, j) is the mixed second-order partial derivative of the Gaussian filtered image in the x and y directions at position (i, j); A deformable convolution kernel is generated according to the direction field, and crack edge enhancement convolution is performed based on the deformable convolution kernel to obtain morphological adaptive features. The expression is: Where y(i,j) is the morphological adaptive feature of the output feature map at coordinate (i,j), w k is the sampling point weight factor, The sampled image is at coordinate The pixel value at They are the convolution kernel coordinate increments Δp k The components in x and y, is the gradient operator of the pixel value of the sampled image at coordinate (i, j), γ is the learnable sharpening coefficient, λ1 is the learnable scaling parameter, R(θ) is the rotation matrix generated according to the direction field θ, is the standard convolution coordinate; The corrected concrete crack image is input into the attention crack recognition model to obtain the concrete crack feature image.

4. The method for assessing long cracks in concrete structures based on deep learning according to claim 1, characterized in that: The method for obtaining feature point descriptions comprises the following steps: The Sift algorithm is used to process the concrete crack feature image to obtain feature point description. The specific steps are as follows: The original concrete crack feature image is used as the first image group, the first image group is downsampled to obtain the second image group, the second image group is downsampled to obtain the third image group, and the three image groups are Gaussian filtered using different standard deviations; Perform differential processing on each set of images one by one to obtain the DOG of each set of images, screen the local mutation part of each layer in each set of DOG, compare the neighborhood of the local mutation part to obtain the undetermined feature points, and use the Hessian matrix to eliminate the edge response points in the undetermined feature points to obtain the image feature points; Take the nearest Gaussian layer neighborhood of the image feature point, calculate the gradient modulus and gradient angle of all points in the neighborhood, accumulate the gradient modulus and sum it according to the gradient angle and perform smoothing filtering, determine the direction group corresponding to the maximum value position as the direction of the image feature point, use the image feature point direction, group information, layer information and XY coordinate information as feature point information, and normalize the feature point information to obtain the feature point description.

5. The method for evaluating long cracks in concrete structures based on deep learning according to claim 1, characterized in that: The method for obtaining an image of a long concrete crack comprises: According to the description of the feature points, the distance ratio between the feature points and the surrounding feature points in the feature image of the same layer of concrete crack is calculated, and the registration information is composed of the distance ratio between the feature points and the surrounding feature points and the direction of the feature points; Calculate the cosine similarity of the registration information of the feature points of the same layer of concrete crack feature images in image groups of different sizes, and define three image feature points with a cosine similarity greater than 0.95 as the same feature point; Calculate the comprehensive similarity between the feature points of the reference image and the feature points of the image to be registered in different size groups, calculate the average comprehensive similarity of the three groups at different sizes, and take the group of feature points with an average comprehensive similarity greater than 0.95 as a group of registration points; The pixel distance between the reference image and the registration point of the image to be registered is calculated as the translation parameter, the directional angle difference between the reference image and the registration point of the image to be registered is calculated as the direction parameter, and the ratio of the pixel distance between the reference image and the registration point to the surrounding registration points is calculated as the scaling parameter. The translation parameter, the direction parameter, and the scaling parameter form the affine parameter. A global affine transformation is performed according to the radial parameters to convert the image to be registered into the reference image coordinate system, and the overlapping area between the two images is determined to generate an overlapping image of concrete crack characteristics. The radial basis function is used to detect the cracks in the overlapping area and determine the crack area. The local deformation of the crack area and the non-crack area is then partitioned and elastically optimized. The expression is: in is the partition elasticity optimization target, b=[b2,b2,…,b k ,…,b m ] is the coefficient vector, m is the number of crack key points, b k is the radial basis function φ k The coefficient of (x,y) at (x,y) coordinates, d i is the actual observation value, λ2 is the regularization parameter, (x k ,y k ) is the coordinate of the kth crack key point, τ k is the adaptive parameter at k crack key points, Wide k is the actual width of the crack at k key points of the crack, is the average value of all crack widths, μ base is the influence range parameter; The binary weights of the crack areas of the overlapping images of concrete crack features after partition elastic optimization are calculated, and a multi-scale fusion is performed using a multi-scale pyramid to obtain a concrete crack fusion image. The concrete crack fusion image is then subjected to edge protection processing to obtain a concrete long crack image. The edge protection processing specifically includes forcibly using the main image data when the crack edge pixels do not exceed the corresponding threshold, and performing gradual fusion when the transition zone pixels are within the corresponding threshold range.

6. The method for assessing long cracks in concrete structures based on deep learning according to claim 1, characterized in that: The method for obtaining a long crack pixel image comprises: The fuzzy segmentation model specifically includes an input layer, an encoder, a decoder, an edge refinement module and an output layer; the fuzzy segmentation model is a weighted hybrid loss function of binary cross entropy, Dice loss and edge enhancement loss; the fuzzy segmentation model uses AdamW optimizer to optimize model parameters; The encoder is embedded with a crack-guided adaptive fuzzy kernel algorithm, which gradually extracts deep semantic features of the image through hierarchical downsampling operations, captures the global morphological features of cracks, and outputs crack pixel optimized features. The crack-guided adaptive fuzzy kernel algorithm optimizes the crack features extracted by the encoder based on the crack probability distribution and local texture complexity. The specific steps are as follows: Calculate the crack probability prediction and local variance respectively, and determine the adaptive fuzzy kernel according to the crack probability prediction and local variance to perform fuzzy convolution operation. The expression is: Among them F blur (x,y) is the eigenvalue of the (x,y) position after the fuzzy convolution operation, is the adaptive fuzzy kernel with the internal offset (i, j) at the (x, y) position, β1 and β2 are the fuzzy weights, is the fuzzy constant, P crack (x,y) is the crack probability prediction at the (x,y) position, σ 2 (x,y) is the local variance at the (x,y) position, F(x+i,y+j) is the eigenvalue at the (x+i,y+j) position, is the radius coefficient, k is the neighborhood radius calculated based on the variance, a and b are the radius of the convolution kernel in the row and column dimensions respectively, and W p (m,n) is the weight value of the convolution kernel at the position (m,n), and F(x+m,y+n) is the eigenvalue at the position (x+m,y+n); The decoder gradually restores the spatial resolution of the crack pixel optimization features through upsampling and feature fusion operations, accurately locates the crack boundary and outputs the fuzzy long crack pixel features; The edge refinement module performs crack edge enhancement on the blurred long crack pixel features through deformable convolution to obtain a long crack pixel image; The concrete long crack image is input into the fuzzy segmentation model to obtain the long crack pixel image.

7. The method for assessing long cracks in concrete structures based on deep learning according to claim 1, characterized in that: The method for performing quantitative evaluation to obtain crack quantitative pixel results includes: Length quantization is performed to obtain the crack pixel length. The specific steps include: binarizing the long crack pixel image, extracting the crack skeleton line using the cv2.distanceTransform function, and calculating the crack pixel length; The crack pixel width is obtained by performing width quantization, and the specific steps include: selecting a local crack pixel image, traversing all points to sum the crack pixel points to obtain the local crack pixel area, extracting the skeleton line of the local crack pixel image to calculate the local crack pixel length, and calculating the ratio of the local crack pixel area to the local crack pixel length to obtain the local crack pixel width; Angle quantification is performed to obtain the crack angle. The specific steps include: traversing all pixel points of the long crack pixel image, storing all crack coordinates in a crack point set, taking the first coordinate in the crack point set as the starting coordinate of the crack and the last coordinate as the ending coordinate of the crack, and calculating the angle of the crack using the inverse tangent function based on the starting coordinate point and the ending coordinate point to obtain the crack angle; The crack pixel length, crack pixel width and crack angle are combined to form the crack quantization pixel results.

8. The method for assessing long cracks in concrete structures based on deep learning according to claim 1, characterized in that: The method for obtaining the true result of crack quantification includes: The camera's internal and external parameters are calibrated respectively to obtain the camera model parameters. The camera coordinate system is transformed into the pixel coordinate system according to the camera model parameters, and the conversion coefficient between the camera distance and the pixel distance is determined. The crack quantification pixel results are converted into the real crack quantification results according to the conversion coefficient. The real crack quantification results are used as the long crack assessment results of the concrete structure.

Images