A bridge crack detection method based on image recognition

Abstract

The present invention discloses a bridge crack detection method based on image recognition, which belongs to the field of image recognition technology. The present invention first splits the bridge image into three monochrome channels of red, green and blue, and performs clustering and scanning processing on each spectral image to accurately locate the suspected crack points. By calculating the channel contrast value and the regional trajectory distortion value, the red, green and blue channel weight matrices are constructed, and then the bridge image is gray-scale processed and the light and dark are adjusted to enhance the crack area. Subsequently, the adjusted image is texture enhanced using the red, green and blue channel weight matrices to generate three texture-enhanced images. Finally, a three-channel crack recognition neural network is used to process the enhanced image to obtain bridge cracks, thereby improving the accuracy of crack detection.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a bridge crack detection method based on image recognition. Background Art

[0002] At present, the academic and engineering communities have tried to use a variety of image processing techniques to identify bridge cracks. The main methods include threshold segmentation, edge detection, texture analysis, and machine learning-based recognition methods. These methods attempt to automatically identify crack features by performing grayscale processing, filtering, feature extraction, and other operations on bridge images. However, existing image recognition technology still faces serious limitations in bridge crack detection: existing methods lack deep mining and effective use of multi-channel spectral information, resulting in insufficient crack feature extraction and inability to accurately capture the subtle texture and structural features of cracks, especially for bridge structures with different surface states, which makes it difficult to accurately identify cracks. Summary of the invention

[0003] In view of the above-mentioned deficiencies in the prior art, the present invention provides a bridge crack detection method based on image recognition to solve the problem of low accuracy of bridge crack detection in the prior art.

[0004] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is: a bridge crack detection method based on image recognition, comprising the following steps:

[0005] S1. Split the bridge image into three monochrome channels of red, green and blue to obtain three spectral images, wherein the three spectral images are specifically: red, green and blue spectral images;

[0006] S2, clustering the spectral image to find the background points, horizontally and vertically scanning the spectral image to find the candidate points, and eliminating the background points from each candidate point to obtain the suspected points;

[0007] S3, according to each suspected point on each spectral image, calculate the channel contrast value and regional trajectory distortion value of each suspected point, and construct a red channel weight matrix, a green channel weight matrix and a blue channel weight matrix;

[0008] S4, grayscale processing is performed on the bridge image to obtain a grayscale image, and brightness adjustment is performed on the grayscale image to obtain an adjusted image;

[0009] S5, using a red channel weight matrix, a green channel weight matrix, and a blue channel weight matrix to perform texture enhancement processing on the adjusted image respectively, to obtain three texture enhanced images;

[0010] S6. Use a three-channel crack recognition neural network to process three texture-enhanced images and obtain bridge cracks.

[0011] Further, S2 includes the following sub-steps:

[0012] S21, performing clustering processing on pixel points on the spectral image according to channel values ​​on the spectral image to obtain multiple clusters;

[0013] S22, finding the cluster containing the most pixels, and marking the pixels in the cluster as background points;

[0014] S23, scanning the spectral image horizontally to find out the horizontal texture points, and scanning the spectral image vertically to find out the vertical texture points;

[0015] S24, when the same pixel point on the spectral image is both a horizontal texture point and a vertical texture point, the pixel point is a candidate point;

[0016] S25. Discard candidate points that are background points from among the candidate points to obtain suspected points.

[0017] Further, S22 includes the following sub-steps:

[0018] S231, calculating a lateral deviation value of two adjacent pixel points in the lateral direction of the spectral image, and when the lateral deviation value is greater than a deviation threshold, both of the two adjacent pixel points are marked as lateral texture points;

[0019] S232, calculating a longitudinal deviation value of two adjacent pixel points in the longitudinal direction of the spectral image, and when the longitudinal deviation value is greater than a deviation threshold, both of the two adjacent pixel points are marked as longitudinal texture points.

[0020] Furthermore, the formula for calculating the lateral deviation value in S231 is: , the formula for calculating the longitudinal deviation value in S232 is: , where d i,x is the lateral deviation value of the ith row of the spectral image, d j,y is the longitudinal deviation value of the jth column of the spectral image, C i,(x+1,y) is the channel value of the pixel at (x+1,y) on the i-th row of the spectral image, C i,(x,y) is the channel value of the pixel at (x, y) on the i-th row of the spectral image, C j,(x,y+1) is the channel value of the pixel at (x, y+1) on the jth column of the spectral image, C j,(x,y) is the channel value of the pixel at (x, y) on the jth column of the spectral image, i is the row number, j is the column number, x is the horizontal coordinate of the pixel, and y is the vertical coordinate of the pixel.

[0021] Furthermore, S3 includes the following sub-steps:

[0022] S31, classifying the suspected points that are in contact with each other into one area on the spectral image;

[0023] S32, calculating the channel contrast value of each suspected point in a region;

[0024] S33, in a region, calculating a regional trajectory distortion value;

[0025] S34. Calculate the weight value of the suspected point according to the regional trajectory distortion value and the channel contrast value: , where w n is the weight value of the nth suspected point in the region, γ n is the channel contrast value of the nth suspicious point, and ε is the regional trajectory distortion value of the area;

[0026] S35. On the spectral image, the pixel value of the suspected point is replaced by the weight value, and the weight values ​​of other pixel points are 0, to obtain a channel weight matrix, wherein, when the spectral image in S31 is a red spectral image, S35 obtains a red channel weight matrix; when the spectral image in S31 is a green spectral image, S35 obtains a green channel weight matrix; when the spectral image in S31 is a blue spectral image, S35 obtains a blue channel weight matrix.

[0027] Furthermore, the formula for calculating the channel comparison value in S32 is: , where γ n is the channel contrast value of the nth suspected point, C p,n is the channel value of the nth suspected point, C Avg is the channel mean of all background points on the spectral image, and n is the number of the suspected point.

[0028] Furthermore, the formula for calculating the regional trajectory distortion value in S33 is: , where ε is the regional trajectory distortion value, α1 is the tortuosity of the region, and α2 is the curvature change value of the region.

[0029] Further, S4 includes the following sub-steps:

[0030] S41, grayscale processing is performed on the bridge image to obtain a grayscale image;

[0031] S42, calculating the brightness adjustment value for the grayscale value of each pixel in the grayscale image: ,in, is the brightness adjustment value of the mth pixel, H Avg is the grayscale mean of the background points on the grayscale image, h m is the gray value of the mth pixel;

[0032] S43, replacing the grayscale value of the corresponding pixel point on the grayscale image with the brightness adjustment value to obtain an adjusted image.

[0033] Furthermore, the formula for texture enhancement in S5 is: , , , where V1 is the first texture enhanced image, V2 is the second texture enhanced image, V3 is the third texture enhanced image, R matrix is the red channel weight matrix, G matrix is the green channel weight matrix, B matrix is the blue channel weight matrix, To multiply the corresponding elements, Z is the adjusted image.

[0034] Furthermore, the three-channel crack recognition neural network in S6 includes: 3 point-wise convolution channels, texture feature weighting layer, channel attention module CAM, residual space attention module RSAM, first CNN network, second CNN network and fully connected layer;

[0035] The input end of each point-by-point convolution channel is used to input a spectral image; the input end of the texture feature weighted layer is respectively connected to the output ends of the three point-by-point convolution channels, and its output end is respectively connected to the input end of the channel attention module CAM and the input end of the residual space attention module RSAM; the input end of the first CNN network is connected to the output end of the channel attention module CAM; the input end of the second CNN network is connected to the output end of the residual space attention module RSAM; the input end of the fully connected layer is respectively connected to the output end of the first CNN network and the output end of the second CNN network, and its output end serves as the output end of the three-channel crack recognition neural network.

[0036] The beneficial effects of the present invention are:

[0037] 1. The present invention splits the bridge image into three monochrome channels of red, green and blue to obtain three spectral images, processes the spectral images corresponding to each channel respectively, extracts the suspected points on each spectral image, and then calculates the channel contrast value and regional trajectory distortion value of the suspected points according to each suspected point on each spectral image, reflecting the subtle texture and structural characteristics of the cracks, and then constructs the red, green and blue channel weight matrices to characterize the weights of the cracks on the three channels. The present invention quantifies the weights of the pixels on the cracks through the red, green and blue channel weight matrices, which can more accurately distinguish real cracks from background noise, greatly reduce the false detection rate, and improve the reliability of crack detection.

[0038] 2. The present invention performs grayscale processing on the bridge image to obtain a grayscale image, adjusts the brightness of the grayscale image to make the pixel value of the crack area more significant, lowers the pixel value of the non-crack area, and then uses red, green and blue channel weight matrices to perform texture enhancement processing on the adjusted image respectively, so that the characteristics of the crack area are further enhanced, the recognizability of the cracks is improved, and the accuracy of bridge crack detection is improved.

[0039] 3. The three-channel crack recognition neural network is used to process three texture-enhanced images, which can comprehensively process the enhanced images of the three channels, greatly improving the intelligence level and accuracy of crack recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 The figure is a flow chart of a bridge crack detection method based on image recognition.

[0041] Figure 2 Schematic diagram of the structure of the three-channel crack recognition neural network. DETAILED DESCRIPTION

[0042] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0043] like Figure 1 As shown, a bridge crack detection method based on image recognition includes the following steps:

[0044] S1. Split the bridge image into three monochrome channels of red, green and blue to obtain three spectral images, wherein the three spectral images are specifically: red, green and blue spectral images;

[0045] S2, clustering the spectral image to find the background points, horizontally and vertically scanning the spectral image to find the candidate points, and eliminating the background points from each candidate point to obtain the suspected points;

[0046] S3, according to each suspected point on each spectral image, calculate the channel contrast value and regional trajectory distortion value of each suspected point, and construct a red channel weight matrix, a green channel weight matrix and a blue channel weight matrix;

[0047] S4, grayscale processing is performed on the bridge image to obtain a grayscale image, and brightness adjustment is performed on the grayscale image to obtain an adjusted image;

[0048] S5, using a red channel weight matrix, a green channel weight matrix, and a blue channel weight matrix to perform texture enhancement processing on the adjusted image respectively, to obtain three texture enhanced images;

[0049] S6. Use a three-channel crack recognition neural network to process three texture-enhanced images and obtain bridge cracks.

[0050] In the present invention, the bridge image is an RGB image.

[0051] In this embodiment, S2 includes the following sub-steps:

[0052] S21, performing clustering processing on pixel points on the spectral image according to channel values ​​on the spectral image to obtain multiple clusters;

[0053] S22, finding the cluster containing the most pixels, and marking the pixels in the cluster as background points;

[0054] S23, scanning the spectral image horizontally to find out the horizontal texture points, and scanning the spectral image vertically to find out the vertical texture points;

[0055] S24, when the same pixel point on the spectral image is both a horizontal texture point and a vertical texture point, the pixel point is a candidate point;

[0056] S25. Discard candidate points that are background points from among the candidate points to obtain various suspected points.

[0057] The present invention clusters the pixels on the spectral image according to the channel values ​​on the spectral image, that is, the pixels with similar channel values ​​on the spectral image are classified into a cluster, and the cluster containing the most pixels is classified as the background area, and then the spectral image is scanned horizontally and vertically to find the horizontal texture points and the vertical texture points. The present invention adopts a method combining horizontal and vertical scanning, which can fully capture the directional characteristics of the crack. By considering the horizontal and vertical texture points at the same time, the boundary area between the crack and the background can be more accurately identified. The present invention then discards the candidate points marked as background points from each candidate point, and retains the candidate points belonging to the crack, that is, the suspected points.

[0058] In this embodiment, S22 includes the following sub-steps:

[0059] S231, calculating a lateral deviation value of two adjacent pixel points in the lateral direction of the spectral image, and when the lateral deviation value is greater than a deviation threshold, both of the two adjacent pixel points are marked as lateral texture points;

[0060] S232, calculating a longitudinal deviation value of two adjacent pixel points in the longitudinal direction of the spectral image, and when the longitudinal deviation value is greater than a deviation threshold, both of the two adjacent pixel points are marked as longitudinal texture points.

[0061] In this embodiment, the formula for calculating the lateral deviation value in S231 is: , the formula for calculating the longitudinal deviation value in S232 is: , where d i,x is the lateral deviation value of the ith row of the spectral image, d j,y is the longitudinal deviation value of the jth column of the spectral image, C i,(x+1,y) is the channel value of the pixel at (x+1,y) on the i-th row of the spectral image, Ci,(x,y) is the channel value of the pixel at (x, y) on the i-th row of the spectral image, C j,(x,y+1) is the channel value of the pixel at (x, y+1) on the jth column of the spectral image, C j,(x,y) is the channel value of the pixel at (x, y) on the jth column of the spectral image, i is the row number, j is the column number, x is the horizontal coordinate of the pixel, and y is the vertical coordinate of the pixel.

[0062] The present invention determines whether the channel values ​​at the two positions are located in the transition area between the crack and the background by calculating the lateral deviation value and the longitudinal deviation value. In the transition area, the channel values ​​at the two positions vary greatly. When the lateral deviation value is greater than the deviation threshold, and when the longitudinal deviation value is greater than the deviation threshold, the lateral texture point and the longitudinal texture point are obtained. When a pixel point is both a lateral texture point and a longitudinal texture point, the pixel point is a candidate point to ensure the accuracy of obtaining the suspected point. At the same time, since S231 and S232 in the present invention mark a pair of pixel points as texture points, the texture point includes the pixel points in the background area, and the pixel points in the background area need to be eliminated to ensure that the selected suspected point is a texture point on the crack.

[0063] The deviation threshold is a threshold set for the lateral deviation value and the longitudinal deviation value.

[0064] In this embodiment, S3 includes the following sub-steps:

[0065] S31, classifying the suspected points that are in contact with each other into one area on the spectral image;

[0066] S32, calculating the channel contrast value of each suspected point in a region;

[0067] S33, in a region, calculating a regional trajectory distortion value;

[0068] S34. Calculate the weight value of the suspected point according to the regional trajectory distortion value and the channel contrast value: , where w n is the weight value of the nth suspected point in the region, γ n is the channel contrast value of the nth suspicious point, and ε is the regional trajectory distortion value of the area;

[0069] S35. On the spectral image, the pixel value of the suspected point is replaced by the weight value, and the weight values ​​of other pixel points are 0, to obtain a channel weight matrix, wherein, when the spectral image in S31 is a red spectral image, S35 obtains a red channel weight matrix; when the spectral image in S31 is a green spectral image, S35 obtains a green channel weight matrix; when the spectral image in S31 is a blue spectral image, S35 obtains a blue channel weight matrix.

[0070] The channel weight matrix has the same size as the spectral image.

[0071] Step S31 is to classify the suspected point pixels in the same area into one area, and then calculate the channel contrast value and regional trajectory distortion value of the suspected point to characterize the possibility that the area is a crack. The larger the weight value, the greater the possibility that the suspected point is a texture point on the crack.

[0072] In this embodiment, the specific implementation process of S31 is:

[0073] S311, taking a suspected point that is not partitioned on the spectral image as the center;

[0074] S312, judge the center Whether there are other unpartitioned suspected points within the neighborhood, if so, the suspected point at the center and the unpartitioned suspected points are classified into one area, and the process jumps to step S313; if not, the suspected point at the center is discarded, and the process jumps to step S311;

[0075] S313, selecting a suspected point that has not been a center in the area as a new center;

[0076] S314, determine the new center Are there other unpartitioned suspected points within the neighborhood? If so, classify the other unpartitioned suspected points into the area and jump to step S313 until all the suspected points in the area are through the center and a region classification is completed, then jump to step S311. If not, jump directly to step S311 until all the suspected points on the spectral image are traversed.

[0077] The present invention searches within the neighborhood of a suspected point to find out a suspected point in contact with the suspected point, and then finds the next suspected point in contact from the new suspected point in contact, thereby classifying the texture lines formed by the suspected points into one area.

[0078] In this embodiment, the formula for calculating the channel comparison value in S32 is: , where γ n is the channel contrast value of the nth suspected point, C p,n is the channel value of the nth suspected point, C Avg is the channel mean of all background points on the spectral image, and n is the number of the suspected point.

[0079] The channel value of the pixel point at the crack in the crack is close to 0. Therefore, the present invention calculates the ratio of the channel value of the suspected point to the channel mean of the background point to reflect the difference between the suspected point and the background. The smaller the ratio, the greater the channel contrast.

[0080] In this embodiment, the formula for calculating the regional trajectory distortion value in S33 is: , where ε is the regional trajectory distortion value, α1 is the tortuosity of the region, and α2 is the curvature change value of the region.

[0081] In this embodiment, , , where x n is the horizontal coordinate of the nth suspected point in the region, x n+1 is the horizontal coordinate of the n+1th suspected point in the region, y n is the ordinate of the nth suspected point in the region, y n+1 is the ordinate of the n+1th suspected point in the region, x s is the horizontal coordinate of the suspected point at the starting point of the area, y s is the ordinate of the suspected point at the starting point of the area, x e is the horizontal coordinate of the suspected point at the end point of the region, y e is the ordinate of the suspected point at the end point of the region, μ n is the curvature of the nth suspected point in the region, L is the number of suspected points in the region, and n is the number of the suspected point.

[0082] The present invention classifies the suspected points that are in contact with each other into an area, which is equivalent to classifying the crack contours one by one. Calculating the regional trajectory distortion value is to evaluate the distortion of the crack contours formed by each suspected point in the area. The tortuosity α1 of the area is the ratio of the length of the crack formed by each suspected point in the area to the length of the straight line between the two farthest suspected points in the area. The curvature change value α2 of the area reflects the fluctuation of the curvature. The present invention reflects the irregularity of the area through the tortuosity α1 of the area and the curvature change value α2 of the area, which is convenient for eliminating the structure of the bridge itself.

[0083] In this embodiment, S4 includes the following sub-steps:

[0084] S41, grayscale processing is performed on the bridge image to obtain a grayscale image;

[0085] S42, calculating the brightness adjustment value for the grayscale value of each pixel in the grayscale image: ,in, is the brightness adjustment value of the mth pixel, H Avg is the grayscale mean of the background points on the grayscale image, h m is the gray value of the mth pixel;

[0086] S43, replacing the grayscale value of the corresponding pixel point on the grayscale image with the brightness adjustment value to obtain an adjusted image.

[0087] The present invention performs grayscale processing on the bridge image to obtain a grayscale image, and then calculates the brightness adjustment value for the grayscale value of each pixel in the grayscale image, so that the pixel value of the crack area is enhanced. m Greater than H AvgWhen h is , the gray value of the pixel is large, so it does not belong to the crack area. m Less than H Avg When m The smaller it is, the more likely it is that the pixel is on a crack.

[0088] In this embodiment, the formula for texture enhancement in S5 is: , , , where V1 is the first texture enhanced image, V2 is the second texture enhanced image, V3 is the third texture enhanced image, R matrix is the red channel weight matrix, G matrix is the green channel weight matrix, B matrix is the blue channel weight matrix, To multiply the corresponding elements, Z is the adjusted image.

[0089] The present invention uses red, green and blue channel weight matrices to enhance the adjusted image, further enhancing the characteristic value at the crack, making the crack features significant and improving the crack recognition accuracy.

[0090] In this embodiment, if Figure 2 As shown, the three-channel crack recognition neural network in S6 includes: 3 point-by-point convolution channels, texture feature weighting layer, channel attention module CAM, residual space attention module RSAM, the first CNN network, the second CNN network and the fully connected layer;

[0091] The input end of each point-by-point convolution channel is used to input a spectral image; the input end of the texture feature weighted layer is respectively connected to the output ends of the three point-by-point convolution channels, and its output end is respectively connected to the input end of the channel attention module CAM and the input end of the residual space attention module RSAM; the input end of the first CNN network is connected to the output end of the channel attention module CAM; the input end of the second CNN network is connected to the output end of the residual space attention module RSAM; the input end of the fully connected layer is respectively connected to the output end of the first CNN network and the output end of the second CNN network, and its output end serves as the output end of the three-channel crack recognition neural network.

[0092] In this embodiment, the expression of the texture feature weighted layer is: , where Y is the output of the texture feature weighted layer, X k is the output of the kth point-wise convolution channel, W k is the weighted weight, and k is a positive integer.

[0093] In the present invention, the point-by-point convolution channel adopts The convolution layer of different sizes does not change the size of the input image, and then the texture feature weighting layer is used to weight the outputs of the three point-by-point convolution channels so that the feature values ​​of the same pixel are weighted. Then, the channel-wise attention module (CAM) and the residual spatial attention module (RSAM) are used to pay attention to different features. Then, two CNN networks are used to extract two aspects of features. The fully connected layer combines the two aspects of features to improve the accuracy of crack detection.

[0094] In this embodiment, the channel attention module CAM can be replaced by an average pooling layer, and the residual space attention module RSAM can be replaced by a maximum pooling layer, but the channel attention module CAM and the residual space attention module RSAM can adaptively focus on key features and have greater flexibility.

[0095] The present invention splits the bridge image into three monochrome channels of red, green and blue to obtain three spectral images, processes the spectral images corresponding to each channel respectively, extracts the suspected points on each spectral image, and then calculates the channel contrast value and regional trajectory distortion value of the suspected points according to each suspected point on each spectral image, reflecting the subtle texture and structural characteristics of the cracks, and then constructs the red, green and blue channel weight matrices to characterize the weights of the cracks on the three channels. The present invention quantifies the weights of the pixels on the cracks through the red, green and blue channel weight matrices, which can more accurately distinguish real cracks from background noise, greatly reduce the false detection rate, and improve the reliability of crack detection.

[0096] The present invention performs grayscale processing on the bridge image to obtain a grayscale image, performs brightness adjustment on the grayscale image to make the pixel value of the crack area more significant, lowers the pixel value of the non-crack area, and then uses red, green and blue channel weight matrices to perform texture enhancement processing on the adjusted image respectively, so that the characteristics of the crack area are further enhanced, the recognizability of the cracks is improved, and the accuracy of bridge crack detection is improved.

[0097] The present invention adopts a three-channel crack recognition neural network to process three texture enhanced images, and can comprehensively process the enhanced images of the three channels, thereby greatly improving the intelligence level and accuracy of crack recognition.

[0098] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A bridge crack detection method based on image recognition, characterized in that: The following steps are involved: S1. Split the bridge image into three monochrome channels of red, green and blue to obtain three spectral images, wherein the three spectral images are specifically: red, green and blue spectral images; S2, clustering the spectral image to find the background points, horizontally and vertically scanning the spectral image to find the candidate points, and eliminating the background points from each candidate point to obtain the suspected points; S3. According to each suspected point on each spectral image, the channel contrast value and the regional trajectory distortion value of each suspected point are calculated, and the red channel weight matrix, the green channel weight matrix and the blue channel weight matrix are constructed, where the regional trajectory distortion value is: , where α1 is the tortuosity of the region, α2 is the curvature change value of the region, α1 is the ratio of the length of the crack formed by each suspected point in the region to the length of the straight line between the two farthest suspected points in the region, and α2 is the standard deviation of the curvature of each suspected point in the region; S4, grayscale processing is performed on the bridge image to obtain a grayscale image, and brightness adjustment is performed on the grayscale image to obtain an adjusted image; S5, using a red channel weight matrix, a green channel weight matrix, and a blue channel weight matrix to perform texture enhancement processing on the adjusted image respectively, to obtain three texture enhanced images; S6. Use a three-channel crack recognition neural network to process three texture-enhanced images and obtain bridge cracks.

2. The bridge crack detection method based on image recognition according to claim 1 is characterized in that: The S2 comprises the following sub-steps: S21, performing clustering processing on pixel points on the spectral image according to channel values ​​on the spectral image to obtain multiple clusters; S22, finding the cluster containing the most pixels, and marking the pixels in the cluster as background points; S23, scanning the spectral image horizontally to find out the horizontal texture points, and scanning the spectral image vertically to find out the vertical texture points; S24, when the same pixel point on the spectral image is both a horizontal texture point and a vertical texture point, the pixel point is a candidate point; S25. Discard candidate points that are background points from among the candidate points to obtain suspected points.

3. The bridge crack detection method based on image recognition according to claim 2 is characterized in that: The S23 comprises the following sub-steps: S231, calculating a lateral deviation value of two adjacent pixel points in the lateral direction of the spectral image, and when the lateral deviation value is greater than a deviation threshold, both of the two adjacent pixel points are marked as lateral texture points; S232, calculating a longitudinal deviation value of two adjacent pixel points in the longitudinal direction of the spectral image, and when the longitudinal deviation value is greater than a deviation threshold, both of the two adjacent pixel points are marked as longitudinal texture points.

4. The bridge crack detection method based on image recognition according to claim 3 is characterized in that: The formula for calculating the lateral deviation value in S231 is: , the formula for calculating the longitudinal deviation value in S232 is: , where d i,x is the lateral deviation value of the ith row of the spectral image, d j,y is the longitudinal deviation value of the jth column of the spectral image, C i,(x+1,y) is the channel value of the pixel at (x+1,y) on the i-th row of the spectral image, C i,(x,y) is the channel value of the pixel at (x, y) on the i-th row of the spectral image, C j,(x,y+1) is the channel value of the pixel at (x, y+1) on the jth column of the spectral image, C j,(x,y) is the channel value of the pixel at (x, y) on the jth column of the spectral image, i is the row number, j is the column number, x is the horizontal coordinate of the pixel, and y is the vertical coordinate of the pixel.

5. The bridge crack detection method based on image recognition according to claim 1 is characterized in that: The S3 comprises the following sub-steps: S31, classifying the suspected points that are in contact with each other into one area on the spectral image; S32, calculating the channel contrast value of each suspected point in a region; S33, in a region, calculating a regional trajectory distortion value; S34. Calculate the weight value of the suspected point according to the regional trajectory distortion value and the channel contrast value: , where w n is the weight value of the nth suspected point in the region, γ n is the channel contrast value of the nth suspected point, ε is the regional trajectory distortion value of the area, and n is the number of the suspected point; S35. On the spectral image, the pixel value of the suspected point is replaced by the weight value, and the weight values ​​of other pixel points are 0, to obtain a channel weight matrix, wherein, when the spectral image in S31 is a red spectral image, S35 obtains a red channel weight matrix; when the spectral image in S31 is a green spectral image, S35 obtains a green channel weight matrix; when the spectral image in S31 is a blue spectral image, S35 obtains a blue channel weight matrix.

6. The bridge crack detection method based on image recognition according to claim 5 is characterized in that: The formula for calculating the channel comparison value in S32 is: , where C p,n is the channel value of the nth suspected point, C Avg is the channel mean of all background points on the spectral image.

7. The bridge crack detection method based on image recognition according to claim 1 is characterized in that: The S4 comprises the following sub-steps: S41, grayscale processing is performed on the bridge image to obtain a grayscale image; S42, calculating the brightness adjustment value for the grayscale value of each pixel in the grayscale image: ,in, is the brightness adjustment value of the mth pixel, H Avg is the grayscale mean of the background points on the grayscale image, h m is the grayscale value of the mth pixel, and m is a positive integer; S43, replacing the grayscale value of the corresponding pixel on the grayscale image with the brightness adjustment value to obtain an adjusted image.

8. The bridge crack detection method based on image recognition according to claim 1, characterized in that: The formula for texture enhancement in S5 is: , , , where V1 is the first texture enhanced image, V2 is the second texture enhanced image, V3 is the third texture enhanced image, R matrix is the red channel weight matrix, G matrix is the green channel weight matrix, B matrix is the blue channel weight matrix, To multiply the corresponding elements, Z is the adjusted image.

9. The bridge crack detection method based on image recognition according to claim 1, characterized in that: The three-channel crack recognition neural network in S6 includes: 3 point-by-point convolution channels, a texture feature weighted layer, a channel attention module CAM, a residual space attention module RSAM, a first CNN network, a second CNN network and a fully connected layer; The input end of each point-by-point convolution channel is used to input a spectral image; the input end of the texture feature weighted layer is respectively connected to the output ends of the three point-by-point convolution channels, and its output end is respectively connected to the input end of the channel attention module CAM and the input end of the residual space attention module RSAM; the input end of the first CNN network is connected to the output end of the channel attention module CAM; the input end of the second CNN network is connected to the output end of the residual space attention module RSAM; the input end of the fully connected layer is respectively connected to the output end of the first CNN network and the output end of the second CNN network, and its output end serves as the output end of the three-channel crack recognition neural network.

Images