CN116485719A - Self-adaptive canny method for crack detection - Google Patents

Abstract

The invention discloses a self-adaptive canny method for crack detection, which solves the problems of more noise points, more dirty blocks and unstable detection effect in the detection of a tunnel wall crack. The algorithm firstly realizes effective improvement of the brightness of the dark part of the image by a method of carrying out illumination compensation on the gray level image; in order to remove the influence of partial isolated bright spot noise in the image, median filtering processing is carried out on the image, so that the impulse noise is reduced while the edge characteristics of cracks are maintained in the image; the method searches the threshold value by a method of maximizing the variance among gradient amplitude classes, and carries out edge detection by a Canny operator based on the threshold value; and finally, automatically detecting and removing noise and dirty blocks in the image by combining closed operation and connectivity to obtain a cleaner crack detection result. The invention can be widely applied to the problem of detecting tunnel cracks and other diseases.

Description

An Adaptive Canny Method for Crack Detection

technical field

The invention relates to the field of crack image recognition, and more particularly, relates to an adaptive canny method for crack detection.

Background technique

Tunnel cracks are one of the important factors affecting the safety of tunnel interiors. The formation of cracks will be affected by the deformation and shedding of adjacent rocks. Construction methods and technologies will also cause cracks in the tunnel, such as the excavation method of the tunnel, the proportion of various concrete raw materials, and the vibration method. After the completion of the repair, the uneven shrinkage of the concrete in the tunnel, bias pressure, frost heaving and other factors are also important reasons for the cracks on the tunnel surface. If the width of the crack exceeds a certain limit, it will have a great impact on the tunnel and seriously affect the safety of subway operation. Therefore, it is very important to regularly detect whether there are cracks in the tunnel and determine the width of the cracks. In addition, the appearance and development of tunnel cracks is a gradual process. There are few and scattered cracks in the tunnel at the beginning, but as time goes by, smaller cracks will gradually develop. Therefore, while detecting cracks, we should also pay close attention to issues such as the width and deformation of tunnel cracks.

The identification and detection of tunnel cracks and other diseases has been studied for quite a while at home and abroad, and contact detection methods are generally used, but this method is inconvenient and has large measurement errors. Foreign research started earlier, mostly based on lidar, ultrasonic, shock elastic wave, image processing and other methods.

With the rapid development of image processing technology, its fast, accurate and non-destructive characteristics are favored by the majority of researchers, and it is widely used in the field of industrial production and research in the detection of cracks in subway tunnels and other diseases. Due to continuous breakthroughs in the field of computer vision, cutting-edge traditional image processing plus deep learning can be used in the image field to quickly acquire images of tunnel cracks and detect the location of cracks on the surface of the tunnel, providing an accurate and fast method for identifying cracks on the surface of subway tunnels. The image processing method overcomes the subjectivity and low efficiency of traditional manual inspections and greatly improves the flexibility of detection. However, in the two-dimensional image collection of tunnel cracks, the two-dimensional information of cracks may be indistinguishable from brick cracks similar to cracks in image recognition, or subtle changes in cracks in key parts cannot be detected. monitoring issues.

In recent years, many methods and technologies have been used in the research of image processing methods to detect tunnel cracks at home and abroad. For example, Ahmed Mahgoub Ahmed Talab et al. proposed a concrete structure crack detection algorithm based on image edges. This method firstly changes the image into a grayscale image, and then utilizes the Sobel method to detect crack edges. Use the appropriate threshold to classify the pixels into two categories, foreground and background, and extract the original image of the suspected fracture area. Then use the Sobel operator again to filter out the residual noise, and use the 0tsu algorithm for optimal segmentation. The main idea is to iteratively use a filter to sift through fractured regions. This method can handle high-quality images better and can filter part of the noise. This method still has the problem of unstable detection effect.

Contents of the invention

The invention provides an adaptive canny method for crack detection, which solves the problems of many noise points, many dirty blocks and unstable detection effect encountered in the crack detection of tunnel walls.

In order to solve the problems of the technologies described above, the technical solution of the present invention is as follows:

An adaptive canny method for crack detection, comprising the following steps:

S1: acquiring an image including cracks;

S2: performing light compensation according to the size of the image to obtain an image with uniform brightness;

S3: performing median filtering on the image with uniform brightness to obtain a noise-reduced image;

S4: Calculate the gradient magnitude of the noise-reduced image, and automatically generate the optimal segmentation threshold th by maximizing the variance between classes of the gradient magnitude;

S5: Obtain the lowest threshold and the highest threshold of the Canny edge detection algorithm according to the optimal segmentation threshold th, and use the lowest threshold and the highest threshold to perform edge detection to obtain the first image;

S6: Using eight connected domains to search and calculate connected components on the first image, remove connected components whose area is smaller than a preset minimum value, and obtain a second image;

S7: Using a closed operation method on the second image to splice the disconnected cracks and fill the holes of the dirty block to obtain a third image;

S8: Using eight connected domains to search and calculate the connected components on the third image, if the area of the length multiplied by the width of the connected components is greater than the first preset value of the area of the original image and the actual area is greater than the second preset value of the area of the original image, remove the connected components to obtain the final crack image.

Preferably, step S2 is specifically:

S2.1: Convert the image into a grayscale image, use f(x, y)=(x, y) to represent the grayscale value after grayscale, and R, G, and B respectively represent the red, green, and blue components in the original true color image, and there are:

f(x,y)=0.299R+0.587G+0.114B

S2.2: Complement the illumination of the grayscale image, and perform brightness compensation for different brightness areas in the image:

Solve the average gray level of the grayscale image I, and record the width w and height h of the grayscale image I;

Take m=min(w,h)/20, and divide the grayscale image I into squares, find the average value of each block, and obtain the brightness matrix D of the sub-block;

Subtract the average grayscale of the grayscale image I from each element of the matrix D to obtain the brightness difference matrix E of the sub-block;

Using the bicube difference method, the matrix E is differenced into a brightness distribution matrix R of the same size as the grayscale image I;

An image with uniform brightness result=I-R is obtained.

Preferably, step S3 is specifically:

S3.1: Determine the size of the convolution kernel;

S3.2: Correspond the center of the convolution kernel to the first pixel on the upper left of the image, sort the gray value of each pixel in the area covered by the convolution kernel, select the median of the gray sequence, and assign the median value to the pixel corresponding to the center position of the convolution kernel;

S3.3: Repeat step S3.2 for each pixel on the image to obtain a noise-reduced image.

Preferably, step S4 automatically generates the optimal segmentation threshold th by maximizing the variance between classes of gradient magnitudes, specifically:

S4.1: Set the optimal segmentation threshold th to divide the image gradient amplitude into two categories C1 and C2, the image gradient amplitude in C1 is smaller than th, and the image gradient amplitude in C2 is greater than th, and the respective mean values of these two types of gradient amplitude values are m1 and m2, and the global mean value of the gradient amplitude value is mG. At the same time, the probability that the gradient amplitude value is divided into C1 and C2 classes is p1 and p2 respectively, as follows:

p1*m1+p2*m2=mG

p1+p2=1

S4.2: According to the concept of variance, the expression of variance between classes is:

σ ² =p1(m1-mG) ² +p2(m2-mG) ²

S4.3: Normalize the gradient magnitude matrix to the range of [0, 255], and then iterate 256 times to find the value that makes the inter-class variance σ ² reach the maximum value, which is the required threshold th.

Preferably, step S5 is specifically:

S5.1: Calculate the image gradient;

S5.2: Perform non-maximum suppression on the gradient amplitude;

S5.3: Perform double-threshold screening, set the pixels whose gray scale change is greater than the highest threshold as strong edge pixels, and remove the pixels below the lowest threshold; set the pixels between the lowest threshold and the highest threshold as weak edge pixels, if there are strong edge pixels in the area of the weak edge pixels, keep the weak edge pixels, and remove the weak edge pixels if there are no strong edge pixels in the area of the weak edge pixels.

Preferably, the lowest threshold is 0.66th, and the highest threshold is 0.9h.

Preferably, step S6 is specifically:

S6.1: Scan each pixel of the first image, divide the pixels with the same pixel value and connected with each other in 8 neighborhoods into the same group, and finally obtain the connected components of all pixels in the first image;

S6.2: Arrange all the connected components in ascending order of area, remove the connected components whose area is smaller than the preset minimum value, and obtain the second image.

Preferably, the preset minimum value in step S6.2 is the 90% quantile of the areas of all connected components.

Preferably, step S7 is specifically:

S7.1: Use convolution kernel B to perform expansion operation on the second image:

Define the convolution kernel B, which can be of any shape and size, and has a separately defined reference point-anchor point;

Convolve the convolution kernel B with the second image, and calculate the maximum value of pixels in the area covered by the convolution kernel B;

assigning said maximum value to the pixel specified by the reference point;

S7.2: Corrosion operation using convolution kernel C:

Define the convolution kernel C, which can be of any shape and size, and has a separately defined reference point-anchor point;

Convolve the convolution kernel C with the expanded second image, and calculate the minimum pixel value of the area covered by the convolution kernel C;

Assign the minimum value to the pixel specified by the reference point to obtain the third image.

Preferably, step S8 is specifically:

S8.1: Scan each pixel of the third image, divide the pixels with the same pixel value and connected with each other in the 8 neighborhoods into the same group, finally obtain all the pixel connected components in the image, and calculate the area, length and width of each connected component;

S8.2: Calculate the length-by-width area of each connected component;

S8.3: If the area of length multiplied by width of the connected component is greater than the first preset value of the area of the original image and the actual area is greater than the second preset value of the area of the original image, remove the connected component to obtain the final crack image.

Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

The present invention uses the gradient amplitude of the target image to realize an adaptive Canny edge detection operator, and realizes adaptively removing block noises according to connectivity properties, and can be widely applied to tunnel crack detection problems in various dark environments. Compared with commonly used image crack detection methods, the detection effect is more stable, and more parts can be detected.

Description of drawings

Fig. 1 is a schematic flow chart of the method of the present invention.

Fig. 2 is an image of cracks in the tunnel arch collected by a real digital camera provided in the embodiment.

Figure 3 is a schematic diagram of labeling Figure 2 using labelme.

Figure 4 is a schematic diagram of the detection results of Figure 2 using the OTSU algorithm.

FIG. 5 is a schematic diagram of a result after light compensation is performed on FIG. 2 .

FIG. 6 is a schematic diagram of the result of performing median filtering on FIG. 5 .

FIG. 7 is a schematic diagram of an edge detection result using an adaptive Canny edge detection operator in FIG. 6 .

FIG. 8 is a schematic diagram of the result of using noise removal on FIG. 7 .

Fig. 9 is a schematic diagram of the result after splicing the cracks in Fig. 8 and filling the dirty blocks.

FIG. 10 is a schematic diagram of the result after removing the surface dirt in FIG. 9 .

FIG. 11 is a schematic diagram of the crack marking in FIG. 2 .

Fig. 12 is a schematic diagram of crack marks detected based on OTSU algorithm.

Fig. 13 is a schematic diagram of crack marks detected by the method of the present invention.

Detailed ways

The accompanying drawings are for illustrative purposes only and cannot be construed as limiting the patent;

In order to better illustrate this embodiment, some parts in the drawings will be omitted, enlarged or reduced, and do not represent the size of the actual product;

For those skilled in the art, it is understandable that some well-known structures and descriptions thereof may be omitted in the drawings.

The technical solutions of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

Example 1

An adaptive canny method for crack detection, comprising the following steps:

S1: acquiring an image including cracks;

S2: performing light compensation according to the size of the image to obtain an image with uniform brightness;

S3: performing median filtering on the image with uniform brightness to obtain a noise-reduced image;

S4: Calculate the gradient magnitude of the noise-reduced image, and automatically generate the optimal segmentation threshold th by maximizing the variance between classes of the gradient magnitude;

S5: Obtain the lowest threshold and the highest threshold of the Canny edge detection algorithm according to the optimal segmentation threshold th, and use the lowest threshold and the highest threshold to perform edge detection to obtain the first image;

S6: Using eight connected domains to search and calculate connected components on the first image, remove connected components whose area is smaller than a preset minimum value, and obtain a second image;

S7: Using a closed operation method on the second image to splice the disconnected cracks and fill the holes of the dirty block to obtain a third image;

S8: Using eight connected domains to search and calculate the connected components on the third image, if the area of the length multiplied by the width of the connected components is greater than the first preset value of the area of the original image and the actual area is greater than the second preset value of the area of the original image, remove the connected components to obtain the final crack image.

Example 2

On the basis of Embodiment 1, this embodiment continues to disclose the following content:

Step S2 is to obtain a picture with a more concentrated brightness range, specifically:

S2.1: Convert the image into a grayscale image, and convert the original image into a grayscale image. On the one hand, it facilitates subsequent faster image processing, and on the other hand, it also unifies the processing of multiple color cracks. Moreover, some subsequent processing methods also require the image to be a grayscale image, so the crack image is first processed in grayscale. Use f(x,y)=(x,y) to represent the grayscale value after grayscale, R, G, and B respectively represent the red, green, and blue components in the original true color image, there are:

f(x,y)=0.299R+0.587G+0.114B

Among them, 0.299, 0.587 and 0.114 are the most reasonable grayscale image weights proved by theoretical and experimental derivation.

S2.2: Complement the illumination of the grayscale image, and perform brightness compensation for different brightness regions in the image, so that the brightness range of the entire image is more concentrated, which is conducive to information extraction. The main steps are:

Solve the average gray level of the grayscale image I, and record the width w and height h of the grayscale image I;

Take m=min(w,h)/20, and divide the grayscale image I into squares, find the average value of each block, and obtain the brightness matrix D of the sub-block;

Subtract the average grayscale of the grayscale image I from each element of the matrix D to obtain the brightness difference matrix E of the sub-block;

Using the bicube difference method, the matrix E is differenced into a brightness distribution matrix R of the same size as the grayscale image I;

An image with uniform brightness result=I-R is obtained.

Step S3 Median filtering removes noise, removes some isolated bright spot noise in the image, and retains the edge details of the cracks to facilitate subsequent processing. The main principle is to perform convolution calculation on the image, and each convolution calculation is to sort the pixels in a certain pixel neighborhood (the neighborhood range depends on the size of the convolution kernel) according to the gray value, select the median in the sequence and assign it to the pixel corresponding to the neighborhood, so that the pixel with a large difference in the gray value of the surrounding pixels can be changed to a value close to the surrounding pixel value, so that isolated noise points can be eliminated. Specifically:

S3.1: Determine the size of the convolution kernel. In this embodiment, a convolution kernel with a size of 5×5 is used;

S3.2: Correspond the center of the convolution kernel to the first pixel on the upper left of the image, sort the gray value of each pixel in the area covered by the convolution kernel, select the median of the gray sequence, and assign the median value to the pixel corresponding to the center position of the convolution kernel;

S3.3: Repeat step S3.2 for each pixel on the image to obtain a noise-reduced image.

Step S4 automatically generates the optimal segmentation threshold th by maximizing the inter-class variance of the gradient amplitude. The determination of the image threshold usually requires manual parameter adjustment, which is not conducive to the large-scale application of the method. A threshold is solved by using the method of maximizing the inter-class variance of the gradient amplitude to realize the adaptive process, specifically:

S4.1: Set the optimal segmentation threshold th to divide the image gradient amplitude into two categories C1 and C2, the image gradient amplitude in C1 is smaller than th, and the image gradient amplitude in C2 is greater than th, and the respective mean values of these two types of gradient amplitude values are m1 and m2, and the global mean value of the gradient amplitude value is mG. At the same time, the probability that the gradient amplitude value is divided into C1 and C2 classes is p1 and p2 respectively, as follows:

p1*m1+p2*m2=mG

p1+p2=1

S4.2: According to the concept of variance, the expression of variance between classes is:

σ ² =p1(m1-mG) ² +p2(m2-mG) ²

S4.3: Normalize the gradient magnitude matrix to the range of [0, 255], and then iterate 256 times to find the value that makes the inter-class variance σ ² reach the maximum value, which is the required threshold th.

Step S5 is specifically:

S5.1: Calculate the image gradient. The edge of the image is the place where the gray value changes significantly, and the size of the change can be measured by the image gradient amplitude, so the Sobel operator is used to calculate the image gradient value and direction;

S5.2: Suppress the non-maximum value of the gradient amplitude. After the global gradient is obtained by calculation, it is necessary to identify the part with obvious gradient changes. In order to determine the edge, the local gradient maximum point must be retained. The processing method is to replace the non-maximum value point with zero, so that the obtained edge will be a thinned edge, and a set of all possible "edges" is also obtained;

S5.3: Perform double-threshold screening, set the pixels whose grayscale change is greater than the highest threshold as strong edge pixels, and remove pixels lower than the lowest threshold; set the pixels between the lowest threshold and the highest threshold as weak edge pixels, if there are strong edge pixels in the area of the weak edge pixels, then keep the weak edge pixels, and if there are no strong edge pixels in the area of the weak edge pixels, then reject the weak edge pixels. Points are added so that the edges are as closed as possible.

The lowest threshold is 0.66th, and the highest threshold is 0.9h.

Step S6 denoises according to the area of connected components, specifically:

S6.1: Scan each pixel of the first image, divide the pixels with the same pixel value and connected with each other in 8 neighborhoods into the same group, and finally obtain the connected components of all pixels in the first image;

S6.2: Arrange all the connected components in ascending order of area, and the connected components whose area is smaller than the preset minimum value are regarded as noise and removed to obtain the second image.

In step S6.2, the preset minimum value is the 90% quantile of the areas of all connected components.

Step S7 is specifically:

S7.1: Use the convolution kernel B to expand the second image, connect the disconnected cracks and fill the dirty block holes:

Define the convolution kernel B. The convolution kernel B can be of any shape and size, and has a separately defined reference point-anchor point, which is usually a square or a disk with a reference point. The kernel can be called a template or mask;

Convolve the convolution kernel B with the second image, and calculate the maximum value of pixels in the area covered by the convolution kernel B;

assigning said maximum value to the pixel specified by the reference point;

S7.2: Corrosion operation using convolution kernel C:

Define the convolution kernel C. The convolution kernel C can be of any shape and size, and has a separately defined reference point-anchor point. Usually, the kernel is a square or disc with a reference point. The kernel can be called a template or a mask;

Convolve the convolution kernel C with the expanded second image, and calculate the minimum pixel value of the area covered by the convolution kernel C;

Assign the minimum value to the pixel specified by the reference point to obtain the third image.

Step S8 is specifically:

S8.1: Scan each pixel of the third image, divide the pixels with the same pixel value and connected with each other in the 8 neighborhoods into the same group, finally obtain all the pixel connected components in the image, and calculate the area, length and width of each connected component;

S8.2: Calculate the length-by-width area of each connected component;

S8.3: If the area of the length times the width of the connected component is greater than the first preset value of the area of the original image and the actual area is greater than the second preset value of the area of the original image, it is considered to be a dirty block and not a crack, and the connected component is removed to obtain the final crack image.

In this embodiment, the first preset value is 1/25, and the second preset value is 1/100.

Example 3

The present embodiment uses the method of embodiment 1 and embodiment 2 to provide the following specific examples:

Figure 2 is the image of cracks in the tunnel arch collected by a real digital camera. The image resolution is 4019*1000 pixels. The image has the problem of different brightness and darkness of the same color object in the real world caused by uneven illumination, too much noise interferes with crack recognition, and there are also dirty iron pieces that are not cracks in the middle of the image. After using labelme to mark the original image, it is shown in Figure 3.

At present, the conventional processing method in the industry is to obtain the threshold th through the OTSU algorithm, and use th and 2h as the Canny threshold for edge detection respectively. The effect is shown in Figure 4. There are three serious problems. Although the algorithm detects some edges, there is a lot of snowflake noise. The cracks in the dark part of the right tail cannot be detected completely, and the dirty iron block in the middle cannot be filtered. Therefore, new detection methods are needed to solve it.

Next, the crack detection method of the present invention is used to detect cracks in the image of this embodiment. The specific steps are as follows:

1) 1000 and 4019 are the width and length of the input image respectively, 50=min(4019, 1000)/20, then the image is divided into 80×20 grids, and light compensation is performed to obtain a picture showing more information. As shown in Figure 5, the brightness of the dark part of the picture is effectively improved, and the gray scale range of the picture is more concentrated;

2) Obtain a picture with uniform brightness according to step 1), and perform a median filter with a kernel of 5×5 on the picture to remove some isolated bright spot noise in the image, while retaining edge details to facilitate subsequent processing, as shown in Figure 6, after the processing of step 2), the image becomes smoother in other parts while retaining the edge characteristics of the crack, which greatly helps to reduce noise;

3) According to step 2) to obtain the noise-reduced image, calculate the gradient magnitude of the image, automatically generate the optimal segmentation threshold th by maximizing the variance between classes of the gradient magnitude, and use 0.66h and 0.9h as the lowest threshold and the highest threshold of the Canny operator respectively for edge detection, and obtain crack edges with noise but richer detail information.

4) Obtain a binary image with edges and noises according to step 3), and search and calculate connected components through eight connected domains, wherein the 90% quantile of the area of all connected components is 33, and the connected components with an area less than 33 are considered to be noise and directly removed, as shown in Figure 8;

5) According to step 4) to obtain a relatively clean but relatively sparse image of the cracks, use the closed operation method to splice the disconnected cracks and fill the holes of the dirty block to prepare for step 6). The effect is shown in Figure 9. After the processing of step 5), the fractures of the cracks identified by the binary image in the previous step are effectively connected without changing the overall characteristics of the cracks, and the connection of the holes and holes of the dirty block of the iron sheet becomes a whole;

6) Follow step 5) to obtain images of cracks that are further mosaiced, and use eight-connected domain search to calculate connected components. If the length by width of the part is larger than the area of the original image and the actual area is larger than the area of the original image, it is considered not to be a crack but a surface stain. After removing these noise components, the final clean crack detection map is obtained, as shown in Figure 10. After the processing in step 6), many noises, including the entire noise of the metal in the middle, are effectively removed, and the overall connectivity of the crack is good.

Comparing Fig. 2 with Fig. 10, using the methods of Embodiment 1 and Embodiment 2, the noise is greatly reduced, which is beneficial to identify cracks. Secondly, cracks in dark places are effectively identified and have better connectivity.

SSIM (Structural Similarity), structural similarity, is an index to measure the similarity of two images. Given two images and , the structural similarity of the two images can be found as follows:

where μ _x is the mean value of x, μ _y is the mean value of y, is the variance of x, /> is the variance of y and σ _xy is the covariance of x and y. c ₁ =(k ₁ L) ² , c ₂ =(k ₂ L) ² are constants for maintaining stability. L is the dynamic range of pixel values. k ₁ =0.01, k ₂ =0.03.

The range of structural similarity is -1 to 1. The closer the SSIM value is to 1, the closer the two images are. When the two images are exactly the same, the SSIM value is 1. The detection results and the SSIM value of the original image label can reflect the accuracy of crack detection to a certain extent. After calculation, the SSIM value of Figure 10 and Figure 2 is 0.963, while the SSIM value of Figure 4 and Figure 3 is 0.921. It can be seen that the method of this embodiment is significantly better than the commonly used method based on OTSU in terms of SSIM indicators.

As shown in Figures 11 to 13, in order to compare the length of the detected cracks, the crack mark pictures are obtained by manually marking the original picture and the test result picture. There are two cracks on the left and right in the experimental picture, and the length and width information is calculated based on the connectivity.

Through calculation, it can be found that both algorithms can identify the cracks on the left relatively completely, which is very similar to the original image; for the cracks on the right, the method of this embodiment is obviously better than the commonly used method based on OTSU, and more parts can be detected.

The same or similar reference numerals correspond to the same or similar components;

The terms describing the positional relationship in the drawings are only for illustrative purposes and cannot be interpreted as limitations on this patent;

Apparently, the above-mentioned embodiments of the present invention are only examples for clearly illustrating the present invention, rather than limiting the implementation of the present invention. For those of ordinary skill in the art, other changes or changes in different forms can be made on the basis of the above description. It is not necessary and impossible to exhaustively list all the implementation manners here. All modifications, equivalent replacements and improvements made within the spirit and principles of the present invention shall be included within the protection scope of the claims of the present invention.

Claims (10)

1. An adaptive canny method for crack detection, comprising the steps of:

s1: acquiring an image including a crack;

s2: performing illumination compensation according to the size of the image to obtain an image with uniform brightness;

s3: performing median filtering on the image with uniform brightness to obtain a noise reduction image;

s4: calculating the gradient amplitude of the noise reduction image, and automatically generating an optimal segmentation threshold th by a method of maximizing the variance among gradient amplitude classes;

s5: obtaining a lowest threshold and a highest threshold of a Canny edge detection algorithm according to the optimal segmentation threshold th, and carrying out edge detection by using the lowest threshold and the highest threshold to obtain a first image;

s6: searching and calculating communication components by utilizing eight communication domains on the first image, removing the communication components with the area smaller than a preset minimum value, and obtaining a second image;

s7: the broken cracks are spliced together by using a closed operation method on the second image, holes of dirty blocks are filled, and a third image is obtained;

s8: and searching and calculating a communication component by using the eight communication domains for the third image, and removing the communication component to obtain a final crack image if the length multiplied by the width area of the communication component is larger than a first preset value of the original image area and the actual area is larger than a second preset value of the original image area.

2. The adaptive canny method for crack detection according to claim 1, wherein step S2 is specifically:

s2.1: the image is converted into a gray scale image, the gray scale values after gray scale are expressed by f (x, y) = (x, y), and R, G, B respectively represent red, green and blue components in the original true color image, and the method comprises the following steps:

f(x,y)＝0.299R+0.587G+0.114B

s2.2: performing illumination complementation on the gray level image, and performing brightness compensation on different brightness areas in the image:

solving the average gray level of the gray level diagram I, and recording the width w and the height h of the gray level diagram I;

taking m=min (w, h)/20, dividing the gray scale image I intoEach block is used for calculating the average value of each block to obtain a brightness matrix D of each sub-block;

subtracting the average gray level of the gray level map I from each element of the matrix D to obtain a brightness difference matrix E of the sub-block;

the matrix E is differentiated into a brightness distribution matrix R with the same size as the gray level diagram I by a double-cube difference method;

an image result=i-R of uniform brightness is obtained.

3. The adaptive canny method for crack detection according to claim 2, wherein step S3 is specifically:

s3.1: determining a convolution kernel size;

s3.2: the center of the convolution kernel corresponds to the first pixel at the upper left of the image, the gray values of all pixels in the area covered by the convolution kernel are ordered, the median of the gray sequence is selected, and the median is assigned to the pixel corresponding to the center of the convolution kernel;

s3.3: and repeating the step S3.2 for each pixel on the image to obtain a noise reduction image.

4. The adaptive canny method for crack detection of claim 3, wherein step S4 automatically generates an optimal segmentation threshold th by maximizing the inter-class variance of the gradient magnitude, specifically:

s4.1: setting an optimal segmentation threshold th to divide the image gradient amplitude into two types C1 and C2, wherein the image gradient amplitude in C1 is smaller than th, the image gradient amplitude in C2 is larger than th, the respective average value of the two types of gradient amplitude is m1 and m2, the global average value of the gradient amplitude is mG, and the probability that the gradient amplitude is divided into the C1 and C2 type is p1 and p2 respectively, and the method comprises the following steps:

p1*m1+p2*m2＝mG

p1+p2＝1

s4.2: based on the concept of variance, the inter-class variance expression is:

σ ² ＝1(m1-mG) ² +2(m2-G) ²

s4.3: normalize gradient magnitude matrix to [0,255]Is then iterated 256 times to solve for such a way that the inter-class variance sigma ² The value at which the maximum value is reached is the threshold th.

5. The adaptive canny method for crack detection of claim 4, wherein step S5 is specifically:

s5.1: calculating an image gradient;

s5.2: performing non-maximum suppression on the gradient amplitude;

s5.3: performing double-threshold screening, setting pixels with gray level change larger than the highest threshold value as strong edge pixels, and removing pixels lower than the lowest threshold value; the setting between the lowest threshold and the highest threshold is a weak edge pixel, if there is a strong edge pixel in the field of the weak edge pixel, the weak edge pixel is reserved, and if there is no strong edge pixel in the field of the weak edge pixel, the weak edge pixel is rejected.

6. The adaptive canny method for fracture detection of claim 5, wherein the minimum threshold is 0.66th and the maximum threshold is 0.9h.

7. The adaptive canny method for crack detection of claim 6, wherein step S6 is specifically:

s6.1: scanning each pixel point of the first image, and dividing pixels which have the same pixel value and are mutually communicated in 8 adjacent domains into the same group to finally obtain communication components of all pixels in the first image;

s6.2: and arranging all the communication components from small to large in area, and removing the communication components with the areas smaller than the preset minimum value to obtain a second image.

8. The adaptive canny method for crack detection of claim 7, wherein the preset minimum in step S6.2 is 90% quantile of the area of all connected components.

9. The adaptive canny method for crack detection of claim 8, wherein step S7 is specifically:

s7.1: performing expansion operation on the second image by using a convolution kernel B:

defining a convolution kernel B, which may be of any shape and size, and has a separately defined reference point-anchor point;

convolving the convolution kernel B with the second image, and calculating the maximum value of the pixel points of the coverage area of the convolution kernel B;

assigning the maximum value to the pixel designated by the reference point;

s7.2: corrosion operation using convolution kernel C:

defining a convolution kernel C, which may be of any shape and size, and having a separately defined reference point-anchor point;

convolving the convolution kernel C with the second image subjected to expansion operation, and calculating a minimum value of pixel points of a coverage area of the convolution kernel C;

and assigning the minimum value to the pixel designated by the reference point to obtain a third image.

10. The adaptive canny method for crack detection of claim 9, wherein step S8 is specifically:

s8.1: scanning each pixel point of the third image, dividing pixels which have the same pixel value and are mutually communicated in 8 adjacent domains into the same group, finally obtaining all pixel communication components in the image, and calculating the area, the length and the width of each communication component;

s8.2: calculating the length by width area of each communication component;

s8.3: and if the length-width area of the communication component is larger than the first preset value of the original image area and the actual area is larger than the second preset value of the original image area, removing the communication component to obtain a final crack image.