Crack identification method and device based on multi-scale infrared thermal feature fusion

Abstract

The invention discloses a crack identification method and device based on multi-scale infrared thermal feature fusion, and the method comprises the steps: obtaining the gray images of cracks and noise for a single-frame infrared thermal image; histogram equalization is adopted to adjust image gray level distribution, and a gray level enhanced image is obtained; setting a larger local feature sensitivity weight coefficient for the background dominant window and the noise dominant window, and setting a smaller local feature sensitivity weight coefficient for the target composite window; counting the number of foreground pixels in each window, and calculating the ratio of the number of foreground pixels to the total number of pixels; determining a background dominant window, a noise dominant window and a target composite window; setting local feature sensitivity weight coefficients of a background dominant window, a noise dominant window and a target composite window; segmenting the grayscale enhanced image to obtain a binary image, and extracting a connected domain of foreground pixels; for each foreground pixel connected domain, calculating the total number of pixels, randomly selecting a plurality of seed points, determining corresponding maximum inscribed circles, taking the average value of the diameters of all the maximum inscribed circles as the width, taking the ratio of the total number of pixels to the width as the length, and screening out connected domains with the length-width ratio smaller than 10 as noise; and optimizing the foreground pixel connected domain, removing fine noise, smoothing the contour, and regarding the finally obtained foreground pixel connected domain as a crack.

Description

Technical Field

[0001] The present invention relates to the technical field of defect detection, and in particular to a crack identification method and device based on multi-scale infrared thermal feature fusion. Background Art

[0002] Cracks are one of the most common forms of structural surface damage. Their type, density, width and other characteristics can reflect the degree of structural damage and are key indicators for evaluating the structural state. Therefore, crack detection is an important part of structural operation and maintenance. Due to the interference of the microenvironment at the crack on the heat flow distribution, temperature anomalies will occur. Infrared thermal imaging is a non-destructive testing technology that infers the surface state and internal structure of an object by detecting the temperature difference on the surface of the object caused by different thermal characteristics. Therefore, infrared thermal imaging has gradually attracted attention in the field of crack detection.

[0003] Although infrared thermal imaging crack detection technology has been applied to a certain extent, the automation and intelligence of crack identification are still low, making it difficult to achieve fast and efficient infrared thermal image crack identification. Threshold segmentation is a common image segmentation algorithm that has been widely used in crack identification. According to the segmentation strategy, threshold segmentation algorithms can be divided into global threshold segmentation algorithms and local threshold segmentation algorithms. However, due to the low signal-to-noise ratio of crack infrared thermal images, the noise interference is serious, and the global threshold segmentation algorithm is difficult to achieve crack identification. In comparison, the local threshold segmentation algorithm has stronger noise resistance, but the local threshold still needs to be determined empirically, which cannot meet the needs of crack identification automation and intelligence, limiting their large-scale application. In recent years, deep learning has gradually attracted attention in infrared thermal image defect recognition due to its powerful feature learning ability and environmental adaptability. However, deep learning uses image color, shape, texture and other features to identify cracks, while the low resolution of infrared thermal images and heavy noise interference make the crack feature information insufficient, limiting the accuracy of crack identification. In addition, the lack of large-scale and diverse infrared thermal image data sets further restricts the application of deep learning in infrared thermal image crack identification. It can be seen that infrared thermal image crack recognition still needs an automated and intelligent recognition method. Summary of the invention

[0004] The present invention aims to overcome the above problems existing in the prior art and provides a crack identification method and device based on multi-scale infrared thermal feature fusion, thereby realizing rapid and automatic crack identification.

[0005] The key of the present invention is to utilize the local statistical features of the binarized image obtained by the global threshold segmentation algorithm, establish an adaptive threshold determination criterion for the local threshold segmentation algorithm, overcome the common methods' reliance on manual labor and models, and have the advantages of strong adaptability, low computing power requirements, and high robustness.

[0006] In order to achieve the above technical objectives, the technical solution adopted by the present invention is:

[0007] A crack identification method based on multi-scale infrared thermal feature fusion includes the following steps:

[0008] (1) For a single-frame infrared thermal image, weighted average grayscale conversion is performed. If the crack is a low-temperature "cold spot", grayscale inversion is further performed to obtain a grayscale image in which the crack and noise have high grayscale values ​​and the background has low grayscale values.

[0009] (2) For the grayscale image obtained by step (1), histogram equalization is used to adjust the grayscale distribution of the image, enhance the contrast of the cracks in the image, and obtain a grayscale enhanced image.

[0010] (3) For the grayscale enhanced image obtained by step (2), a sliding window is used to traverse the image, and local statistical features such as the mean and standard deviation of the grayscale in each window are calculated. A larger local feature sensitivity weight coefficient is set for the background dominant window and the noise dominant window, and a relatively smaller local feature sensitivity weight coefficient is set for the target composite window.

[0011] (4) For the grayscale enhanced image obtained by step (2), the Otsu algorithm is used to obtain a binary image, and a sliding window of the same size as in step (3) is used to traverse the image, and the number of foreground pixels in each window is counted, and the ratio of the number of foreground pixels to the total number of pixels, i.e., the foreground pixel ratio, is calculated.

[0012] (5) Calculate the quartiles for the foreground pixel ratios of all windows obtained in step (4). The windows corresponding to the foreground pixel ratios in the lower quartile are identified as background-dominated windows, the windows corresponding to the foreground pixel ratios in the upper quartile are identified as noise-dominated windows, and the windows corresponding to the foreground pixel ratios in the middle quartile are identified as target composite windows.

[0013] (6) According to the window type identified in step (5), the background-dominated window uses the background pixel ratio as the local feature sensitivity weight coefficient, the noise-dominated window uses the foreground pixel ratio as the local feature sensitivity weight coefficient, and the target composite window also uses the foreground pixel ratio as the local feature sensitivity weight coefficient.

[0014] (7) The grayscale mean value of the sliding window obtained by step (3) plus the grayscale standard deviation is multiplied by the local feature sensitivity weight coefficient obtained by step (6) to obtain the grayscale threshold of each window, and the grayscale enhanced image obtained by step (2) is segmented to obtain a binary image, and the connected domain of the foreground pixels is extracted.

[0015] (8) For each connected domain of foreground pixels obtained by step (7), calculate the total number of pixels, randomly select a number of seed points, determine the corresponding maximum inscribed circle, take the average of all the maximum inscribed circle diameters as the width, take the ratio of the total number of pixels to the width as the length, and consider the connected domain with a length-to-width ratio less than 10 as noise and screen it out.

[0016] (9) For the binary image processed in step (8), the morphological opening operation is used to optimize the foreground pixel connected domain, remove small noise, and smooth the contour. The final foreground pixel connected domain is regarded as the crack.

[0017] Furthermore, in step (3), when the size of the sliding window is between 1 / 8 and 1 / 12 of the corresponding side length of the original image, there is no significant difference in the binary image obtained by segmentation. Therefore, a sliding window with a size of 1 / 10 of the corresponding side length of the original image can be used to traverse the image. In addition, when the background ratio in the window is high, the weight coefficient is small and the threshold is reduced, which is conducive to extracting small targets; when the noise ratio in the window is high, the weight coefficient is large and the threshold is increased, which is conducive to suppressing noise. Therefore, the crack recognition method based on multi-scale infrared thermal feature fusion also has excellent small target segmentation and noise suppression capabilities.

[0018] Furthermore, in step (4), there are differences in the foreground pixel ratios of the three types of windows in the Otsu binary image. The foreground pixel ratio of the background-dominated window is extremely small, the foreground pixel ratio of the noise-dominated window is extremely large, and the foreground pixel ratio of the target composite window is between the two. Therefore, the foreground pixel ratios of each sliding window can be used for window classification later.

[0019] Furthermore, in step (5), for the Otsu binary image, its foreground pixel ratio histogram will show an obvious U-shaped distribution. Therefore, the foreground pixel ratio quartiles can be used to divide it into the left high-frequency area (lower quartile area), the middle low-frequency area (middle quartile area) and the right high-frequency area (upper quartile area), which correspond to the background dominant window, the target composite window and the noise dominant window respectively.

[0020] Furthermore, in step (6), the expected value of the local feature sensitivity weight coefficient of each sliding window is basically consistent with the distribution characteristics of the sliding window pixel ratio. Therefore, the local feature sensitivity weight coefficient of each sliding window can be adaptively determined according to the determination criterion of step (6). Specifically, it can be calculated by the following formula:

[0021]

[0022] Where k is the weight coefficient of local feature sensitivity, r is the foreground pixel ratio in the Otsu binary image window, Q1 is the lower quartile, and Q3 is the upper quartile.

[0023] The second aspect of the present invention relates to a crack identification device based on multi-scale infrared thermal feature fusion, comprising a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the crack identification method based on multi-scale infrared thermal feature fusion of the present invention.

[0024] A third aspect of the present invention relates to a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the crack identification method based on multi-scale infrared thermal feature fusion of the present invention.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] (1) Based on the infrared thermal images of cracks, an adaptive determination criterion for the local threshold is developed by utilizing their multi-scale infrared thermal characteristics. It can adapt to a variety of complex working conditions and has excellent environmental adaptability.

[0027] (2) The present invention realizes fully automatic crack recognition of single-frame infrared thermal images through an adaptive threshold determination mechanism, without the need for human intervention throughout the process. It has the advantages of strong adaptability, low computing power requirements, and high robustness.

[0028] (3) The results of the present invention are highly readable, user-friendly, and efficient in recognition. They can be used as edge computing modules for infrared thermal imaging technology, avoiding the transmission of large-scale image data, and are particularly suitable for edge computing scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 is a flow chart of the steps of the present invention.

[0030] Figure 2 It is a schematic diagram of the device of the present invention. DETAILED DESCRIPTION

[0031] The specific implementation of the present invention will be further described below through specific examples in conjunction with the accompanying drawings.

[0032] Example 1

[0033] This embodiment relates to a crack identification method based on multi-scale infrared thermal feature fusion, comprising the following steps:

[0034] (1) Since infrared thermal images use RGB three-channel pseudo-color to represent the temperature information in the image, in order to simplify the subsequent image processing process and restore the one-to-one correspondence between temperature and color, the grayscale algorithm is used to restore the infrared thermal image to a grayscale image, that is, an image composed of pixels with grayscale values ​​ranging from 0 to 255. First, extract the RGB three-channel components of each pixel in the infrared pseudo-color image. Then, use the weighted average method to calculate the grayscale value of each pixel, that is:

[0035] I(x,y)=αR(x,y)+βG(x,y)+γB(x,y) (2)

[0036] Where I(x,y) is the pixel grayscale value, R(x,y), G(x,y), B(x,y) are the red, green, and blue channel components, respectively, and α, β, and γ are the corresponding weighting coefficients. The present invention adopts the commonly used weighting coefficients, which are 0.30, 0.59, and 0.11, respectively. Finally, a grayscale image is obtained in which the grayscale value completely reflects the temperature. For images of cracks with low-temperature "cold spots", grayscale inversion is further used to convert the target (crack) and noise into high grayscale areas, and the background into low grayscale areas.

[0037] (2) If the RGB color distribution of the crack infrared thermal image is too concentrated, the grayscale value difference between the pixels of the corresponding grayscale image is not significant, which will make it difficult to distinguish the target crack from the background in the image, which is not conducive to subsequent crack identification. Therefore, for the grayscale image processed by step (1), histogram equalization is used to enhance the grayscale image. Specifically, histogram equalization is an image processing method that redistributes the grayscale levels of the image to make the grayscale value distribution of the image more uniform, thereby improving the image contrast. First, calculate the grayscale histogram of the image, count the frequency of occurrence of each grayscale level and normalize the histogram:

[0038]

[0039] Where p(k) represents the probability of gray level k, and N is the total number of pixels in the image. Then, the Cumulative Distribution Function (CDF) is calculated:

[0040]

[0041] Where C(k) represents the cumulative probability of gray level k. As a key step in histogram equalization, it is used to accumulate the probabilities of each gray level to determine the new gray value of each pixel. Then, a new gray value is calculated for each original gray level through the cumulative distribution function. The new gray value is linearly transformed according to the cumulative distribution function, the purpose is to stretch the gray level distribution of the original image so that all gray levels are evenly distributed over the entire range [0,255]. The transformation function S(v) of histogram equalization is calculated by the following formula:

[0042] S(v)=round(255C(v)) (5)

[0043] In the formula, the round operator is rounded to the nearest integer. The transformation function S(v) maps the grayscale value in the input image to the grayscale value after equalization. Finally, the output image is obtained through the following transformation relationship:

[0044] I e (x,y)=S(I f (x,y)) (6)

[0045] In the formula, I e (x, y) is the gray value after histogram equalization processing, I f (x, y) is the grayscale value of the input image. After histogram equalization, the grayscale distribution of the output image is more uniform and the contrast is enhanced.

[0046] (3) For the grayscale enhanced image obtained by step (2), in the noise-dominated window, the grayscale mean is larger and the standard deviation is smaller; in the target composite window, the grayscale mean is smaller and the standard deviation is larger. Taking advantage of this difference in local statistical features, a sliding window with a size of 1 / 10 of the corresponding side length of the original image is used to traverse the grayscale enhanced image, and local statistical features such as the mean and standard deviation of each sliding window in the grayscale enhanced image are obtained.

[0047] For each type of window, the window with a larger background ratio is the background-dominant window, which presents the statistical characteristics of low grayscale mean and small standard deviation. A larger weight coefficient is set to increase the threshold and thus separate the background; the window with a larger noise ratio is the noise-dominant window, which presents the statistical characteristics of high grayscale mean but small standard deviation. Similarly, a larger weight coefficient is set to obtain a higher threshold and thus suppress the noise; the window containing the target, background, and noise is the target composite window, whose grayscale mean is between the background-dominant window and the noise-dominant window and the standard deviation is greater than these two types of windows. Therefore, a relatively small weight coefficient is set to obtain a smaller threshold to identify small targets. Based on this feature, the local threshold segmentation method can be used to segment cracks in the future to separate cracks from the target and background.

[0048] (4) In order to adaptively set the segmentation thresholds of the three types of windows in step (3). For the grayscale enhanced image obtained by step (2), the Otsu algorithm is first used to segment the low grayscale area (background) and the high grayscale area (target and noise) to obtain the Otsu binary image. The Otsu algorithm is a threshold segmentation algorithm based on global statistical features. The algorithm adopts the maximization of the inter-class variance criterion and uses the statistical information of the image grayscale histogram to determine the optimal segmentation threshold to segment the grayscale image into foreground (high grayscale category) and background (low grayscale category). The inter-class variance is defined as

[0049] σ 2 (T) = ω0(T)[μ0(T)-μ T ] 2 +ω1(T)[μ1(T)-μ T ] 2 (7)

[0050] In the formula, σ 2 (T) is the inter-class variance when the segmentation threshold is T, ω0(T) and ω1(T) are the pixel ratios of foreground and background when the segmentation threshold is T, μ0(T) and μ1(T) are the average grayscale values ​​of foreground and background pixels when the segmentation threshold is T, and μ T is the average gray value of the input image. The algorithm takes the segmentation threshold from 0 to 255 in turn, calculates the inter-class variance corresponding to each segmentation threshold, and the segmentation threshold with the largest inter-class variance is the optimal segmentation threshold. Then, the Otsu binary image is traversed using a sliding window with a size of 1 / 10 of the corresponding side length of the original image, and the foreground pixel ratio of each window is calculated.

[0051] (5) For the foreground pixel ratio of all windows obtained in step (4), the foreground pixel ratio histogram shows an obvious U-shaped distribution. The quartiles of the foreground pixel ratio are calculated and divided into four regions. The quartiles of the foreground pixel ratio can be used to divide it into a high-frequency region on the left, a low-frequency region in the middle, and a high-frequency region on the right. The high-frequency regions on both sides correspond to the foreground pixel ratios of the background-dominated window and the noise-dominated window in the grayscale enhanced image, respectively, and the middle continuous low-frequency region corresponds to the foreground pixel ratio of the target composite window. Accordingly, the foreground pixel ratio of the lower quartile area is extremely small, and it is identified as the background-dominated window; the foreground pixel ratio of the upper quartile area is extremely large, and it is identified as the noise-dominated window; the foreground pixel ratio of the middle quartile area (including the lower middle quartile area and the lower middle quartile area) is between the background-dominated window and the noise-dominated window, and it is identified as the target composite window.

[0052] (6) According to the window type identified in step (5), it can be seen that the expected value of the local feature sensitivity weight coefficient of each sliding window is basically consistent with the distribution characteristics of the sliding window pixel ratio. Therefore, the local feature sensitivity weight coefficient of each sliding window can be adaptively determined according to the determination criteria of step (5). The background-dominant window uses the background pixel ratio as its local feature sensitivity weight coefficient, the noise-dominant window uses the foreground pixel ratio as its local feature sensitivity weight coefficient, and the target composite window also uses the foreground pixel ratio as its local feature sensitivity weight coefficient. The specific setting method is shown in formula (1)

[0053] (7) The grayscale mean value of the sliding window obtained by step (3) plus the grayscale standard deviation is multiplied by the local feature sensitivity weight coefficient obtained by step (6) to obtain the segmentation threshold of each window, that is,

[0054] T=μ+k×σ (8)

[0055] Where, T is the segmentation threshold, μ is the grayscale mean in the sliding window, σ is the grayscale standard deviation in the sliding window, and k is the weight coefficient for adjusting the sensitivity of the local features in the sliding window. The grayscale enhanced image obtained by step (2) is segmented using the local threshold to obtain a binary image. For targets with high grayscale values, in order to avoid misidentifying the target as background due to excessively high thresholds, the weight coefficient is usually set to 0-1. When the background accounts for a high proportion in the target composite window, the weight coefficient is small and the threshold is reduced, which is conducive to extracting small targets; when the noise accounts for a high proportion in the target composite window, the weight coefficient is large and the threshold is increased, which is conducive to suppressing noise. Therefore, the crack recognition method based on multi-scale infrared thermal feature fusion can also achieve small target segmentation and noise suppression in the target composite window, and has excellent adaptive performance. Therefore, for the binary image segmented by this step, the crack part is completely retained, the uniform background and noise parts are almost completely divided into background, and the non-uniform noise and background parts are divided into isolated noise points.

[0056] (8) For the binary image obtained by step (7), only cracks and isolated noise points exist in the image. For the isolated noise points after being broken and segmented, they often do not have slender morphological features and their shapes are completely irregular. At this time, there are obvious differences in morphological features between noise and cracks. This difference is used to traverse all the foreground pixel connected areas in the image, extract the length and width corresponding to each foreground pixel connected domain, and use the size ratio of length and width as the screening condition for cracks to perform noise discrimination on the target area. Among them, the width of the foreground pixel connected domain is extracted using the maximum inscribed circle. For each foreground pixel connected domain obtained by step (7), the total number of pixels is calculated, and several seed points are randomly selected to determine the corresponding maximum inscribed circle. The average of all the maximum inscribed circle diameters is used as the width, and the ratio of the total number of pixels to the width is used as the length. The connected domain with a length-to-width ratio less than 10 is regarded as noise and screened out.

[0057] (9) After the denoising process of the binary image in step (8), only the target area corresponding to the crack remains in the binary image. However, the edge contour of the area has jagged features due to the characteristics of the local traversal segmentation algorithm. This is because the thresholds used for the pixels near the edge contour line are not consistent. In order to smooth the crack edge contour line, morphological processing is performed on the binary image after denoising to make the recognition result of the crack recognition algorithm more user-friendly and readable.

[0058] Morphological processing is a mathematical algorithm based on set theory, which is used to process the shape, structure and topological relationship in the image. It changes and extracts the features of the image through basic algorithms such as expansion and erosion. In morphological processing, a structural element is first defined. If the expansion algorithm is used, the structural element is overlapped with the image, the maximum pixel value in the image is found, and it is assigned to the pixel value of the corresponding position in the result image. The expansion operation can expand the area in the image and fill the hole. If the erosion algorithm is used, the structural element is overlapped with the image, the minimum pixel value in the image is found, and it is assigned to the pixel value of the corresponding position in the result image. The erosion operation can reduce the area in the image and remove isolated pixels. In addition, the opening and closing operations can be performed by combining the erosion and expansion operations. Since the boundary discontinuity of the crack binary image after denoising is low, the jagged features of the boundary are mainly due to the small area protruding foreground pixels, so the crack boundary contour is smoothed by the opening operation morphological processing of first erosion and then expansion to obtain the binary image of the crack and further judge the cracking situation based on it.

[0059] Example 2

[0060] Reference Figure 2This embodiment relates to a crack identification device based on multi-scale infrared thermal feature fusion, including a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the crack identification method based on multi-scale infrared thermal feature fusion of Example 1.

[0061] Example 3

[0062] This embodiment relates to a computer-readable storage medium on which a program is stored. When the program is executed by a processor, the crack identification method based on multi-scale infrared thermal feature fusion of Example 1 is implemented.

[0063] The specific description of the present invention in the above embodiments is only used to further illustrate the present invention and cannot be understood as limiting the scope of protection of the present invention. Technical engineers in this field may make some non-essential improvements and adjustments to the present invention based on the contents of the above invention, which fall within the scope of protection of the present invention.

Claims

1. A crack identification method based on multi-scale infrared thermal feature fusion, comprising the following steps: (1) For a single-frame infrared thermal image, weighted average grayscale conversion is performed. If the crack is a low-temperature "cold spot", grayscale inversion is further performed to obtain a grayscale image in which the crack and noise have high grayscale values ​​and the background has low grayscale values; (2) for the grayscale image obtained by step (1), histogram equalization is used to adjust the grayscale distribution of the image, enhance the contrast of the cracks in the image, and obtain a grayscale enhanced image; (3) For the grayscale enhanced image obtained by step (2), use a sliding window to traverse the image, calculate the local statistical features such as the mean and standard deviation of the grayscale in each window, set a larger local feature sensitivity weight coefficient for the background dominant window and the noise dominant window, and set a relatively smaller local feature sensitivity weight coefficient for the target composite window; (4) For the grayscale enhanced image obtained by step (2), the Otsu algorithm is used to obtain a binary image, and a sliding window of the same size as that in step (3) is used to traverse the image, and the number of foreground pixels in each window is counted, and the ratio of the number of foreground pixels to the total number of pixels is calculated, i.e., the foreground pixel ratio; (5) Calculating quartiles for the foreground pixel ratios of all windows obtained in step (4); the windows corresponding to the foreground pixel ratios in the lower quartile are identified as background-dominated windows; the windows corresponding to the foreground pixel ratios in the upper quartile are identified as noise-dominated windows; and the windows corresponding to the foreground pixel ratios in the middle quartile are identified as target composite windows; (6) According to the window type identified in step (5), the background dominant window uses the background pixel ratio as the local feature sensitivity weight coefficient, the noise dominant window uses the foreground pixel ratio as the local feature sensitivity weight coefficient, and the target composite window also uses the foreground pixel ratio as the local feature sensitivity weight coefficient; (7) adding the grayscale mean value of the sliding window obtained by step (3) to the grayscale standard deviation and multiplying it by the local feature sensitivity weight coefficient obtained by step (6) to obtain the grayscale threshold of each window, segmenting the grayscale enhanced image obtained by step (2) to obtain a binary image, and extracting the connected domain of the foreground pixels; (8) For each connected domain of foreground pixels obtained by step (7), the total number of pixels is calculated, several seed points are randomly selected, the corresponding maximum inscribed circle is determined, the average of the diameters of all the maximum inscribed circles is taken as the width, the ratio of the total number of pixels to the width is taken as the length, and the connected domain with an aspect ratio less than 10 is regarded as noise and is screened out; (9) For the binary image processed in step (8), the morphological opening operation is used to optimize the foreground pixel connected domain, remove small noise, and smooth the contour. The final foreground pixel connected domain is regarded as the crack.

2. A crack identification method based on multi-scale infrared thermal feature fusion as claimed in claim 1, characterized in that: In step (3), a sliding window with a size of 1 / 10 of the corresponding side length of the original image is used to traverse the image. When the background ratio in the window is high, the weight coefficient is small and the threshold is reduced, which is conducive to extracting small targets; when the noise ratio in the window is high, the weight coefficient is large and the threshold is increased, which is conducive to suppressing noise.

3. A crack identification method based on multi-scale infrared thermal feature fusion as claimed in claim 1, characterized in that: In step (4), there are differences in the foreground pixel ratios of the three types of windows in the Otsu binary image. The foreground pixel ratio of the background-dominated window is extremely small, the foreground pixel ratio of the noise-dominated window is extremely large, and the foreground pixel ratio of the target composite window is between the two. Therefore, the foreground pixel ratios of each sliding window are used for window classification in the subsequent step.

4. A crack identification method based on multi-scale infrared thermal feature fusion as claimed in claim 1, characterized in that: In step (5), for the Otsu binary image, its foreground pixel ratio histogram will show an obvious U-shaped distribution, and the foreground pixel ratio quartiles are used to divide it into a high-frequency area on the left, a low-frequency area in the middle, and a high-frequency area on the right, which correspond to the background dominant window, the target composite window, and the noise dominant window, respectively.

5. The crack identification method based on multi-scale infrared thermal feature fusion according to claim 1, characterized in that: In step (6), the expected value of the local feature sensitivity weight coefficient of each sliding window is consistent with the distribution characteristics of the sliding window pixel ratio, and the local feature sensitivity weight coefficient of each sliding window is adaptively determined according to the determination criterion of step (6): Where k is the weight coefficient of local feature sensitivity, r is the foreground pixel ratio in the Otsu binary image window, Q1 is the lower quartile, and Q3 is the upper quartile.

6. A crack identification device based on multi-scale infrared thermal feature fusion, characterized in that: It comprises a memory and one or more processors, wherein the memory stores executable codes, and when the one or more processors execute the executable codes, they are used to implement the crack identification method based on multi-scale infrared thermal feature fusion as described in any one of claims 1 to 5.

7. A computer-readable storage medium, characterized in that: A program is stored thereon, and when the program is executed by a processor, the crack identification method based on multi-scale infrared thermal feature fusion described in any one of claims 1 to 5 is implemented.

Images