A machine vision-based method for detecting defects in concrete

Abstract

The present application relates to the field of image data processing, and more particularly to a concrete defect detection method based on machine vision. The content includes: collecting original images and preprocessing to obtain normalized gradient intensity map; the normalized gradient intensity map is processed by a local-global entropy collaborative adaptive threshold segmentation algorithm to generate a binary defect segmentation image; independent defect regions in the binary defect segmentation image are distinguished to obtain a label map, and feature extraction is performed on the label map to obtain a feature vector; based on the feature vector, the concrete defects are classified to obtain the final classification result of the defects. The technical problem of the prior art is solved, which is that the traditional defect detection method relies on fixed threshold for image segmentation, and the threshold cannot be flexibly adjusted; during defect calibration, especially when the defect regions are close to each other or the morphology is complex, there is a risk of missing detection or false detection; when facing different types of defects, accurate distinction may not be possible, resulting in low classification accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image data processing, and particularly relates to a concrete defect detection method based on machine vision. BACKGROUND

[0002] As one of the most widely used materials in modern construction engineering, the quality of concrete is directly related to the safety, durability and service life of the building structure. However, due to construction technology, environmental factors, material aging and other reasons, defects such as cracks, cavities, peeling and honeycomb may occur in the concrete structure during use, which not only affects the appearance of the building, but also may lead to a decrease in structural strength, and even cause safety accidents. Therefore, timely detection and evaluation of concrete defects have become an important research direction in the field of construction engineering.

[0003] Traditional concrete defect detection methods mainly rely on manual visual inspection or simple non-destructive testing techniques such as ultrasonic testing, infrared thermal imaging and knocking method. Although these methods can find defects to some extent, they have low efficiency, strong subjectivity and limited detection accuracy. Especially in the detection of large-scale infrastructure such as bridges, tunnels and high-rise buildings, it is difficult to meet the demand for high efficiency and high precision. With the rapid development of computer vision and artificial intelligence technology, concrete defect detection methods based on machine vision have gradually emerged and become a research hotspot in recent years.

[0004] Concrete defect detection methods based on machine vision have overcome many limitations of traditional methods due to their non-contact, high efficiency and high precision, and have become an important technical means in the field of modern construction engineering. With the continuous progress of image acquisition devices, image processing algorithms and deep learning technology, the application prospect in the fields of bridges, tunnels and high-rise buildings is increasingly broad.

[0005] However, the above methods have the following problems: they mainly rely on fixed threshold for image segmentation, which is easily affected by background complexity and local changes, resulting in the inability to flexibly adjust the threshold; in defect labeling, especially when defect regions are close to each other or have complex shapes, there may be missed detection or false detection; when facing different types of defects (such as cracks, holes and peeling), accurate differentiation may not be possible, resulting in low classification accuracy. SUMMARY

[0006] The present application provides a concrete defect detection method based on machine vision to solve the technical problems that traditional defect detection methods mainly rely on fixed threshold for image segmentation, which is easily affected by background complexity and local changes, resulting in the inability to flexibly adjust the threshold; in defect labeling, especially when defect regions are close to each other or have complex shapes, there may be missed detection or false detection; when facing different types of defects (such as cracks, holes and peeling), accurate differentiation may not be possible, resulting in low classification accuracy.

[0007] The application discloses a machine vision-based concrete defect detection method.

[0008] The machine vision-based concrete defect detection method comprises the following steps:

[0009] S1, collecting an original image and performing preprocessing to obtain a normalized gradient intensity graph; a local-global entropy cooperative adaptive threshold segmentation algorithm is used to process the normalized gradient intensity graph to generate a binary defect segmentation image;

[0010] S2, distinguishing independent defect regions in the binary defect segmentation image to obtain a label graph, performing feature extraction on the label graph to obtain a feature vector, and classifying the concrete defects based on the feature vector to obtain a final classification result of the defects.

[0011] Preferably, S1 specifically comprises:

[0012] In the implementation process of the preprocessing, Gaussian filtering is applied to the original image for denoising processing to output a filtered image; a weighted average method is used to convert the filtered image into a gray-scale image; the gray-scale image is normalized to output a normalized gray-scale image; based on the normalized gray-scale image, the gradient in the horizontal direction and the gradient in the vertical direction are calculated respectively, the gradient intensity value of the pixel is calculated based on the gradient in the horizontal direction and the gradient in the vertical direction, and a gradient intensity graph is generated; the gradient intensity graph is normalized to obtain a normalized gradient intensity graph.

[0013] Preferably, S1 specifically comprises:

[0014] In the implementation process of the local-global entropy cooperative adaptive threshold segmentation algorithm, the normalized gradient intensity graph is divided into windows, and the local features of each local window are counted, and the average value of the gradient intensity of all pixels in the local window and the local standard deviation are calculated.

[0015] Preferably, S1 specifically comprises:

[0016] In the implementation process of the local-global entropy cooperative adaptive threshold segmentation algorithm, the distribution of the gradient intensity values of the entire normalized gradient intensity graph is counted to obtain a probability distribution; based on the probability distribution, the global information entropy is calculated.

[0017] Preferably, S1 specifically comprises:

[0018] Based on the average value of the gradient intensity of the local window and the local standard deviation, the weight coefficient of the local standard deviation and the adjustment coefficient of the global entropy constraint are introduced in combination with the global information entropy to adaptively adjust the local segmentation threshold.

[0019] Preferably, S1 specifically includes:

[0020] Iterate through each pixel in the normalized gradient intensity map, compare the gradient intensity value of the pixel with the local segmentation threshold, and determine the binary label of the current pixel. When the gradient intensity value of the pixel is higher than the local segmentation threshold, it means that the current pixel is in the defect region; otherwise, the current pixel is in the background region. Based on the binary labels of the pixels in the normalized gradient intensity map, a binary defect segmentation image is formed.

[0021] Preferably, S2 specifically includes:

[0022] The binary defect segmentation image is traversed pixel by pixel. Pixels belonging to the same connected region are assigned the same label number by a connected component labeling algorithm to obtain a label map.

[0023] Preferably, S2 specifically includes:

[0024] Based on the label map, the area and perimeter of each connected region are extracted, and the shape factor is calculated; the area, perimeter, and shape factor are combined to form a feature vector.

[0025] Preferably, S2 specifically includes:

[0026] The feature vectors are input into a pre-trained random forest classification model to classify concrete defects and obtain the final classification result of the defects.

[0027] The beneficial effects of the technical solution of the present invention are:

[0028] 1. By using a high-resolution industrial camera to acquire RGB images of the concrete surface, the original image is ensured to be rich in detail, providing high-quality input for subsequent processing. In the preprocessing stage, Gaussian filtering is used to effectively remove noise and generate a clear filtered image. A weighted average method is used, combined with human eye color sensitivity, to convert the filtered image into a grayscale image, retaining key texture information. Normalization processing is used to eliminate the influence of uneven illumination, so as to output a stable normalized grayscale image, enhance image quality, lay the foundation for subsequent edge feature extraction and defect segmentation, and improve detection accuracy and robustness.

[0029] 2. By applying the Sobel operator to the normalized grayscale image, the horizontal and vertical gradients are calculated, and the gradient intensity map is synthesized, which significantly enhances the edge features of defect areas such as cracks and holes. By further normalizing the gradient intensity map, the output range is ensured to be consistent, the adaptability to different lighting and background conditions is improved, the defect area is effectively highlighted, a reliable basis is provided for subsequent segmentation, and the identification effect of defects in complex backgrounds is improved.

[0030] 3. Based on the normalized gradient intensity map, a local-global entropy collaborative adaptive threshold segmentation algorithm is proposed. By dividing the normalized gradient intensity map into multiple windows, the local gradient mean and local standard deviation are combined to accurately capture the local defect characteristics. At the same time, the texture complexity of the entire normalized gradient intensity map is evaluated by the global information entropy; based on the local gradient mean and local standard deviation, and combined with the global information entropy, the local segmentation threshold is dynamically adjusted to realize adaptive segmentation, which overcomes the problems of excessive averaging and isolation of traditional global threshold algorithm, significantly reduces the missed detection and misjudgment rate, and generates high-quality binary defect segmentation image to adapt to the defect detection requirements under complex background.

[0031] 4. The binary defect segmentation image is processed by using the connected domain labeling algorithm to distinguish the independent defect regions, generate a label map, extract the area and perimeter of each connected region, and calculate the shape factor to form a feature vector. The feature vector comprehensively describes the geometric and morphological properties of the defect, providing rich information support for subsequent classification, and ensuring accurate differentiation and feature expression of the defect region.

[0032] 5. The extracted feature vector is input into a pre-trained random forest classification model to automatically classify defects into multiple categories such as cracks, holes, and peeling; the random forest classification model significantly improves the accuracy and stability of defect classification through joint determination of multiple decision trees, and is suitable for complex scenarios of multiple defect types, achieving efficient and reliable defect classification and expanding the application range of the detection method. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 A flowchart of a concrete defect detection method based on machine vision according to the present application. DETAILED DESCRIPTION

[0034] In order to further illustrate the technical means and effects taken by the present application to achieve the predetermined object, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0035] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs.

[0036] The specific scheme of the concrete defect detection method based on machine vision provided by the present application will be described below with reference to the drawings.

[0037] Referring to the drawings Figure 1 It shows a flow chart of a machine vision-based concrete defect detection method provided by an embodiment of the application, and the method comprises the following steps:

[0038] S1, collect an original image and perform preprocessing to obtain a normalized gradient intensity map; a local-global entropy collaborative adaptive threshold segmentation algorithm is used to process the normalized gradient intensity map to generate a binary defect segmentation image;

[0039] A high-resolution industrial camera is used to collect a concrete surface RGB image, which is denoted as an original image, and the original image is preprocessed to obtain a normalized gradient intensity map; the preprocessing step is as follows:

[0040] First, Gaussian filtering is applied to the original image for denoising processing to output a filtered image;

[0041] Second, a weighted average method (based on the sensitivity of the human eye to color) is used to convert the filtered image into a grayscale image;

[0042] Next, the grayscale image is normalized to eliminate the influence of light, and a normalized grayscale image is output;

[0043] Finally, in order to extract the edge features of the concrete surface image in the image processing process, the Sobel operator is applied to the normalized grayscale image, the horizontal and vertical direction gradients are calculated respectively, the gradient intensity value of the pixel is calculated based on the horizontal and vertical direction gradients, and the gradient intensity map is generated to enhance the edge features of the defect area. In order to improve the robustness, the gradient intensity map is normalized to obtain a normalized gradient intensity map, and the output range is ensured to be consistent;

[0044] The Gaussian filtering, weighted average method, Sobel operator and normalization processing method are all well-known technical means to those skilled in the art, and will not be described here;

[0045] In order to solve the problem of concrete defect detection in complex background, a local-global entropy collaborative adaptive threshold segmentation algorithm is used to process the normalized gradient intensity map to generate a binary defect segmentation image;

[0046] The local-global entropy collaborative adaptive threshold segmentation algorithm divides the normalized gradient intensity map into several local windows of the same size. The size of the local window ranges from 16x16 to 64x64 pixels, and can be set according to the specific implementation scenario, without limitation here. For each divided local window, local feature statistics are performed, and the feature distribution of all pixels within the local window is statistically analyzed. Specifically, the average gradient intensity of all pixels in the local window is calculated, and the average gradient intensity reflects the overall gradient distribution level of the local window. At the same time, the standard deviation between the gradient intensity values ​​of all pixels in the same local window and the average gradient intensity of the local window is calculated, i.e., the local standard deviation, which represents the fluctuation range based on the mean. In areas where the local background is relatively uniform and there are no obvious defects, the local standard deviation value will be lower. If there are defects with high gradient changes such as cracks and holes, the local standard deviation value will be relatively higher.

[0047] While completing the statistics of all the local features mentioned above, the texture complexity is evaluated from a global perspective. Specifically, the distribution of gradient intensity values ​​in the entire normalized gradient intensity map is statistically analyzed to obtain the frequency of different gradient intensity values ​​appearing globally, and thus obtain the probability distribution. Based on the probability distribution, the global information entropy is further calculated to reflect the texture complexity of the entire normalized gradient intensity map. The higher the global information entropy, the richer the gradient distribution of the entire normalized gradient intensity map, the more texture fluctuations and the more complex the background. If the global information entropy is low, the entire normalized gradient intensity map tends to be uniform and flat, and the defects and background have a strong contrast.

[0048] Global information entropy H global The calculation formula is:

[0049]

[0050] Where g represents the gradient intensity value of a pixel in the normalized gradient intensity map; p(g) represents the probability that the gradient intensity value in the normalized gradient intensity map is g;

[0051] By dynamically combining local feature statistics with global information entropy, and by fusing global information entropy based on the local features of different windows, the local segmentation threshold is adaptively adjusted.

[0052] Traverse each pixel of the normalized gradient intensity map, compare the gradient intensity value of the pixel with the local segmentation threshold. If the gradient intensity value of the pixel is higher than the local segmentation threshold, the pixel is determined to be in the defect region; otherwise, the pixel is determined to be in the background region.

[0053] After the above traversal and judgment, each pixel of the normalized gradient intensity map is assigned a binary label of defect or background, forming a binary defect segmentation image.

[0054] The determination formula of the binary label is as follows:

[0055]

[0056] Wherein, B(x, y) represents the binary label of the pixel point (x, y) of the normalized gradient intensity image, which represents whether the current pixel point is in the defect area or in the background area, 1 is determined as the defect area, and 0 is determined as the background area; G n (x, y) represents the gradient intensity value of the normalized gradient intensity image at the pixel point (x, y), which is used to measure the local change degree at the pixel point, and the greater the gradient intensity value represents the stronger the gradient (edge, defect indication); μ i,j + represents the local partition threshold; μ i,j represents the average value of the gradient intensity in the local window W i,j , which reflects the average gradient level of the pixels in the current local window, and is used as the local partition baseline, and the calculation formula of the average value of the gradient intensity is as follows:

[0057]

[0058] Wherein, |W i,j | represents the number of pixels in the local window; W i,j represents the local window indexed by the i-th row and the j-th column, and the local partition threshold can be adapted to the local by performing the blocking operation on the entire normalized gradient intensity image;

[0059] α represents the weight coefficient of the local standard deviation, which is used to adjust the influence weight of the local standard deviation on the local partition threshold, decides the sensitivity of the elastic partition, and the value range is 0.5 to 3.0, which can be 1.5, and can be set according to the specific implementation scene, which is not limited here; σ i,j is the local standard deviation, which represents the standard deviation of the gradient intensity in the local window W i,j , which is used to measure the fluctuation of the gradient intensity of the pixels in the local window, and the greater the local standard deviation indicates that the local texture is more complex and uneven, which is used to adjust the sensitivity of the local partition threshold to the local change, and the calculation method of the standard deviation is a technical means familiar to those skilled in the art, which is not described here; β represents the adjustment coefficient of the global entropy constraint, and the value range is 0.1 to 1.0, which can be 0.5, and can be set according to the specific implementation scene, which is not limited here; H global represents the global information entropy, which reflects the texture (gradient distribution) complexity of the entire normalized gradient intensity image, and the high texture complexity (global information entropy) indicates that the whole is easy to misjudge, and the local partition threshold needs to be moderately improved to better distinguish the defect and the complex background; The global entropy ratio is represented, and the denominator is used to normalize the global information entropy, so as to ensure that the global information entropy is related to the area MxN of the normalized gradient intensity diagram, and the influence proportion of the entropy value is consistent under different sizes of pictures, so as to increase the universality of the determination formula; M represents the width (pixel number) of the normalized gradient intensity diagram; and N represents the height (pixel number) of the normalized gradient intensity diagram.

[0060] Relying on the cooperative regulation of local feature statistics and global information entropy, the local segmentation threshold can be dynamically self-adapted to the actual detection scene, effectively solving the problems of local missed detection and misjudgment caused by the "over-averaging" of traditional global threshold algorithm and the "isolated one-sidedness" of local algorithm.

[0061] S2, independent defect regions in the binary defect segmentation image are distinguished to obtain a label image, and feature extraction is performed on the label image to obtain a feature vector; based on the feature vector, the concrete defect is classified to obtain a final classification result of the defect;

[0062] All independent defect regions in the binary defect segmentation image are distinguished by a connected domain marking algorithm, and the specific implementation process is as follows: the binary defect segmentation image is traversed pixel by pixel, and 8-connected or 4-connected mode is used for pixel adjacency judgment, pixels belonging to the same connected region are assigned the same label number, and a label image is obtained, each connected region corresponds to a unique label;

[0063] Further feature extraction is performed on the label image to obtain a feature vector; the specific implementation process is as follows: all pixels inside each connected region are traversed by a pixel counting method, and the number of pixels is counted, that is, the area; all boundary pixels of each connected region are found by a boundary tracking algorithm, and the boundary pixels are counted to obtain the perimeter; based on the area and the perimeter, a shape factor is directly calculated by a shape descriptor-circularity, which is used to distinguish the shape attribute of the defect, for example, the closer the shape factor is to 1, the closer the shape of the defect is to a circle, and the shape factor is much smaller than 1, which indicates that the shape of the defect is elongated or irregular, which is a commonly used discrimination standard; the area, the perimeter and the shape factor extracted by the feature extraction are combined to form a feature vector F k ; The pixel counting method and the boundary tracking algorithm used in feature extraction are well-known technical means to those skilled in the art, and will not be described here;

[0064] The feature vector is input into a pre-trained random forest classification model to obtain a final classification result of the defect; the random forest is a machine learning algorithm based on the idea of ensemble learning and improves the overall classification ability by joint judgment of a large number of decision trees, each decision tree independently makes a class judgment on the input features according to its own splitting rule and data training result, and the final voting result of all trees is weighted and counted according to their respective weights, and the class with the most votes is the final classification result, which can be expressed as follows:

[0065]

[0066] wherein C k represents the final classification result of the kth defect, such as crack, hole, peeling, etc. represents that after the summation expression is calculated for all possible defect categories c, the defect category c with the highest score (weighted voting) is selected as the output;∑ a∈A represents that summation is performed on each decision tree a in the decision tree set A; a represents the ath decision tree in the random forest model; A represents the set of all decision trees in the random forest, |A|∈[100, 500], which can be 200, |A| represents the number of all decision trees in the random forest; w a represents the weight of the ath decision tree, which can be in the range of which can be in the range of which can be set according to the specific implementation scenario, and is not limited here; I(·) represents an indicator function, which takes the value of 1 if the content in the parentheses is true, and 0 otherwise; tree a (F k ) represents the prediction result of the ath decision tree on the feature vector F k , that is, to determine which defect category (such as crack, hole, etc.) the kth defect belongs to; c represents the label of the candidate defect category; F k represents the feature vector of the kth defect, which includes area, perimeter and shape factor;

[0067] With the strong learning ability of the random forest classification model, high-precision automatic classification of multi-category (such as crack, hole, peeling, etc.) defects is realized, which significantly improves the reliability and application range of the detection system.

[0068] In summary, a concrete defect detection method based on machine vision is completed.

[0069] The order of the embodiments is only for description, and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are also possible or can be advantageous.

[0070] Each embodiment in the specification is described in a progressive manner, and the same or similar parts between each embodiment can be referred to each other. Each embodiment focuses on the differences from other embodiments.

[0071] The above examples are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can be modified, or some technical features can be replaced by equivalent features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A method for detecting defects in concrete based on machine vision, characterized in that, The method comprises the following steps: S1, collect original image and pre-process, get normalized gradient intensity image; through local-global entropy collaborative adaptive threshold segmentation algorithm, divide window to normalized gradient intensity image, for each local window divided, count local feature, calculate average value of gradient intensity of all pixels of local window and local standard deviation; based on average value of gradient intensity of local window and local standard deviation, combined with global information entropy, introduce weight coefficient of local standard deviation and adjustment coefficient of global information entropy constraint, adaptively adjust local segmentation threshold; represents local segmentation threshold; represents average value of gradient intensity in local window , calculation formula is: ， wherein, represents the number of pixels within the local window; represents the local window indexed by the row and the column; represents the gradient intensity value of the normalized gradient intensity image at pixel point ; a weight coefficient representing a local standard deviation; is a local standard deviation; is a regulation coefficient representing a global information entropy constraint; is a global information entropy; is a width of a normalized gradient strength map; is a height of a normalized gradient strength map; Each pixel of the normalized gradient intensity image is traversed, the gradient intensity value of the pixel is compared with a local segmentation threshold, the binary label of the current pixel is determined, and a binary defect segmentation image is generated; S2, the independent defect regions in the binary defect segmentation image are distinguished, a label image is obtained, feature extraction is performed on the label image, a feature vector is obtained, and the concrete defects are classified based on the feature vector to obtain a final classification result of the defects.

2. The method for detecting concrete defects based on machine vision according to claim 1, characterized in that, The S1 specifically comprises: In the implementation process of the preprocessing, Gaussian filtering is applied to the original image for denoising processing to output a filtered image; a weighted average method is used to convert the filtered image into a grayscale image; the grayscale image is normalized to output a normalized grayscale image; based on the normalized grayscale image, horizontal and vertical direction gradients are calculated, the gradient intensity value of the pixel is calculated based on the horizontal direction gradient and the vertical direction gradient, and a gradient intensity image is generated; the gradient intensity image is normalized to obtain a normalized gradient intensity image. 3.The machine vision-based concrete defect detection method of claim 1, wherein, The S1 specifically comprises: In the implementation process of the local-global entropy collaborative adaptive threshold segmentation algorithm, the distribution of the gradient intensity values of the entire normalized gradient intensity image is counted to obtain the frequency of different gradient intensity values appearing globally, thereby obtaining a probability distribution; based on the probability distribution, a global information entropy is calculated.

4. The method for detecting concrete defects based on machine vision according to claim 1, characterized in that, The S1 specifically comprises: When the gradient intensity value of the pixel is higher than the local segmentation threshold, it indicates that the current pixel is in a defect region, otherwise, the current pixel is in a background region; based on the binary label of the pixel of the normalized gradient intensity image, a binary defect segmentation image is generated.

5. The machine vision-based concrete defect detection method of claim 1, wherein, The S2 specifically comprises: The binary defect segmentation image is traversed pixel by pixel, and through a connected domain marking algorithm, pixels belonging to the same connected region are assigned the same label number to obtain a label image, each connected region corresponding to a unique label.

6. The machine vision-based concrete defect detection method of claim 5, wherein, The S2 specifically comprises: Based on the label image, the area, perimeter and shape factor of each connected region are extracted; the area, perimeter and shape factor are combined to form a feature vector.

7. The machine vision-based concrete defect detection method of claim 6, wherein, The S2 specifically comprises: The feature vector is input into a pre-trained random forest classification model to classify the concrete defects and obtain a final classification result of the defects.

Images