AU2020101011A4

AU2020101011A4 - Method for identifying concrete cracks based on yolov3 deep learning model

Info

Publication number: AU2020101011A4
Application number: AU2020101011A
Authority: AU
Inventors: Yonggang SHEN; Zuolin WEN; Zhenwei Yu; Yiping Zhang
Original assignee: Zhejiang University ZJU
Current assignee: Zhejiang University ZJU
Priority date: 2019-06-26
Filing date: 2020-06-15
Publication date: 2020-07-23
Anticipated expiration: 2028-06-15
Also published as: CN110175658A

Abstract

The present invention belongs to the technical field of concrete structure damage detection, and discloses a method for identifying concrete cracks based on a YOLOv3 deep learning model. Crack images are imported into the YOLOv3 model and automatically compressed to a 416x416 pixel resolution; the original images are each divided into SxS grids according to the scale of a feature map by up-sampling and feature fusion methods similar to FPN; an Intersection over Union (IoU) of a candidate bounding box and a ground truth bounding box is taken as an evaluation standard, and all crack target annotation boxes in an image training set are subjected to K-means clustering analysis to obtain the size of the candidate bounding box; and a probability that each bounding box contains targets is predicted through logistic regression. The present invention simplifies the complexity of network training and reduces the computing cost, quickly and accurately identifies multiple targets, has a much better accuracy rate than other models while quickly detecting the targets, has stronger robustness and generalization capability, and is more suitable for engineering application environment. 14

Description

METHOD FOR IDENTIFYING CONCRETE CRACKS BASED ON YOLOV3 DEEP LEARNING MODEL TECHNICAL FIELD The present invention belongs to the technical field of concrete structure damage detection, and particularly relates to a method for identifying concrete cracks based on a YOLOv3 deep learning model. BACKGROUND Concrete is a building material most widely used in the construction of roads, bridges, houses, tunnels, dams and other infrastructure. Due to the low tensile strength of concrete and the joint influence of internal and external factors such as shrinkage and creep, external temperature change and foundation deformation, various cracks often occur in the process of construction and operation. Crack propagation is the initial stage of structural failure. With the continuous development of cracks, once the width of cracks exceeds a certain limit, it not only affects the appearance of infrastructure, but also may cause leakage, durability reduction, protective layer falling off, reinforcement corrosion, concrete carbonization, etc., or even have a great impact on traffic and pedestrian safety. Therefore, it is necessary to regularly detect the cracks on the surface of concrete structures and to prevent and control the cracks in advance according to test results. The meso-scale study of concrete strength and engineering practice experience show that cracking of the concrete structure is inevitable. If the requirements for crack evaluation of the concrete structure are too high, the maintenance process will be complicated and the cost will be high. The scientific approach should be to set a limit. The crack width on the same crack is generally uneven. The controlling width of the crack refers to controlling the average width of a wider section (within a range of 10%-15% of the crack length), and the average crack width determined in this way is the maximum width of the crack. Similarly, the average width of a narrower section (within a range of %-15% of the crack length) of the crack is the minimum crack width. The average crack width is between the maximum and the minimum. The minimum crack width visible to naked eyes is generally 0.05 mm. Cracks with a width less than 0.05 mm are usually called microcracks, cracks with a width greater than or equal to 0.05 mm are called macrocracks, and macrocracks are the result of microcrack propagation. Cracks with a width less than 0.05 mm in general concrete structures are not dangerous to use, and thus structures with microcracks having a width less than 0.05 mm can be considered as crack-free structures. Therefore, the structure that does not allow cracking in the design can only be a structure without cracks greater than 0.05 mm. The control standard for the maximum crack width of concrete is rough as follows: the maximum crack width is 0.3-0.4 mm if there is no erosion medium and no waterproof requirement; it is 0.2-0.3 mm if there is slight erosion and no waterproof requirements; and it is 0.1-0.2 mm if there is serious erosion and waterproof requirements. China's Technical Codefor Test and Evaluation ofcity bridges stipulates that the width of bridge cracks shall not exceed 0.3 mm. According to shape, cracks fall into surface cracks, through cracks, longitudinal cracks, transverse cracks, cracks with a wide top and a narrow bottom, cracks with a narrow top and a wide bottom, date pit-shaped cracks, diagonal cracks, oblique cracks, cracks with a wide exterior and a narrow interior, deep cracks (with a depth up to 1/2 of the thickness), etc. The shape of the crack is directly related to its stress state. The directions of most cracks are perpendicular to the direction of principal tensile stress, while the direction of pure shear cracks is parallel to the direction of shear stress. Accurately identifying the lengths, directions and widths of cracks in concrete structures is of great significance for determining the disease degree and operation status of structures, and is also a great issue for health inspection of concrete structures. The crack testing method adopted in the early stage is mainly manual testing, requiring maintenance personnel to carry out on-site investigation, marking and measurement and record test results. The manual testing method features high working intensity, low efficiency and insecurity, and it is required to approach the surface of the structure with the aid of a detecting auxiliary device, thus requiring higher professional knowledge and experience of testing personnel. In addition, the results of manual testing are subjectively affected, resulting in inconsistent measurement results, low measurement accuracy, etc. Besides, it is sometimes necessary to close the site for manual testing on bridges, tunnels and other traffic arteries. This greatly affects the normal traffic, bringing inconvenience to the normal passage of vehicles and pedestrians, or even causing traffic accidents in serious cases. Advanced nondestructive testing methods, such as an ultrasonic method, a thermal imaging method, computed tomography and an electromagnetic acoustic emission sensor testing method, have the disadvantages of expensive instruments, small measurement range, inability to completely realize non-contact measurement, etc. In about 2000, image processing techniques (IPTs) based on computer vision technology began to be used for identification of concrete surface cracks. The IPTs have an obvious advantage of identifying almost all surface defects (such as cracks and corrosion). However, the light intensity, light and shade changes, image distortion and other factors of an image seriously affect detection results, and a large amount of noise is generated during the processing. This makes it difficult to accurately and efficiently identify concrete crack targets in the image in the conventional computer vision technology. An edge detection method closest to the present invention is the most commonly used method in the IPTs, and its commonly used operators generally include first derivatives such as Sobel operator and Canny operator, and second derivatives such as Laplacian operator. The core principle of the edge detection method is to detect a set of pixel points with sharp changes in the gray level of surrounding pixels. In the image with uniform gray level changes, a boundary obtained by using only the first derivative operator is relatively coarse, or even no boundary can be found. However, the second derivative operator based on zero-crossing detection is sensitive to noise. Even if the edge can be detected, the number of edge points obtained is relatively small. Since the concrete image background has a small overall gray difference and a crack edge has a low pixel gradient and is usually a weak edge, the edge detection method has a poor effect when applied in the crack identification field. With the rapid development of artificial intelligence, a deep learning algorithm is widely used in all aspects of image processing. An image processing technology based on deep learning provides a good solution for concrete crack detection. Some scholars have proposed some methods in concrete crack identification by applying deep learning theory, such as detecting cracks without calculating defect features, proposing a deep learning framework for detecting cracks in each video frame, proposing a novel data fusion solution to aggregate information, and applying principal component analysis (PCA) to classify cracks. Such methods are generally based on a sliding window technology and deep learning, and the sliding window technology is not targeted. Due to the small widths of target cracks, most of the windows generated by the sliding window technology are usually redundant windows, actual effective windows account for only a small part, so that a large amount of computation resources are wasted, the computation cost is high, and the detection efficiency is low. The deep learning method requires a large amount of data, which is also the biggest bottleneck in the development of deep learning. In order to solve this problem, some deep learning frameworks that have their own data sets and have some target features extracted to facilitate use have been widely used. How to better realize the detection and identification of specific targets under these built frameworks is an urgent problem to be solved in thefield of concrete structure damage detection at present. In summary, the prior art has the following problems: At present, the commonly used edge detection method is not suitable for the field of crack identification; and the deep learning method based on the sliding window technology needs a large amount of data, lacks pertinence and has high computing cost and low detection efficiency. It is difficult to solve the foregoing technical problems for the following reasons. There is a lack of universally recognized and effective crack data sets in the world, while manual production of data sets requires a lot of manpower and material resources and is high in cost; and simplifying the complex model framework to maintain the accuracy and speed of detection while reducing the intermediate output of the network requires a lot of time for adjustment, trial calculation and training of the network. The solving of the foregoing technical problems has the significance that accurate identification results can be obtained under limited data sets, and all targets to be detected in an image can be obtained after a single detection of the input image, thus realizing end-to-end object detection and identification, i.e., achieves the input end--single neural network--output end, and reduces generated redundant data and computing costs. SUMMARY In view of the problems existing in the prior art, the present invention provides a method for identifying concrete cracks based on a YOLOv3 deep learning model. The present invention is implemented in this way: a method for identifying concrete cracks based on a YOLOv3 deep learning model, where the method takes a Darknet-53 network in the YOLOv3 model as a feature extractor and treats object detection as a regression problem, and target area positioning and target class prediction are directly performed based on a single end-to-end neural network model; Crack images are imported into the YOLOv3 model and automatically compressed to a 416x416 pixel resolution; the original images are each divided into SxS grids according to the scale of a feature map by up-sampling and feature fusion methods similar to Feature Pyramid Networks (FPN); an Intersection over Union (IoU) of a candidate bounding box and a ground truth bounding box is taken as an evaluation standard, and all crack target annotation boxes in an image training set are subjected to K-means clustering analysis to obtain the size of the candidate bounding box; and the YOLOv3 model predicts a probability that each bounding box contains targets through logistic regression. Further, the method for identifying concrete cracks based on a YOLOv3 deep learning model specifically includes: step 1: obtaining crack target images with a crack width of about 0.05 mm or more; cutting the images into a uniform size, and manually adding rectangular labels to targets in the images as ground truth bounding boxes to form a database; step 2: importing the target images into the YOLOv3 model, and automatically compressing the images to a 416x416 pixel resolution, and dividing the original images into SxS grids according to a feature map by up-sampling and feature fusion methods similar to FPN; step 3: determining the number and widths and heights of anchors, learning features from data of crack target annotation boxes to find statistical rules, then using a K-means clustering algorithm for clustering analysis of the target boxes, and taking a K value as the number of the anchors; taking an IoU of a candidate bounding box and a ground truth bounding box as an evaluation standard, and subjecting all crack target annotation boxes in an image training set to K-means clustering analysis to obtain the size of the candidate bounding box; step 4: using three anchor boxes, i.e., prior boxes, for each grid to predict a target boundary, where bounding box information is an offset (t, ty) of a center position of an object relative to an upper left corner of the grid where the point is located, and a width t, and a height th; if a targetcenter has an offset (cx, cy) in a cell relative to an upper left corner of the image, and the candidate bounding box has a width p, and a height ph; step 5: predicting classes of the targets, and predicting, by the YOLOv3 model, a probability that each bounding box contains targets through logistic regression, where if the predicted bounding box overlaps with the ground truth bounding box in a large area and has a higher coincidence rate than other predicted bounding boxes, the probability is 1; if the coincidence rate does not reach 0.5 of a YOLOv3 threshold or is between the threshold and the maximum, the predicted bounding box will be ignored and will not affect a loss function; and step 6: taking the predicted results: a high-dimensional feature vector, a target ground truth bounding box and a class label as inputs, setting a learning rate, a learning rate decay rate, a network momentum parameter, a weight decay regular term, a batch size, batch size divisions and the threshold, and training a regression model for fine regression of the target bounding box and the class. Further, a distance function of K-means clustering in step 3 is: d(B,C)=1-Riou(B,C); where B is the size of a rectangular box, C is the center of the rectangular box, and Rou (B,C) represents an IoU of a candidate bounding box and a ground truth bounding box. Further, in step 4, the bounding box information is an offset (tx, t,) of a center position of a target relative to an upper left corner of the grid where the point is located, and a width t, and a height th; and if a target center has an offset (cx, cy) in a cell relative to an upper left corner of the image, and the candidate bounding box has a widthp, and a heightph, then the corrected bounding boxes are: bx=a(tx)+cx b,=a(ty)+cy bw=pwetw bh=pheth

Further, in step 5, the classes of the targets are predicted, and the feature size obtained by each prediction task is: Sx Sx [3 x (4+1+B)]; where S is the grid size, 3 is the number of prior boxes of each grid, 4 is the bounding box size and coordinates, 1 is the confidence level, and B is the number of classes. Another objective of the present invention is to provide a concrete structure damage detection and control system applying the method for identifying concrete cracks based on a YOLOv3 deep learning model. In summary, the present invention has the advantages that the present invention takes a Darknet 53 network in a YOLOv3 model as a feature extractor and treats object detection as a regression problem, and target area positioning and target class prediction are directly performed based on a single end-to-end neural network model; and the deeper network is used for multi-scale prediction. This simplifies the complexity of network training, reduces the computing cost, achieves a higher accuracy rate while detecting targets quickly, and is more suitable for an engineering application environment. Deep learning-based object detection generally involves two-stage and one-stage models. When the two-stage model is used to detect targets, a series of sample candidate bounding boxes need to be generated, and then sample classification is performed by a convolution neural network, which has high computing costs and low efficiency. The YOLOv3 model converts the problem of locating the target frame into the regression problem for processing by using the one-stage model without generating candidate bounding boxes, and the algorithm has a quick speed. By a comparison with other three representative object detection models Faster-rcnn, SSD and CNN+SD, it is found that the present invention has obvious advantages in detection speed and detection effect. A detection speed comparison diagram is shown in FIG. 8 and a detection effect comparison diagram is shown in FIG. 9. It is proven that the present invention is more suitable for the field of crack identification than the existing deep learning model. At the same time, the actually detected objects in the present invention include structural cracks and non-structural cracks, as shown in FIGs. 11 and 13. The structural cracks are generally caused by structural cracking resulting from the structural stress reaching the limit, and are the beginning of structural damage, while the non-structural cracks are generally formed due to the influence of various factors such as concrete materials, pouring, curing and use environment. The appearance form of the non-structural cracks is quite different from that of the structural cracks. The present invention has better detection effect on the two kinds of cracks, and solves the problem that the general models only detect structural cracks occurring in a laboratory. Compared with IPTs, the YOLOv3 model of the present invention directly performs target area positioning and target class prediction through the single neural network model, effectively reduces the interference of the external environment and improves the accuracy rate. BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a flowchart of a method for identifying concrete cracks based on a YOLOv3 deep learning model according to an example of the present invention; FIG. 2 is an implementation flowchart of a method for identifying concrete cracks based on a YOLOv3 deep learning model according to an example of the present invention; FIG. 3 is a schematic diagram of a variation curve between an average IoU and the number of candidate bounding boxes according to an example of the present invention, where when the number of the candidate bounding boxes increases, the average IoU increases, but the growth speed gradually decreases; FIG. 4 is a schematic diagram of candidate bounding box positions in an example of the present invention; FIG. 5 is a schematic diagram of predicted bounding boxes of a YOLOv3 model in 13x13 cells according to an example of the present invention; FIG. 6 is a schematic diagram of a learning rate change curve according to an example of the present invention; FIG. 7 is a schematic diagram of a PR change curve according to an example of the present invention; FIG. 8 is a diagram of a comparison between detection speeds of the present invention and other models according to an example of the present invention; FIG. 9 is a diagram of a comparison between detection effects of the present invention and other models according to an example of the present invention; FIG. 10 is a schematic diagram of a span size ofJiangling Bridge in Wujiang District, Suzhou City, provided by an example of the present invention; FIG. 11 shows an image detection result of structural cracks ofJiangling Bridge according to an example of the present invention; FIG. 12 is a schematic diagram of the acquisition of images by an unmanned aerial vehicle in an example of the present invention; FIG. 13 shows an image detection result of non-structural cracks ofJiangling Bridge according to an example of the present invention; and FIG. 14 is a principle flowchart of a method for identifying concrete cracks based on a YOLOv3 deep learning model according to an example of the present invention. DETAILED DESCRIPTION In order to make the objectives, technical solutions and advantages of the present invention clearer and more comprehensible, the present invention is further described in detail below with reference to the examples. It should be understood that the specific examples described herein are merely illustrative of the present invention and are not intended to limit the present invention. In view of the problems that at present, a commonly used approximately exhaustive type deep learning method needs a large amount of data, lacks pertinence and has a high computing cost and low detection efficiency, the present invention provides a crack detection technology with strong robustness, good generalization capability and higher detection efficiency and accuracy rate, which has originality and is more suitable for the engineering field. The application principle of the present invention is described in detail below with reference to the accompanying drawings. As shown in FIG. 1, a method for identifying concrete cracks based on a YOLOv3 deep learning model according to an example of the present invention includes the following steps. S101: Obtain crack target images with a crack width of about 0.05 mm or more; cut the images into a uniform size, and manually add rectangular labels to targets in the images as ground truth bounding boxes to form a database. S102: Import the target images into the YOLOv3 model, and automatically compress the images to a 416x416 pixel resolution, divide the original images into SxS grids according to a feature map by up-sampling and feature fusion methods similar to FPN. S103: Determine the number and widths and heights of anchors, learn features from data of crack target annotation boxes to find statistical rules, then use a K-means clustering algorithm for clustering analysis of the target boxes, and finally take a K value as the number of the anchors; take an IoU (represented by Riou) of a candidate bounding box and a ground truth bounding box as an evaluation standard, and subject all crack target annotation boxes in an image training set to K-means clustering analysis to obtain the size of the candidate bounding box. S104: Use three anchor boxes, i.e., prior boxes, for each grid to predict a target boundary, where bounding box information is an offset (t, ty) of a center position of an object relative to an upper left comer of the grid where the point is located, and a width t, and a height th; if a target center has an offset (c, cy) in a cell relative to an upper left corner of the image, and the candidate bounding box has a width p, and a height ph; S105: Predict classes of the targets, and predict, by the YOLOv3 model, a probability that each bounding box contains targets through logistic regression, where if the predicted bounding box overlaps with the ground truth bounding box in a large area and has a higher coincidence rate than other predicted bounding boxes, the probability is 1; if the coincidence rate does not reach 0.5 of a YOLOv3 threshold or is between the threshold and the maximum, the predicted bounding box will be ignored and will not affect a loss function. S106: Take the predicted results: a high-dimensional feature vector, a target ground truth bounding box and a class label as inputs, set a learning rate, a learning rate decay rate, a network momentum parameter, a weight decay regular term, a batch size, batch size divisions, the threshold and other parameters, and train a regression model for fine regression of the target bounding box and the class. The application principle of the present invention is further described below with reference to the accompanying drawings. As shown in FIG. 2, a method for identifying concrete cracks based on a YOLOv3 deep learning model according to an example of the present invention includes the following steps. (1) A database was obtained. Based on human eye accuracy, crack target images with a crack width of about 0.05 mm or more were obtained. In order to ensure better robustness and stability of the trained model, the development directions and definitions of targets were different. The structural background surfaces on which the targets were located were also disturbed by stains, template traces or water stains, etc. In this example, 242 images were obtained, and the original images were uniformly cut into smaller images with a resolution of 1000x1000 pixels; the non-concrete background parts were removed, and cracks in the images were manually annotated by LabellImg and used as a total database, totaling 3821 images. A training set, a verification set and a test set were randomly generated in the database.

(2) The original images were segmented. The crack images were imported into YOLOv3 and automatically compressed to a 416x416 pixel resolution. The original images were each divided into SxS grids according to the sizes of a feature map (13x13, 26x26, 52x52) by up-sampling and feature fusion methods similar to FPN. (3) The number and widths and heights of anchors were determined. Crack ground truth bounding boxes were subjected to clustering analysis by using a K-means algorithm. The number of anchors and the width-height dimension of the YOLOv3 model were obtained by clustering VOC20 data sets, there was a big difference between the target shapes in the data sets, and there was a greater difference from concrete crack targets. Therefore, it is necessary to perform clustering analysis on the crack ground truth bounding boxes to re-determine the number and widths and heights of anchors. An IoU (represented by Riou) of a candidate bounding box and a ground truth bounding box was taken as an evaluation standard, and K-means clustering analysis was performed on all crack target annotation boxes in an image training set to obtain the size of the candidate bounding box. A distance function of K-means clustering is: d(B,C)=1-Riou (B,C); where B is the size of a rectangular box, C is the center of the rectangular box, and Riou (B,C) represents an IoU of a candidate bounding box and a ground truth bounding box. A variation curve between an average IoU and the number of candidate bounding boxes is shown in FIG. 3. It can be seen that as the number of candidate bounding boxes increases, the average IoU increases more and more slowly. In this example, after the relationship between the average IoU and the number of candidate bounding boxes was balanced, nine candidate bounding boxes were taken, namely (33,46), (122,43), (50,148), (73,362), (144,155), (350,74), (153,381), (364,75), (322,365). Every three anchor corresponded to three scales, as shown in FIG. 4, where small dots were annotation boxes of crack data sets, large dots represented clustering candidate bounding boxes of crack data sets, and triangular dots were candidate bounding boxes of VOC20 data sets of original YOLOv3. (4) Bounding boxes were predicted. For each grid, three anchor boxes, i.e. prior boxes, were used to predict the target boundary. A prediction example is shown in FIG. 5. The boundary boxes were predicted in the 13x13 cells, where gray boxes were prior boxes obtained by clustering, white boxes are ground truth bounding boxes, and dark gray boxes were grids where center points of objects were located. Bounding box information was an offset (t, t,) of a center position of an object relative to an upper left corner of the grid where the point was located, and a width tw and a height th. If a target center had an offset (c, cy) in a cell relative to an upper left corner of the image, and the candidate bounding box had a width p, and a height ph, then the corrected bounding boxes were: bx=a-(tx)+cx b,=oa(ty)+cy bw=pwe'" bh=pheth (5) The classes of the targets were predicted, and the feature size obtained by each prediction task was: Sx Sx [3 x (4+1+B)]; where S is the grid size, 3 is the number of prior boxes of each grid, 4 is the bounding box size and coordinates, 1 is the confidence level, and B is the number of classes. The YOLOv3 model predicted a probability that each bounding box contains targets through logistic regression. If the predicted bounding box overlapped with the ground truth bounding box in a large area and had a higher coincidence rate than other predicted bounding boxes, the probability was 1; if the coincidence rate did not reach 0.5 of a YOLOv3 threshold or was greater than the threshold and was not the maximum, the predicted bounding box would be ignored and would not affect a loss function. In the training process, the YOLOv3 model assigned only one prediction box to each object. (6) A model was trained. The predicted results: a high-dimensional feature vector, a target ground truth bounding box and a class label were taken as inputs. A learning rate was set to 0.001; a learning rate decay rate was 10 times when iterations reached 40000 times, and was 100 times when iterations reached 45000 times, and a change curve is shown in FIG. 6; a network momentum parameter was 0.9; a weight decay regular term was 0.005; a batch size was 64, there were 16 batch size divisions; the threshold was 0.5, and there were 50200 iterations, and a regression model for fine regression of the target bounding box and the class was trained. After the training was completed, the identification ability of the YOLOv3 model was verified with the samples of the test set, the AP value reached 89.4, and a PR curve is shown in FIG. 7. By a comparison with other three representative object detection models Faster-rcnn, SSD, CNN+SD, the present invention has obvious advantages in detection speed and detection effect. A detection speed comparison diagram is shown in FIG. 8 and a detection effect comparison diagram is shown in FIG. 9. In view of the problem of the example, the present invention has more prominent detection speed and detection effect, and the detection speed can meet the requirement for real-time detection, which proves the advantage of the present invention in the field of concrete structure damage detection. The method was applied to crack detection and identification ofJiangling Bridge in Wujiang District, Suzhou City. The bridge is a three-span prestressed concrete rigid frame bridge with a continuous box girder, and has a span of 50 m+86 m+50 m = 186 m, as shown in FIG. 10. Cracks appeared after bridge operation for 15 years. Crack images of a main span baseplate above the canal were acquired by an unmanned aerial vehicle and identified. The results show that the concrete crack detection model based on a YOLOv3 deep learning model can quickly and accurately identify multiple targets such as structural cracks and non-structural cracks. The identification speed is much faster than those of other models, and even meets the requirements of real-time detection of concrete crack targets. The object detection model has much better concrete crack identification accuracy than other models, and has stronger robustness and generalization capability. The above are only the preferred examples of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements and improvements and the like made within the spirit and principles of the present invention should fall within the protection scope of the present invention.

Claims

What is claimed is: 1. A method for identifying concrete cracks based on a YOLOv3 deep learning model, wherein the method takes a Darknet-53 network in the YOLOv3 model as a feature extractor and treats object detection as a regression problem, and target area positioning and target class prediction are directly performed based on a single end-to-end neural network model; crack images are imported into the YOLOv3 model and automatically compressed to a 416x416 pixel resolution; the original images are each divided into SxS grids according to the scale of a feature map by up-sampling and feature fusion methods similar to FPN; an Intersection over Union (IoU) of a candidate bounding box and a ground truth bounding box is taken as an evaluation standard, and all crack target annotation boxes in an image training set are subjected to K-means clustering analysis to obtain the size of the candidate bounding box; and the YOLOv3 model predicts a probability that each bounding box contains targets through logistic regression.
2. The method for identifying concrete cracks based on a YOLOv3 deep learning model according to claim 1, specifically comprising: step 1: obtaining crack target images with a crack width of about 0.05 mm or more; cutting the images into a uniform size, and manually adding rectangular labels to targets in the images as ground truth bounding boxes to form a database; step 2: importing the target images into the YOLOv3 model, and automatically compressing the images to a 416x416 pixel resolution; dividing the original images into SxS grids according to a feature map by up-sampling and feature fusion methods similar to FPN; step 3: determining the number and widths and heights of anchors, learning features from data of crack target annotation boxes to find statistical rules, then using a K-means clustering algorithm for clustering analysis of the target boxes, and taking a K value as the number of the anchors; taking an IoU of a candidate bounding box and a ground truth bounding box as an evaluation standard, and subjecting all crack target annotation boxes in an image training set to K-means clustering analysis to obtain the size of the candidate bounding box; step 4: using three anchor boxes, i.e., prior boxes, for each grid to predict a target boundary, wherein bounding box information is an offset (t, ty) of a center position of an object relative to an upper left corner of the grid where the point is located, and a width t, and a height th; if a targetcenter has an offset (c, cy) in a cell relative to an upper left corner of the image, and the candidate bounding box has a width p, and a height ph; step 5: predicting classes of the targets, and predicting, by the YOLOv3 model, a probability that each bounding box contains targets through logistic regression, wherein if the predicted bounding box overlaps with the ground truth bounding box in a large area and has a higher coincidence rate than other predicted bounding boxes, the probability is 1; if the coincidence rate does not reach 0.5 of a YOLOv3 threshold or is between the threshold and the maximum, the predicted bounding box will be ignored and will not affect a loss function; and step 6: taking the predicted results: a high-dimensional feature vector, a target ground truth bounding box and a class label as inputs, setting a learning rate, a learning rate decay rate, a network momentum parameter, a weight decay regular term, a batch size, batch size divisions and the threshold, and training a regression model for fine regression of the target bounding box and the class.
3. The method for identifying concrete cracks based on a YOLOv3 deep learning model according to claim 2, wherein a distance function of K-means clustering in step 3 is: d(B,C)=1-Rou (B,C) wherein B is the size of a rectangular box, C is the center of the rectangular box, and Rou (B,C) represents an IoU of a candidate bounding box and a ground truth bounding box.
4. The method for identifying concrete cracks based on a YOLOv3 deep learning model according to claim 2, wherein the bounding box information is an offset (tx, ty) of a center position of a target relative to an upper left corner of the grid where the point is located, and a width t" and a height th; and if a target center has an offset (cx, cy) in a cell relative to an upper left corner of the image, and the candidate bounding box has a width p, and a heightph, then the corrected bounding boxes are: b= otx)+cx b,= a(ty)+cy bw=pwe'" bh-pheth wherein in step 5, the classes of the targets are predicted, and the feature size obtained by each prediction task is: Sx Sx [3 x (4+1+B)]; wherein S is the grid size, 3 is the number of prior boxes of each grid, 4 is the bounding box size and coordinates, 1 is the confidence level, and B is the number of classes.
5. A concrete structure damage detection and control system applying the method for identifying concrete cracks based on a YOLOv3 deep learning model according to any one of claims 1 to 4.