Pavement crack rapid identification method based on high-speed digital image acquisition

Abstract

The invention provides a pavement crack rapid identification method based on high-speed digital image acquisition. The method comprises the following steps: S1, obtaining an original pavement picture; s2, dividing the original pictures into a training set and a test set; s3, carrying out manual labeling on the training set pictures in the S2, and carrying out digital processing to generate an XML (Extensible Markup Language) file; s4, taking the XML file obtained in the S3 as an input sample to train a YOLO algorithm to obtain algorithm parameters; s5, converting the pictures of the test set in the S2 into binary digital pictures, and inputting the binary digital pictures into the trained YOLO algorithm to perform crack detection on the pictures of the test set; and S6, performing automatic analysis on crack information detected in S5. According to the method, by training the YOLO algorithm, the length, width, position, type and area information of the pavement crack can be quickly and accurately obtained only by inputting the pavement picture, so that the problems of time and labor waste, strong subjectivity and low efficiency of the current pavement crack manual detection are solved.

Description

Technical Field

[0001] The invention relates to the technical field of image recognition, and in particular to a method for quickly identifying road cracks based on high-speed digital image acquisition. Background Art

[0002] With the rapid development of the expressway industry, the focus of expressway work will gradually shift from the construction period to the maintenance period, and the maintenance work is arduous and long. In March 2016, the General Office of the Ministry of Transport issued the "Transportation Science and Technology Development Plan", which emphasized that major scientific and technological research and development should be carried out in the field of infrastructure, focusing on highway construction and disease treatment under complex natural conditions, including rapid non-destructive testing technology and equipment for the pavement conditions of in-service highways. In road maintenance, one of the most important indicators of road quality is pavement cracks. At present, the means of detecting highway conditions in China mainly rely on manual measurement methods, which can no longer meet the detection requirements of highways, and there are disadvantages such as large manpower requirements, high time cost, limited accuracy of detection data, serious interference with highway vehicle traffic, high risk of detection work, and high detection cost.

[0003] With the rapid development of computer image processing technology and computer hardware equipment, the development of pavement crack recognition systems based on image processing technology has made great progress. At present, the methods of pavement crack detection systems at home and abroad mainly include: pavement crack detection algorithm based on traditional digital image processing; pavement image crack detection based on texture analysis; pavement image crack detection based on fuzzy sets; pavement image crack detection based on neural network algorithm; pavement image crack detection based on deep learning algorithm. The advantages of these existing methods are mainly: first, compared with traditional manual detection, the inspection efficiency is greatly improved; second, mobile vehicles are used to obtain image information, which is convenient and practical; third, image processing technology is introduced to process the image, making the target easier to identify; fourth, lighting equipment is partially used to increase the clarity of the image; fifth, handheld detection equipment is easy to carry, accurate information collection, low cost, easy to operate and strong applicability.

[0004] However, the current road surface detection systems at home and abroad also have some serious defects: First, the cost of vehicle-mounted equipment is high and the maintenance cost is high. Second, the image processing subsystems of these systems almost all use traditional image recognition algorithms (Hand-engineered) such as HOG, Gabor filters, Haar, etc., that is, they all use the method of manually extracting features for image recognition. There is currently no crack detection system that uses deep learning to automatically extract features and quickly process image information; third, only some systems have vehicle display systems, and there is no crack detection system with the function of displaying mobile terminal recognition results. Fourth, handheld detection equipment is time-consuming and labor-intensive, inefficient, and the inspection results are highly subjective and not objective. There is no unified standard for the inspection results, and the risk factor is high.

[0005] At present, most crack detection in China uses training convolutional neural network algorithms to process crack data. Training convolutional neural network algorithms are widely used in computer vision, natural language processing and other fields due to their automatic feature extraction characteristics. However, training convolutional neural network algorithms have the following disadvantages: ① When the network layer is too deep, the use of BP propagation to modify parameters will slow down the change of parameters close to the input layer; ② The use of gradient descent algorithms can easily make the training results converge to the local minimum rather than the global minimum; ③ The pooling layer will lose a lot of valuable information and ignore the correlation between the local and the whole. The training convolutional neural network algorithm generates a large number of possible bounding boxes that may contain the object to be detected through "region pre-selection", and then uses a classifier to determine whether each bounding box contains an object and the possibility of the category to which the object belongs. The detection speed is not fast enough and the detection results are not accurate enough.

[0006] Therefore, it is of great significance to study a method for rapid identification of pavement cracks based on high-speed digital image acquisition technology. Summary of the invention

[0007] The purpose of the present invention is to overcome the above technical shortcomings and deficiencies, and to provide a method for rapid identification of pavement cracks based on high-speed digital image acquisition. This method can quickly and accurately obtain information such as the length, width, location, type and area of ​​pavement cracks by training the YOLO (You Only Look Once) algorithm, and only needs to input a pavement picture, and is committed to studying and solving the current problems of manual detection of pavement cracks being time-consuming, labor-intensive, highly subjective, and inefficient. The technical solutions adopted by the present invention are as follows:

[0008] A method for quickly identifying pavement cracks based on high-speed digital image acquisition of the present invention comprises the following steps:

[0009] S1. High-speed image acquisition: high-speed digital image acquisition equipment is used to obtain high-resolution original road images in real time at different locations, different lighting conditions and different weather conditions. The image acquisition rate is at least 30 frames per second, and dynamic exposure adjustment technology is used to optimize the imaging effect;

[0010] S2. Dataset division: The original images obtained in S1 are divided into a training set and a test set according to a predetermined ratio; preliminary image enhancement processing including noise removal and resolution improvement is performed on the test set images to optimize the image quality;

[0011] S3. Manual labeling: The training set in S2 was manually labeled, and the crack locations were labeled using the LabelImg software developed by Tzutalin. Automated auxiliary tools were used in the labeling process to improve the labeling accuracy and generate an XML file containing the labeled objects and the bounding box information of each image.

[0012] S4. Improved YOLO algorithm training: using the training set XML file generated in S3 as an input sample, training the improved YOLO algorithm, wherein the YOLO algorithm includes specific network structure optimization and loss function adjustment to improve the accuracy and speed of crack identification;

[0013] S5. Crack detection and quantification: Crack detection is performed on the test set images in S2. First, the images are sampled, quantized, and encoded to be converted into binary digital images. Then, the images are input into the YOLO algorithm trained in S4 to output the length, width, and location information of the cracks.

[0014] S6. Crack type identification and analysis: Based on the crack information detected in S5, the crack type is automatically analyzed. By combining historical data and image patterns, different crack types including transverse cracks, longitudinal cracks and fissures can be accurately distinguished, and a crack report can be generated.

[0015] Preferably, the division of the training set and the test set in step S2 is combined with an adaptive data enhancement strategy to ensure the diversity and completeness of the training set and the test set; in the data division process, not only the type of road cracks is considered, but also the weight distribution of data collected under different lighting and weather conditions is added; the training set is labeled by combining automated assisted calibration and manual review, and a local crack morphology tracking algorithm is used in the labeling process to improve the accuracy and efficiency of labeling; the test set is used to evaluate the generalization ability of the model, and the algorithm performance is comprehensively verified through a multi-index evaluation method including crack detection accuracy, recall rate, and false alarm rate.

[0016] Preferably, in the annotation work of step S3, a dynamic adaptive bounding box generation algorithm is adopted, which automatically adjusts the shape and size of the bounding box according to the morphological characteristics of the crack, so that the annotation box can more accurately wrap the crack; the XML file generated by the annotation contains not only the bounding box information, but also the description of the crack morphological characteristics including width, direction, and curvature, so as to enhance the recognition ability of the detection model; the data set format is based on the VOC2007 data set format, but is expanded for the crack detection task, and a label design for complex crack features is added, such as multi-level crack type classification labels and detailed morphological description labels.

[0017] Preferably, in step S3, an adaptive bounding box optimization algorithm is used when annotating data. The algorithm automatically adjusts the size of the bounding box according to the complexity, morphology and environmental background of the crack to ensure that the bounding box can accurately wrap the crack. At the same time, a bounding box overlap optimization strategy is adopted in the annotation process, and a multi-level overlap control algorithm is used to dynamically adjust the overlap ratio between adjacent crack annotation boxes to reduce annotation errors and improve the training accuracy of the detection model.

[0018] Preferably, in step S4, the training algorithm adopts an optimization strategy based on adaptive data enhancement, which automatically performs adaptive data enhancement on the training images after placing the manually calibrated training set images and XML documents in the JPEGImages folder and Annotation folder under the VOC2007 folder of the working file, including random rotation, scaling, contrast adjustment and crack edge detail enhancement processing, so as to improve the generalization ability of the model for different crack morphologies; in addition, the annotated images are not only converted into videos, but also the parameter training speed of the YOLO algorithm is further accelerated through the inter-frame difference optimization technology.

[0019] Preferably, in step S5, the length of the linear crack is no longer simply characterized based on the diagonal distance, but the actual length of the crack is calculated by the skeleton line of the crack; the skeleton line extracts the central axis of the crack through an image processing algorithm, and uses the vertical distance from each crack pixel to the skeleton line as a reference to accurately measure the nonlinear shape of the crack; the lengths of the transverse crack, the longitudinal crack and the oblique crack are calculated by the cumulative distance of the skeleton line points to avoid errors;

[0020] l l =u*(x1-x2),l w =u*(y1-y2),

[0021] Among them, l l , l w , l d They represent the lengths of horizontal, vertical and oblique cracks respectively, and u represents the actual length of each pixel.

[0022] Preferably, in step S6, the determination of the type of crack depends not only on the aspect ratio of the crack, but also on the morphological characteristics of the crack skeleton, including: A. If the ratio of the length to the width of the crack skeleton line is greater than 2:1, the crack is a transverse crack; if the ratio of the length to the width of the skeleton line is less than 1:2, the crack is a longitudinal crack; cracks between the two are oblique cracks and reticular cracks; B. On this basis, the ratio of the gelatinized area of ​​the crack to the area of ​​the outer region of the crack skeleton line is combined for judgment. If the gelatinized area accounts for more than 1 / 2, the crack is a reticular crack, otherwise it is an oblique crack.

[0023] Preferably, the area calculation method of the mesh cracks in step S6 is a comprehensive calculation through the crack skeleton line and the crack outer frame information output by the YOLO algorithm; the specific method is: the crack boundary box coordinates identified by the YOLO algorithm are combined with the key point coordinates on the crack skeleton line to form a polygon representing the actual area of ​​the crack, and then the real area of ​​the mesh cracks is calculated by the ratio of the image pixel to the actual image size; the position, size and other information are output in the YOLO algorithm, and the coordinates of the four corners are (x1, y1), (x1, y2), (x2, y2), (x2, y1), respectively, and the length of the rectangular box is x1-x2, and the width of the rectangular box is y1-y2; the area calculation formula of the mesh cracks is as follows:

[0024] s=u 2 *(x1-x2)*(y1-y2)

[0025] Among them, u represents the actual length of each pixel, u 2 Represents the actual area of ​​each pixel.

[0026] The advantages and technical effects of the present invention are:

[0027] The present invention proposes a method for rapid identification of pavement cracks based on high-speed digital image acquisition. The YOLO algorithm used in the method is an object identification and positioning algorithm based on a deep neural network. The two stages of candidate area and object identification are creatively merged, the network structure is adjusted, and multi-scale features are used for object detection. Only one calculation is required to simultaneously give the position and category probability of the bounding box, and the performance is excellent in object identification and positioning. The present invention adjusts the network structure and uses multi-scale features for object detection, and the accuracy is significantly improved. The advantages are: 1. It can detect quickly, improving the current situation of time-consuming and labor-intensive manual detection of pavement cracks; 2. It can detect cracks objectively and accurately, which can well avoid the problem of excessive subjectivity of manual detection and low accuracy of other algorithms; 3. It has strong generalization ability and can be applied to pavement crack detection of roads of all levels. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a flow chart of the analysis method of the present invention;

[0029] Figure 2 This is the artificial annotation effect diagram of the present invention;

[0030] Figure 3 This is a schematic diagram of YOLO input and output of the present invention;

[0031] Figure 4 This is a diagram showing the calculation effect of the horizontal, vertical and oblique crack lengths of the present invention;

[0032] Figure 5 This is a diagram showing the effect of calculating the mesh crack area of ​​the present invention. DETAILED DESCRIPTION

[0033] The embodiments of the present invention are described in detail below, wherein the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions. The following embodiments are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0034] The present invention is a method for rapid identification of pavement cracks based on high-speed digital image acquisition. By training the YOLO algorithm, only the pavement picture needs to be input to quickly and accurately obtain information such as the length, width, location, type and area of ​​pavement cracks. It is committed to studying and solving the current problems of manual detection of pavement cracks, which is time-consuming, labor-intensive, highly subjective and inefficient. Figure 1 , including the following steps:

[0035] S1. High-speed image acquisition: high-speed digital image acquisition equipment is used to obtain high-resolution original road images in real time at different locations, different lighting conditions and different weather conditions. The image acquisition rate is at least 30 frames per second, and dynamic exposure adjustment technology is used to optimize the imaging effect;

[0036] S2. Dataset division: The original images obtained in S1 are divided into a training set and a test set according to a predetermined ratio; preliminary image enhancement processing including noise removal and resolution improvement is performed on the test set images to optimize the image quality;

[0037] S3. Manual labeling: The training set in S2 was manually labeled, and the crack locations were labeled using the LabelImg software developed by Tzutalin. Automated auxiliary tools were used in the labeling process to improve the labeling accuracy and generate an XML file containing the labeled objects and the bounding box information of each image.

[0038] S4. Improved YOLO algorithm training: using the training set XML file generated in S3 as an input sample, training the improved YOLO algorithm, wherein the YOLO algorithm includes specific network structure optimization and loss function adjustment to improve the accuracy and speed of crack identification;

[0039] S5. Crack detection and quantification: Crack detection is performed on the test set images in S2. First, the images are sampled, quantized, and encoded to be converted into binary digital images. Then, the images are input into the YOLO algorithm trained in S4 to output the length, width, and location information of the cracks.

[0040] S6. Crack type identification and analysis: Based on the crack information detected in S5, the crack type is automatically analyzed. By combining historical data and image patterns, different crack types including transverse cracks, longitudinal cracks and fissures can be accurately distinguished, and a crack report can be generated.

[0041] Preferably, the division of the training set and the test set in step S2 is combined with an adaptive data enhancement strategy to ensure the diversity and completeness of the training set and the test set; in the data division process, not only the type of road cracks is considered, but also the weight distribution of data collected under different lighting and weather conditions is added; the training set is labeled by combining automated assisted calibration and manual review, and a local crack morphology tracking algorithm is used in the labeling process to improve the accuracy and efficiency of labeling; the test set is used to evaluate the generalization ability of the model, and the algorithm performance is comprehensively verified through a multi-index evaluation method including crack detection accuracy, recall rate, and false alarm rate.

[0042] Preferably, in the annotation work of step S3, a dynamic adaptive bounding box generation algorithm is adopted, which automatically adjusts the shape and size of the bounding box according to the morphological characteristics of the crack, so that the annotation box can more accurately wrap the crack; the XML file generated by the annotation contains not only the bounding box information, but also the description of the crack morphological characteristics including width, direction, and curvature, so as to enhance the recognition ability of the detection model; the data set format is based on the VOC2007 data set format, but is expanded for the crack detection task, and a label design for complex crack features is added, such as multi-level crack type classification labels and detailed morphological description labels.

[0043] Preferably, in step S3, an adaptive bounding box optimization algorithm is used when annotating data. The algorithm automatically adjusts the size of the bounding box according to the complexity, morphology and environmental background of the crack to ensure that the bounding box can accurately wrap the crack. At the same time, a bounding box overlap optimization strategy is adopted in the annotation process, and a multi-level overlap control algorithm is used to dynamically adjust the overlap ratio between adjacent crack annotation boxes to reduce annotation errors and improve the training accuracy of the detection model.

[0044] Preferably, in step S4, the training algorithm adopts an optimization strategy based on adaptive data enhancement, which automatically performs adaptive data enhancement on the training images after placing the manually calibrated training set images and XML documents in the JPEGImages folder and Annotation folder under the VOC2007 folder of the working file, including random rotation, scaling, contrast adjustment and crack edge detail enhancement processing, so as to improve the generalization ability of the model for different crack morphologies; in addition, the annotated images are not only converted into videos, but also the parameter training speed of the YOLO algorithm is further accelerated through the inter-frame difference optimization technology.

[0045] Preferably, in step S5, the length of the linear crack is no longer simply characterized based on the diagonal distance, but the actual length of the crack is calculated by the skeleton line of the crack; the skeleton line extracts the central axis of the crack through an image processing algorithm, and uses the vertical distance from each crack pixel to the skeleton line as a reference to accurately measure the nonlinear shape of the crack; the lengths of the transverse crack, the longitudinal crack and the oblique crack are calculated by the cumulative distance of the skeleton line points to avoid errors;

[0046]

[0047] Among them, l l , l w , l d They represent the lengths of horizontal, vertical and oblique cracks respectively, and u represents the actual length of each pixel.

[0048] Preferably, in step S6, the determination of the type of crack depends not only on the aspect ratio of the crack, but also on the morphological characteristics of the crack skeleton, including: A. If the ratio of the length to the width of the crack skeleton line is greater than 2:1, the crack is a transverse crack; if the ratio of the length to the width of the skeleton line is less than 1:2, the crack is a longitudinal crack; cracks between the two are oblique cracks and reticular cracks; B. On this basis, the ratio of the gelatinized area of ​​the crack to the area of ​​the outer region of the crack skeleton line is combined for judgment. If the gelatinized area accounts for more than 1 / 2, the crack is a reticular crack, otherwise it is an oblique crack.

[0049] Preferably, the area calculation method of the mesh cracks in step S6 is a comprehensive calculation through the crack skeleton line and the crack outer frame information output by the YOLO algorithm; the specific method is: the crack boundary box coordinates identified by the YOLO algorithm are combined with the key point coordinates on the crack skeleton line to form a polygon representing the actual area of ​​the crack, and then the real area of ​​the mesh cracks is calculated by the ratio of the image pixel to the actual image size; the position, size and other information are output in the YOLO algorithm, and the coordinates of the four corners are (x1, y1), (x1, y2), (x2, y2), (x2, y1), respectively, and the length of the rectangular box is x1-x2, and the width of the rectangular box is y1-y2; the area calculation formula of the mesh cracks is as follows:

[0050] s=u 2 *(x1-x2)*(y1-y2)

[0051] Among them, u represents the actual length of each pixel, u 2 Represents the actual area of ​​each pixel.

[0052] In order to more clearly describe the specific implementation mode of the present invention, an embodiment is provided below:

[0053] The present invention is dedicated to studying and solving the problems of manual detection of road cracks in my country, which is time-consuming, labor-intensive, subjective, and inefficient. It provides a more accurate scientific basis for the highway maintenance department, so that higher highway performance can be obtained with the lowest maintenance cost. The specific research contents are as follows:

[0054] 1) Research on digital processing of road damage images. Image digitization is the process of converting continuous-tone analog images into digital images after sampling and quantization. Image digitization uses computer graphics and image technology. Only after the image is converted into numerical values ​​can it be recognized by the computer.

[0055] 2) Research on road damage recognition algorithms. Briefly introduce existing image recognition algorithms, such as back propagation algorithm, convolutional neural network algorithm, YOLOv3 algorithm, etc. Compare the advantages and disadvantages of various algorithms and select a suitable algorithm for crack detection.

[0056] 3) Research on road damage detection. First, the original road damage images are divided into training sets and test sets, and the cracks in the training set images are manually calibrated. Then, the calibrated images are used as samples as the input of the damage recognition algorithm, and the algorithm is trained to obtain the parameters of the algorithm. Finally, based on the research on road damage classification, the road damage recognition algorithm with determined parameters is used to detect cracks in the test set images, and the road damage type is identified based on the detection results, and the recognition results are output.

[0057] 4) Construction of road damage identification platform. Based on the Python language development environment Spyder, a road damage identification platform was built. The identification platform interface includes operation options such as "crack photo folder path selection", "YOLO path selection", "target detection", and "result path selection", which can realize functions such as selecting the location of photos to be detected, switching the location to save the detection results, and starting damage detection.

[0058] 5) Test results and analysis: Based on the road disease automatic identification platform, the test results are classified and sorted, and relevant statistical indicators are proposed to achieve comparative analysis of the results.

[0059] In addition, it should be noted that Figure 3 As shown in the figure, the input of the YOLOv3 model is an image, and the output is the model recognition result displayed on the mobile terminal. Without considering the details of the neural network structure, for an input image, YOLOv3 maps it to an output tensor of three scales, representing the probability of various objects existing at various positions in the image.

[0060] For a 416*416 input image, 3 prior boxes are set in each grid of the feature map at each scale, with a total of 13*13*3+26*26*3+52*52*3=10647 predictions. In comparison, YOLOv2 uses 13*13*5=845 predictions. YOLOv3 attempts to predict the number of bounding boxes by more than 10 times, and it is performed at different resolutions, so the average precision has a certain improvement in the recognition ability of small objects.

[0061] Each output of the YOLOv3 model is a vector, including the coordinates of the bounding box, the confidence of the bounding box, and the object classification. The coordinates of the bounding box include the x and y coordinates of the center point of the bounding box and the width and height of the bounding box; the confidence of the bounding box indicates the probability that the bounding box is an object; the object classification is the object to which the bounding box belongs, that is, the detected target.

[0062] Finally, the pavement crack detection results collected and counted using the method of the present invention are as follows:

[0063] Table 1 Statistical results

[0064] Disease type Image length (m / piece) Number of pictures Recognition speed (pictures / second) Proportion (%) Horizontal cracks 2 122000 25 25.83 Vertical cracks 2 122000 25 21.75 Oblique cracks 2 122000 25 20.05 Web-like cracks 2 122000 25 20.93

[0065] Table 2 Crack detection results

[0066]

[0067] Table 3 Recognition rate verification

[0068]

[0069]

[0070] It can be seen from Tables 1-3 that YOLOv3 has a high correct recognition rate for transverse cracks, longitudinal cracks and oblique cracks, and the difference with the manual recognition results is small. The accuracy of the recognition of mesh cracks and no cracks is low. The reason is that some mesh cracks are mistakenly identified as transverse cracks or longitudinal cracks, especially the mesh cracks at the edge of the picture. There is a considerable probability that a mesh crack is divided into two pictures. In addition, there are many interferences on the road, which also affect the accuracy of the algorithm's crack recognition. Despite this, the correct recognition rate of the YOLOv3 algorithm designed by the present invention is still over 94%, which exceeds the recognition rate of most road disease recognition algorithms.

[0071] Finally, all the parts not described in the present invention adopt mature products and mature technical means in the prior art.

[0072] In the description of this specification, reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in the embodiments or examples of the present invention.

[0073] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A method for rapid identification of pavement cracks based on high-speed digital image acquisition, characterized in that: The following steps are involved: S1. High-speed image acquisition: high-speed digital image acquisition equipment is used to obtain high-resolution original road images in real time at different locations, different lighting conditions and different weather conditions. The image acquisition rate is at least 30 frames per second, and dynamic exposure adjustment technology is used to optimize the imaging effect; S2. Dataset division: The original images obtained in S1 are divided into a training set and a test set according to a predetermined ratio; preliminary image enhancement processing including noise removal and resolution improvement is performed on the test set images to optimize the image quality; S3. Manual labeling: The training set in S2 was manually labeled, and the crack locations were labeled using the LabelImg software developed by Tzutalin. Automated auxiliary tools were used in the labeling process to improve the labeling accuracy and generate an XML file containing the labeled objects and the bounding box information of each image. S4. Improved YOLO algorithm training: using the training set XML file generated in S3 as an input sample, training the improved YOLO algorithm, wherein the YOLO algorithm includes specific network structure optimization and loss function adjustment to improve the accuracy and speed of crack identification; S5. Crack detection and quantification: Crack detection is performed on the test set images in S2. First, the images are sampled, quantized, and encoded to be converted into binary digital images. Then, the images are input into the YOLO algorithm trained in S4 to output the length, width, and location information of the cracks. S6. Crack type identification and analysis: Based on the crack information detected in S5, the crack type is automatically analyzed. By combining historical data and image patterns, different crack types including transverse cracks, longitudinal cracks and fissures can be accurately distinguished, and a crack report can be generated.

2. The method for rapid identification of pavement cracks based on high-speed digital image acquisition according to claim 1 is characterized in that: The division of the training set and the test set in step S2 is combined with an adaptive data enhancement strategy to ensure the diversity and completeness of the training set and the test set; in the data division process, not only the type of road cracks is considered, but also the weight distribution of data collected under different lighting and weather conditions is added; the training set is labeled by combining automatic auxiliary calibration and manual review, and a local crack morphology tracking algorithm is used in the labeling process to improve the accuracy and efficiency of labeling; the test set is used to evaluate the generalization ability of the model, and the algorithm performance is comprehensively verified through a multi-index evaluation method including crack detection accuracy, recall rate, and false alarm rate.

3. The method for rapid identification of pavement cracks based on high-speed digital image acquisition according to claim 1 is characterized in that: In the annotation work of step S3, a dynamic adaptive bounding box generation algorithm is adopted. The algorithm automatically adjusts the shape and size of the bounding box according to the morphological characteristics of the crack, so that the annotation box can more accurately wrap the crack; the XML file generated by the annotation contains not only the bounding box information, but also the description of the crack morphological characteristics including width, direction, and curvature, so as to enhance the recognition ability of the detection model; the data set format is based on the VOC2007 data set format, but is targeted at the crack detection task It has been expanded to include label designs for complex crack features, such as multi-level crack type classification labels and detailed morphological description labels.

4. The method for rapid identification of pavement cracks based on high-speed digital image acquisition according to claim 1 is characterized in that: In step S3, an adaptive bounding box optimization algorithm is used when annotating data. The algorithm automatically adjusts the size of the bounding box according to the complexity, morphology and environmental background of the crack to ensure that the bounding box can accurately wrap the crack. At the same time, a bounding box overlap optimization strategy is adopted in the annotation process, and a multi-level overlap control algorithm is used to dynamically adjust the overlap ratio between adjacent crack annotation boxes to reduce annotation errors and improve the training accuracy of the detection model.

5. The method for rapid identification of pavement cracks based on high-speed digital image acquisition according to claim 1 is characterized in that: In step S4, the training algorithm adopts an optimization strategy based on adaptive data enhancement. After placing the manually calibrated training set images and XML documents in the JPEGImages folder and the Annotation folder under the VOC2007 folder of the working file respectively, the strategy automatically performs adaptive data enhancement on the training images, including random rotation, scaling, contrast adjustment and crack edge detail enhancement processing, so as to improve the generalization ability of the model for different crack morphologies; in addition, the annotated images are not only converted into videos, but also the parameter training speed of the YOLO algorithm is further accelerated through the inter-frame difference optimization technology.

6. The method for rapid identification of pavement cracks based on high-speed digital image acquisition according to claim 1 is characterized in that: In step S5, the length of the linear crack is no longer simply characterized by the diagonal distance, but the actual length of the crack is calculated by the skeleton line of the crack; the skeleton line extracts the central axis of the crack through an image processing algorithm, and uses the vertical distance from each crack pixel to the skeleton line as a reference to accurately measure the nonlinear shape of the crack; the lengths of the transverse crack, the longitudinal crack and the oblique crack are calculated by the cumulative distance of the skeleton line points to avoid errors; the l =u*(x1-x2),l w =u*(y1-y2), Among them, l l , l w , l d They represent the lengths of horizontal, vertical and oblique cracks respectively, and u represents the actual length of each pixel.

7. The method for rapid identification of pavement cracks based on high-speed digital image acquisition according to claim 1 is characterized in that: In step S6, the determination of the type of crack not only depends on the length-width ratio of the crack, but also combines the morphological characteristics of the crack skeleton, including: A. If the ratio of the length to the width of the crack skeleton line is greater than 2:1, the crack is a transverse crack; if the ratio of the length to the width of the skeleton line is less than 1:2, the crack is a longitudinal crack; cracks between the two are oblique cracks and reticular cracks; B. On this basis, the ratio of the gelatinized area of ​​the crack to the area of ​​the outer region of the crack skeleton line is combined for judgment. If the gelatinized area accounts for more than 1 / 2, the crack is a reticular crack, otherwise it is an oblique crack.

8. The method for rapid identification of pavement cracks based on high-speed digital image acquisition according to claim 1 is characterized in that: The area calculation method of the mesh cracks in step S6 is to comprehensively calculate the crack skeleton line and the crack outer frame information output by the YOLO algorithm; the specific method is: the crack boundary box coordinates identified by the YOLO algorithm are combined with the key point coordinates on the crack skeleton line to form a polygon representing the actual area of ​​the crack, and then the real area of ​​the mesh cracks is calculated by the ratio of the image pixel to the actual image size; the position, size and other information are output in the YOLO algorithm, and the coordinates of the four corners are (x1, y1), (x1, y2), (x2, y2), (x2, y1), respectively, so the length of the rectangular box is x1-x2, and the width of the rectangular box is y1-y2; the area calculation formula of the mesh cracks is as follows: s=u 2 *(x1-x2)*(y1-y2) Among them, u represents the actual length of each pixel, u 2 Represents the actual area of ​​each pixel.

Images