Concrete apparent disease high-precision detection method based on improved YOLOR

Abstract

The invention discloses a concrete apparent disease high-precision detection method based on improved YOLOR, and relates to the technical field of target detection, in particular to the technical field of concrete disease detection. In order to solve the problems that in the prior art, the multi-disease identification precision is low, coverage is incomplete, the anchor frame size of an original YOLOR model is not matched with an actual object, the performance of a loss function is poor and the like, and the detection effect is influenced, the following scheme is provided: constructing a concrete bridge apparent disease data set, improving a YOLOR-BDIM model and training; the trained model is deployed on a local or cloud server, a detection platform based on web is developed and deployed on the server, a user uploads a to-be-detected image for high-precision disease detection, and finally the trained model is called to automatically mark the position and type of a disease in the to-be-detected image. By adopting the method and the device, the convenience of the detection process and the accuracy of the detection result are improved, so that the safety of detection personnel is guaranteed, and the detection work cost is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of target detection, and in particular to the technical field of concrete disease detection. Background Art

[0002] As an important part of the transportation network, the safety of bridges has a great impact on transportation. Due to the coupling of multiple factors such as severe vehicle overloading, aging of concrete materials, and harsh operating environment, bridges in service will inevitably suffer from cracking, falling of concrete protective layers, water seepage, alkali seepage, exposed reinforcement and other apparent bridge defects, which are a great test for the safety and durability of bridges in service. Bridge accidents are often major accidents. If the problems of the bridges are not discovered in time or maintenance is not carried out in time, it may pose a huge threat to people's lives and the economy. Therefore, in the context of the large number of bridges in my country, strict requirements must be imposed on the inspection and maintenance of bridges to meet the requirements of preventing accidents. Today's R&D personnel have realized the necessity and importance of bridge inspection technology research, and actively promoted the assessment of bridge structure conditions, preventive maintenance, and repair and reinforcement methods and technologies, and carried out research on rapid intelligent assessment technology for bridge bearing capacity. The existing technologies are as follows:

[0003] Traditional manual inspection is highly subjective, inefficient, and difficult to ensure the safety of inspectors, making it difficult to adapt to the increasing number of inspection tasks in my country. Traditional bridge surface disease inspection is mainly based on manual on-site inspections. Not only does it require professional knowledge training for inspectors, but when the bridge is in a complex environment, inspectors need to observe the surface of the component with the naked eye by building scaffolding, climbing bridges, using inspection platforms (bridge inspection vehicles, aerial work vehicles, kayaks, etc.) or remote observation. The inspectors then judge the health of the component based on the observation results and their own professional knowledge and work experience, and finally determine the type and scope of the diseased area and record it. In addition, the types of apparent diseases of concrete bridges are complex, and long-term and large-scale observations are a severe test for the physical strength and judgment of the inspectors.

[0004] The continuous development of computer vision technology in recent years has made intelligent and automated bridge detection in complex environments possible.

[0005] In the field of single disease detection:

[0006] (1) An author proposed a study on bridge crack detection based on Mask R-CNN. The author added a semantic enhancement module SEM to the FPN network structure to enhance the information preservation capability during feature extraction. In addition, the output feature layer was upsampled and superimposed on the same layer to increase the amount of information in the feature map. After training with the data set, the improved network increased the mAP by 6.66%, verifying the effectiveness of the improved method. The author also explored the impact of crack subclassification on the recognition results, defining cracks with more complex distribution as damage, and the results showed that the mAP increased by 5.7%. It can be seen that subclassification has a certain impact on the identification of diseases and can provide more detailed results for maintenance.

[0007] (2) Some authors first processed the image to enhance the crack features, used the Mask homogenization algorithm to subtract the original image from the background image to obtain an image with uniform illumination, and then used the generalized intersection-union algorithm to improve the loss function of YOLOv3 to improve positioning accuracy and achieve better crack identification. This study started with data set processing and reduced environmental interference by preprocessing the photos.

[0008] (3) A real-time crack detection model for drones or other fast capture equipment is proposed. A target detection model based on YOLOv4 is proposed. The accuracy and speed of the original model are improved through loss function optimization, pruning and dataset improvement. On the test set, the mAP is 97.6% and can reach 119FPS, which can basically meet the requirements of real-time crack detection.

[0009] In the field of multi-disease detection:

[0010] (1) Some authors first extracted the texture features, grayscale features and color moment features in the pictures, and then used support vector machines (SVM) to classify the pictures. This study is based on the differences in texture, grayscale and color of the defects. Relying on the above eigenvalue parameters, it can realize the classification of cracks, exposed steel bars, rust and defects, which provides ideas for further research on the classification and feature research of defect pictures.

[0011] (2) Some authors mainly consider how to distinguish healthy areas from diseased areas on the bridge surface, including template marks, water stains, shadows and other situations. The authors mainly classify the bridge surface images by sequentially slicing and using advanced classifiers, and then merge them into risk areas. This method can greatly reduce the area that inspectors need to inspect, allowing inspectors to focus more on risky areas, but the effectiveness of the classifier is difficult to guarantee when different diseases are adjacent or even superimposed. In addition, a hierarchical classification method consisting of three image classifiers is proposed. Three independent deep neural pre-trained networks are fine-tuned based on multi-source datasets, and multiple categories of labels can be assigned to an image. Its average score on the dataset is 85%, which can provide more reasonable auxiliary information to inspectors than single-label classifiers.

[0012] In summary, with the rapid development of industrial level and computer industry, it has become feasible to detect the apparent structure of bridges with the help of computer vision technology. There are currently three major research directions in computer vision technology, namely image classification, target detection and image segmentation. The existing technology mainly focuses on the identification of crack diseases, while other diseases are also important factors affecting the normal operation of bridges, and further research is needed on the simultaneous identification of multiple diseases. However, the existing multi-disease recognition technology not only has poor recognition accuracy, but also has certain limitations. It does not fully cover the types of diseases required by the specifications. It is necessary to increase the types of diseases to better serve the detection work on the real site. In addition, the size of the anchor frame of the original YOLOR model is different from the size of the actual detection object, which affects the detection accuracy and efficiency. The loss function of the original model is not comprehensive enough when considering factors such as the overlapping area of ​​the real frame and the predicted frame, which also has a negative impact on the detection effect. Summary of the invention

[0013] In order to solve the problems in the prior art that affect the detection effect, such as low multi-disease recognition accuracy, incomplete coverage, mismatch between the anchor frame size of the original YOLOR model and the actual object, and poor performance of the loss function, the present invention provides the following solution.

[0014] A high-precision detection method for concrete surface defects based on improved YOLOR, the method comprising the following steps:

[0015] S1. Construct the concrete bridge apparent disease dataset DOBD V1.0;

[0016] S2. Construct an improved YOLOR-BDIM model, wherein the improved YOLOR-BDIM model is optimized through loss function optimization and anchor box optimization;

[0017] S3. Training the improved YOLOR-BDIM model based on the concrete bridge apparent disease dataset DOBD V1.0 to obtain a trained improved YOLOR-BDIM model;

[0018] S4. Deploy the trained improved YOLOR-BDIM model on a local or cloud server;

[0019] S5. Developing a web-based detection platform based on the local or cloud server;

[0020] S6. Upload the image to be tested on the web-based detection platform for high-precision disease detection;

[0021] S7. Automatically annotate the location and type of the disease in the image to be tested by calling the trained improved YOLOR-BDIM model.

[0022] Furthermore, in S1, the concrete bridge apparent disease dataset DOBD V1.0 includes: an initial dataset and a dataset after image augmentation;

[0023] The initial data set includes data on nine types of damage: honeycomb, peeling, holes, whitening, mesh cracks, single cracks, rust expansion, exposed reinforcement and water stains;

[0024] The image-augmented data set is obtained by amplifying the initial data set using an image augmentation technique.

[0025] Furthermore, in S2, the loss function is optimized as follows: SIoU_loss is used to replace the original loss function in the model, and SIoU_loss is defined as follows:

[0026]

[0027] SIoU_loss=1-SIoU

[0028] Among them, IoU is the intersection over union ratio of the predicted box and the real box; Δ is the distance loss; Ω is the shape loss.

[0029] Furthermore, in S2, the anchor frame is optimized as follows: using a K-means clustering algorithm, adopting partial cluster replacement, calculating the area ratio of the clustered anchor frame and the original anchor frame, and replacing the anchor frame whose area ratio differs by less than 1% using the clustering result.

[0030] Further, in S5, the web-based detection platform provides an API interface.

[0031] The present invention also provides a computer device, which includes a memory and a processor, wherein a computer program is stored in the memory, and when the processor runs the computer program stored in the memory, the processor executes the above detection method.

[0032] The present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the above-mentioned detection method is implemented.

[0033] The beneficial effects of the present invention are:

[0034] 1. The present invention can reduce the cost of bridge maintenance. Compared with the traditional concrete bridge surface defect detection, the present invention no longer relies on large-scale detection equipment such as bridge inspection vehicles, while ensuring the personal safety of workers, avoiding the subjectivity of detection personnel in judgment, and can quickly and accurately identify defects, thereby reducing the cost of bridge maintenance.

[0035] 2. The present invention makes up for the shortcoming that the apparent disease detection model for concrete bridges can detect few types of diseases. In combination with regulatory requirements and actual engineering needs, the present invention established a concrete bridge apparent disease dataset (DOBD V1.0) using more than 7,000 photos, which includes 9 types of diseases: honeycombed surface, peeling, voids, whitening, mesh cracks, single cracks, rust, exposed reinforcement and water stains. The YOLOR-BDIM model trained on the DOBD V1.0 dataset can accurately and quickly identify 9 types of diseases, meeting the requirements for the types of diseases specified in the current national standards.

[0036] 3. The present invention realizes visualization of the results based on computer vision technology. The YOLOR deep learning model used in the present invention can realize visualization of the detection results based on computer vision technology. The present invention will automatically mark the detected disease types and locations in the image with rectangular boxes of different colors, making the detection results more intuitive.

[0037] 4. The present invention has a wide range of applications. The web-based detection platform developed based on the trained YOLOR-BDIM model deployed in a local or cloud server can provide a channel for image upload detection for inspectors. In addition, the web-based detection platform developed on the cloud server can also provide an API interface for third-party use and platform call. Compared with traditional concrete bridge surface disease detection, this method greatly facilitates the detection work. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is a flow chart of a high-precision detection method for concrete surface defects based on improved YOLOR;

[0039] Figure 2 9 examples of diseases determined in combination with current domestic and international standards and actual engineering needs;

[0040] Figure 3 Schematic diagram of the calculation parameters of the IoU_loss formula (the red box is the ground truth box, and the green box is the prediction box);

[0041] Figure 4 The test results of the YOLO-BDIM model on the surface damage images of exposed reinforcement and peeling corner concrete bridges;

[0042] Figure 5 The figure shows the test results of the YOLO-BDIM model on the apparent disease image of the concrete bridge with holes;

[0043] Figure 6 The test results of the YOLO-BDIM model on the surface disease images of rust expansion and exposed reinforcement concrete bridges;

[0044] Figure 7 The figure shows the test results of the YOLO-BDIM model on the surface disease image of the white concrete bridge;

[0045] Figure 8 The figure shows the test results of the YOLO-BDIM model on the surface damage image of the water-stained concrete bridge;

[0046] Fig. 9 The figure shows the test results of the YOLO-BDIM model on the surface disease image of the honeycombed concrete bridge. DETAILED DESCRIPTION

[0047] Example: Combination Figure 1-9 This embodiment describes a high-precision detection method for concrete surface defects based on improved YOLOR, the method comprising the following steps:

[0048] S1. Construct the concrete bridge apparent disease dataset DOBD V1.0;

[0049] S2. Construct an improved YOLOR-BDIM model, wherein the improved YOLOR-BDIM model is optimized through loss function optimization and anchor box optimization;

[0050] S3. Training the improved YOLOR-BDIM model based on the concrete bridge apparent disease dataset DOBD V1.0 to obtain a trained improved YOLOR-BDIM model;

[0051] S4. Deploy the trained improved YOLOR-BDIM model on a local or cloud server;

[0052] S5. Developing a web-based detection platform based on the local or cloud server;

[0053] S6. Upload the image to be tested on the web-based detection platform for high-precision disease detection;

[0054] S7. Automatically annotate the location and type of the disease in the image to be tested by calling the trained improved YOLOR-BDIM model.

[0055] Specifically, Figure 1 The flowchart of the prediction method is given. First, a data set containing images of apparent defects on concrete bridges is constructed, and then an improved YOLOR-BDIM model is constructed to improve the defect detection performance. During the training process, the model learns to identify different types of defects from images. The trained model is deployed on the server. Users upload images through the Web platform. The system uses the model to detect defects and automatically annotate them, achieving accurate identification and location annotation of concrete bridge defects. The entire process is trained through deep learning algorithms and combined with Web technology to achieve a user-friendly interactive interface.

[0056] Combination Figure 2 Note that in S1, the concrete bridge apparent disease dataset DOBD V1.0 includes: an initial dataset and a dataset after image augmentation;

[0057] The initial data set includes data on nine types of damage: honeycomb, peeling, holes, whitening, mesh cracks, single cracks, rust expansion, exposed reinforcement and water stains;

[0058] The image-augmented data set is obtained by amplifying the initial data set using an image augmentation technique.

[0059] Specifically: For the currently undisclosed large-scale concrete bridge apparent disease dataset, more than 5,000 bridge disease photos were collected through cooperation with maintenance units and personal outdoor photography. In addition, image augmentation technology was used to amplify the small sample data with a small number (due to the different conditions for the generation of various diseases, the number of 9 diseases is unbalanced. With the help of image augmentation technology, some of the smaller types of diseases are amplified, such as cropping, rotating, scaling and other operations to balance the number of various types of diseases in the dataset). Then, combined with the current domestic and international specifications and actual engineering needs, the concrete bridge apparent disease dataset Date ofbridge diseases Version1.0 (DOBDV1.0) was finally constructed. This dataset contains 9 common apparent diseases of concrete bridges (see Table 1 for details, and see the attached examples for examples). Figure 2), a total of more than 7,000 disease photos, and there are no strict regulations for the shooting of all pictures in the DOBD V1.0 dataset. It includes different shooting conditions (focal length, object distance, shooting equipment, etc.) and different resolutions to ensure the diversity of disease pictures, thereby enhancing the robustness of the detection model.

[0060] Table 1 Classification and definition of apparent defects of concrete bridges

[0061]

[0062] At the same time, in the process of building the data set, it is necessary to annotate the pictures and divide them into validation and training sets. The disease pictures are annotated in the form of annotated boxes according to the 9 disease types that have been divided. After the annotation is completed, it will be checked by professional inspection personnel to ensure the accuracy of the annotation. The annotated pictures are randomly divided into training and validation sets in a ratio of 8:2, and finally there are 5607 training sets and 1402 validation sets.

[0063] Combination Figure 3 Note that in S2, the loss function is optimized as follows: SIoU_loss is used to replace the original loss function in the model. SIoU_loss is defined as follows:

[0064]

[0065] SIoU_loss=1-SIoU

[0066] Among them, IoU is the intersection over union ratio of the predicted box and the real box; Δ is the distance loss; Ω is the shape loss.

[0067] Specifically, in the target detection task, classification and regression of detection boxes are the core issues. The choice of loss function has a certain impact on the performance of the model. An inappropriate loss function will lead to slow convergence and low efficiency.

[0068] IoU_loss is the most widely used detection box loss function. IoU is the ratio of the intersection and union of the Prediction Box and the Ground Truth Box. The smaller the IoU_loss, the larger the IoU, which means the closer the Prediction Box is to the Ground Truth Box, and the more accurate the detection result. The calculation method of IoU_loss can be seen in Figure 3 And formulas (1)~(6):

[0069] Inter_area=[min(x 12 , x 22 )-max(x 11 , x 21 )]×[min(y12 ,y 22 )-max(y 11 ,y 21 )](1)

[0070] W1=x 12 -x 11 , h1=y 12 -y 11 (2)

[0071] W2=x 22 -x 21 , h2=y 22 -y 21 (3)

[0072] Union_area=w1×h1+W2×h2-Inter_area(4)

[0073]

[0074] IoU_loss = 1 - IoU (6)

[0075] Among them, x 11 is the x coordinate of the lower left corner of the real box; y 11 is the y coordinate of the lower left corner of the real box; x 12 is the x coordinate of the upper right corner of the real box; y 12 is the y coordinate of the upper right corner of the real box; x 21 is the x coordinate of the lower left corner of the prediction box; y 21 is the y coordinate of the lower left corner of the prediction box; x 22 is the x coordinate of the upper right corner of the prediction box; y 22 is the y coordinate of the upper right corner of the prediction box; w1 is the width of the real box; w2 is the width of the prediction box; h1 is the height of the real box; h2 is the height of the prediction box; Inter_area is the intersection of the prediction box and the real box; Union_area is the union of the prediction box and the real box.

[0076] However, when the two boxes have no overlapping area, IoU is 0, and there is no gradient backpropagation at this time, so learning and training cannot be performed. To address this problem, YOLOR uses CIoU_loss instead of IoU_loss by default, thus solving the problem that IoU cannot be optimized without overlap. This is because CIoU_loss takes into account the overlapping area, center point distance and aspect ratio of the Prediction Box and the Ground Truth Box, compared to IoU_loss, which can speed up the regression speed of the PredictionBox to a certain extent. CIoU_loss is defined by formulas (7) to (15):

[0077] c w =max(x 12 , x 22 )-min(x 11 , x 21 )(7)

[0078] c h =max(y 12 ,y 22 )-min(y 11 ,y 21 )(8)

[0079]

[0080] CIoU_loss = 1 - CIoU (15)

[0081] Where: c w is the width of the rectangle surrounding the two bounding boxes; c h is the height of the circumscribed rectangle of the two bounding boxes; c is the diagonal length of the circumscribed rectangle of the two bounding boxes; ρ is the Euclidean distance between the two center points; ρ w is the width difference between the real box and the predicted box; ρ h is the height difference between the real box and the predicted box; α is the weight function, and v is used to measure the similarity of aspect ratios.

[0082] However, there is still an obvious problem. During the prediction box regression process, when the aspect ratio of the prediction box and the ground truth box is in linear proportion, the w and h of the prediction box cannot be increased or decreased at the same time, and the regression optimization cannot be continued. To address this problem, EIoU_loss separates the aspect ratio of the prediction box and the ground truth box on the basis of CIoU_loss, and calculates the length and width of the prediction box and the ground truth box respectively to solve the problem of CIoU_loss. SIoU_loss, on the basis of EIoU_loss, further considers the vector angle between the real box and the prediction box, so it is finally proposed to use SIoU_loss. EIoU_loss is defined by formulas (16) to (17).

[0083]

[0084] EIoU_loss = 1 - EIoU (17)

[0085] SIoU_loss involves distance loss, shape loss and angle loss. SIoU_loss is defined using distance loss and shape loss, and angle loss is used in the distance loss formula. It is defined by formulas (18) to (26).

[0086] Angle loss formula:

[0087]

[0088] Among them, α is the smaller of the two angles, and the other angle is named β.

[0089] Distance loss formula:

[0090]

[0091] γ=2-Λ#(22)

[0092]

[0093] Among them, ρ x Represents the horizontal distance between the center point of the predicted box and the center point of the real box; ρ y Indicates the vertical distance between the center point of the predicted box and the center point of the real box; c w is the difference between the predicted box width and the true box width; c h is the difference between the height of the measured frame and the height of the real frame.

[0094] Shape loss formula:

[0095]

[0096] Among them, ω w is the difference measure between the width of the predicted box and the real box; ω h It is a measure of the height difference between the predicted box and the true box.

[0097] Specifically, in S2, the anchor frame is optimized as follows: using a K-means clustering algorithm, adopting partial clustering replacement, calculating the area ratio of the clustered anchor frame and the original anchor frame, and replacing the anchor frame whose area ratio differs by less than 1% using the clustering result.

[0098] In target detection, the bounding box is continuously optimized to make the model detect the target more accurately. In order to improve the performance of the model, the closer the size and aspect ratio of the anchor box to the detected object, the better the performance of the model. YOLOR uses the 12 prior boxes obtained in advance by the clustering algorithm for reference in the calculation of the prediction box. Table 2 shows the anchor box size of the original model and the YOLOR-BDIM model and the area ratio of each anchor box. The last column shows the difference in area ratio. As can be seen from the table below, the YOLOR-BDIM model selected items 3, 6, 8, 9, and 10 for replacement, and selected sizes close to them in the model.

[0099] Table 2 Anchor box replacement

[0100]

[0101] In this embodiment, in S5, the web-based detection platform provides an API interface.

[0102] Specifically, by providing an API interface for third-party use and platform call, it greatly facilitates the detection work.

[0103] Combination Figure 4-Figure 9 The effect of this embodiment is described as follows:

[0104] After building the improved YOLOR-BDIM model, the model deployment environment is based on the Pytorch 1.7.0 deep learning framework, the computer operating system is Ubuntu 20.04LTS, and Python 3.8.11, CUDA 11.0, cuDNN8.4.0 are installed, and the CPU is AMD Ryzen TM 9 5900X processor, GPU is GeForce RTX TM 3080Ti-12G, 32G memory, use Anaconda to install the environment required for model deployment, and use Tensorboard to view the training process in real time during the experiment.

[0105] Then prepare for training, set the number of training rounds to 300 rounds, set the batch size to 4 based on the video memory limit, and then start training. Use Tensorboard to view the training progress in real time. The specific parameters of random data enhancement are: image HSV hue enhancement (score): 0.015; image HSV-saturation enhancement (score): 0.7; image HSV-value enhancement (score): 0.4; image rotation (+ / - degrees): 360.0; image horizontal and vertical translation (+ / - fraction): 0.5; image scale (+ / - gain): 0.25; image shear (+ / - degrees): 10.0; image upside down flip (probability): 0.5; image left and right flip (probability): 0.5. As the number of training rounds increases, the model training is overfitting, no longer learning new target features, and can no longer learn new information from the current dataset.

[0106] The model performance was tested on the test set. The test parameters were set in the same way as in the previous article, and YOLOR-BDIM was tested on the test set. The experimental results are shown in Table 3 and Table 4. Figure 4-Figure 9 .

[0107] Table 3 BDIM experimental results

[0108]

[0109] It can be seen that after Anchor optimization and loss function adjustment, YOLOR-BDIM has improved the detection and recognition effect of apparent defects of concrete bridges on the basis of YOLOR, and has a certain improvement in average detection accuracy. Its mAP value has increased by 1.9%, making it more suitable for the detection of apparent defects of concrete bridges in complex scenarios. At the same time, its detection speed FPS is basically unchanged compared with the original YOLOR model, and it can maintain real-time detection and more accurately identify a variety of apparent defects of concrete bridges. Figure 4-Figure 9 The recognition results of apparent defects of some concrete bridges are displayed. The detection results will mark the type and location of the defects in the picture, which is more intuitive.

[0110] Embodiment 2: The computer device of the present invention may be a device including a processor and a memory, such as a single chip microcomputer including a central processing unit, etc. Moreover, the processor is used to implement the steps of the above detection method when executing a computer program stored in the memory.

[0111] Embodiment 3: Computer readable storage medium embodiment.

[0112] The computer-readable storage medium of the present invention can be any form of storage medium that can be read by a processor of a computer device, including but not limited to non-volatile memory, volatile memory, ferroelectric memory, etc. A computer program is stored on the computer-readable storage medium. When the processor of the computer device reads and executes the computer program stored in the memory, the steps of the above-mentioned detection method can be implemented.

[0113] Although the present invention has been described according to a limited number of embodiments, it will be apparent to those skilled in the art, with the benefit of the above description, that other embodiments may be envisioned within the scope of the invention thus described. In addition, it should be noted that the language used in this specification is selected primarily for readability and didactic purposes, rather than for explaining or defining the subject matter of the present invention. Therefore, many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the appended claims. The disclosure of the present invention is illustrative, not restrictive, with respect to the scope of the present invention, which is defined by the appended claims.

Claims

1. A high-precision detection method for concrete surface defects based on improved YOLOR, characterized in that: The method comprises the following steps: S1. Construct the concrete bridge apparent disease dataset DOBD V1.0; S2. Construct an improved YOLOR-BDIM model, wherein the improved YOLOR-BDIM model is optimized through loss function optimization and anchor box optimization; S3. Training the improved YOLOR-BDIM model based on the concrete bridge apparent disease dataset DOBD V1.0 to obtain a trained improved YOLOR-BDIM model; S4. Deploy the trained improved YOLOR-BDIM model on a local or cloud server; S5. Developing a web-based detection platform based on the local or cloud server; S6. Upload the image to be tested on the web-based detection platform for high-precision disease detection; S7. Automatically annotate the location and type of the disease in the image to be tested by calling the trained improved YOLOR-BDIM model.

2. According to claim 1, a high-precision detection method for concrete surface defects based on improved YOLOR is characterized in that: In S1, the concrete bridge apparent disease dataset DOBD V1.0 includes: an initial dataset and a dataset after image augmentation; The initial data set includes data on nine types of damage: honeycomb, peeling, holes, whitening, mesh cracks, single cracks, rust expansion, exposed reinforcement and water stains; The image-augmented data set is obtained by amplifying the initial data set using an image augmentation technique.

3. According to claim 1, a high-precision detection method for concrete surface defects based on improved YOLOR is characterized in that: In S2, the loss function is optimized as follows: SIoU_loss is used to replace the original loss function in the model. SIoU_loss is defined as follows: SIoU_loss=1-SIoU Among them, IoU is the intersection over union ratio of the predicted box and the real box; Δ is the distance loss; Ω is the shape loss.

4. According to claim 1, a high-precision detection method for concrete surface defects based on improved YOLOR is characterized in that: In S2, the anchor frame is optimized as follows: using a K-means clustering algorithm, adopting partial cluster replacement, calculating the area ratio of the clustered anchor frame and the original anchor frame, and replacing the anchor frames whose area ratios differ by less than 1% using the clustering results.

5. According to claim 1, a high-precision detection method for concrete surface defects based on improved YOLOR is characterized in that: In S5, the web-based detection platform provides an API interface.

6. A computer device comprising a memory and a processor, characterized in that A computer program is stored in the memory. When the processor runs the computer program stored in the memory, the processor executes the detection method according to any one of claims 1 to 5.

7. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the detection method according to any one of claims 1 to 5 is implemented.

Images