A method for detecting apparent defects in lightweight bridges based on deep learning

Abstract

The present invention discloses a method for detecting apparent defects in lightweight bridges based on deep learning, and relates to the technical field of concrete building safety assessment. The present invention improves the small target detection layer network in the YOLOv8 neural network, discards the large target detection layer of the original detection network, and introduces LMConv convolution in the C2f layer of the YOLOv8 neural network, thereby reducing the number of parameters while enhancing the ability to extract defect features and improving accuracy. The two convolutions on the BBox Loss branch and the Cls Loss branch on the YOLOv8 detection head are merged and replaced with group convolutions, and parallel calculations are performed to reduce the amount of computation. Inner-IoU is introduced, and auxiliary bounding boxes are used to break the limitations of the original loss function, accelerate network convergence, and enhance detection capabilities. Compared with the original model, the method proposed by the present invention significantly reduces the number of parameters and computational complexity, and improves detection accuracy, which is beneficial for bridge defect detection deployed on edge computing devices with limited computing power.

Description

Technical Field

[0001] The present invention relates to the technical field of concrete building safety assessment, and specifically to a method for detecting apparent defects of lightweight bridges based on deep learning. Background Art

[0002] Currently, deep learning-based bridge defect detection primarily focuses on various types of defects, including cracks, breakage, exposed reinforcement, pitting, cavities, and water erosion. These defects not only directly impact the service life of bridges but can also lead to structural collapse, severely impacting transportation and socioeconomic activities.

[0003] In the early days of bridge health monitoring, manual inspection was one of the primary methods. Engineers and technicians regularly inspected bridges through visual inspection and instrumental measurements. However, this method is not only extremely inefficient and hazardous, but also susceptible to subjective factors, leading to the possibility of missed detections and false positives. Most importantly, manual inspection cannot provide comprehensive and continuous monitoring of bridge structures, making it difficult to promptly detect potential hazards. To address these challenges, digital image processing technology has been introduced to bridge defect detection. By utilizing high-resolution digital cameras and advanced image processing algorithms, it enables automated, real-time monitoring of bridge structures, significantly improving detection accuracy. Early detection of superficial structural defects focused primarily on cracks in concrete structures. Image processing algorithms, such as grayscale processing, binarization, sharpening, image enhancement, and edge detection, were applied to defect digital images to extract and isolate crack features, enabling crack identification. However, digital image processing technology still faces challenges in complex scenarios and specific defects. For example, image processing algorithms may not accurately identify defects in conditions with uneven lighting or significant shadows. In addition, for some minor diseases, the sensitivity of digital image processing technology needs to be further improved.

[0004] To further improve the accuracy and efficiency of defect detection, object detection algorithms within computer vision have become a hot topic in bridge defect detection in recent years. Object detection algorithms offer high accuracy and robustness, with strong recognition capabilities for complex scenes and subtle defects. However, deploying bridge defect detection algorithms on mobile devices or edge devices still presents challenges, particularly striking a balance between speed and accuracy. Limited by the memory size of edge devices, existing detection algorithm models are large and difficult to deploy. Reducing the size of baseline models and making the algorithm lightweight while maximizing accuracy and minimizing false detections are key challenges in bridge defect detection. Furthermore, due to limited computing power on edge systems, traditional deep learning-based defect detection methods rely on drones to collect data before performing defect detection, resulting in cumbersome and inefficient operations. Therefore, improving algorithm detection speed and achieving real-time detection presents a significant challenge, necessitating a new solution to address these issues. Summary of the Invention

[0005] In response to the above-mentioned shortcomings of the existing technology, the purpose of the present invention is to provide a lightweight bridge apparent defect detection method based on deep learning, which can achieve the lightweighting of the existing detection model while keeping the accuracy of bridge apparent defect detection unchanged or even improving it, and can be deployed on the edge with insufficient computing power and limited memory size.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for detecting apparent defects of lightweight bridges based on deep learning, comprising at least the following steps:

[0007] Step 1: Collect a dataset of apparent concrete bridge defects, preprocess the dataset, and then perform data enhancement.

[0008] Step 2: Discard the detection head of YOLOv8 for detecting larger objects to reduce the computational complexity of the model;

[0009] Step 3: Use LMConv to replace the Conv module of the C2f layer in YOLOv8. The multi-scale feature extraction and point-by-point convolution in LMConv fully utilize redundant features and use the minimum amount of computation to exchange for richer feature information.

[0010] Step 4: Merge the two convolutions on the BBox Loss branch and the Cls Loss branch on the detection head respectively and replace them with group convolution, which effectively reduces the size of the model.

[0011] Step 5: Introduce Inner-IoU and combine it with CIOU in the original YOLOv8. Use auxiliary bounding boxes to speed up the model convergence, thereby improving the network accuracy without changing the network structure or increasing the network computation load.

[0012] Step 6: Conduct ablation experiments and comparative experiments on the improved lightweight network to verify the effectiveness of the improved network.

[0013] Preferably, the specific process of data set preprocessing in step 1 is as follows:

[0014] Manual screening based on image quality ultimately resulted in 996 images of concrete bridge surface defects of five types, including damage, exposed reinforcement, pitting, voids, and water erosion, which were manually annotated using Labelimg software.

[0015] These images also contain different lighting conditions, occlusions, shadows, and shooting angles, which can ensure the model's adaptability to different complex environments;

[0016] In order to alleviate the overfitting phenomenon during the experiment, 996 images of concrete bridges with multiple defects were processed. The original images and annotated images were augmented using the ImgAug method. The data was expanded to 3018 images by rotating, adding noise, left and right mirroring, and cropping.

[0017] In order to ensure the accuracy of the experiment, the expanded data set was divided into network training set, validation set and test set in a ratio of 8:1:1 for subsequent experiments.

[0018] Preferably, the step 2 at least includes the following steps:

[0019] The original model structure outputs a 640×640 image into 20×20 detection units, with each unit cell being 32×32 in size. However, the image of the apparent bridge disease is relatively small, and larger units will cause background information to cause stronger interference, reducing detection accuracy.

[0020] In the neck, the Conv, Concat, and C2f modules connected to the SPPF layer are deleted, and in the head, a Detect module is deleted;

[0021] The improved model structure parameters are smaller, and the influence of background information on the detection results is reduced, thereby improving the detection accuracy.

[0022] Preferably, the LMConv acquisition process in step 3 is as follows:

[0023] Divide the input feature count into 3 parts, occupying 1 / 2c_out, 1 / 4c_out and 1 / 4c_out respectively;

[0024] The two feature maps of 1 / 4c_out are respectively subjected to 3×3 convolution and 5×5 convolution and stacked into a new tensor;

[0025] Perform catcat operation with the original feature with 1 / 2c_out channel;

[0026] Through 1×1 point-by-point convolution, cross-channel information interaction is performed, and the feature information of different channels is integrated without increasing excessive computational overhead, thereby improving the detection accuracy of bridge defects.

[0027] Preferably, the specific inference process of step 4 is as follows:

[0028] Before replacement, the number of output channels is, the number of output channels is, and the convolution kernel is k×k in size, so the parameter size before replacement is ××k×k;

[0029] If the convolution is replaced by group convolution and the input channels of the original convolution are divided into 16 parts for parallel calculation, then the parameter size after replacement is ××k×k×16;

[0030] Therefore, the parameter size after replacement is 1 / 16 of the parameter size of the original convolution, which greatly reduces the parameter size and model size.

[0031] Preferably, the calculation method of introducing Inner-IoU and applying it to CIOU in YOLOv8 in step 5 is as follows:

[0032]

[0033]

[0034]

[0035]

[0036]

[0037] union=(w gt *h gt )*(ratio) 2 +(w*h)*(ratio) 2 -inter

[0038]

[0039] L Inner-CIOU =L CIOU +IOU-IOU inner .

[0040] The ground truth box and the predicted box are denoted as B gt and B, b l ,b r ,b t and b b They are the left boundary, right boundary, top boundary and bottom boundary of the prediction box respectively;

[0041] and They are the left boundary, right boundary, top boundary and bottom boundary of the real box respectively;

[0042] The center point of the real frame and the inner GT frame is Indicates that (x c ,y c ) represents the center point of the prediction box and the inner prediction box;

[0043] The width and height of the GT box are represented by w gt and h gt express;

[0044] The width and height of the prediction box are denoted by w and h.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] This method improves the small target detection layer network within the YOLOv8 neural network, discarding the large target detection layer of the original detection network. It also introduces LMConv convolution within the C2f layer of the YOLOv8 neural network, reducing the number of parameters while enhancing the ability to extract defect features and improving accuracy. It also merges the two convolutions on the BBox Loss branch and the Cls Loss branch of the YOLOv8 detection head and replaces them with group convolutions, allowing for parallel computation to reduce the amount of computation. It also introduces Inner-IoU, utilizing auxiliary bounding boxes to overcome the limitations of the original loss function, accelerating network convergence and enhancing detection capabilities. Compared to the original model, the proposed method significantly reduces the number of parameters and computational complexity, while improving detection accuracy. This makes it advantageous for bridge defect detection deployed on edge computing devices with limited computing power. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0048] Figure 1 This is an abstract of a method for detecting apparent defects in lightweight bridges based on deep learning;

[0049] Figure 2 This is the network structure diagram after discarding the large target detection layer;

[0050] Figure 3 This is the introduced LMConv convolution structure diagram;

[0051] Figure 4 It is the original detection head of YOLOv8;

[0052] Figure 5 Improved detection head for YOLOv8;

[0053] Figure 6 Schematic diagram of the Inne-IoU calculation method at larger scale and smaller scale;

[0054] Figure 7 This is the overall network structure diagram of this method. DETAILED DESCRIPTION

[0055] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0056] See also Figure 1 - Figure 7 A method for detecting apparent defects of lightweight bridges based on deep learning includes at least the following steps:

[0057] Step 1: Collect a dataset of apparent concrete bridge defects, preprocess the dataset, and then perform data enhancement.

[0058] Step 2: Discard the detection head for detecting larger objects in YOLOv8, namely the P5 feature layer, to reduce the computational complexity of the model;

[0059] Step 3: Use LMConv to replace the Conv module of the C2f layer in YOLOv8. The multi-scale feature extraction and point-by-point convolution in LMConv fully utilize redundant features and use the minimum amount of computation to exchange for richer feature information.

[0060] Step 4: Merge the two convolutions on the BBox Loss branch and the Cls Loss branch on the detection head respectively and replace them with group convolution, which effectively reduces the size of the model.

[0061] Step 5: Introduce Inner-IoU and combine it with CIOU in the original YOLOv8. Use auxiliary bounding boxes to speed up the model convergence, thereby improving the network accuracy without changing the network structure or increasing the network computation load.

[0062] Step 6: Conduct ablation experiments and comparative experiments on the improved lightweight network to verify the effectiveness of the improved network.

[0063] The specific process of data set preprocessing in step 1 is as follows:

[0064] Manual screening based on image quality ultimately resulted in 996 images of concrete bridge surface defects of five types, including damage, exposed reinforcement, pitting, voids, and water erosion, which were manually annotated using Labelimg software.

[0065] These images also contain different lighting conditions, occlusions, shadows, and shooting angles, which can ensure the model's adaptability to different complex environments;

[0066] In order to alleviate the overfitting phenomenon during the experiment, 996 images of concrete bridges with multiple defects were processed. The original images and annotated images were augmented using the ImgAug method. The data was expanded to 3018 images by rotating, adding noise, left and right mirroring, and cropping.

[0067] In order to ensure the accuracy of the experiment, the expanded data set was divided into network training set, validation set and test set in a ratio of 8:1:1 for subsequent experiments.

[0068] The overall apparent defect image of the bridge is relatively small, so how to improve the accuracy of defect detection has always been one of the main research directions. The defect image is relatively small and requires higher-resolution feature image detection, and these features are often provided by the underlying network. YOLOv8 detects small targets through shallow networks and convolution to extract feature information. The branch of the P5 layer outputs a 20×20 feature map with low resolution, which is suitable for detecting larger targets and greatly increases the burden on the network. Therefore, the present invention abandons the feature network of the P5 layer for predicting large targets and redesigns the feature fusion layer, such as Figure 2 shown.

[0069] Step 2 includes at least the following steps:

[0070] The original model structure outputs a 640×640 image into 20×20 detection units, with each unit cell being 32×32 in size. However, the image of the apparent bridge disease is relatively small, and larger units will cause background information to cause stronger interference, reducing detection accuracy.

[0071] In the neck, the Conv, Concat, and C2f modules connected to the SPPF layer are deleted, and in the head, a Detect module is deleted;

[0072] The improved model structure parameters are smaller, and the influence of background information on the detection results is reduced, thereby improving the detection accuracy.

[0073] In neural networks, feature redundancy refers to the existence of some features or neurons that provide redundant information in the learning and prediction of the model, and this information is highly similar or repetitive with the information already provided by other features or neurons. These redundancies are caused by the correlation between different features. It not only increases the computational cost and wastes computing resources, but also does not fully utilize these redundant features. In response to the above problems, the present invention proposes to integrate LMConv (Light Multi-Scale Conv) into the C2f layer of YOLOv8, denoted as C2f-LM. This structure uses multi-scale fusion to fully utilize redundant features, improves accuracy while further reducing the number of parameters and computational complexity. The MConv convolution structure is shown in the figure below. Figure 3 As shown,

[0074] The LMConv acquisition process in step 3 is as follows:

[0075] Divide the input feature count into 3 parts, occupying 1 / 2c_out, 1 / 4c_out and 1 / 4c_out respectively;

[0076] The two feature maps of 1 / 4c_out are respectively subjected to 3×3 convolution and 5×5 convolution and stacked into a new tensor;

[0077] Perform catcat operation with the original feature with 1 / 2c_out channel;

[0078] Through 1×1 point-by-point convolution, cross-channel information interaction is carried out to fuse feature information from different channels without increasing excessive computational overhead, thereby improving the detection accuracy of bridge defects.

[0079] When LMconv is placed in the shallow layer, the number of input and output channels is small, the number of feature maps is small, the convolution effect is not obvious, and the computational cost is increased. Therefore, this paper places LMConv in the deeper layers of the YOLOv8 network, embedded in the C2f layer of the backbone network and the neck network respectively, which has the best feature extraction effect and lower computational cost.

[0080] The number of parameters for a regular convolution is 3×3×(c_out)×(c_out)=9(c_out)^2, while the number of parameters for an LMConv convolution is 0.25×c_out×0.25×c_out×[3×3+5×5]+1×1×c_out×c_out=2.125(c_out)^2+(c_out)^2=3.125c_out^2. This shows that the number of parameters for LMConv is nearly 1 / 3 of that of a regular convolution. Therefore, its multi-scale information fusion feature not only fully utilizes redundant features to improve the accuracy of bridge multi-defect detection and reduce the false detection rate, but also has fewer parameters than a regular convolution, greatly improving the performance of bridge defect detection.

[0081] Each of the three detection heads of YOLOv8 has a Bbox Loss branch and a Cls Loss branch. Each branch has two ordinary convolutions. In order to reduce the parameters of the network and the size of the model, the two ordinary convolutions on each branch are merged and replaced with group convolutions. Group convolutions can divide each convolution into 16 parts for parallel calculation, which greatly reduces the amount of calculation and speeds up the operation. The original detection head of YOLOv8 is as follows: Figure 4 As shown, the improved detection head of YOLOv8 is as follows Figure 5 As shown;

[0082] The specific inference process of step 4 is as follows:

[0083] Before replacement, the number of output channels is, the number of output channels is, and the convolution kernel is k×k in size, so the parameter size before replacement is ××k×k;

[0084] If the convolution is replaced by group convolution and the input channels of the original convolution are divided into 16 parts for parallel calculation, then the parameter size after replacement is ××k×k×16;

[0085] Therefore, the parameter size after replacement is 1 / 16 of the parameter size of the original convolution, which greatly reduces the parameter size and model size.

[0086] Each detection head of the original YOLOv8 is divided into a BBox Loss branch and a Cls Loss branch. Each branch has two conv, a conv2d and a loss term. Its original structure is as follows Figure 4 As shown in Figure 2. Such a structure cannot fully utilize feature information and also increases the computational cost;

[0087] Merge the two convs on the BBox Loss branch and the Cls Loss branch on the detection head respectively, which saves computing resources and reduces the computing burden;

[0088] After merging, group convolution is used to replace it. Both group convolutions are divided into 16 parts for parallel calculation, which improves the calculation efficiency.

[0089] Nowadays, most IoU functions are based on the existing IoU function and further add new loss terms to speed up the convergence speed, while ignoring the convergence ability of the IoU function itself. Therefore, the author proposed the Inner-IoU loss function in 2023. Its core idea is to introduce a scaling factor to control the scale of the auxiliary bounding box for calculating the loss. Unlike the previous IoU function, it can be used on the existing IoU function to speed up the convergence speed.

[0090] Inner-IoU is introduced and combined with CIOU in the original YOLOv8 to accelerate the convergence of the model using auxiliary bounding boxes. The schematic diagram of the Inne-IoU calculation method for larger and smaller scales is as follows Figure 6 shown.

[0091] The original YOLOv8 model uses DFL Loss+CIoU Loss as regression loss;

[0092] The calculation method of CIoU Loss is as follows:

[0093]

[0094] Among them, v is the consistency of the measured aspect ratio. Since most of the diseases detected in this dataset are small targets with small length and width, CIoU may not be able to predict the actual situation and will optimize the detection results in an inappropriate way;

[0095] In order to solve the above limitations, the present invention introduces the Inner-IoU loss function based on CIoU, and expands the proportion of the auxiliary bounding box with a proportional factor ratio, which can accelerate the regression process of low IoU samples, speed up convergence, and optimize the detection effect.

[0096] Step 5 introduces Inner-IoU and applies it to the calculation method of CIOU combination in YOLOv8 as follows:

[0097]

[0098]

[0099]

[0100]

[0101]

[0102] union=(wgt *h gt )*(ratio) 2 +(w*h)*(ratio) 2 -inter

[0103]

[0104] L Inner-CIOU =L CIOU +IOU-IOU inner

[0105] The ground truth (GT) box and the predicted box are denoted as B gt and B, such as Figure 6 As shown. l ,b r ,b t and b b They are the left boundary, right boundary, top boundary and bottom boundary of the prediction box respectively. and They are the left boundary, right boundary, top boundary and bottom boundary of the real box. The center point of the real box (GT box) and the inner GT box is Indicates that (x c ,y c ) represents the center point of the prediction box and the inner prediction box. The width and height of the GT box are w gt and h gt The width and height of the predicted box are represented by w and h. The variable "ratio" corresponds to the scale factor, which is usually in the range of [0.5, 1.5]. The value of this paper is 1.13. The intersection and union are defined to calculate the inner IoU, so the following formula:

[0106]

[0107] The calculation is the intersection area (inner) of the real box and the predicted box;

[0108] formula:

[0109] union=(w gt *h gt )*(ratio) 2 +(w*h)*(ratio) 2 -inter

[0110] Their union area is calculated.

[0111] The specific operations of the series of ablation experiments and comparative experiments in step 6 are as follows:

[0112] To verify the effectiveness of the improved network, we conducted a series of ablation experiments on the improved modules to validate the role and effect of each module. Because discarding the P5 layer is a fundamental improvement to the entire network and has a fundamental impact on the network, in addition to verifying the effectiveness of individual modules, all other ablation experiments were conducted without the P5 layer.

[0113] The present invention also uses a series of other classic networks such as Faster-rcnn, Centernet and YOLO series networks, such as YOLOv5 and YOLOv7-tiny, to conduct comparative experiments to verify the effect of the improved network.

[0114] Most of the current IoU Loss optimizes the calculation method of the loss function by adding loss terms, and focuses more on making the predicted box closer to the real box. Inner-IoU focuses on the center of the bounding box. It can control the generation of auxiliary bounding boxes by setting different scale factors. It can also combine different loss functions to adapt to the needs of different tasks, with better generalization and robustness. The improved complete network structure is as follows Figure 7 shown.

[0115] In order to verify the effectiveness of each module, it is necessary to add each module one by one to compare the experimental results. The specific acquisition process of the ablation experiment is as follows:

[0116]

[0117] As shown in the table above, each module reduces computational complexity and parameter requirements to a certain extent, facilitating the model's deployment. The improved model achieves a 1.5% improvement in mAP@.5 and a 1.08% improvement in F1, significantly improving accuracy. With only 1.49M parameters, a 50.5% reduction, and 4.7 GFLOPs, a 42% reduction in computational complexity, the improved model achieves 400 frames per second (FPS). Therefore, the improved model outperforms the original YOLOv8 model in both accuracy and size, while also achieving real-time detection speed.

[0118] In order to compare the improved model with other classic network models, comparative experiments are conducted on various classic models on this dataset to compare the performance of each model. The specific process of obtaining the comparative experiment is as follows:

[0119]

[0120] From this we can see that compared with many classic networks, our improved network experimental performance is also the best, and good results are achieved with the lowest number of parameters and computational complexity.

[0121] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

Claims

1. A method for detecting apparent defects in lightweight bridges based on deep learning, characterized by: At least the following steps are included: Step 1: Collect a dataset of apparent concrete bridge defects, preprocess the dataset, and then perform data enhancement. Step 2: Discard the detection head for detecting large objects in YOLOv8 to reduce the computational complexity and size of the model. Step 3: Use LMConv to replace the Conv module of the C2f layer in YOLOv8. The multi-scale feature extraction and point-by-point convolution in LMConv fully utilize redundant features and use the minimum amount of computation to exchange for richer feature information. Step 4: Merge the two convolutions on the BBox Loss branch and the Cls Loss branch on the detection head respectively and replace them with group convolution, which effectively reduces the size of the model. Step 5: Introduce Inner-IoU and combine it with CIOU in the original YOLOv8. Use auxiliary bounding boxes to speed up the model convergence, thereby improving the network accuracy without changing the network structure or increasing the network computation load. The calculation method of introducing Inner-IoU and applying it to CIOU in YOLOv8 in step 5 is as follows: union=(w gt *h gt )*(ratio) 2 +(w*h)*(ratio) 2 -inter L Inner-CIOU =L CIOU +IOU-IOU inner The ground truth box and the predicted box are denoted as B gt and B, b l ,b r ,b t and b b They are the left boundary, right boundary, top boundary and bottom boundary of the prediction box respectively; and They are the left boundary, right boundary, top boundary and bottom boundary of the real box respectively; The center point of the real frame and the inner GT frame is Indicates that (x c ,y c ) represents the center point of the prediction box and the inner prediction box; The width and height of the GT box are represented by w gt and h gt express; The width and height of the prediction box are represented by w and h; Step 6: Conduct ablation experiments and comparative experiments on the improved lightweight network to verify the effectiveness of the improved network.

2. The method for detecting apparent defects of lightweight bridges based on deep learning according to claim 1, characterized in that: The specific process of data set preprocessing in step 1 is as follows: Manual screening based on image quality ultimately resulted in 996 images of five types of concrete bridge surface defects, including damage, exposed reinforcement, pitting, voids, and water erosion. These images were then manually annotated using Labelimg software. These images also include different lighting conditions, occlusions, shadows, and shooting angles to ensure the model's adaptability to different complex environments; In order to alleviate the overfitting phenomenon during the experiment, 996 images of concrete bridges with multiple defects were processed. The original images and annotated images were augmented using the ImgAug method. The data was expanded to 3018 images by rotating, adding noise, left and right mirroring, and cropping. In order to ensure the accuracy of the experiment, the expanded data set was divided into network training set, validation set and test set in a ratio of 8:1:1 for subsequent experiments.

3. The method for detecting apparent defects of lightweight bridges based on deep learning according to claim 1, characterized in that: The second step at least includes the following steps: The original model structure outputs a 640×640 image into 20×20 detection units, with each unit cell being 32×32 in size. However, the image of the apparent bridge disease is small, and large units will cause strong interference from background information, reducing detection accuracy. In the neck, the Conv, Concat, and C2f modules connected to the SPPF layer are deleted, and in the head, a Detect module is deleted; The improved model structure has small parameters and reduces the impact of background information on detection results, thereby improving detection accuracy.

4. The method for detecting apparent defects of lightweight bridges based on deep learning according to claim 1, characterized in that: The LMConv acquisition process in step 3 is as follows: Divide the input feature count into 3 parts, occupying 1 / 2c_out, 1 / 4c_out and 1 / 4c_out respectively; The two feature maps of 1 / 4c_out are respectively subjected to 3×3 convolution and 5×5 convolution and stacked into a new tensor; Perform catcat operation with the original feature with 1 / 2c_out channel; Through 1×1 point-by-point convolution, cross-channel information interaction is performed, and the feature information of different channels is integrated without increasing excessive computational overhead, thereby improving the detection accuracy of bridge defects.

5. The method for detecting apparent defects of lightweight bridges based on deep learning according to claim 1, characterized in that: The specific inference process of step 4 is as follows: Before replacement, the number of output channels is, the number of output channels is, and the convolution kernel is k×k in size, so the parameter size before replacement is ××k×k; If the convolution is replaced by group convolution and the input channels of the original convolution are divided into 16 parts for parallel calculation, then the parameter size after replacement is ××k×k×16; Therefore, the parameter size after replacement is 1 / 16 of the parameter size of the original convolution, which greatly reduces the parameter size and model size.

Images