Lightweight detection model and detection method for concrete cracks

Abstract

The application discloses a concrete crack lightweight detection model and a detection method, and the detection model comprises three parts of a feature extraction backbone network, a feature fusion network and a regression prediction network. The backbone network and the feature fusion network are alternately connected by four LGSE modules and four C3-LGSE modules respectively, so that the feature extraction of the concrete crack and the information fusion of the multi-dimensional features are realized; and the regression prediction network is responsible for outputting the detection result of the concrete crack. The lightweight convolution module LGS can balance the network parameters and the detection accuracy, enhance the inter-group information exchange of the feature maps, obtain more features by using fewer feature maps, the channel attention module ECA can further improve the detection accuracy without increasing the parameter amount, and the LGSE module with the simultaneous action of the two can realize the lightweight and high-precision detection at the same time. The application is helpful to realize the lightweight and high-precision detection of the concrete crack, and has important significance for realizing the real-time detection embedded in a low-computing-power platform.

Description

Technical Field

[0001] This invention belongs to the field of target detection and relates to a lightweight detection model, method, equipment and storage medium for concrete crack detection. Background Technology

[0002] Concrete cracks can occur due to factors such as design flaws, uneven stress distribution, thermal expansion and contraction, and material selection. These cracks not only affect the aesthetics of concrete buildings but also severely impact their load-bearing capacity, durability, and even structural stability, safety, and lifespan. Therefore, timely detection and treatment of these cracks are crucial for the stability and safety of buildings and for protecting public safety and property.

[0003] Currently, commonly used methods for detecting concrete cracks include manual inspection, ultrasonic testing, and penetrant testing. However, these methods all suffer from drawbacks such as reliance on manual labor, low efficiency, and high costs. With the development of the internet and the continuous improvement of computer hardware capabilities, early researchers used image processing methods for crack defect detection, such as threshold segmentation, Canny edge detection, and morphological algorithms. However, these algorithms have low detection accuracy and poor adaptability. In recent years, the research and application of deep learning and convolutional neural networks have developed rapidly, with related achievements shining brightly in the field of computer vision. Consequently, more and more researchers are applying deep learning to crack detection.

[0004] Typical achievements in improving network detection accuracy include Hamed et al.'s application of single-stage Yolov2 and two-stage Faster-RCNN for intelligent crack identification. Experiments show that the Yolov2 network model has higher identification accuracy, but both networks have a large number of parameters. See "HAMED M, JIN P, YAW AG, et al. Pavement image datasets: a new benchmark dataset to classify and densify pavement distresses[J]. Transportati-on Research Record, 2020, 2674(2):328-339." Zhang et al. proposed an improved Yolov3 network for bridge defect detection. By introducing a focus loss function and a new transfer learning method, the defect identification accuracy was effectively improved. See "ZHANG C B, CHANG CC, JAMSHIDI M. Concrete bridge surface damage detection using a single-stage detector[J]. Computer-Aided Civil and Infrastructure Engineering, 2020, 35(4):389-409." Shi et al. improved the feature fusion structure of the YOLOv4 network to enhance the detection capability of fine cracks and achieve high-precision detection of concrete cracks. See "Shi Jie, Ma Wenqi, Wu Hongjie. Improved YOLOv4 algorithm for concrete building crack detection [J]. Microelectronics & Computer, 2023, 40(3):56-66". Although the above improvements have improved the network's detection accuracy, the number of model parameters and computational load are too large, posing a huge challenge to mobile low-computing-power platforms and failing to meet the requirements of portable real-time detection.

[0005] Sun et al., representing the achievements in lightweight detection, proposed a method for identifying pavement cracks based on a convolutional neural network fusion model. They used an SSD network model for crack detection and a U-Net network model for crack segmentation. For details, see "Sun Chaoyun, Ma Zhidan, Li Wei, et al. A method for identifying pavement cracks based on a deep convolutional neural network fusion model. Journal of Chang'an University (Natural Science Edition), 2020, 40(4):1–13". Song et al. reduced the number of network parameters by simplifying the YOLOv4-tiny network structure, enabling it to be implanted into low-computing platforms such as Raspberry Pi for real-time detection. For details, see "Song Libo, Fei Yanqiong. Comparison of the crack detection performance of two types of simplified YOLOv4-tiny networks [J]. Journal of Tongji University (Natural Science Edition), 2022, 50(01):129-137". The networks selected and improved above reduced the number of parameters, but it is difficult to achieve a balance between the number of parameters and accuracy.

[0006] The inventors found that existing deep learning-based concrete crack detection methods are not effective and have the following shortcomings: (1) The network model is huge, which requires high computing power and memory, making it difficult to deploy on low computing power platforms for real-time detection. (2) It is difficult to achieve a balance between lightweight and high-precision detection, and to achieve both lightweight and high-precision crack detection at the same time. Summary of the Invention

[0007] To address the aforementioned issues, this invention provides a lightweight concrete crack detection model. While reducing the number of model parameters and computational load, it can increase the accuracy of the network in detecting concrete crack images. This solves the problem that existing network models are too large and difficult to deploy on low-computing platforms for high-precision real-time detection.

[0008] The second objective of this invention is to provide a lightweight method for detecting cracks in concrete.

[0009] A third objective of this invention is to provide an electronic device.

[0010] A fourth objective of this invention is to provide a computer storage medium.

[0011] The technical solution adopted in this invention is a lightweight concrete crack detection model, which includes three parts: a backbone network, a feature fusion network, and a prediction output.

[0012] The backbone network consists of 4 groups of LGSE modules and C3-LGSE modules connected alternately. Each group outputs tensors of 160×160×128, 80×80×256, 40×40×512 and 20×20×1024 respectively. It is mainly responsible for reducing model parameters, enhancing the fusion of shallow and deep features, and improving feature extraction capabilities.

[0013] The feature fusion network is also composed of 4 LGSE modules and 4 C3-LGSE modules connected alternately. It adopts the FPN (Feature Pyramid Network) feature pyramid structure and fuses feature maps at different levels together through upsampling and downsampling operations to generate a multi-scale feature pyramid.

[0014] The lightweight convolutional module LGSE fuses point convolutions and 3x3 group convolutions. As the network deepens, the number of channels is gradually increased by multiplying the number of convolutions, thereby increasing the complexity and expressive power of the model. Subsequently, a channel shuffle module is used to promote information exchange between groups and get rid of information constraints between channels. Finally, a channel attention (ECA) module is used to enhance the channel feature representation of the feature map and improve the feature extraction capability of the model.

[0015] The predicted output is responsible for providing the detection results of concrete cracks.

[0016] Furthermore, the LGSE module of the backbone network is concatenated with one 1×1 convolution and two 3×3 group convolutions, and each convolution is processed by BN and SiLU activation. The three are then concatenated to obtain the output, achieving lightweight feature extraction. Then, a channel shuffle module is used to enhance the inter-group channel information exchange. Finally, a channel attention (ECA) module is used to enhance the channel feature representation of the feature map and improve the feature extraction capability of the model.

[0017] The LGSE module is used in the feature extraction backbone network. The number of channels in the four LGSE modules used in sequence are 128, 256, 512 and 1024 respectively. The number of output channels is increased by multiplying the convolutions, thereby increasing the complexity and expressive power of the model.

[0018] The LGSE module is composed of three modules connected in series: LG, ChannelShuffle, and ECA. LG (LightGhost), compared to GhostConv, aims to transform what was originally a high-level, low-cost linear computation into a more efficient and cost-effective solution. k (1≤k≤C), replaced by two low-level, inexpensive linear operations Ф 1i and Ф 2jThe concatenated structure of (1≤1i≤C, 1≤2j≤C) replaces the group convolution with 5 kernels and 1 / 2 input channels in the original GhostConv with a concatenated structure of two group convolutions with 3 kernels and 1 / 4 input channels. The stack of two 3*3 convolutions has the same effect and receptive field as a single 5*5 convolution, but with less computation and more nonlinearity to prevent gradient vanishing or exploding. The LG module can obtain more features with fewer feature maps, further removing feature redundancy in the network and reducing the number of parameters in the network.

[0019] The ChannelShuffle module in the LGSE module arranges the channels sequentially into a matrix of a specified dimension, then transposes the matrix, and finally flattens it to obtain a channel-shuffled feature map. Channel Shuffle cleverly uses conventional tensor operations to distribute the information of each group to other groups, so that the output feature map can evenly contain the features of each group, promoting information exchange between groups and thus improving the expressive power of the feature model.

[0020] When extracting feature map information, the LGSE module leverages the channel attention ECA module to enhance the expressive power of the feature map. First, it performs global average pooling on the input feature map, compressing the H×W×C dimension feature map into a 1×1×C dimension feature map. Then, it performs channel feature learning through 1×1 convolution to obtain the weights for each feature point. Finally, it combines channel attention by multiplying the channel attention weights with the original input feature map channel by channel, ultimately obtaining a channel-attention feature map. This allows for channel feature enhancement of the input feature map without changing the output feature map dimension and with almost no increase in parameters. To address the extraction of features from different input feature maps and different ranges, a dynamic convolution kernel is used to determine the kernel size. The dynamic convolution kernel formula is shown below:

[0021]

[0022] Where k represents the kernel size; C represents the number of channels; ||odd means that k can only be an odd number; γ and b are set to 2 and 1, respectively, to change the ratio between the number of channels C and the kernel size.

[0023] Furthermore, the feature extraction module C3-LGSE is designed to fully extract concrete crack features;

[0024] Each C3-LGSE module consists of m (values ​​3, 6, 9, 3) Bottleneck-LGSE modules and two LGSE (LG+ChannelShuffle+ECA) convolutional modules, and its number of channels is the same as the number of input channels of the C3-LGSE module.

[0025] The Bottleneck-LGS module consists of two LGSE modules connected in series with the input residual. The residual connection between the input and output helps to fuse shallow and deep features in the feature extraction backbone network, and can also alleviate the gradient explosion problem caused by network depth. In the feature fusion neck network, the series connection method can preserve the fused feature information to a greater extent.

[0026] The feature extraction backbone network contains four C3-LGSE modules, which in turn contain three, six, nine, and three Bottleneck-LGSE modules. The feature fusion neck network contains four C3-LGSE modules, each of which contains three Bottleneck-LGSE modules.

[0027] The number of Bottleneck-LGSE modules is automatically determined by the product of the configuration file .yaml and the parameter depth_multiple;

[0028] The Bottleneck-LGSE module, by replacing the traditional convolution module with the LGSE module, has about half the number of parameters of the traditional convolution module, which can reduce the number of parameters while ensuring the extraction of more features.

[0029] A method for detecting cracks in lightweight concrete, characterized by the following steps:

[0030] S1. Input the concrete crack image into the feature extraction backbone network to complete the information extraction of crack features, and send the generated feature map tensor into the feature fusion neck network.

[0031] S2. Feature Fusion: After fusing the multi-dimensional feature information extracted by the neck network with the backbone network, the information is fed into the prediction output.

[0032] S3. Prediction Output: The prediction is completed, and the detection results of concrete cracks are output.

[0033] 5. The calculation process of the convolutional module LGSE involved in the backbone network in step S1 of the lightweight concrete crack detection method is as follows:

[0034] Y0 = Conv p (K i )

[0035] Y1 = Concat(Conv) g2 (Conv g1 (Y0)),Conv g1 (Y0),Y0)

[0036] Y2 = ChannelShuffle(Y1)

[0037] K i+1 =ECA(Y2)

[0038] Where Y0 represents the condition for input K i (i takes the values ​​1, 2, ..., n) via Conv p The feature maps extracted by point convolution, Y1 represents the features extracted from Y0 after two and one Convolution operations respectively. g The concatenated result of the group convolutions, Y2 represents the result of channel shuffling of Y1, and finally K i+1 This represents the feature map finally extracted after passing through the Channel Attention (ECA) module.

[0039] 6. An electronic device, characterized in that it uses the method described in claim 4 to detect concrete cracks.

[0040] 7. A computer storage medium, characterized in that the storage medium stores at least one program instruction, the at least one program instruction being loaded and executed by a processor to implement the concrete crack detection method as described in claim 4.

[0041] The beneficial effects of this invention are:

[0042] The lightweight detection network YOLOv5s-LGSE, proposed for concrete crack detection, addresses challenges such as channel semantic loss, insufficient target feature extraction, and the large size of the network model, making it difficult to deploy on computational platforms. Ablation and comparative experiments demonstrate that, during feature extraction, combining the LG module with channel shuffling reduces network parameters and computational load, compresses model size, and promotes channel semantic information exchange, balancing parameter quantity and detection accuracy. Introducing the channel attention (ECA) module enhances the expression of channel features, further improving network detection accuracy with almost no increase in network parameters. This is of great significance for achieving real-time and portable concrete crack detection. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a schematic diagram of the lightweight detection model according to an embodiment of the present invention.

[0045] Figure 2 This is a schematic diagram of the lightweight LGSE convolution module of the lightweight detection model in an embodiment of the present invention.

[0046] Figure 3 This is a schematic diagram of the C3-LGSE and Bottleneck-LGSE modules of the lightweight detection model in an embodiment of the present invention.

[0047] Figure 4 This is a comparison chart of the detection methods of this invention and the original algorithm for detecting different concrete cracks.

[0048] Figure 5 This is a physical image of the low-computing-power embedded platform of the lightweight detection model in this embodiment of the invention.

[0049] Figure 6 This is a real-time detection effect diagram of the lightweight detection model of this invention on a low-computing-power embedded platform. Detailed Implementation

[0050] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0051] Example 1,

[0052] A lightweight concrete crack detection model, the structure of which is as follows: Figure 1 As shown, it consists of three parts: the feature extraction backbone network, the feature fusion network Neck, and the regression prediction network Head.

[0053] The backbone network consists of 4 groups of LGSE modules and C3-LGSE modules connected alternately. Each group outputs tensors of 160×160×128, 80×80×256, 40×40×512 and 20×20×1024 respectively. It is mainly responsible for reducing model parameters, enhancing the fusion of shallow and deep features, and improving feature extraction capabilities.

[0054] The feature fusion network is also composed of 4 LGSE modules and 4 C3-LGSE modules connected alternately. It adopts the FPN (Feature Pyramid Network) feature pyramid structure and fuses feature maps at different levels together through upsampling and downsampling operations to generate a multi-scale feature pyramid.

[0055] The predicted output is responsible for providing the detection results of concrete cracks.

[0056] like Figure 2 As shown,

[0057] The LGSE module of the backbone network is achieved by concatenating a 1×1 convolution and two 3×3 group convolutions, with each convolution undergoing BN and SiLU activation, and then concatenating the three to obtain the output, thus realizing lightweight feature extraction. Then, a channel shuffle module is used to enhance the inter-group channel information exchange. Finally, a channel attention (ECA) module is used to enhance the channel feature representation of the feature map and improve the feature extraction capability of the model.

[0058] The LGSE module is used in the feature extraction backbone network. The number of channels in the four LGSE modules used in sequence are 128, 256, 512 and 1024 respectively. The number of output channels is increased by multiplying the convolutions, thereby increasing the complexity and expressive power of the model.

[0059] When extracting feature map information, the LGSE module uses the channel attention ECA module to improve the expressive power of the feature map, which can enhance the channel features of the input feature map without changing the dimension of the output feature map and with almost no increase in the number of parameters.

[0060] like Figure 3 As shown,

[0061] Each C3-LGSE module consists of n Bottleneck-LGSE modules and two LGSE (LG+ChannelShuffle+ECA) convolutional modules, and its number of channels is the same as the number of input channels of the C3-LGSE module.

[0062] The Bottleneck-LGS module consists of two LGSE modules connected in series with the input residual. The residual connection between the input and output helps to fuse shallow and deep features in the feature extraction backbone network, and can also alleviate the gradient explosion problem caused by network depth. In the feature fusion neck network, the series connection method can preserve the fused feature information to a greater extent.

[0063] The feature extraction backbone network contains 4 C3-LGSE modules, which in turn contain 3, 6, 9, and 3 Bottleneck-LGSE modules. The feature fusion neck network contains 4 C3-LGSE modules, each of which contains 3 Bottleneck-LGSE modules.

[0064] The Bottleneck-LGSE module replaces the traditional convolutional module with the LGSE module, resulting in approximately half the number of parameters compared to traditional convolution. This allows for a significant reduction in the number of parameters while ensuring the extraction of more features.

[0065] Example 2,

[0066] A method for detecting cracks in lightweight concrete, characterized by the following steps:

[0067] S1. Input the concrete crack image into the feature extraction backbone network to complete the information extraction of crack features, and send the generated feature map tensor into the feature fusion neck network.

[0068] The calculation process of the convolutional module LGSE is as follows:

[0069] Y0 = Conv p (K i )

[0070] Y1 = Concat(Conv) g2 (Conv g1 (Y0)),Conv g1 (Y0),Y0)

[0071] Y2 = ChannelShuffle(Y1)

[0072] K i+1 =ECA(Y2)

[0073] Where Y0 represents the condition for input K i (i takes the values ​​1, 2, ..., n) via Conv p The feature maps extracted by point convolution, Y1 represents the features extracted from Y0 after two and one Convolution operations respectively. g The concatenated result of the group convolutions, Y2 represents the result of channel shuffling of Y1, and finally K i+1 This represents the feature map finally extracted after passing through the channel attention (ECA) module;

[0074] S2. Feature Fusion: After fusing the multi-dimensional feature information extracted by the neck network with the backbone network, the information is fed into the prediction output.

[0075] S3. Prediction Output: The prediction is completed, and the detection results of concrete cracks are output.

[0076] To verify the detection performance of YOLOv5-LGSE in this invention, a comparative experiment was conducted. A custom dataset containing 4058 crack images was created for training and validation, using images of concrete wall cracks. These cracks were categorized into two types: 2006 ordinary cracks (gap) and 2534 leaky cracks (leaky_gap) (used to detect pipe ruptures inside the wall). All images in the dataset were manually labeled using LabelImg, and the training and validation sets were randomly partitioned at a ratio of 8:2. The final partitioning result was 3678 images in the training set and 980 images in the validation set.

[0077] The eight classic object detection networks compared are SSD (see "LIU W, ANGUELOV D, ERHAN D, et al.").

[0078] Ssd: Single shot multibox detector[C].European Conference on ComputerVision.Sprin ger,Cham,2016:21-37."), Faster-RCNN (for details, see "REN S,HE K,GIRSHICK R,etal.Faster R-CNN:Towards real-time object detection with re-gion proposal networks[J].Advance s in Neural Information Processing Systems, 2015, 28:91-99.") and the YOLO series (for details, see "RED MON J, FARHADI A.YOLOv3: An inc-remental improvement[J].arXivpreprint arXiv:1804.02767,2018.", "BOCHKOVSKIY A,WANG CY,LIAO H YM.YOLOv4: Optimal sp eed andaccuracy of obj-ect detection[J].arXiv “Preprint arXiv:2004.10934,2020.”, “Chien-Yao, Wang, Alexey Bochkovskiy, Hong-Yuan MarkLiao YOLOv7:Trainable bag-of-freebies sets new state-of-the-art for real-time object determinants [doi:10.48550 / arXiv.2207.02696.2022].”) Comparative experimental results are shown in Table 1, where bold text represents the best performance.

[0079] Table 1 Comparison of experimental results

[0080]

[0081] As shown in Table 1, in terms of detection accuracy, YOLOv4-Tiny may experience missed detections in multi-target and small-target detection, resulting in an mAP@0.5 of only 88.3%, which is significantly lower than other mainstream target detection accuracies. The YOLOv5s-LGSE network shows improvements of 1.2% and 2.1% compared to the lightweight networks YOLOv5s and YOLOv7-Tiny, respectively; improvements of 1.3%, 0.5%, 4.4%, and 0.2% compared to the one-stage classic networks SSD, YOLOv3, YOLOv4, and the newer YOLOv7; and a 0.2% improvement compared to the two-stage classic network Faster-RCNN. It can be seen that the models of this invention outperform traditional target detection networks and meet the requirements for high-precision detection of concrete cracks.

[0082] In terms of lightweight design, the YOLOv5s-LGSE network has 3.7M parameters, 8.2G of computation, and a model size of only 7.6Mb. Compared with large networks such as SSD, Faster-RCNN, YOLOv3, YOLOv4, and YOLOv7, it represents a significant reduction in parameters, computation, and model size. Compared with lightweight networks YOLOv4-Tiny, YOLOv5s, and YOLOv7-Tiny, the number of network parameters is reduced by 37.3%, 47.1%, and 38.3%, respectively; the computation is increased by 20.6%, decreased by 48.8%, and reduced by 37.9%, respectively; and the model size is reduced by 66%, 44.5%, and 35%, respectively.

[0083] In terms of detection speed, the YOLOv5s-LGSE network has an FPS of 35, which is faster than large networks such as SSD, Faster-RCNN, YOLOv3, YOLOv4 and YOLOv7, but slower than lightweight networks such as YOLOv4-Tiny, YOLOv5s and YOLOv7-Tiny, but basically meets the requirements for real-time detection.

[0084] like Figure 4 As shown,

[0085] The YOLOv5s model achieved detection accuracies of 87% and 73% for the two images of seepage cracks, and 77% and 36% for the images of ordinary cracks, respectively. In contrast, the YOLOv5s-LGSE model achieved detection accuracies of 89%, 83%, 90%, and 43% for the same images. As the detection results show, the YOLOv5s-LGSE model has higher detection accuracy and is more precise in predicting crack locations.

[0086] like Figure 5 As shown,

[0087] The low-computing embedded platform selected for the network in this invention is the Jetson TX2 development platform manufactured by NVIDIA. The TX2 is a powerful AI edge computing device mainly developed by NVIDIA, which is very suitable as a development platform for edge devices and can deploy tasks such as image processing.

[0088] like Figure 6 As shown,

[0089] As can be seen, the YOLOv5s-LGSE model proposed in this invention can perform normal detection functions on a low computing power platform while still maintaining high detection accuracy.

[0090] If the defect detection method described in this embodiment of the invention is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the image classification method described in this embodiment of the invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, external hard drives, ROM, RAM, magnetic disks, or optical disks.

[0091] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A lightweight detection model for concrete cracks, characterized in that, It consists of three parts: the backbone network, the feature fusion network Neck, and the prediction network Head. The core structure of the backbone network consists of four groups of LGSE modules and C3-LGSE modules connected alternately. Each group outputs tensors of 160×160×128, 80×80×256, 40×40×512 and 20×20×1024 respectively. It is mainly responsible for reducing model parameters, enhancing the fusion of shallow and deep features, and improving feature extraction capabilities. The LGSE module uses a 1×1 convolution and two 3×3 group convolutions concatenated together. Each convolution is processed by BN and SiLU activation, and then the three are concatenated to obtain the output, achieving lightweight feature extraction. Then, a channel shuffle module is used to enhance the inter-group channel information exchange. Finally, a channel attention (ECA) module is used to enhance the channel feature representation of the feature map and improve the feature extraction capability of the model. The LGSE module is used in the feature extraction backbone network. The number of channels in the four LGSE modules used in sequence are 128, 256, 512 and 1024 respectively. The number of output channels is increased by multiplying the convolutions, thereby increasing the complexity and expressive power of the model. The LGSE consists of LG + ChannelShuffle + ECA; When extracting feature map information, the LGSE module uses the channel attention ECA module to improve the expressive power of the feature map, which can enhance the channel features of the input feature map without changing the dimension of the output feature map and with almost no increase in the number of parameters. The feature fusion network consists of 4 LGSE modules and 4 C3-LGSE modules connected alternately. It adopts the FPN (Feature Pyramid Network) feature pyramid structure and fuses feature maps at different levels together through upsampling and downsampling operations to generate a multi-scale feature pyramid. The prediction network consists of three parallel layers and is responsible for providing the detection results of concrete cracks.

2. The lightweight concrete crack detection model according to claim 1, characterized in that, The C3-LGSE module of the backbone network is designed to fully extract concrete crack features. Each backbone network's C3-LGSE module consists of m Bottleneck-LGSE and two LGSE convolutional modules, with the number of channels being the same as the number of input channels of the C3-LGSE module, where m takes values ​​of 3, 6, 9, and 3 respectively. The Bottleneck-LGS module consists of two LGSE modules connected in series with the input residual. The residual connection between the input and output helps to fuse shallow and deep features in the feature extraction backbone network, and can also alleviate the gradient explosion problem caused by network depth. In the feature fusion neck network, the series connection method can preserve the fused feature information to a greater extent. The feature extraction backbone network contains four C3-LGSE modules, which in turn contain three, six, nine, and three Bottleneck-LGSE modules. The feature fusion neck network contains four C3-LGSE modules, each of which contains three Bottleneck-LGSE modules. The number of Bottleneck-LGSE modules is automatically determined by the product of the configuration file .yaml and the parameter depth_multiple; The Bottleneck-LGSE module, by replacing the traditional convolution module with the LGSE module, has about half the number of parameters of the traditional convolution module, which can reduce the number of parameters while ensuring the extraction of more features.

3. A lightweight detection method for concrete cracks, characterized in that, Based on the detection model described in any one of claims 1 to 2, the following steps are performed: S1. Input the concrete crack image into the feature extraction backbone network to complete the information extraction of crack features, and send the generated feature map tensor into the feature fusion neck network. The calculation process of the convolutional module LGSE involved in the backbone network is as follows: Y0=Conv p (K i ) Y1=Concat(Conv g2 (Conv g1 (Y0)),Conv g1 (Y0),Y0) Y2 = ChannelShuffle(Y1) K i+1 =ECA(Y2) Where Y0 represents the condition for input K i， i = 1, 2, ... n via Conv p The feature map extracted by point convolution is processed by Conv2 on Y0 and Y0. g Group convolution, Y0 twice Conv g The concatenated result of the group convolutions is Y1, Y2 represents the result of channel shuffling of Y1, and finally K i+1 This represents the feature map finally extracted after passing through the channel attention (ECA) module; S2. Feature Fusion: After fusing the multi-dimensional feature information extracted by the neck network with the backbone network, the information is fed into the prediction output. S3. Prediction Output: The prediction is completed, and the detection results of concrete cracks are output.

4. A computer storage medium, characterized in that, The storage medium stores at least one program instruction, which is loaded and executed by a processor to implement the concrete crack detection method as described in claim 3.

Images