FRP strip erosion detection method and device based on multi-scale target detection model

Abstract

Embodiments of the present disclosure provide a FRP strip erosion detection method and device based on a multi-scale target detection model, which comprises the following steps: introducing a SA-C2f attention mechanism module and a BiFPN-HMC feature fusion module into a YOLOv8n network model to construct a multi-scale target detection model; obtaining an erosion detection image sample set of an FRP strip, the sample set being labeled with a damage area and a damage category; training the multi-scale target detection model based on the sample set to obtain a trained multi-scale target detection model; and obtaining an FRP strip image to be detected, and using the trained multi-scale target detection model to perform erosion detection on the FRP strip image and output the position and characteristic information of the erosion defect.

Description

TECHNICAL FIELD

[0001] Embodiments of the present disclosure relate to the field of civil material detection, and in particular, to a FRP strip erosion detection method and device based on a multi-scale target detection model and a computer readable storage medium storing a computer program. BACKGROUND

[0002] FRP (Fiber Reinforced Polymer) has significant advantages in structural reinforcement due to its high strength, corrosion resistance, low weight, and other physical and chemical properties, which can significantly improve the load-bearing capacity, seismic performance, and durability of structures. In particular, in the field of civil engineering reinforcement, FRP is used to reinforce and repair infrastructure such as buildings, bridges, and tunnels. Although FRP has significant advantages in civil engineering, if the FRP material itself is damaged, it will directly affect the safety and stability of the engineering structure. In extreme environments, especially in desert areas, FRP materials face the challenge of wind and sand erosion. The fine particles contained in the wind and sand cause friction and scratches on the surface of FRP materials, leading to gradual erosion of the surface and damage to the internal structure, affecting the stability of the engineering structure and even causing significant economic losses. Therefore, it is necessary to detect and evaluate the damage to existing FRP materials.

[0003] The erosion damage of FRP materials in desert areas is often very subtle and complex. Traditional detection equipment and sensors may not be suitable for strong winds, sandstorms, and drastic temperature changes, resulting in unstable or distorted detection results. The damage caused by wind and sand erosion has multi-scale characteristics, which may cause slight surface damage over a large area, while causing more serious damage in local areas. The degree and morphology of the damage vary greatly, making it difficult for traditional damage detection methods such as manual inspection or simple visual inspection to effectively capture these small and widely distributed damages. Unlike traditional structural damage such as cracks and breakage, the damage caused by wind and sand erosion often manifests as subtle surface damage or micro-cracks. These damages are not obvious in the early stage and are usually difficult to accurately determine through conventional methods.

[0004] With the continuous development of deep learning technology, damage detection methods based on computer vision and neural networks have become a research hotspot. Compared with traditional manual inspection methods, deep learning algorithms have obvious advantages in handling complex, ambiguous, and multi-scale damage features. Especially in wind and sand erosion detection, deep learning algorithms can learn the characteristics of different types of damage through a large amount of training data, thereby achieving accurate detection and positioning of small damage. Due to the problems of extreme environment, small size of erosion range, and non-obvious erosion features in FRP material erosion detection, traditional deep learning algorithms cannot effectively detect both large and small erosion areas at the same time when dealing with such multi-scale erosion patterns, thereby missing some small damage or failing to accurately locate the damage position. SUMMARY

[0005] The embodiments described herein provide an FRP strip erosion detection method and device based on a multi-scale target detection model, and a computer readable storage medium storing a computer program.

[0006] According to a first aspect of the present disclosure, an FRP strip erosion detection method based on a multi-scale target detection model is provided, including: introducing a SA-C2f attention mechanism module and a BiFPN-HMC feature fusion module into a YOLOv8n network model to construct a multi-scale target detection model; obtaining an erosion detection image sample set of an FRP strip, the sample set being labeled with damage regions and damage categories; training the multi-scale target detection model based on the sample set to obtain a trained multi-scale target detection model; and obtaining an FRP strip image to be detected, and using the trained multi-scale target detection model to perform erosion detection on the FRP strip image and output the position and feature information of the erosion defect.

[0007] In some embodiments of the present disclosure, introducing the SA-C2f attention mechanism module and the BiFPN-HMC feature fusion module into the YOLOv8n network model to construct the multi-scale target detection model includes: introducing the SA-C2f attention mechanism module into the backbone network of the YOLOv8n network model to add spatial and channel attention to the input feature map; and integrating the BiFPN-HMC feature fusion module in the neck network of the YOLOv8n network model to perform feature refinement, weighting, and fusion on the feature map extracted by the backbone network to obtain a multi-scale feature map.

[0008] In some embodiments of the present disclosure, the SA-C2f attention mechanism module includes a spatial attention submodule, a channel attention submodule, and a feature reorganization unit.

[0009] The spatial attention sub-module includes a horizontal direction pooling layer, a vertical direction pooling layer, a splicing module, a horizontal direction convolution, a vertical direction convolution, an Fc linear transformation module and sigmoid activation functions in the vertical direction, a horizontal direction Sigmoid activation function and a vertical direction Sigmoid activation function, wherein the horizontal direction pooling layer is configured to perform a maximum pooling operation on an input feature map in a horizontal direction to obtain a horizontal direction feature map; the vertical direction pooling layer is configured to perform a maximum pooling operation on the input feature map in a vertical direction to obtain a vertical direction feature map; the splicing module is configured to splice the horizontal direction feature map and the vertical direction feature map to obtain a fused feature map; the horizontal direction convolution and the vertical direction convolution are configured to perform convolution operations on the fused feature map respectively to generate attention weights; and the Fc linear transformation module and the sigmoid activation functions in the vertical direction are configured to map the convolution output to a range of [0, 1] to be applied to the input feature map as the attention weights.

[0010] The channel attention sub-module is configured to perform weighted processing on channel features through a multi-layer perception to obtain dynamically adjusted channels; and the feature recombination unit is configured to combine spatial and channel information through channel rearrangement and feature fusion operations.

[0011] In some embodiments of the present disclosure, the SA-C2f attention mechanism module is configured to divide an input feature map with a dimension of CxHxW into g groups along a channel dimension to form grouped feature maps, where H and W represent the height and width of the feature map, and C represents the number of channels. Each sub-group obtains a sub-feature group through a spatial attention module and a corresponding weight coefficient; and the sub-feature group applies a stochastic unit to simultaneously construct channel attention and spatial attention through two branches of the spatial and channel branches to be applied to the input feature map as the attention weights.

[0012] In some embodiments of the present disclosure, the BiFPN-HMC feature fusion module includes a bidirectional feature pyramid network, a multi-convolution module, a feature weighting mechanism and a residual connection; the bidirectional feature pyramid network is configured to fuse feature maps of different scales through bidirectional feature flow in uplink and downlink paths; the multi-convolution module adopts a depth separable convolution and a variable dilated rate convolution to obtain eroded feature details; and the feature weighting mechanism is configured to assign dynamic weights to the feature maps of different scales.

[0013] ​In some embodiments of the present disclosure, the BiFPN-HMC feature fusion module is used to obtain feature maps from the backbone network; different scale feature maps are extracted through upsampling and downsampling operations; deep separable convolution and variable dilated rate convolution are used to refine the features of different scale feature maps; channel attention modules and spatial attention modules are used to respectively perform channel weighting adjustment and spatial information optimization; the feature maps after channel weighting adjustment and spatial information optimization are reorganized to output multi-scale feature maps.

[0014] In some embodiments of the present disclosure, the erosion detection image sample set of the FRP strip is obtained, and the sample set is labeled with damage areas and damage categories, including: obtaining FRP strip images simulating sand and wind erosion damage under different shooting angles and lighting conditions; labeling damage areas and damage categories in the FRP strip images.

[0015] In some embodiments of the present disclosure, the multi-scale target detection model is trained based on the sample set to obtain the trained multi-scale target detection model, including: dividing the labeled images into a training set and a validation set, the training set being used to adjust parameters of the multi-scale target detection model, and the validation set being used to evaluate performance of the multi-scale target detection model; inputting the training set into the multi-scale target detection model, updating model parameters through back propagation to minimize the BCEWithLogitsLoss loss function, and obtaining the trained multi-scale target detection model, the BCEWithLogitsLoss loss function being:

[0016] ;

[0017] In the formula, n is the number of samples, is the number of samples, is the true label of the i th sample, and takes a value of 0 or 1, is the prediction value of the model for the i th sample without an activation function.

[0018] ​​According to a second aspect of the present disclosure, a FRP strip erosion detection device based on a multi-scale target detection model is provided, comprising at least one processor; and at least one memory storing a computer program. When the computer program is executed by the at least one processor, the device is caused to: introduce a SA-C2f attention mechanism module and a BiFPN-HMC feature fusion module in a YOLOv8n network model, and construct a multi-scale target detection model; obtain an erosion detection image sample set of an FRP strip, the sample set being labeled with a damage area and a damage category; train the multi-scale target detection model based on the sample set, and obtain a trained multi-scale target detection model; and obtain an FRP strip image to be detected, and use the trained multi-scale target detection model to perform erosion detection on the FRP strip image, and output a position and feature information of an erosion defect.

[0019] According to a third aspect of the present disclosure, a computer-readable storage medium storing a computer program is provided, wherein the computer program, when executed by a processor, implements the steps of the FRP strip erosion detection method based on a multi-scale target detection model according to the first aspect of the present disclosure.

[0020] The FRP strip erosion detection method based on a multi-scale target detection model according to the embodiments of the present disclosure effectively enhances the model's ability to capture local features by introducing an attention mechanism, effectively combines spatial attention mechanisms and channel attention mechanisms using a SA-C2f module, thereby more comprehensively capturing the relationship between features and enhancing the channel and spatial information of feature maps; based on the BiFPN-HMC feature fusion module, effective information transmission and fusion can be performed between features of different scales, and large-scale and small-scale erosion damage can be more accurately detected. Compared with existing model detection methods, the method has higher precision, robustness, multi-scale detection capability, and can also provide accurate position and damage classification information, significantly improving the efficiency and accuracy of FRP strip erosion detection. BRIEF DESCRIPTION OF DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings of the embodiments will be briefly described below. It should be known that the drawings described below only relate to some embodiments of the present disclosure, and are not a limitation on the present disclosure, wherein:

[0022] Figure 1 is a structural schematic diagram of a Yolov8n model according to an embodiment of the present disclosure;

[0023] Figure 2 shows an exemplary flowchart of the FRP strip erosion detection method 200 based on a multi-scale target detection model according to an embodiment of the present disclosure;

[0024] Figure 3is a structural schematic diagram of an SA-C2f attention mechanism module according to an embodiment of the present disclosure.

[0025] Figure 4 is a structural schematic diagram of a BiFPN-HMC feature fusion module according to an embodiment of the present disclosure.

[0026] Figure 5 is an application environment diagram of FRP strip erosion detection according to an embodiment of the present disclosure.

[0027] Figure 6 is an FRP strip schematic diagram according to an embodiment of the present disclosure.

[0028] Figure 7 is a schematic block diagram of an FRP strip erosion detection device 700 based on a multi-scale target detection model according to an embodiment of the present disclosure.

[0029] It should be noted that the elements in the drawings are schematic and not drawn to scale. DETAILED DESCRIPTION

[0030] In order to make the purposes, technical solutions and advantages of the embodiments of the present disclosure clearer, the technical solutions of the embodiments of the present disclosure will be described clearly and completely below with reference to the drawings. Obviously, the described embodiments are part of the embodiments of the present disclosure, rather than all the embodiments. Based on the described embodiments of the present disclosure, all other embodiments obtained by a person of ordinary skill in the art without any inventive effort also belong to the scope of protection of the present disclosure.

[0031] Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this present subject matter belongs. It will be further understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the specification and relevant art and will not be interpreted in an idealized or overly formal sense unless expressly so defined herein.

[0032] The Yolov8n model is a lightweight version in the YOLO (You Only Look Once) series, which consists of four main parts: input, backbone network, neck network, and detection head. Figure 1 is a structural schematic diagram of a Yolov8n model according to an embodiment of the present disclosure. Referring to Figure 1As shown, the Input input end and the Backbone backbone network introduce SPPF (Spatial Pyramid Pooling), which aims to capture features of different scales through different size pooling operations, helping the model to aggregate information at different spatial scales to improve the model's detection ability for multi-scale objects. The Backbone backbone network also includes multiple CBS convolution modules (Convolution + BatchNorm + SiLU) and C2f modules (CSP + 2 Feedforward Networks), and the C2f module combines CSP (Cross Stage Partial connections) and two feedforward neural networks (Feedforward Networks) for further feature extraction and processing. The Neck neck network is responsible for further processing and fusion of features from the backbone network, using a feature pyramid network (FPN) to fuse information at multiple scales to enhance the model's detection ability for objects of different sizes. The Head detection head is responsible for the final object classification, positioning and confidence scoring. Object detection is achieved by directly predicting bounding boxes, class labels and confidence.

[0033] To solve the problem of low multi-scale erosion detection precision caused by extreme environment, small erosion range size, and non-obvious erosion features in FRP material erosion detection, the embodiments of the present disclosure improve the target detection model based on YOLOv8n, by introducing SA-C2f attention mechanism modules and BiFPN-HMC feature fusion modules in the feature extraction network of the YOLOv8n model to cope with multi-scale erosion detection application scenarios.

[0034] Figure 2 An exemplary flowchart of an FRP strip erosion detection method 200 based on a multi-scale target detection model according to an embodiment of the present disclosure is shown.

[0035] In Figure 2 At block S202, SA-C2f attention mechanism modules and BiFPN-HMC feature fusion modules are introduced into the YOLOv8n network model to construct a multi-scale target detection model.

[0036] In some embodiments of the present disclosure, to improve the precision of multi-scale target detection, first, the C2f module of the backbone network of the YOLOv8n deep learning model is replaced with the SA-C2f module to add spatial and channel attention to the input feature map. The BiFPN-HMC feature fusion module is integrated into the neck network of the YOLOv8n network model to refine, weight and fuse the feature maps extracted by the backbone network, and obtain multi-scale feature maps.

[0037] The SA-C2f module combines a spatial attention mechanism and a channel attention mechanism, so that the network is more intelligent when extracting features. The spatial attention mechanism helps the network to automatically focus on important spatial regions in the image, and to strengthen the features of these regions while suppressing unimportant background parts. The channel attention mechanism helps the model to focus on the most representative feature channels by assigning weights to each channel, and to suppress irrelevant channels. In the BiFPN module, a HMC (Hybrid Multi-scale Cross Attention) module is integrated. This feature fusion method further optimizes the fusion of information of different scales by introducing a multi-scale cross attention mechanism (Cross Attention). Specifically, HMC uses a multi-level cross attention mechanism to weight and refine features of different scales, to ensure that the features of each scale can be properly represented in the final output.

[0038] In some embodiments of the present disclosure, the SA-C2f module includes a spatial attention sub-module, a channel attention sub-module, and a feature reorganization unit. The spatial attention sub-module includes a horizontal direction pooling layer, a vertical direction pooling layer, a concatenation module, a horizontal direction convolution, a vertical direction convolution, an Fc linear transformation module, and sigmoid activation functions in the horizontal and vertical directions.

[0039] The horizontal direction pooling layer is used to perform a maximum pooling operation on the input feature map in the horizontal direction to obtain a horizontal direction feature map. Similarly to the horizontal direction pooling, the vertical direction pooling layer is used to perform a maximum pooling operation on the input feature map in the vertical direction to obtain a vertical direction feature map. The concatenation module is used to concatenate the horizontal direction feature map and the vertical direction feature map to obtain a fused feature map. The horizontal direction convolution and the vertical direction convolution perform convolution operations on the fused feature map, respectively, to learn the attention weights in the horizontal and vertical directions, helping the model to focus on key information in different directions. The Fc linear transformation module and sigmoid activation function are used to map the convolution output to the range [0, 1] as the final attention weight applied to the input feature map. The horizontal direction Sigmoid activation function and the vertical direction Sigmoid activation function are used to control the emphasis degree of spatial features, and their role is to generate the final spatial attention map according to the attention in the horizontal and vertical directions.

[0040] The channel attention submodule uses a multilayer perceptron to weight channel features, resulting in dynamically adjusted channels. The multilayer perceptron weights the outputs of each channel to generate channel attention weights. Finally, the channel attention mechanism assigns an importance value to each channel, allowing the network to focus more on specific, key channel features and enhancing the model's discriminative ability. The feature reorganization unit combines spatial and channel information through channel rearrangement and feature fusion operations. Rearranging the channels changes their order, allowing for better integration of information between channels and resulting in a more comprehensive output feature containing more information.

[0041] Figure 3 This is a schematic diagram of the SA-C2f attention mechanism module structure according to an embodiment of this disclosure. (Refer to...) Figure 3 As shown, for a given input feature map X of dimension C×H×W, where H and W represent the height and width of the feature map, and C represents the number of channels, the SA-C2f attention mechanism module can divide the feature map X into g groups along the channel dimension, forming grouped feature maps. ,in Each subgroup The number of channels is C / g; each subgroup Sub-feature groups are obtained through a Spatial Attention (SA) module and corresponding attention weight coefficients. These sub-feature groups utilize a random unit to simultaneously construct channel and spatial attention through both spatial and channel branches, and then combine this attention information from both dimensions to weight the input feature map. This weighting mechanism allows the model to better focus on important spatial regions and key pathways, thereby improving the model's performance on specific tasks.

[0042] In some embodiments of the present disclosure, the BiFPN-HMC feature fusion module includes a bidirectional feature pyramid network, a multi-convolution module, a feature weighting mechanism, and a residual connection. The bidirectional feature pyramid network BiFPN is used to fuse feature maps of different scales through bidirectional feature flow in the uplink and downlink paths. This means that the feature maps can not only be fused between different scales, but also can strengthen feature expression at multiple levels from fine-grained to coarse-grained. The multi-convolution module uses depth separable convolution and variable dilation rate convolution to obtain eroded feature details. Using depth separable convolution to decompose the regular convolution operation into depth convolution and pointwise convolution can reduce the amount of calculation, thereby greatly improving the calculation efficiency. Using the dilation rate convolution, the receptive field is increased by adjusting the dilation rate of the convolution kernel, which helps to capture more context information, especially for fine-grained features such as erosion features. The feature weighting mechanism is used to assign dynamic weights to feature maps of different scales. By adaptively adjusting the weights of features of different scales, it can ensure that key features are fully utilized, thereby optimizing the information flow in the feature fusion process. The residual connection in the BiFPN module can effectively alleviate the gradient vanishing problem, ensuring that information can be smoothly transmitted in the deep network, thereby improving the stability of training.

[0043] Figure 4 is a structural diagram of the BiFPN-HMC feature fusion module according to the embodiments of the present disclosure. Referring to Figure 4 As shown in the figure, the feature maps extracted from the Backbone backbone network part are input into the BiFPN-HMC feature fusion module, and different scale feature maps (C3, C4, C5, C6, C7) are extracted through upsampling and downsampling operations, which helps to ensure that features of different scales can be fully fused and avoid feature loss due to scale difference.

[0044] The multi-convolution module DDS uses depth separable convolution (k=1, d=1, s=1) and variable dilation rate convolution (k=3, d=3, s=1) to refine features of different scales; through the channel attention module (CA module) and the spatial attention module (SAM module), the channel weighting adjustment and the spatial information optimization are performed respectively, so as to enhance the expression of key features, optimize the spatial information of the FRP strip erosion image, highlight the key area and suppress the background noise, thereby improving the detection accuracy of the erosion features. The feature maps after channel weighting adjustment and spatial information optimization are reorganized, the channel invariance and spatial specificity of involution are used, and a convolution kernel is generated for each position. In order to reduce parameters, grouping is performed by channel, and the channel parameters within the group are shared. Finally, multi-scale feature maps (P3, P4, P5, P6, P7) are output.

[0045] return Figure 2 As shown, then in box S204, an erosion detection image sample set of FRP strips is obtained, and the sample set is labeled with the damage area and damage type.

[0046] The images in the sample set simulate the damage caused by wind and sand erosion to glass fiber reinforced plastic (FRP) strips. Images of FRP strips simulating wind and sand erosion damage can be acquired under different shooting angles and lighting conditions. Then, the damaged areas and damage types are labeled in the FRP strip images. For example, specialized labeling tools are used to annotate the samples, ensuring that the erosion damage in each image is accurately identified. Typically, these labeling tools allow users to draw boxes, outlines, etc., on the images and label the damaged areas. The labeling tools will mark the erosion damage areas involved in the images, such as cracks, scratches, localized corrosion, etc. Each damaged area will be assigned a corresponding label indicating the damage type and location.

[0047] Subsequently, in box S206, the multi-scale target detection model is trained based on the sample set to obtain the trained multi-scale target detection model.

[0048] In some embodiments of this disclosure, the labeled images are divided into a training set and a validation set. The training set is used to adjust the parameters of the multi-scale object detection model, and the validation set is used to evaluate the performance of the multi-scale object detection model. To train and validate the model, the sample set is randomly divided in an 8:2 ratio. That is, 80% of the samples are used for the training set, and 20% of the samples are used for the validation set.

[0049] The training set is input into a multi-scale object detection model. The BCEWithLogitsLoss loss function is minimized through backpropagation, and the model parameters are updated to obtain the trained multi-scale object detection model. BCEWithLogitsLoss is a loss function used for binary classification tasks, suitable for handling class prediction for each pixel or region in object detection. It combines the Sigmoid activation function and binary cross-entropy loss. The BCEWithLogitsLoss loss function is:

[0050] ;

[0051] In the formula, For the sample size, For the first The true label of each sample, with a value of either 0 or 1. For the model to the first The predicted value of each sample before activation. It is the output after Sigmoid activation.

[0052] After the training is completed, the trained multi-scale target detection model can be evaluated using the validation set to check its performance in the actual sand erosion damage detection task. The evaluation indicators usually include accuracy, recall, F1 score, mean average precision (mAP), etc., which are selected according to the task requirements. Save the trained model as a file, for example, duoopti_yolo_model.pth, so that the model can be loaded in real-time detection of sand erosion damage in practical applications.

[0053] Finally, in block S208, the FRP strip image to be detected is obtained, and the trained multi-scale target detection model is used to detect the erosion of the FRP strip image, and the position and feature information of the erosion defect are output.

[0054] The FRP strip erosion detection based on the multi-scale target detection model provided by the embodiments of the present application can be applied to the application environment as shown in the figure. Figure 5 Figure 5 is an application environment diagram of the FRP strip erosion detection according to the embodiments of the present disclosure. The terminal 102 image information is sent to the server 104 by means of unmanned aerial vehicle shooting, etc. The data storage system can be separately set, or can be integrated on the server 104, or can be placed on the cloud or other servers. After the server 104 receives the FRP material image, for the FRP strip image, the server 104 uses the trained multi-scale target detection model to detect the erosion of the FRP material. In addition, in some embodiments, the FRP erosion damage detection method using the trained multi-scale target detection model can also be realized by the server 104 alone, and the server 104 can obtain the FRP strip image from the data storage system and perform erosion detection on the FRP strip image. The terminal 102 can be, but is not limited to, a ship, an airplane, a bridge, a wind turbine blade, etc. mechanical or building structure.

[0055] Figure 6 is an FRP strip diagram according to an embodiment of the present disclosure. The FRP strip is a strip-shaped structure made of fiber reinforced plastic (Fiber Reinforced Polymer, FRP for short). FRP is a composite material mainly composed of high-strength fibers (such as glass fibers, carbon fibers or aramid fibers) and polymer matrix (such as epoxy resin, polyester resin, etc.). In addition, it also includes reinforcing sheets and steel plates. This combination makes FRP materials have excellent properties such as high strength, light weight, corrosion resistance, and high temperature resistance.

[0056] ​The FRP strip erosion detection method based on the multi-scale target detection model provided by the embodiment can be applied in mechanical or civil engineering structure inspection scenarios. The mechanical or building structure inspection scenario includes an image acquisition link and an image recognition link. The FRP material erosion detection method based on the DuoOpti-YOLO deep learning algorithm provided by the embodiment belongs to the image recognition link.

[0057] In some embodiments of the present disclosure, first, the FRP strip images to be detected are acquired. These images can come from various types of sensors or devices, such as high-definition cameras, industrial detection systems, or unmanned aerial vehicles, etc. In this stage, it is ensured that the image quality is high and can clearly show the surface features of the FRP strip, so as to facilitate the subsequent erosion detection. Before processing, certain preprocessing operations are usually performed on the images, such as denoising, image enhancement, contrast adjustment, etc., to improve the clarity and detail information of the images, and to ensure that the erosion defects can be more prominent.

[0058] Then, the trained multi-scale target detection model is loaded. This model has been trained based on the labeled image sample set, and has learned the features of different types of erosion defects (such as cracks, corrosion, delamination, etc.) and their features at different scales. The model includes a SA-C2f attention mechanism module and a BiFPN-HMC feature fusion module, which enables it to effectively extract feature information at different scales in the image to be detected and focus on important areas.

[0059] The FRP strip image to be detected is input into the trained model. After processing the image, the model automatically performs feature extraction, target detection, and analyzes the potential damage features of different regions. The position of the erosion region (upper left and lower right coordinates) is marked in the form of a bounding box. These frames surround the erosion defects and mark the region where they are located. In addition to the position, the model also identifies different types of erosion defects and provides corresponding feature information. These feature information may include: damage type: such as corrosion, crack, delamination, etc. Damage degree: such as mild corrosion, moderate corrosion, severe corrosion, etc., which is usually judged by the size and shape of the defect. Defect area or size: the area of the detection region, which is usually proportional to the damage degree. Damage morphology: such as the length and width of the crack, etc., which provides a detailed description of the damage morphology. These information can be presented in a visual way, for example, the position of the erosion defect is marked with a bounding box, and the damage type and severity are marked on the image.

[0060] In order to verify the effectiveness and superiority of the multi-scale target detection model provided by the embodiments of the present disclosure, the improved model is compared with the detection effect of the current YOLOv8n model. It is verified that the algorithm proposed in the present application is superior to the YOLOv8n model in detection accuracy, and the detection reaches an average precision of 95.2%, an increase of 3.3%, a recall of 3.7%, an F1 score of 3.5%, and an mAP of 3.1%.

[0061] Figure 7 is a schematic block diagram of an FRP strip erosion detection device 700 based on a multi-scale target detection model according to an embodiment of the present disclosure. As shown in Figure 7 , the device 700 can include a processor 710 and a memory 720 storing a computer program. When the computer program is executed by the processor 710, the device 700 can perform the steps of the FRP strip erosion detection method 200 based on a multi-scale target detection model as shown in Figure 2 In one example, the device 700 can be a computer device or a cloud computing node. The device 700 can: introduce a SA-C2f attention mechanism module and a BiFPN-HMC feature fusion module in a YOLOv8n network model, construct a multi-scale target detection model; obtain an erosion detection image sample set of an FRP strip, the sample set being labeled with a damage area and a damage category; train the multi-scale target detection model based on the sample set, and obtain a trained multi-scale target detection model; and obtain an FRP strip image to be detected, and use the trained multi-scale target detection model to perform erosion detection on the FRP strip image, and output the position and feature information of the erosion defect.

[0062] In embodiments of the present disclosure, the processor 710 can be, for example, a central processing unit (CPU), a microprocessor, a digital signal processor (DSP), a processor based on a multi-core processor architecture, etc. The memory 720 can be any type of memory implemented using data storage technology, including but not limited to random access memory, read only memory, semiconductor-based memory, flash memory, disk storage, etc.

[0063] In addition, in embodiments of the present disclosure, the device 700 can also include an input device 730, such as a keyboard, a mouse, etc., for inputting the erosion detection image sample set of the FRP strip and the FRP strip image to be detected. In addition, the device 700 can also include an output device 740, such as a display, etc., for outputting the position and feature information of the erosion defect.

[0064] In other embodiments of the present disclosure, a computer readable storage medium storing a computer program is also provided, wherein the computer program can implement the FRP strip erosion detection method 200 based on a multi-scale target detection model as shown in Figure 2Steps of the FRP strip erosion detection method 200 based on a multi-scale target detection model shown.

[0065] In summary, according to the FRP strip erosion detection method based on a multi-scale target detection model of the embodiments of the present disclosure, the attention mechanism is introduced to effectively enhance the model's ability to capture local features, the SA-C2f module effectively combines the spatial attention mechanism and the channel attention mechanism, thereby more comprehensively capturing the relationship between features and enhancing the channel and spatial information of the feature map; based on the BiFPN-HMC feature fusion module, effective information transmission and fusion can be performed between features of different scales, and large-scale and small-scale erosion damage can be more accurately detected. Compared with the existing model detection method, it has higher precision, robustness, multi-scale detection capability, and can also provide accurate position and damage classification information, significantly improving the efficiency and accuracy of FRP strip erosion detection.

[0066] The flow and block diagrams in the drawings show the architectural, functional, and operational aspects of possible implementations of apparatuses and methods according to various embodiments of the present disclosure. In this regard, each block in the flow and block diagrams can represent a module, a segment, or a portion of code that comprises one or more executable instructions for implementing the specified logical functions. In some alternative implementations, the functions noted in the blocks can occur in a different order than that noted in the figures. For example, two blocks shown in succession can in fact be executed substantially concurrently or in the opposite order, depending on the functionality involved. It is also noted that each block in the block and / or flow diagrams and combinations of blocks in the block and / or flow diagrams can be implemented by dedicated hardware-based systems that perform the specified functions or combinations of dedicated hardware and computer instructions.

[0067] Unless specifically stated otherwise, as used herein and in the appended claims, the singular form "a," "an" and "the" include plural references unless the context clearly dictates otherwise. Accordingly, when a singular form is used, it is intended to include a plural form as well. Similarly, the terms "comprises", "comprising", "includes", "including" and the like are to be construed in an inclusive, rather than an exclusive, sense unless explicitly indicated to be such. Similarly, the terms "include", "comprise", and "or" should be construed to be inclusive rather than exclusive, unless explicitly indicated to be such. Where the term "example" is used in the following disclosure, particularly in the context of a series of terms, the term "example" is merely an example of and is not to be construed as a limitation of the scope of the disclosure.

[0068] Further aspects and ranges of adaptation become apparent from the description provided herein. It should be understood that various aspects of the application can be implemented alone or in combination with one or more other aspects. It should also be understood that the description and specific examples herein are intended to be illustrative only and are not intended to limit the scope of the present application.

[0069] The above detailed description of several embodiments of the disclosure has been presented for the purposes of illustration and description. It is apparent to those skilled in the art that various modifications and variations can be made to the embodiments of the disclosure without departing from the spirit and scope of the disclosure. The scope of protection of the disclosure is defined by the appended claims.

Claims

1. A method for detecting FRP strip erosion based on a multi-scale target detection model, characterized in that, include: A multi-scale object detection model is constructed by introducing the SA-C2f attention mechanism module and the BiFPN-HMC feature fusion module into the YOLOv8n network model. This includes: introducing the SA-C2f attention mechanism module into the backbone network of the YOLOv8n network model to add spatial and channel attention to the input feature map; and integrating the BiFPN-HMC feature fusion module into the neck network of the YOLOv8n network model to refine, weight, and fuse the feature maps extracted by the backbone network to obtain multi-scale feature maps. The SA-C2f attention mechanism module includes a spatial attention submodule, a channel attention submodule, and a feature reconstruction unit. The spatial attention submodule includes a horizontal pooling layer, a vertical pooling layer, a concatenation module, a horizontal convolution, a vertical convolution, an Fc linear transformation module, a sigmoid activation function, a horizontal sigmoid activation function, and a vertical sigmoid activation function. The horizontal pooling layer performs max pooling on the input feature map in the horizontal direction to obtain a horizontal feature map; the vertical pooling layer performs max pooling on the input feature map in the vertical direction to obtain a vertical feature map. The feature map is constructed by: a concatenation module that concatenates the horizontal and vertical feature maps to obtain a fused feature map; a horizontal and vertical convolution that performs convolution operations on the fused feature map to generate attention weights; an Fc linear transformation module and a sigmoid activation function that map the convolution output to the range [0,1] to be used as attention weights applied to the input feature map; a channel attention submodule that performs weighted processing on the channel features through a multilayer perceptron to obtain dynamically adjusted channels; and a feature recombination unit that combines spatial and channel information through channel rearrangement and feature fusion operations. The BiFPN-HMC feature fusion module includes a bidirectional feature pyramid network, a multi-convolutional module, a feature weighting mechanism, and residual connections. The bidirectional feature pyramid network is used to fuse feature maps of different scales through bidirectional feature flow in the up and down paths. The multi-convolutional module uses depthwise separable convolution and variable dilation rate convolution to obtain erosion feature details. The feature weighting mechanism is used to assign dynamic weights to feature maps of different scales. Obtain an image sample set of erosion detection images of FRP strips, wherein the sample set is labeled with damage areas and damage categories; The multi-scale object detection model is trained based on the sample set to obtain a trained multi-scale object detection model; and The FRP strip image to be detected is obtained, and the trained multi-scale target detection model is used to perform erosion detection on the FRP strip image, outputting the location and feature information of the erosion defects.

2. The FRP strip erosion detection method based on a multi-scale target detection model according to claim 1, characterized in that, The SA-C2f attention mechanism module is used to divide the input feature map of dimension C×H×W into g groups along the channel dimension, forming grouped feature maps X=[X,...,X], where H and W represent the height and width of the feature map, and C represents the number of channels. k ,...,X g ], where k∈[1,g]; each subgroup X k Sub-feature groups are obtained through a spatial attention module and corresponding weight coefficients; the sub-feature groups are constructed simultaneously through spatial and channel branches, using a random unit to construct channel attention and spatial attention, which are then applied as attention weights to the input feature map.

3. The FRP strip erosion detection method based on a multi-scale target detection model according to claim 1, characterized in that, The BiFPN-HMC feature fusion module is used to obtain feature maps from the backbone network; extract feature maps of different scales through upsampling and downsampling operations; refine the feature maps of different scales by using depthwise separable convolution and variable dilation rate convolution; and perform channel weighting adjustment and spatial information optimization through channel attention module and spatial attention module respectively, and reorganize the feature maps after channel weighting adjustment and spatial information optimization to output multi-scale feature maps.

4. The FRP strip erosion detection method based on a multi-scale target detection model according to claim 1, characterized in that, The acquisition of the erosion detection image sample set of FRP strips, wherein the sample set is labeled with damage areas and damage categories, includes: FRP strip images simulating wind and sand erosion damage were obtained under different shooting angles and lighting conditions; The damaged areas and damage types are marked in the FRP strip image.

5. The FRP strip erosion detection method based on a multi-scale target detection model according to claim 1, characterized in that, The step of training the multi-scale target detection model based on the sample set to obtain a trained multi-scale target detection model includes: The labeled images are divided into a training set and a validation set. The training set is used to adjust the parameters of the multi-scale object detection model, and the validation set is used to evaluate the performance of the multi-scale object detection model. The training set is input into the multi-scale object detection model, and the BCEWithLogitsLoss loss function is minimized through backpropagation to update the model parameters. The model performance is evaluated using the validation set to obtain the trained multi-scale object detection model. The BCEWithLogitsLoss loss function is: In the formula, N is the sample size. For the first i The true label of each sample, with a value of either 0 or 1. For the model to the first The predicted value of each sample before activation.

6. An FRP strip erosion detection device based on a multi-scale target detection model, characterized in that, include: At least one processor; as well as At least one memory storing a computer program; When the computer program is executed by the at least one processor, the device performs the steps of the FRP strip erosion detection method based on the multi-scale target detection model according to any one of claims 1 to 5.

7. A computer-readable storage medium storing a computer program, characterized in that, The computer program, when executed by a processor, implements the steps of the FRP strip erosion detection method based on a multi-scale target detection model as described in any one of claims 1 to 5.

Images