Carbon fiber reinforced polymer concrete surface defect detection method and system based on neural network

Abstract

The present application discloses a method and system for detecting surface defects of carbon fiber reinforced polymer concrete based on neural network. The method comprises: using an image acquisition device to acquire an image of the surface of carbon fiber reinforced polymer concrete to input into an efficient backbone network; inputting the output feature vectors of multiple target network layers in the efficient backbone network into a PANet structure network, the PANet structure network is used to perform up- and down-sampling on the output feature vectors of the multiple target network layers to fuse multi-scale features; and using a head network including one or more decoupling heads based on an Anchor‑free strategy to process the output of the PANet structure network, the head network performs reasoning by using decoupled classification branches and regression branches to locate the position of the surface defects of the carbon fiber reinforced polymer concrete. The scheme of the present application can improve detection accuracy, reduce detection costs, and enhance adaptability to different detection scenarios.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence, and in particular to a method and system for detecting surface defects of carbon fiber reinforced polymer concrete based on a neural network. Background Art

[0002] CFRP (carbon fiber reinforced polymer) concrete is a new type of composite material with advantages such as high strength, high toughness and corrosion resistance. It is widely used in buildings, bridges, tunnels and other fields, especially in the reinforcement, repair and transformation of concrete structures, showing great potential and advantages. However, CFRP concrete may also produce various surface defects during use, such as cracks, peeling, bubbles, etc. These defects may be caused by construction quality, environmental factors, load effects, material aging and other reasons, which will affect the bonding properties, mechanical properties and durability of CFRP concrete, and even lead to structural failure and accidents.

[0003] Traditional CFRP concrete surface defect detection methods are mainly divided into two categories: manual visual inspection and instrument detection based on physical signals. Manual visual inspection is the most commonly used method. It manually observes, judges, and records the defects on the surface of CFRP concrete. This method requires a lot of manpower, time, and resources, and it is difficult to meet the needs of large-scale and high-frequency detection. The results of manual visual inspection are affected by factors such as the experience, skills, vision, and fatigue of the inspectors, and lack objectivity and consistency. Manual visual inspection is difficult to detect small, hidden, and complex defects, and is prone to problems such as missed detection, false detection, and repeated detection. Instrument detection based on physical signals is another commonly used method. It detects the defects on the surface of CFRP concrete by using physical signals such as ultrasound, infrared, and electromagnetic waves. This method requires the use of professional instruments and equipment, which is costly, difficult to maintain, and susceptible to environmental interference. Instrument detection based on physical signals requires professional operators, is complex to operate, and requires preprocessing, analysis, and interpretation of physical signals, which is prone to misjudgment. Instrument detection based on physical signals is difficult to adapt to different detection scenarios, such as different defect types, shapes, sizes, and locations.

[0004] Therefore, there is an urgent need for a technical solution that can improve detection accuracy, reduce detection costs, and enhance adaptability to different detection scenarios. Summary of the invention

[0005] In order to solve the shortcomings of the prior art, the embodiment of the present application provides a method and system for detecting surface defects of carbon fiber reinforced polymer concrete based on a neural network. The present application solves the technical problems of the prior art such as low detection accuracy, high detection cost, and poor adaptability to different detection scenarios.

[0006] A method for detecting surface defects of carbon fiber reinforced polymer concrete based on a neural network comprises: using an image acquisition device to obtain an image of the surface of carbon fiber reinforced polymer concrete to input into an efficient backbone network; wherein the efficient backbone network comprises one or more CSPDarknet network structures based on a C2f module; wherein a method for calculating the number of CSPDarknet network structures based on the C2f module comprises: Among them, X' represents the number after calculation optimization, N represents the matrix vector of the image dataset, X represents the number that is not calculated and optimized, and T represents the expected time parameter. Represents the Frobenius norm; inputs the output feature vectors of multiple target network layers in the efficient backbone network into the PANet structure network, the PANet structure network is used to up-sample and down-sample the output feature vectors of the multiple target network layers to fuse multi-scale features; and uses a head network including one or more decoupled heads based on an Anchor-free strategy to process the output of the PANet structure network, the head network performs reasoning by using decoupled classification branches and regression branches to locate the positions of surface defects of the carbon fiber reinforced polymer concrete.

[0007] In a possible implementation, the CSPDarknet network structure based on the C2f module includes one or more CNN models and multiple RepViT modules; when using an image acquisition device to obtain an image of the surface of carbon fiber reinforced polymer concrete to input into an efficient backbone network, it includes: using a CNN model to extract local features of the image, and the local features are used for segmentation to obtain a first local feature and a second local feature; inputting the first local feature into a first RepViT module, and the output of the first RepViT module is used to perform a first residual connection with the second local feature to input into a second RepViT module; inputting the output of the second RepViT module into a third RepViT module to perform a second residual connection; and using the result of the second residual connection output as the input of a CNN model in another CSPDarknet network structure based on the C2f module.

[0008] In one possible implementation, the attention mechanism in the RepViT module includes: SE(X)=σ(FC2(δ(FC1(X)))), wherein X represents the input feature map, SE(X) represents the Squeeze-and-Excitation operation on the feature map X, FC1 and FC2 represent the first fully connected layer and the second fully connected layer respectively, δ represents the attention activation function, and σ represents the Sigmoid function.

[0009] In one possible implementation, the attention mechanism in the RepViT module includes: Among them, EMA represents the exponential moving average, t represents the current time, and θ t-i It represents the value at the past ti moments, and β represents the attenuation factor.

[0010] In a possible implementation, the RepViT module further includes: connecting the input and output of multiple 1x1 convolutions for adjusting the channel dimension through residual connections to form a feedforward network.

[0011] In a possible implementation, before using an image acquisition device to acquire an image of the surface of carbon fiber reinforced polymer concrete for input into an efficient backbone network, it also includes: using an image acquisition device to acquire an image of the surface of carbon fiber reinforced polymer concrete, wherein the image is used for classification and numbering to generate an initial image set; importing the initial image set into a label making software to annotate the bounding boxes of defects in the image and the corresponding categories; and integrating the annotation information including the position, size and category label of the bounding box with the initial image set into a complete data set, wherein the complete data set is used to be divided into a training set and a validation set according to a ratio for training the model network.

[0012] A possible implementation method further includes: Wherein, a represents the distribution ratio of positive samples in the training set, t represents the number of training iterations, t0 represents the number of initial iterations, and k represents a parameter for controlling the dynamic adjustment speed.

[0013] A possible implementation method further includes: Wherein, LOSS represents the value of the loss function during the training process, λ represents the control coefficient, q represents the probability distribution parameter, and L PIoU Indicates the pixel IoU value.

[0014] The embodiment of the present application also provides a carbon fiber reinforced polymer concrete surface defect detection system based on a neural network, comprising: an acquisition module, the acquisition module is used to acquire an image of the surface of the carbon fiber reinforced polymer concrete for input into an efficient backbone network; a feature extraction module, the feature extraction module is used to input the output feature vectors of multiple target network layers in the efficient backbone network into a PANet structure network, the PANet structure network is used to up-sample and down-sample the output feature vectors of the multiple target network layers to fuse multi-scale features; and a detection module, the detection module is used to process the output of the PANet structure network using a head network including one or more decoupled heads based on an anchor-free strategy, the head network performs reasoning by using decoupled classification branches and regression branches to locate the position of the surface defects of the carbon fiber reinforced polymer concrete.

[0015] An embodiment of the present application also provides a carbon fiber reinforced polymer concrete surface defect detection device based on a neural network, comprising: a processor, a memory, and a system bus; wherein the processor and the memory are connected via the system bus; the memory is used to store one or more programs, and the one or more programs include instructions, and when the instructions are executed by the processor, the processor executes any one of the methods described above.

[0016] Through the neural network-based carbon fiber reinforced polymer concrete surface defect detection method and system provided above, the embodiment of the present application can improve detection accuracy, reduce detection costs, and enhance adaptability to different detection scenarios by combining artificial intelligence image recognition technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] Figure 1 A schematic diagram of a method for detecting surface defects of carbon fiber reinforced polymer concrete based on a neural network provided in an embodiment of the present application;

[0019] Figure 2 A schematic diagram of another method for detecting surface defects of carbon fiber reinforced polymer concrete based on a neural network provided in an embodiment of the present application;

[0020] Figure 3 An exemplary architecture diagram of a neural network-based carbon fiber reinforced polymer concrete surface defect detection model provided in an embodiment of the present application. DETAILED DESCRIPTION

[0021] Various exemplary embodiments of the present disclosure will now be described in detail with reference to the accompanying drawings. It should be noted that the relative arrangement of components and steps, numerical expressions and numerical values ​​set forth in these embodiments do not limit the scope of the present disclosure unless otherwise specifically stated.

[0022] Those skilled in the art can understand that the terms "first", "second" and the like in the embodiments of the present disclosure are only used to distinguish different steps, devices or modules, etc., and neither represent any specific technical meaning nor represent the necessary logical order between them. It should also be understood that in the embodiments of the present disclosure, "multiple" can refer to two or more, and "at least one" can refer to one, two or more. It should also be understood that for any component, data or structure mentioned in the embodiments of the present disclosure, in the absence of explicit limitation or contrary revelation given in the context, it can generally be understood as one or more. In addition, the term "and / or" in the present disclosure is only a description of the association relationship of the associated objects, indicating that there can be three relationships, for example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in the present disclosure generally indicates that the associated objects before and after are an "or" relationship. It should also be understood that the description of each embodiment in the present disclosure emphasizes the differences between the embodiments, and the same or similar parts can refer to each other. For the sake of brevity, they will not be repeated one by one.

[0023] At the same time, it should be understood that, for ease of description, the sizes of the various parts shown in the drawings are not drawn according to the actual proportional relationship. The following description of at least one exemplary embodiment is actually only illustrative and is by no means intended to limit the present disclosure and its application or use. The techniques, methods and devices known to ordinary technicians in the relevant fields may not be discussed in detail, but where appropriate, the techniques, methods and devices should be considered part of the specification. It should be noted that similar numbers and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further discussed in subsequent drawings.

[0024] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0025] Figure 1 A schematic diagram of a method for detecting surface defects of carbon fiber reinforced polymer concrete based on a neural network provided in an embodiment of the present application. Figure 1As shown, at step S101, an image acquisition device is used to acquire an image of the surface of carbon fiber reinforced polymer concrete to input into an efficient backbone network. The efficient backbone network includes one or more CSPDarknet network structures based on C2f modules; the number calculation method of the CSPDarknet network structure based on C2f modules includes:

[0026]

[0027] Among them, X' represents the number after calculation optimization, N represents the matrix vector of the image dataset, X represents the number that is not calculated and optimized, and T represents the expected time parameter. Represents the Frobenius norm. Specifically, the algorithm of this application uses an efficient backbone network, based on the CSPDarknet structure of the C2f module. The number of channels of each BottleNeck input Tensor in the C2f module is only 0.5 times that of the previous level, so the amount of calculation is significantly reduced. On the other hand, the increase in gradient flow can also significantly improve the convergence speed and convergence effect, providing powerful feature extraction capabilities.

[0028] At step S102, the output feature vectors of multiple target network layers in the efficient backbone network are input into the PANet structure network, which is used to upsample and downsample the output feature vectors of the multiple target network layers to fuse multi-scale features. In an implementation scenario, the algorithm's Neck adopts the PANet structure. Unlike the commonly used FPN structure, PANet is a bidirectional path network that introduces a bottom-up path, making it easier for bottom-level information to be transmitted to the top layer. Multi-scale feature fusion and up-down sampling are achieved, enhancing the expressiveness of features.

[0029] At step S103, a head network including one or more decoupling heads based on the Anchor-free strategy is used to process the output of the PANet structure network, and the head network is inferred by using decoupled classification branches and regression branches to locate the position of the surface defects of the carbon fiber reinforced polymer concrete. In the head network of the present application, the decoupling head and the Anchor-free strategy are selected, and there is no longer the previous Objectness branch, only the decoupled classification and regression branches. Compared with the Anchor-based strategy used in the past. The difference lies in whether the prior frame is set in the training process. The Anchor-based strategy is very dependent on a good prior frame design. For different tasks and targets with large deformations (large changes in aspect ratio), it may be necessary to manually design and adjust the prior frame hyperparameters based on experience. Anchor-free is much simpler. The Anchor-Free strategy provides an efficient and flexible solution for the field of surface defect detection of carbon fiber reinforced polymer concrete by simplifying the detection process and reducing the hyperparameters that need to be adjusted. Based on the structure of Anchor-Free and Focal Loss, efficient target detection and classification are achieved, and the influence of background interference and category imbalance is reduced.

[0030] Specifically, in the embodiment of the present application, the Anchor-Free strategy is used to train the model. Furthermore, the Anchor-Free strategy is a target detection method. Unlike the traditional anchor-based strategy, it does not rely on predefined anchor boxes to identify targets in the image. The Anchor-Free method directly predicts the location and category of the target based on the image features, which simplifies the training and reasoning process and improves flexibility.

[0031] During the training process, a deep learning model (such as CNN) is first used to extract the feature map of the input image. This step is the basis of object detection and aims to capture key information in the image. Then the target center is defined, and the Anchor-Free method identifies the center point or key point of the target (for example, the center or corner of an object). This step does not require a predefined anchor box, but directly predicts the center position of the target on the feature map. In addition to the center point, the model also needs to predict the size of the target, including width and height. This allows the model to determine the specific bounding box of the target.

[0032] Next, define the loss function. The loss function used in the training process usually contains two parts, one is the position loss of the center point or key point (such as the center point deviation), and the other is the loss of the target size. This ensures that the model can accurately predict the location and size of the target. Based on the predicted center point and size, the model regresses the bounding box of each target. At the same time, the model also needs to predict the category of each target. This is usually achieved by applying a classifier near each predicted center point.

[0033] During the training and optimization process, the model parameters are adjusted to minimize the loss function through continuous iterative training and optimization algorithms such as back propagation and gradient descent. In addition, in the inference stage, some post-processing steps such as non-maximum suppression (NMS) can be taken to eliminate overlapping detection results and retain the best predictions.

[0034] Figure 2 A schematic diagram of another method for detecting surface defects of carbon fiber reinforced polymer concrete based on a neural network provided in an embodiment of the present application. Combined with the contents of the previous embodiment, the defect detection system is mainly composed of three parts, including image acquisition, data set preparation and surface defect detection. In the image acquisition part, a device with image acquisition function such as a camera and a mobile phone is used to collect defect images, and the collected images are classified and numbered to obtain an initial image set. Further, the initial image set is imported into a label making software, and a bounding box is drawn for each defect in the image and their categories are marked. After the labeling is completed, the labeling information is saved to a specified folder in XML format. The labeling information generally includes the position, size and category label of the bounding box. The original image and the label file are integrated into a complete data set, and the data set is divided into a training set and a validation set according to the ratio. Then the training set is input into the improved YOLOv8 model for training to obtain a trained model, and then the image to be detected is input into the trained defect detection model for detection, and finally the model will display the location and type of the defect.

[0035] It should be understood that the validation set is mainly used to evaluate the performance of the model. In this application, the improved YOLOv8 model is used for surface defect detection. During the training phase, the model can only access the data in the training set, while the validation set provides an independent set of data for testing the model's ability to handle new data. This helps to verify the generalization ability of the model, that is, the model's ability to predict unseen data. When training a model, you need to set and adjust a variety of hyperparameters, such as learning rate, batch size, or the depth and width of the network. The validation set allows you to evaluate the effects of these hyperparameters without touching the final test set. By testing different hyperparameter configurations on the validation set, you can choose the most optimized model configuration. In addition, overfitting is a common problem in model training, which means that the model performs well on the training set but performs poorly on new or unseen data. By monitoring the performance on the validation set, you can find out early when the model starts to overfit. If the error rate on the validation set starts to increase, this is usually a sign of overfitting, and you may need to adjust the complexity of the model or use regularization techniques. At the same time, early stopping is a technique to prevent overfitting, which involves stopping training when the performance on the validation set no longer improves during model training. This not only helps save computing resources, but also further prevents the model from overfitting the training data during the training process. In one embodiment of the present application, the ratio of the training set to the validation set can be set to 3:7 or 2:8.

[0036] Next, we will combine Figure 3 The model architecture in this application is further explained. Figure 3 An exemplary architecture diagram of a carbon fiber reinforced polymer concrete surface defect detection model based on a neural network provided in an embodiment of the present application. It should be understood that the model shown in the figure is exemplary and not restrictive. This means that the model architecture involved is not limited to a specific form or design, but is presented as an example. In other words, the architecture shown in the figure can be regarded as a way of expression to clearly describe related concepts and relationships, and does not exclude other forms of architecture. Therefore, when interpreting the architecture in the picture, it should be understood that the model is flexible and diverse, and its purpose is to provide an exemplary description rather than a restrictive provision on a specific form.

[0037] In one embodiment of the present application, the CSPDarknet network structure based on the C2f module includes one or more CNN models and multiple RepViT modules; when using an image acquisition device to obtain an image of the surface of carbon fiber reinforced polymer concrete to input into an efficient backbone network, it includes: using a CNN model to extract local features of the image, and the local features are used for segmentation to obtain a first local feature and a second local feature; inputting the first local feature into a first RepViT module, and the output of the first RepViT module is used to perform a first residual connection with the second local feature to input into a second RepViT module; inputting the output of the second RepViT module into a third RepViT module to perform a second residual connection; and using the result of the second residual connection output as the input of a CNN model in another CSPDarknet network structure based on the C2f module.

[0038] Specifically, this application introduces RepViT Block (Reparameterized Vision Transformer) as an enhancement of YOLOv8. RepViT Block is a lightweight deep learning model designed for mobile devices. RepViT combines CNN and ViT to form a lightweight model that has both CNN's local perception capabilities and ViT's global abstraction capabilities. This enables RepViT to adapt to the computing and memory resource constraints of mobile devices while maintaining high performance. RepViT advances the deep convolution so that the channel mixer and token mixer can be separated. In addition, a RepViT block is added in the front to further deepen the downsampling layer to increase the network depth and reduce the information loss caused by the reduction in resolution. RepViT adopts a convolutional layer structure similar to CNN, but after each layer of convolution, a self-attention mechanism similar to ViT is added. In addition, RepViT also uses a "position embedding" technology similar to ViT to map each pixel position in the input image to a vector space, thereby better capturing the spatial information in the image.

[0039] In another embodiment, the attention mechanism in the RepViT module includes:

[0040] SE(X)=σ(FC2(δ(FC1(X)))),

[0041] Wherein, X represents the input feature map, SE(X) represents the Squeeze-and-Excitation operation on the feature map X, FC1 and FC2 represent the first fully connected layer and the second fully connected layer respectively, δ represents the attention activation function, and σ represents the Sigmoid function. In the RepViT Block, the SE layer is used as the channel attention module to make up for the limitation of convolution in the lack of data-driven properties, thereby bringing better performance. However, one disadvantage of the SE layer is that it ignores the spatial dimension. Therefore, in another embodiment, the attention mechanism in the RepViT module described in the present application may also include:

[0042]

[0043] Among them, EMA represents the exponential moving average, t represents the current time, and θ t-i Represents the value of the past ti moments, and β represents the decay factor. The EMA attention mechanism is used to optimize the attention allocation process and improve the efficiency of multi-scale feature integration. The EMA attention mechanism is used to replace the original SE layer, so that the model can effectively process and integrate features from different scales of the input data. The EMA attention mechanism processes the input data through convolutional layers to generate feature representations of different scales. These features correspond to different network layers, where the shallow layers capture fine-grained features and the deep layers capture more abstract features. By applying the attention mechanism to the features of each scale separately, the model will learn to assign weights to the features of each scale, allowing the model to dynamically adjust the attention to features of different scales in order to pay more attention to information that is more important to the current task. Subsequently, the model fuses the weighted multi-scale features together to form a comprehensive feature representation. This step is critical because it allows the model to utilize information extracted from different scales, thereby improving overall performance and generalization capabilities. Finally, the fused features are fed into subsequent parts of the model (such as classifiers or regressors) for final task-specific processing and prediction

[0044] It should be noted that the RepViT module described in this application also includes: connecting the inputs and outputs of multiple 1x1 convolutions for adjusting the channel dimension through residuals to form a feedforward network. Specifically, in order to improve performance, structural reparameterization is also introduced to introduce multi-branch topology for deep filters during training. In the channel mixer, the original expansion ratio is reduced from 4 to 2. At the same time, by increasing the width of the network, that is, increasing the number of channels at each stage, parameter redundancy and delay are reduced, and the operation speed and Top-1 accuracy are improved. Use a 1x1 convolution to adjust the channel dimension, and then connect the inputs and outputs of two 1x1 convolutions through residuals to form a feedforward network.

[0045] The above content explains the model architecture, and the present application scheme will be further explained in combination with the training method. Among them, before using the image acquisition device to obtain the image of the surface of the carbon fiber reinforced polymer concrete to input the efficient backbone network, it also includes: using the image acquisition device to obtain the image of the surface of the carbon fiber reinforced polymer concrete, the image is used for classification and numbering to generate an initial image set; importing the initial image set into the label making software to mark the bounding box and corresponding category of the defect in the image; and integrating the annotation information including the position, size and category label of the bounding box with the initial image set into a complete data set, the complete data set is used to be divided into a training set and a validation set according to the ratio for training the model network.

[0046] After obtaining the training set, it is also necessary to divide the proportion of positive samples in the training set. In one embodiment, this may include:

[0047]

[0048] Among them, a represents the distribution ratio of positive samples in the training set, t represents the number of training iterations, t0 represents the number of initial iterations, and k represents the parameter for controlling the dynamic adjustment speed. Specifically, the training introduces a dynamic matching strategy in the algorithm, and adopts the Task-Aligned Assigner positive and negative sample distribution strategy, that is, the alignment assigner, that is, the distribution ratio of positive and negative samples is dynamically adjusted during the training process. Task-Aligned uses a high-order combination of classification scores and IoU to measure the degree of Task-Alignment.

[0049] After obtaining the sample division ratio, it is necessary to calculate the loss value of the loss function based on the output of the model and the labeled data, so as to update the parameters of the model by back propagation. In one embodiment, it includes:

[0050]

[0051] Wherein, LOSS represents the value of the loss function during the training process, λ represents the control coefficient, q represents the probability distribution parameter, and L PIoURepresents the value of pixel IoU. In the original YOLOv8 model, the CIoU loss function loss is used as the position loss term for bounding box regression in target detection tasks. However, the CIoU loss function also has shortcomings such as slow convergence speed, unreasonable penalty factor, and possible problems in calculating the gradient of aspect ratio difference. Therefore, in this algorithm, Powerful-IoU is used to replace the original CIoU as the loss function for bounding box regression. It solves the problem that the existing IoU-based loss function causes the anchor box to expand during the regression process, thereby slowing down the convergence speed, by combining the target size adaptive penalty factor and the gradient adjustment function based on the anchor box quality. The Powerful-IoU loss function guides the anchor box to regress along the effective path, achieving a faster convergence speed than the existing IoU loss function. In addition, by introducing a non-monotonic attention layer and combining it with Powerful-IoU, the ability to focus on medium-quality anchor boxes is enhanced, further improving the performance of the target detector.

[0052] It should be noted that the present invention proposes a surface defect detection method based on computer vision and YOLOv8 algorithm model to address the possible deficiencies of the traditional CFRP concrete surface defect detection system, that is, to use the improved YOLOv8 algorithm to detect and locate CFRP surface defects. The method first collects images of CFRP concrete surface defects, and uses computer vision and deep learning technology to process, analyze and identify the images, thereby locating and classifying the defects. The present invention introduces the RepViT lightweight module, EMA attention mechanism and Powerful-IoU loss function when developing the algorithm model to realize CFRP concrete surface defect detection. After multiple experiments, it has been proved that the method has the advantages of high detection accuracy and lightweight. In addition to the contents described in the above embodiments, the scheme of the present application can also include using Lightweight Shared Convolutional Detection Head as the detection head of the algorithm model, and using Wise-IoU as the loss function of the algorithm. This improved scheme can also achieve the purpose of detecting CFRP concrete surface defects.

[0053] Furthermore, an embodiment of the present application also provides a carbon fiber reinforced polymer concrete surface defect detection device based on a neural network, comprising: a processor, a memory, and a system bus; the processor and the memory are connected via the system bus; the memory is used to store one or more programs, and the one or more programs include instructions, which, when executed by the processor, enable the processor to execute any of the above-mentioned detection methods.

[0054] Furthermore, an embodiment of the present application also provides a computer program product, which, when executed on a terminal device, enables the terminal device to execute any one of the detection methods of the above-mentioned processing methods.

[0055] It can be known from the description of the above implementation mode that those skilled in the art can clearly understand that all or part of the steps in the above-mentioned embodiment method can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product can be stored in a storage medium such as ROM / RAM, a disk, an optical disk, etc., including several instructions for enabling a computer device (which can be a personal computer, a server, or a network communication device such as a media gateway, etc.) to execute the methods described in the various embodiments of the present application or certain parts of the embodiments.

[0056] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same or similar parts between the various embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description.

[0057] It should also be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0058] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for detecting surface defects of carbon fiber reinforced polymer concrete based on neural network, characterized in that: include: Using an image acquisition device to obtain images of the carbon fiber reinforced polymer concrete surface to input into the efficient backbone network; The efficient backbone network includes one or more CSPDarknet network structures based on C2f modules; The method for calculating the number of CSPDarknet network structures based on the C2f module includes: , in, represents the number after calculation optimization, represents the image dataset matrix vector, Indicates the number of optimizations not calculated, represents the expected time parameter, represents the Frobenius norm; Inputting the output feature vectors of the multiple target network layers in the efficient backbone network into the PANet structure network, wherein the PANet structure network is used to upsample and downsample the output feature vectors of the multiple target network layers to fuse multi-scale features; and Processing the output of the PANet structure network using a head network including one or more decoupled heads based on an Anchor-free strategy, wherein the head network performs reasoning using a decoupled classification branch and a regression branch to locate the position of the surface defects of the carbon fiber reinforced polymer concrete; The CSPDarknet network structure based on the C2f module includes one or more CNN models and multiple RepViT modules; when using an image acquisition device to obtain an image of the surface of carbon fiber reinforced polymer concrete to input into an efficient backbone network, the method includes: using a CNN model to extract local features of the image, the local features are used for segmentation to obtain a first local feature and a second local feature; inputting the first local feature into a first RepViT module, and the output of the first RepViT module is used to perform a first residual connection with the second local feature to input into a second RepViT module; inputting the output of the second RepViT module into a third RepViT module to perform a second residual connection; and using the result of the second residual connection output as the input of a CNN model in another CSPDarknet network structure based on the C2f module; The attention mechanism in the RepViT module includes: ,in, represents the input feature map, Represents the feature map Perform a Squeeze-and-Excitation operation. denote the first fully connected layer and the second fully connected layer respectively, represents the attention activation function, represents the Sigmoid function; or, ,in, represents the exponential moving average, Indicates the current moment, Indicates the past The value of the moment, Represents the attenuation factor.

2. The method according to claim 1, characterized in that The RepViT module further comprises: The input and output of multiple 1x1 convolutions used to adjust the channel dimension are connected through residuals to form a feed-forward network.

3. The method according to claim 1, characterized in that in, Before using an image acquisition device to obtain an image of the carbon fiber reinforced polymer concrete surface to input into the efficient backbone network, it also includes: Using an image acquisition device to acquire images of the surface of the carbon fiber reinforced polymer concrete, wherein the images are used for classification and numbering to generate an initial image set; Importing the initial set of images into labeling software to annotate the bounding boxes and corresponding categories of defects in the images; and The annotation information including the position, size and category label of the bounding box is integrated with the initial image set into a complete dataset, and the complete dataset is used to be divided into a training set and a validation set according to a ratio for training the model network.

4. The method according to claim 3, characterized in that: Also includes: ， in, represents the distribution ratio of positive samples in the training set, represents the number of training iterations, represents the initial iteration number, Represents the parameter that controls the dynamic adjustment speed.

5. The method according to claim 4, characterized in that Also includes: ， in, represents the value of the loss function during the training process, represents the control coefficient, represents the probability distribution parameter, Represents pixels The value of .

6. A system using the carbon fiber reinforced polymer concrete surface defect detection method based on neural network as claimed in claim 1, characterized in that: include: An acquisition module, the acquisition module is used to acquire an image of the surface of the carbon fiber reinforced polymer concrete to input into the efficient backbone network; A feature extraction module, wherein the feature extraction module is used to input the output feature vectors of multiple target network layers in the efficient backbone network into a PANet structure network, and the PANet structure network is used to upsample and downsample the output feature vectors of the multiple target network layers to fuse multi-scale features; as well as A detection module, wherein the detection module is used to process the output of the PANet structure network using a head network including one or more decoupled heads based on an Anchor-free strategy, wherein the head network performs reasoning by using a decoupled classification branch and a regression branch to locate the position of the surface defects of the carbon fiber reinforced polymer concrete.

7. A carbon fiber reinforced polymer concrete surface defect detection device based on neural network, characterized in that: include: A processor, a memory, and a system bus; wherein the processor and the memory are connected via the system bus; The memory is used to store one or more programs, wherein the one or more programs include instructions, and when the instructions are executed by the processor, the processor executes the method according to any one of claims 1 to 5.

Images