Bridge image crack identification method and system based on deep neural network

Abstract

The invention discloses a bridge image crack identification method and system based on a deep neural network. The method comprises the following steps: acquiring a bridge B-Scan image through ground penetrating radar equipment; performing preprocessing such as normalization, direct wave weakening and automatic gain enhancement on the image; features are extracted on the basis of an improved OfficientDeet-D3 network, a MoHAttention module is introduced into a backbone network to enhance weak crack features, a BiFPN structure is optimized through deep separable convolution in the feature fusion process to improve the detection capacity of a small-scale crack target, and crack position frames, types and confidence information are output; the system comprises a data acquisition module, a preprocessing module, a neural network detection module and an output display module. According to the method, high-precision and high-robustness identification of the internal hidden crack can be realized, and the method is suitable for rapid detection and intelligent monitoring application in actual bridge inspection engineering.

Description

Technical Field

[0001] The present invention relates to the technical field of structural health monitoring and nondestructive testing, and in particular to a method and system for identifying cracks in bridge images based on a deep neural network. Background Art

[0002] With increasing traffic loads, material aging, and environmental erosion, urban bridges, particularly prestressed concrete bridges, are prone to developing hidden cracks during their service. These cracks are often located deep within the concrete structure, invisible to the naked eye. Initially, they are small, concealed, uncontrollable, and highly destructive. If not detected and intervened promptly, hidden cracks will expand, seriously affecting the durability and safety of the bridge structure and even causing major accidents.

[0003] Ground Penetrating Radar (GPR), a non-contact, non-destructive, and highly efficient detection method, has become an important technology for identifying internal defects in bridges. By emitting high-frequency electromagnetic waves and receiving reflected echoes, GPR can generate A-Scan and B-Scan images, revealing changes in the underground medium and abnormal reflection characteristics. It has significant application potential in detecting hidden internal cracks.

[0004] However, because hidden internal crack signals typically manifest as weak, short, and intermittent linear reflections, GPR image quality fluctuates significantly due to interference from multiple factors, including concrete material inhomogeneity, electromagnetic wave energy attenuation, and environmental noise. Traditional image processing methods, such as threshold segmentation, edge detection, wavelet transform, and morphological filtering, rely on manually designed features and lack robustness, resulting in low accuracy in detecting faint cracks, particularly in complex backgrounds.

[0005] In recent years, convolutional neural networks (CNNs) have been introduced to the field of crack detection in GPR images due to their powerful automatic feature extraction capabilities. Mainstream object detection models such as Faster R-CNN and YOLO have been applied to B-scan images, achieving positive progress in improving detection efficiency and achieving a certain degree of accuracy. However, existing methods still suffer from the following major issues:

[0006] (1) The two-stage detection model (such as FasterR-CNN) has high computational overhead and cannot meet the real-time requirements of large-scale bridge inspection scenarios;

[0007] (2) Although the one-stage detection model (such as the YOLO series) has a high inference speed, its recognition accuracy for small, weak signal hidden cracks is insufficient;

[0008] (3) As a lightweight target detection framework, EfficientDet-D3 combines the EfficientNet backbone with the Bidirectional Feature Pyramid Network (BiFPN) and performs well in natural image detection tasks. However, its native structure is optimized for natural images and is still insufficient in capturing weak crack signals when directly applied to GPR images of prestressed concrete bridges.

[0009] Therefore, it is urgent to propose a deep learning detection method for the characteristics of hidden cracks inside prestressed concrete bridges, and to improve the accuracy and robustness of crack identification by improving the feature extraction and fusion mechanism. At the same time, a detection system integrating data acquisition, intelligent identification and result output should be constructed to meet the engineering needs of efficient, accurate and automated detection of hidden cracks in bridges. Summary of the Invention

[0010] Based on the above content, this application discloses a bridge image crack recognition method and system based on deep neural network; including:

[0011] S1. Collect B-Scan image data of the bridge structure, pre-process the image, unify the image size and format, and form an image dataset;

[0012] S2. Input the images in the image dataset into the deep neural network model built based on the EfficientNet-B3 backbone network, extract image feature maps at different levels, and form a multi-level image feature set;

[0013] S3. Input the multi-level image feature set into the bidirectional feature pyramid network, and use depthwise separable convolution to fuse the image feature maps of different scales in the multi-level image feature set to form a fused feature map;

[0014] S4. Perform target detection on the fused feature map through the detection head, output the crack detection results, and visualize the detection results and output them to the terminal interface to complete the crack identification and positioning tasks.

[0015] Preferably, the B-Scan image data acquired in S1 is preprocessed, and the preprocessing includes normalization processing, noise suppression processing, automatic gain processing and size normalization processing, so as to reduce the imaging quality problems of the B-Scan image data.

[0016] Preferably, the denormalization process is specifically as follows:

[0017] Obtain B-Scan image data and normalize the image pixel values to the interval [-1, 1] using the formula:

[0018]

[0019] Among them, the two-dimensional B-Scan data is composed of N channels of A-Scan signal data, M is the number of sampling points, N is the total number of scan channels, B max 、B min are the minimum and maximum values of the image matrix B, B′ ij is the normalized sampling point value.

[0020] Preferably, the noise suppression process includes zero bias removal and direct wave removal, specifically:

[0021] The de-biasing process removes the DC offset from each A-Scan. The formula is:

[0022]

[0023] Among them, x ij is the jth A-Scan data X j =[x j1 ,x j2 ,...,x jM ] T The i-th sampling point, x′ ij is the data sampling point value after removing the zero bias, and the ground penetrating radar data after removing the zero bias is obtained;

[0024] The channel averaging method is used to reduce the horizontal straight line interference in the direct wave removal process. The formula is:

[0025]

[0026] Among them, x ij is the jth A-Scan data X j =[x j1 ,x j2 ,...,x jM ] T The i-th sampling point, x″ ij is the data sampling point value after removing the direct wave.

[0027] Preferably, the automatic gain processing is specifically:

[0028] Each A-Scan is divided into T overlapping windows, and the average amplitude of each window is calculated using the formula:

[0029]

[0030] The size of each time window is Indicates rounding down, A tj represents the average amplitude of the tth time window in the jth A-Scan data, x ij is the jth A-Scan data X j=[x j1 ,x j2 ,...,x jM ] T The i-th sampling point, G tj Represents the gain value of the tth time window.

[0031] Preferably, in the feature extraction process of the deep neural network model in S2, the MoHAttention module is used to replace the original Squeeze-and-Excitation module, and the response ability to the crack area is improved by the multi-head attention mechanism fused with the shared head and the routing head; the multi-head attention mechanism performs feature transformation by querying Q, key K and value V three sets of learnable weight matrices, and the formula is: Q = W q X,K=W k X,V=W v X, where X∈R B×H×W×C Represents the input feature map, R represents the feature map element, B is the number of images, H is the number of pixels in the feature map in the vertical direction, W is the number of pixels in the feature map in the horizontal direction, C is the number of channels, and W q , W k , W v are the transformation matrices of query, key, and value respectively; the attention weight matrix is calculated by Softmax, and the formula is: The obtained attention weight matrix is then weighted fused and connected to the input through the ReLU activation function and the residual. The formula is: out =σ(W o MohAttention(X)+X), where W o is a learnable linear transformation, and σ represents the ReLU activation function. The combination of routing head and shared head enables the deep neural network model to effectively balance feature aggregation and global attention.

[0032] Preferably, in S3, the image feature maps of different scales in the multi-level image feature set are fused using depthwise separable convolution to form a fused feature map, and the formula is: Among them F i are feature maps of different scales, w i is the trainable weight, w j is the learnable weight of each j-th scale feature map, ∈ is the smoothing coefficient, and N is the total number of feature maps of different scales.

[0033] Preferably, the depthwise separable convolution performs spatial filtering and channel information fusion by depthwise convolution and pointwise convolution to reduce information loss during feature fusion, specifically:

[0034] Depth-wise convolution performs spatial filtering on each channel independently, as follows: where X c is the cth channel of the input feature map, W c is the corresponding convolution kernel, k is the convolution kernel size parameter, m is the horizontal movement value of the convolution kernel, and n is the vertical movement value of the convolution kernel; point-by-point convolution is performed by obtaining the Y c (i, j) performs point-by-point convolution by fusing channel information through 1×1 convolution. The formula is: Where W (p,c) is the 1×1 convolution weight.

[0035] A bridge ground penetrating radar image crack recognition system includes a data acquisition module, a preprocessing module, a neural network detection module and an output display module;

[0036] The data acquisition module includes a mobile platform and a GPR radar device for collecting B-Scan image data;

[0037] The preprocessing module generates an image data set by preprocessing the acquired image data;

[0038] The neural network detection module outputs crack identification information by deploying the EfficientDet-D3 detection model based on MoHAttention and depthwise separable convolution optimization;

[0039] The output display module receives the crack identification information and outputs the detection results, including the crack location frame, category and confidence level.

[0040] Preferably, a bridge ground penetrating radar image crack recognition system also includes a model training and update module, which performs incremental model training and parameter updates through newly added detection data. As the amount of collected bridge ground penetrating radar image data continues to increase, the new data is integrated into the existing model training system. Through retraining, the model learns the features in the new data and optimizes internal parameters.

[0041] Compared with the prior art, the technical solution of this application has the following technical effects:

[0042] This invention enhances the ability to capture weak signal crack features. Inside prestressed concrete bridges, hidden crack signals are weak, making it difficult for traditional methods to accurately identify them. The MoHAttention module uses a multi-head attention mechanism that fuses a shared head with a routing head. This allows it to focus on the crack area, enhance local crack features, and perform global relationship modeling at the same time.

[0043] The present invention optimizes the bidirectional feature pyramid network structure and improves the feature fusion effect through the application of depthwise separable convolution. The present invention independently performs spatial filtering on each channel through depthwise convolution, then uses pointwise convolution to fuse channel information, and adopts learnable weighting coefficients to fuse feature maps of different scales, thereby reducing the loss of feature information and better preserving the characteristics of small-scale hidden cracks.

[0044] While ensuring detection accuracy, the present invention significantly reduces network inference latency, achieving a balance between real-time performance and lightweightness. A deep neural network model is constructed based on the EfficientNet-B3 backbone network, and a series of optimizations are performed on it to reduce computational complexity.

[0045] The technological invention has built a system that integrates data collection, intelligent identification and visual output of detection results, realizing the automation of the entire process from data collection to result output, and improving the intelligence and engineering application feasibility of hidden crack detection inside bridge structures.

[0046] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application so that it can be implemented in accordance with the contents of the specification, and to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the following is a detailed description of the preferred embodiment of the present application in conjunction with the accompanying drawings.

[0047] Based on the detailed description of the specific embodiments of the present application in conjunction with the accompanying drawings below, those skilled in the art will become more aware of the above and other objects, advantages and features of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following will briefly introduce the drawings required for the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings without inventive work. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn according to the actual scale.

[0049] Figure 1 Schematic diagram of the process of detecting hidden cracks in prestressed concrete bridges based on the improved EfficientDet-D3 according to the present invention;

[0050] Figure 2 This is a schematic diagram of the overall functional structure of the bridge GPR image crack detection system of the present invention;

[0051] Figure 3This is a block diagram of the improved EfficientDet-D3 network structure described in the present invention, showing the layout of the backbone network, enhancement modules, and detection head;

[0052] Figure 4 This is a schematic diagram of the MoHAttention multi-head dynamic attention mechanism structure of the present invention;

[0053] Figure 5 This is a partial schematic diagram of the effect of the system of the present invention on detecting hidden cracks in GPR images of prestressed concrete bridges. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions 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, rather than all of the embodiments. In the following description, specific details such as specific configurations and components are provided only to help fully understand the embodiments of the present application. Therefore, it should be clear to those skilled in the art that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present application. In addition, for clarity and brevity, the description of known functions and structures has been omitted in the embodiments.

[0055] It should be understood that references throughout this specification to "one embodiment" or "this embodiment" mean that a particular feature, structure, or characteristic associated with the embodiment is included in at least one embodiment of the present application. Therefore, the appearance of "one embodiment" or "this embodiment" throughout this specification does not necessarily refer to the same embodiment. Furthermore, these particular features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0056] In addition, the present application may repeat reference numerals and / or letters in different examples. This repetition is for the purpose of simplicity and clarity and does not in itself indicate the relationship between the various embodiments and / or settings discussed.

[0057] The term "and / or" in this article is only a description of the association relationship of associated objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist at the same time. The term " / and" in this article describes another type of association object relationship, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the previous and subsequent associated objects are in an "or" relationship.

[0058] The term "at least one" in this article is merely a description of the association relationship between associated objects, indicating that three relationships may exist. For example, at least one of A and B can mean: A exists alone, A and B exist at the same time, and B exists alone.

[0059] It should also be noted that, in this document, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include," "comprises," or any other variations thereof are intended to cover non-exclusive inclusion.

[0060] Example 1

[0061] This embodiment mainly describes a method for identifying cracks in bridge ground penetrating radar images based on a deep neural network. Figure 1 Shown, including:

[0062] S1. Collect B-Scan image data of the bridge structure, pre-process the image, unify the image size and format, and form an image dataset;

[0063] S2. Input the images in the image dataset into the deep neural network model built based on the EfficientNet-B3 backbone network, extract image feature maps at different levels, and form a multi-level image feature set;

[0064] S3. Input the multi-level image feature set into the bidirectional feature pyramid network, and use depthwise separable convolution to fuse the image feature maps of different scales in the multi-level image feature set to form a fused feature map;

[0065] S4. Perform target detection on the fused feature map through the detection head, output the crack detection results, and visualize the detection results and output them to the terminal interface to complete the crack identification and positioning tasks.

[0066] Furthermore, the B-Scan image data acquired in S1 is preprocessed, and the preprocessing includes normalization processing, noise suppression processing, automatic gain processing and size normalization processing, so as to reduce the imaging quality problems of the B-Scan image data.

[0067] Furthermore, the denormalization process is specifically as follows:

[0068] Obtain B-Scan image data and normalize the image pixel values to the interval [-1, 1] using the formula:

[0069]

[0070] Among them, the two-dimensional B-Scan data is composed of N channels of A-Scan signal data, M is the number of sampling points, N is the total number of scan channels, B max 、B min are the minimum and maximum values of the image matrix B, B′ ij is the normalized sampling point value.

[0071] Furthermore, the noise suppression process includes zero bias removal and direct wave removal, specifically:

[0072] The de-biasing process removes the DC offset from each A-Scan. The formula is:

[0073]

[0074] Among them, x ij is the jth A-Scan data X j =[x j1 ,x j2 ,...,x jM ] T The i-th sampling point, x′ ij is the data sampling point value after removing the zero bias, and the ground penetrating radar data after removing the zero bias is obtained;

[0075] The channel averaging method is used to reduce the horizontal straight line interference in the direct wave removal process. The formula is:

[0076]

[0077] Among them, x ij is the jth A-Scan data X j =[x j1 ,x j2 ,...,x jM ] T The i-th sampling point, x″ ij is the data sampling point value after removing the direct wave.

[0078] Furthermore, the automatic gain processing is specifically as follows:

[0079] Each A-Scan is divided into T overlapping windows, and the average amplitude of each window is calculated using the formula:

[0080]

[0081] The size of each time window is Indicates rounding down, A tj represents the average amplitude of the tth time window in the jth A-Scan data, x ij is the jth A-Scan data X j=[x j1 ,x j2 ,...,x jM ] T The i-th sampling point, G tj Represents the gain value of the tth time window.

[0082] Further, if Figure 2 As shown, in the feature extraction process of the deep neural network model in S2, the MoHAttention module is used to replace the original Squeeze-and-Excitation module, and the response ability to the crack area is improved by the multi-head attention mechanism fused with the shared head and the routing head; the multi-head attention mechanism performs feature transformation through three sets of learnable weight matrices of query Q, key K and value V, and the formula is: Q = W q X,K=W k X,V=W v X, where X∈R B×H×W×C Represents the input feature map, R represents the feature map element, B is the number of images, H is the number of pixels in the feature map in the vertical direction, W is the number of pixels in the feature map in the horizontal direction, C is the number of channels, and W q , W k , W v are the transformation matrices of query, key, and value respectively; the attention weight matrix is calculated by Softmax, and the formula is: The obtained attention weight matrix is then weighted fused and connected to the input through the ReLU activation function and the residual. The formula is: out =σ(W o MohAttention(X)+X), where W o is a learnable linear transformation, σ represents the ReLU activation function; Figure 4 As shown, the combination of routing head and shared head enables the deep neural network model to effectively balance feature aggregation and global attention.

[0083] Furthermore, in S3, the image feature maps of different scales in the multi-level image feature set are fused using depthwise separable convolution to form a fused feature map. The formula is: Among them F i are feature maps of different scales, w i is the trainable weight, w j is the learnable weight of each j-th scale feature map, ∈ is the smoothing coefficient, and N is the total number of feature maps of different scales.

[0084] Furthermore, the depthwise separable convolution performs spatial filtering and channel information fusion through depthwise convolution and pointwise convolution to reduce information loss during feature fusion. Specifically:

[0085] Depth-wise convolution performs spatial filtering on each channel independently, as follows: where X c is the cth channel of the input feature map, W c is the corresponding convolution kernel, k is the convolution kernel size parameter, m is the horizontal movement value of the convolution kernel, and n is the vertical movement value of the convolution kernel; point-by-point convolution is performed by obtaining the Y c (i, j) performs point-by-point convolution by fusing channel information through 1×1 convolution. The formula is: Where W (p,c) is the 1×1 convolution weight.

[0086] This example describes in detail how to enhance the crack feature extraction capability through the multi-head dynamic attention mechanism (MoHAttention), and optimize the bidirectional feature pyramid network (BiFPN) structure by combining deep separable convolution (DSConv) to improve the detection accuracy and robustness of small crack targets.

[0087] Example 2

[0088] This embodiment describes in detail a bridge ground penetrating radar image crack identification system, including a data acquisition module, a preprocessing module, a neural network detection module and an output display module;

[0089] The data acquisition module includes a mobile platform and a GPR radar device for collecting B-Scan image data;

[0090] The preprocessing module generates an image data set by preprocessing the acquired image data, wherein the preprocessing includes normalization processing, noise suppression processing, automatic gain processing and size normalization processing;

[0091] The neural network detection module outputs crack identification information by deploying the EfficientDet-D3 detection model based on MoHAttention and depthwise separable convolution optimization;

[0092] The output display module receives the crack identification information and outputs the detection results, including the crack location frame, category and confidence level.

[0093] The output display module receives the crack identification information output by the neural network detection module, converts it into intuitive and easy-to-understand visual content, and displays the crack location box on the terminal interface, accurately marking the area where the crack is located on the corresponding bridge radar image, making it easy for engineering personnel to quickly locate it. It also displays the crack category, clearly stating whether the crack is horizontal, vertical, or oblique, helping to determine the nature of the crack, presenting the confidence level, and expressing the reliability of the detection results as a percentage.

[0094] Furthermore, it also includes a model training and update module, which performs incremental model training and parameter updates through new detection data. As the amount of collected bridge ground penetrating radar image data continues to increase, the new data is integrated into the existing model training system. Through retraining, the model learns the features in the new data and optimizes internal parameters.

[0095] This embodiment describes in detail a crack recognition system for bridge ground-penetrating radar images. The system leverages the MoHAttention module and the EfficientDet-D3 model optimized with deep separable convolution to significantly improve the accuracy of hidden crack detection. It can also accurately identify weak and small-scale cracks. The entire process from data collection to result output is fully automated, offering easy operation and balancing real-time performance with lightweight design. This system can quickly complete detection and is continuously optimized through model training and module updates to adapt to diverse detection needs, effectively ensuring the safety of bridge structures.

[0096] Example 3

[0097] This embodiment, based on Example 1 and Example 2, describes in detail the specific implementation effects of the present application and the technical effects compared with the prior art, specifically:

[0098] This application was fully tested on the TIGPR public dataset, obtaining 325 ground-penetrating radar B-Scan images from different bridges, covering a variety of bridge structures and cracks of varying degrees. The PyTorch 2.3.0 deep learning framework was selected, which provides a wealth of tools and efficient computing power to facilitate model building and training. The optimizer uses SGD (stochastic gradient descent), which can effectively adjust model parameters during training, optimizing the model towards minimizing the loss function. The batch size is set to 8, that is, 8 images are processed simultaneously during each training. This can fully utilize the parallel computing power of the GPU and ensure the stability of the training. The entire training process iterates 300 rounds to ensure that the model can fully learn the features in the data. The RTX 3090 GPU is used on the hardware, and its powerful graphics processing capabilities greatly accelerate the training speed of the model and reduce the time cost required for the experiment.

[0099] The images in the dataset are preprocessed, performing normalization, debiasing, de-direction, automatic gain enhancement, and resizing operations in sequence, converting the original images into high-quality image data suitable for model input. The preprocessed images are then fed into the EfficientDet-D3 detection model, which is optimized based on MoHAttention and depthwise separable convolution, for training and detection.

[0100] The experimental results are evaluated using the AP50, AP75, and AR metrics. AP50 represents the average precision at an IoU (Intersection over Union) threshold of 0.5, and AP75 represents the average precision at an IoU threshold of 0.75. These two metrics are used to measure the accuracy of the model in crack detection under different accuracy requirements. AR (Average Recall) is used to evaluate the model's ability to detect all true cracks, as shown in the following table:

[0101]

[0102] The experimental results show that the model of this invention performs well. Compared with mainstream object detection algorithms such as DeformableDETR, DINO, TOOD, and DDQ, its AP50 index reaches 0.971, which is higher than other algorithms, indicating that it can more accurately detect cracks at lower precision thresholds. The AP75 index reaches 0.839, which is even more significant, indicating that it can still accurately identify cracks under higher precision requirements. The AR is 0.782, which is also better than most compared algorithms, indicating that the model has good recall ability and can effectively detect more real cracks.

[0103] by Figure 5 As shown in the figure, in the detection of a bridge image, other algorithms may miss some cracks or have low confidence levels. However, the model of the present invention not only accurately detects all cracks, but also gives a high confidence level for each crack, such as "Crack: 98.4". This fully demonstrates the high precision and strong robustness of the present invention in crack identification, and provides reliable technical support for bridge safety monitoring.

[0104] This example describes in detail how the optimized model, tested on the TIGPR public dataset, demonstrates excellent performance in the AP50 and AP75 metrics, achieving high accuracy and accurately identifying cracks. It also demonstrates excellent AR performance with minimal missed detections. This significantly outperforms mainstream algorithms, significantly improving the accuracy and reliability of bridge crack detection and effectively safeguarding bridge structures.

[0105] The above are only preferred embodiments of the present invention, which do not limit the scope of protection of the present invention. For those skilled in the art, the present invention can be modified and varied in various ways. Any changes, modifications, replacements, integrations and parameter changes to these embodiments through conventional substitutions or that can achieve the same functions without departing from the principles and spirit of the present invention fall within the scope of protection of the present invention.

Claims

1. A method for identifying cracks in bridge ground penetrating radar images based on deep neural networks, characterized in that: include; S1. Collect B-Scan image data of the bridge structure, pre-process the image, unify the image size and format, and form an image dataset; S2. Input the images in the image dataset into the deep neural network model built based on the EfficientNet-B3 backbone network, extract image feature maps at different levels, and form a multi-level image feature set; S3. Input the multi-level image feature set into the bidirectional feature pyramid network, and use depthwise separable convolution to fuse the image feature maps of different scales in the multi-level image feature set to form a fused feature map; S4. Perform target detection on the fused feature map through the detection head, output the crack detection results, and visualize the detection results and output them to the terminal interface to complete the crack identification and positioning tasks.

2. The method for identifying cracks in bridge ground penetrating radar images based on a deep neural network according to claim 1 is characterized in that: The B-Scan image data acquired in S1 is preprocessed, and the preprocessing includes normalization processing, noise suppression processing, automatic gain processing and size normalization processing, so as to reduce the imaging quality problems of the B-Scan image data.

3. The method for identifying cracks in bridge ground penetrating radar images based on a deep neural network according to claim 2 is characterized in that: The denormalization process is specifically as follows: Obtain B-Scan image data and normalize the image pixel values to the interval [-1, 1] using the formula: Among them, the two-dimensional B-Scan data is composed of N channels of A-Scan signal data, M is the number of sampling points, N is the total number of scan channels, B max 、B min are the minimum and maximum values of the image matrix B, B′ ij is the normalized sampling point value.

4. The method for identifying cracks in bridge ground penetrating radar images based on a deep neural network according to claim 2 or 3, characterized in that: The noise suppression process includes zero bias removal and direct wave removal, specifically: The de-biasing process removes the DC offset from each A-Scan. The formula is: Among them, x ij is the jth A-Scan data X j =[x j1 ,x j2 ,...,x jM ] T The i-th sampling point, x′ ij is the data sampling point value after removing the zero bias, and the ground penetrating radar data after removing the zero bias is obtained; The channel averaging method is used to reduce the horizontal straight line interference in the direct wave removal process. The formula is: Among them, x ij is the jth A-Scan data X j =[x j1 ,x j2 ,...,x jM ] T The i-th sampling point, x″ ij is the data sampling point value after removing the direct wave.

5. The method for identifying cracks in bridge ground penetrating radar images based on a deep neural network according to claim 2 or 3, characterized in that: The automatic gain processing is specifically as follows: Each A-Scan is divided into T overlapping windows, and the average amplitude of each window is calculated using the formula: The size of each time window is Indicates rounding down, A tj represents the average amplitude of the tth time window in the jth A-Scan data, x ij is the jth A-Scan data X j =[x j1 ,x j2 ,...,x jM ] T The i-th sampling point, G tj Represents the gain value of the tth time window.

6. The method for identifying cracks in bridge ground penetrating radar images based on a deep neural network according to claim 1, characterized in that: In the feature extraction process of the deep neural network model in S2, the MoHAttention module is used to replace the original Squeeze-and-Excitation module, and the response ability to the crack area is improved by the multi-head attention mechanism fused with the shared head and the routing head; the multi-head attention mechanism performs feature transformation by querying Q, key K and value V, three sets of learnable weight matrices, and the formula is: Q = W q X,K=W k X,V=W v X, where X∈R B×H×W×C Represents the input feature map, R represents the feature map element, B is the number of images, H is the number of pixels in the feature map in the vertical direction, W is the number of pixels in the feature map in the horizontal direction, C is the number of channels, and W q , W k , W v are the transformation matrices of query, key, and value respectively; the attention weight matrix is calculated by Softmax, and the formula is: The obtained attention weight matrix is then weighted fused and connected to the input through the ReLU activation function and the residual. The formula is: out =σ(W o MohAttention(X)+X), where W o is a learnable linear transformation, and σ represents the ReLU activation function. The combination of routing head and shared head enables the deep neural network model to effectively balance feature aggregation and global attention.

7. The method for identifying cracks in bridge ground penetrating radar images based on a deep neural network according to claim 1, characterized in that: In S3, the image feature maps of different scales in the multi-level image feature set are fused using depthwise separable convolution to form a fused feature map. The formula is: Among them F i are feature maps of different scales, w i is the trainable weight, w j is the learnable weight of each j-th scale feature map, ∈ is the smoothing coefficient, and N is the total number of feature maps of different scales.

8. The method for identifying cracks in bridge ground penetrating radar images based on a deep neural network according to claim 7, characterized in that: The depth-wise separable convolution performs spatial filtering and channel information fusion through depth-wise convolution and point-wise convolution, reducing information loss during feature fusion. Specifically: Depth-wise convolution performs spatial filtering on each channel independently, as follows: where X c is the cth channel of the input feature map, W c is the corresponding convolution kernel, k is the convolution kernel size parameter, m is the horizontal movement value of the convolution kernel, and n is the vertical movement value of the convolution kernel; point-by-point convolution is performed by obtaining the Y c (i, j) performs point-by-point convolution by fusing channel information through 1×1 convolution. The formula is: Where W (p,c) is the 1×1 convolution weight.

9. A bridge ground penetrating radar image crack recognition system, characterized in that: It includes data acquisition module, pre-processing module, neural network detection module and output display module; The data acquisition module includes a mobile platform and a GPR radar device for collecting B-Scan image data; The preprocessing module generates an image data set by preprocessing the acquired image data; The neural network detection module outputs crack identification information by deploying the EfficientDet-D3 detection model based on MoHAttention and depthwise separable convolution optimization; The output display module receives the crack identification information and outputs the detection results, including the crack location frame, category and confidence level.

10. The bridge ground penetrating radar image crack identification system according to claim 9, characterized in that: It also includes a model training and update module, which performs incremental model training and parameter updates through new detection data. As the amount of bridge ground-penetrating radar image data collected continues to increase, the new data is integrated into the existing model training system. Through retraining, the model learns the features in the new data and optimizes internal parameters.

Images