A ground penetrating radar image disease detection method, system and electronic device

Abstract

This invention discloses a method, system, and electronic device for detecting road defects in ground-penetrating radar (GPR) images, belonging to the field of computer vision and road defect detection technology. The method includes: acquiring and preprocessing GPR images; constructing a defect detection model, which includes a multi-scale vertical / horizontal strip convolution module, an adaptive gated routing module, and a bidirectional attention fusion module, used for extracting directional features, dynamically fusing multi-scale features, and adaptively fusing directional features, respectively; training and using the model for defect identification and localization. This invention reduces computational complexity through strip convolution and achieves adaptive feature fusion through gating and attention mechanisms, significantly improving the detection accuracy and efficiency of multi-scale, directional defects in GPR images.

Description

Technical Field

[0001] This invention relates to the fields of computer vision, deep learning and non-destructive testing technology, specifically to a method, system and electronic equipment for detecting defects in ground-penetrating radar images. Background Technology

[0002] Ground-penetrating radar (GPR) is a widely used technology for non-destructive testing of the internal structures of infrastructure such as roads and bridges. By interpreting acquired GPR images, defects such as cracks, interlayer defects, and loosening can be identified. However, automated defect detection using GPR images faces many challenges: defect targets often exhibit significant directionality (such as longitudinal cracks and transverse delamination) and anisotropy; defects of different depths and types vary greatly in image scale; and interference such as electromagnetic noise, clutter, and multiple reflections prevalent in images severely obscure defect features. When existing general-purpose target detection algorithms are directly applied to GPR images, they often suffer from low detection accuracy, poor generalization ability, and high computational cost due to issues such as fixed receptive fields of convolutional kernels, insufficient directional modeling capabilities, and sensitivity to noise, making it difficult to meet the real-time requirements of engineering projects. Summary of the Invention

[0003] The present invention aims to overcome the shortcomings of the prior art and provide a ground-penetrating radar image defect detection method, system and electronic equipment with high detection accuracy, excellent computational efficiency and sensitivity to directional defects.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] In a first aspect, the present invention provides a method for detecting defects in ground-penetrating radar images, comprising the following steps:

[0006] S1. Acquire ground-penetrating radar image data and perform preprocessing;

[0007] S2. Construct a disease detection model, which includes a multi-scale vertical strip convolution module, a multi-scale horizontal strip convolution module, an adaptive gated routing module, and a bidirectional attention fusion module; wherein, the multi-scale vertical strip convolution module is used to extract vertical features; the multi-scale horizontal strip convolution module is used to extract horizontal features; the adaptive gated routing module is used to dynamically weight and fuse the multi-scale features; and the bidirectional attention fusion module is used to adaptively weight and fuse the features in the vertical and horizontal directions.

[0008] S3. Train the disease detection model using the preprocessed data;

[0009] S4. Use the trained model to identify and locate defects in the ground-penetrating radar images to be tested.

[0010] Furthermore, both the multi-scale vertical stripe convolution module and the multi-scale horizontal stripe convolution module are implemented using depthwise separable convolution, with kernel scales including at least three of 3, 7, 15, 19, and 31. This design can capture multi-scale features from local details to macroscopic structures with low parameter count and computational overhead, significantly improving the model's ability to perceive diseases of different sizes.

[0011] Furthermore, the execution of the adaptive gated routing module includes the following steps: stacking the output features of the multi-scale branches into a feature tensor; performing global average pooling on the stacked features to obtain channel-level description vectors; generating weight scores for each branch through two layers of 1×1 convolution; normalizing the weight scores using a Softmax function with a temperature parameter to obtain the fusion weights for each branch; and performing a weighted summation of the multi-scale features based on the fusion weights to obtain the fused features. This mechanism enables the network to adaptively select and fuse the most relevant scale features according to the input image content, enhancing the model's feature selection ability and robustness in complex noise backgrounds.

[0012] Furthermore, the temperature parameter is used to adjust the smoothness of the weight distribution, thereby flexibly controlling the "hardness" of feature fusion.

[0013] Furthermore, the execution of the bidirectional attention fusion module includes the following steps: concatenating the output features of the vertical branch and the output features of the horizontal branch along the channel dimension; performing global average pooling on the concatenated features and generating two weight values ​​through a two-layer convolutional network; normalizing the weight values ​​using the Softmax function to obtain the fusion weights for the vertical and horizontal features; and performing weighted fusion of the vertical and horizontal features according to the fusion weights. This mechanism enables dynamic adjustment of the contribution of vertical and horizontal features, allowing the network to adaptively strengthen the utilization of the dominant directional features in the current input, thereby comprehensively improving the detection accuracy of diseases in any direction.

[0014] Furthermore, the disease detection model is based on the YOLOv8m network structure, and its C2f module is replaced with a bidirectional fusion module that includes the multi-scale vertical stripe convolution module, the multi-scale horizontal stripe convolution module, the adaptive gated routing module, and the bidirectional attention fusion module. This design enables the core modules of the invention to be plug-and-play, facilitating rapid deployment and application migration within existing high-performance detection frameworks.

[0015] Secondly, the present invention provides a ground-penetrating radar image defect detection system, comprising: a data acquisition and preprocessing unit for acquiring and preprocessing ground-penetrating radar image data; a model building unit for building a defect detection model as described above; a training unit for training the defect detection model using the preprocessed data; and a detection unit for using the trained model to identify and locate defects in the image to be detected.

[0016] Thirdly, the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the ground-penetrating radar image defect detection method as described above.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] 1. High computational efficiency and fewer parameters: By decomposing the square convolution into a concatenation of vertical and horizontal strip convolutions and replacing the standard convolution with depthwise separable convolution, the computational complexity of the model is greatly reduced (from O(k²) to O(2k)) and the number of parameters. While maintaining or even expanding the receptive field, it is more conducive to real-time deployment in engineering.

[0019] 2. Strong multi-scale and directional adaptive capabilities: By combining multi-scale strip convolution kernel design with adaptive gating routing, the model can dynamically fuse contextual information at different scales, effectively addressing the problem of large differences in the target scale of diseases; through a bidirectional attention fusion mechanism, the model can adaptively weight features in different directions, significantly improving the detection capability of anisotropic diseases.

[0020] 3. Significantly Improved Detection Accuracy: The aforementioned improvements enable the model to better focus on effective disease features and suppress noise interference in complex GPR images. Experiments show that, on publicly available GPR disease datasets, compared to the baseline model, the method of this invention significantly improves the mean accuracy (mAP), while significantly reducing model size and computational cost, achieving a balance between accuracy and efficiency. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein:

[0022] Figure 1 This is a detailed architecture diagram of the multi-scale vertical strip convolution module (AdaptiveStripConv_V) in an embodiment of the present invention.

[0023] Figure 2This is an architecture diagram of the bidirectional fusion module (C2f_asc) in an embodiment of the present invention.

[0024] Figure 3 This is a schematic diagram of the overall network configuration based on YOLOv8m integration in an embodiment of the present invention.

[0025] Figure 4 These are example images of cracks in the road surface internal structure defect dataset used in embodiments of the present invention.

[0026] Figure 5 These are example images of structural defects in the road surface internal structure defect dataset used in embodiments of the present invention.

[0027] Figure 6 This image shows the results of disease detection on GPR images using the baseline model (YOLOv8m).

[0028] Figure 7 The image shows the effect of applying the method of this invention (ASC) to the detection of defects on GPR images. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0030] This invention provides a method for detecting defects in ground-penetrating radar images, comprising the following steps:

[0031] S1. Acquire ground-penetrating radar image data and perform preprocessing;

[0032] S2. Construct a disease detection model, which includes a multi-scale vertical strip convolution module, a multi-scale horizontal strip convolution module, an adaptive gated routing module, and a bidirectional attention fusion module;

[0033] Specifically, the multi-scale vertical strip convolution module is used to extract vertical features;

[0034] The multi-scale horizontal strip convolution module is used to extract lateral features;

[0035] The adaptive gated routing module is used to dynamically weight and fuse multi-scale features;

[0036] The bidirectional attention fusion module is used to adaptively weight and fuse features in the vertical and horizontal directions;

[0037] S3. Train the disease detection model using the preprocessed data;

[0038] S4. Use the trained model to identify and locate defects in the ground-penetrating radar images to be tested.

[0039] In step S2, both the multi-scale vertical strip convolution module and the multi-scale horizontal strip convolution module are implemented using depthwise separable convolution, and their convolution kernel scales include at least three of the following: 3, 7, 15, 19, and 31.

[0040] The execution of the adaptive gated routing module includes the following steps:

[0041] Stack the output features of the multi-scale branches into a feature tensor;

[0042] Global average pooling is performed on the stacked features to obtain channel-level description vectors;

[0043] The weight scores for each branch are generated by two layers of 1×1 convolution;

[0044] The weight scores are normalized using the Softmax function with a temperature parameter to obtain the fusion weights of each branch.

[0045] The multi-scale features are weighted and summed according to the fusion weights to obtain the fused features.

[0046] Furthermore, the temperature parameter is used to adjust the smoothness of the weight distribution.

[0047] The execution of the bidirectional attention fusion module includes the following steps:

[0048] The vertical branch output features and the horizontal branch output features are concatenated along the channel dimension;

[0049] Global average pooling is performed on the concatenated features, and two weight values ​​are generated through a two-layer convolutional network.

[0050] The Softmax function is used to normalize the weight values ​​to obtain the fused weights for vertical and horizontal features;

[0051] The vertical and horizontal features are weighted and fused according to the fusion weights.

[0052] The disease detection model is based on the YOLOv8m network structure, and the C2f module is replaced with a bidirectional fusion module that includes the multi-scale vertical strip convolution module, the multi-scale horizontal strip convolution module, the adaptive gated routing module, and the bidirectional attention fusion module.

[0053] The present invention also provides a ground-penetrating radar image defect detection system, comprising:

[0054] The data acquisition and preprocessing unit is used to acquire and preprocess ground-penetrating radar image data;

[0055] The model building unit is used to build the disease detection model described above.

[0056] Training unit, used to train the disease detection model using preprocessed data;

[0057] The detection unit is used to identify and locate diseases in the image to be detected using a trained model.

[0058] The multi-scale vertical strip convolution module and the multi-scale horizontal strip convolution module in the disease detection model have their convolution kernel scales configured as adjustable parameters, including at least three of the following: 3, 7, 15, 19, and 31.

[0059] The present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the ground-penetrating radar image defect detection method.

[0060] The method for detecting defects in ground-penetrating radar images provided by the present invention will be described in detail below with specific embodiments.

[0061] Example 1

[0062] The present invention provides a method for detecting defects in ground-penetrating radar images, the specific implementation of which includes the following steps:

[0063] Step 1: Data Preparation

[0064] Obtain a dataset of internal GPR pavement defects, including categories such as cracks, structural defects (cs), bridge deck defects (b), and repair sections (r). Perform preprocessing such as standardization and normalization on the data, and divide it into training, validation, and test sets in a 6:2:2 ratio. Typical defect samples from the dataset can be found in [reference needed]. Figure 4 and Figure 5 As shown, an example of GPR image of cracks and structural defects is presented.

[0065] Step 2: Model Building

[0066] The disease detection model of this invention is constructed using YOLOv8m as a baseline framework, and its overall network integration configuration can be found in [reference needed]. Figure 3 As shown. The core of this invention lies in replacing the original C2f module with the bidirectional fusion module (C2f_asc) designed in this invention. This module is the core of this invention, and its structure is as follows. Figure 2 As shown, it mainly includes two parallel branches and a fusion part:

[0067] Multi-scale vertical strip convolution module (AdaptiveStripConv_V): such as Figure 1 As shown, this module uses a set of vertical strip convolution kernels k (e.g., k∈{3,7,15,31}) with different heights and a width of 1 to perform depthwise separable convolution. For the input feature map F, each scale branch operation is as follows:

[0068] ;

[0069] in, Vertical asymmetric infill is used to maintain dimensions, with an infill amount of [missing information]. ; Indicates the kernel size as The depthwise separable convolution is used. This design is specifically designed for efficiently extracting longitudinally extending lesion features (such as vertical cracks) from images, capturing longitudinal contextual information with a large receptive field in linear computational complexity.

[0070] Multi-scale horizontal strip convolution module (AdaptiveStripConv_H): Symmetrical to the vertical module, it uses a set of horizontal strip convolution kernels with different widths and a height of 1 (e.g., 1×k), which is specifically used to extract lateral disease features (e.g., horizontal layering), and can efficiently model the horizontal extension characteristics of diseases.

[0071] The adaptive gated routing module (Router) is used to dynamically fuse features from multiple scales. First, the output features from the K scale branches are stacked:

[0072] ;

[0073] in, This is a stacked function. It computes gated inputs (e.g., summation). , For the first The output features of each scale branch are obtained, and a global description is obtained through global average pooling. :

[0074] ;

[0075] In the formula, For height; Width;

[0076] Then, a lightweight two-layer 1×1 convolutional network is used to generate the weight scores of each branch. :

[0077] ;

[0078] The fusion weights are obtained by Softmax normalization with a temperature coefficient τ. :

[0079] ;

[0080] The temperature coefficient is used to adjust the smoothness of the weight distribution, with a default value of 1.0. For the first The weight score of each branch.

[0081] The final fusion features are:

[0082] .

[0083] This mechanism enables the network to adaptively assign appropriate weights to features at different scales based on the input content, thereby enhancing its robustness to diseases at varying scales.

[0084] The fused features are then subjected to pointwise convolution (PWConv), batch normalization (BN), and SiLU activation to obtain the output of each branch. , ),in, For vertical branch output, Output for horizontal branches.

[0085] Bidirectional attention fusion module: outputs vertical branches With horizontal branch output Stitching along the channel dimension: ,in This is the concatenation function. The fusion weights are generated through an attention network:

[0086] ;

[0087] in The number of channels was reduced from 2C to C / 4. The output consists of two weight values, α and β, which, after Softmax normalization, satisfy α + β = 1, where α is... The weights, β is The weights. The final output is:

[0088] ;

[0089] This design enables automatic calibration of the contribution of vertical and horizontal features, allowing the network to optimize feature fusion for the dominant direction of disease in a specific image, thereby improving the detection accuracy of diseases in any direction.

[0090] Step 3: Model Training and Evaluation

[0091] Using the training and validation sets prepared in step 1, the constructed model is trained end-to-end with disease bounding boxes and categories as labels. After training, the model performance is evaluated using the test set. Key training parameters are as follows: initial learning rate 0.01, optimizer AdamW, weight decay 0.0005, training epochs 300.

[0092] Step 4: Disease Detection

[0093] By inputting the GPR image to be detected into the trained model, the model can output the category of the disease target in the image and its bounding box location, thus achieving automated disease detection. Figure 6 and Figure 7 As shown, compared with the original YOLOv8m model, the method of this invention (ASC) can detect various diseases more accurately, with fewer missed detections and false detections.

[0094] Experiments show that the present invention achieves an mAP50 of 94.48% on the ground-penetrating radar road defect dataset, which is 1.1 percentage points higher than the baseline model. At the same time, the number of parameters is reduced by about 29% (from 33.93M to 24.65M), and GFLOPs are reduced by about 31% (from 85.3 to 56.1), effectively improving the efficiency and accuracy of identifying internal defects in road structure during ground-penetrating radar data processing.

[0095] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting defects in ground-penetrating radar images, characterized in that, Includes the following steps: S1. Acquire ground-penetrating radar image data and perform preprocessing; S2. Construct a disease detection model, which includes a multi-scale vertical strip convolution module, a multi-scale horizontal strip convolution module, an adaptive gated routing module, and a bidirectional attention fusion module; The multi-scale vertical strip convolution module is used to extract vertical features. The multi-scale vertical strip convolution module uses at least three vertical one-dimensional strip convolution kernels of different scales to extract multi-scale vertical features in parallel. The shape of the vertical one-dimensional strip convolution kernel is (k,1), where k is the scale parameter. The multi-scale horizontal strip convolution module is used to extract lateral features. The multi-scale horizontal strip convolution module uses at least three horizontal one-dimensional strip convolution kernels of different scales to extract multi-scale lateral features in parallel. The shape of the horizontal one-dimensional strip convolution kernel is (1,k), where k is the scale parameter. The adaptive gated routing module dynamically weights and fuses multi-scale features in the vertical direction and multi-scale features in the horizontal direction to generate fused vertical and horizontal features. The adaptive gated routing module performs the following steps: Stack the output features of multi-scale branches in the same direction into a feature tensor; Global average pooling is performed on the stacked features to obtain channel-level description vectors; The weight scores for each branch are generated by two layers of 1×1 convolution; The weight scores are normalized using the Softmax function with a temperature parameter to obtain the fusion weights of each branch. The multi-scale features are weighted and summed according to the fusion weights to obtain the fused features in that direction. The bidirectional attention fusion module adaptively weights and fuses the fused vertical features and the fused horizontal features to generate the final fused features; the bidirectional attention fusion module performs the following steps: The fused vertical features and fused horizontal features are spliced ​​along the channel dimension; Global average pooling is performed on the concatenated features, and two weight values ​​are generated through a two-layer convolutional network. The Softmax function is used to normalize the weight values ​​to obtain the fused weights of the vertical and horizontal features. The vertical and horizontal features are weighted and fused according to the fusion weights to obtain the final fused features; S3. Train the disease detection model using the preprocessed data; S4. Use the trained model to identify and locate defects in the ground-penetrating radar images to be tested.

2. The method according to claim 1, characterized in that, Both the multi-scale vertical strip convolution module and the multi-scale horizontal strip convolution module are implemented using depthwise separable convolution, and their convolution kernel scales include at least three of the following: 3, 7, 15, 19, and 31.

3. The method according to claim 1, characterized in that, The temperature parameter is used to adjust the smoothness of the weight distribution.

4. The method according to claim 1, characterized in that, The disease detection model is based on the YOLOv8m network structure, and the C2f module is replaced with a bidirectional fusion module that includes the multi-scale vertical strip convolution module, the multi-scale horizontal strip convolution module, the adaptive gated routing module, and the bidirectional attention fusion module.

5. A ground-penetrating radar image defect detection system, characterized in that, include: The data acquisition and preprocessing unit is used to acquire and preprocess ground-penetrating radar image data; A model building unit is used to build a disease detection model as described in any one of claims 1 to 4; Training unit, used to train the disease detection model using preprocessed data; The detection unit is used to identify and locate diseases in the image to be detected using a trained model.

6. The system according to claim 5, characterized in that, The multi-scale vertical strip convolution module and the multi-scale horizontal strip convolution module in the disease detection model have their convolution kernel scales configured as adjustable parameters, including at least three of the following: 3, 7, 15, 19, and 31.

7. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and when the processor executes the computer program, it implements the ground-penetrating radar image defect detection method as described in any one of claims 1 to 4.

Images