A geological disease ground penetrating radar detection method based on dynamic gating mixed experts

Abstract

The application discloses a geological disease ground penetrating radar detection method based on a dynamic gating hybrid expert, and belongs to the technical field of ground penetrating radar image processing and deep learning target detection. The method comprises the following steps: constructing a deep learning algorithm model, replacing an A2C2f module in a backbone network with a dynamic gating hybrid expert module based on a YOLO one-stage target detection framework; the module comprises a convolution layer, channel division, a dynamic gating module, an expert selection and weighted fusion module, and a channel fusion and residual output module; training the model by using a ground penetrating radar B-scan image dataset labeled with fissure and loose area diseases; inputting an image to be detected into the trained model, and outputting a detection frame and a category label of the fissure and the loose area. The application can adaptively select a feature extraction mode according to a disease scale, a reflection direction and a background interference degree, and effectively improves the detection precision of fissure and loose area diseases in a ground penetrating radar image.

Description

Technical Field

[0001] This invention belongs to the field of ground-penetrating radar image processing and deep learning target detection technology, and specifically relates to a ground-penetrating radar detection method for geological defects based on dynamic gating hybrid experts. Background Technology

[0002] Ground-penetrating radar (GPR) is a detection technology that uses the propagation, reflection, and scattering patterns of high-frequency electromagnetic waves in underground media to identify underground structures and abnormal areas. In GPR B-scan images, defects such as fissures and loose areas typically appear as areas of enhanced local reflection, altered reflection boundaries, hyperbolic disturbances, or areas of energy attenuation.

[0003] In geological defect detection, fissures and loose areas are two typical and engineering-significant defect targets. Fissures are typically narrow and elongated with weak boundary reflections, easily blending with surrounding background reflections. Loose areas usually exhibit large-scale layered structural anomalies or variations in reflection energy; their boundaries are not always clear, and they are easily superimposed with direct surface waves, interlayer interface reflections, and random noise. Therefore, ground-penetrating radar (GPR) detection of underground defects needs to simultaneously address issues such as unclear defect characteristics, large differences in target scale, complex variations in reflection morphology, and strong background reflection interference.

[0004] Traditional ground-penetrating radar interpretation methods rely on manual interpretation, resulting in low processing efficiency and interpretation results significantly influenced by human experience. Traditional image processing methods typically depend on filtering, thresholding, edge detection, or manual feature extraction. While these methods can highlight some anomalies against simple backgrounds, they struggle to reliably distinguish between real defects and background reflections in complex underground structures, multi-layered media, and under varying acquisition conditions.

[0005] Deep learning-based object detection methods can learn target features through labeled samples, offering advantages such as end-to-end training and automatic feature extraction. The YOLO series of one-stage detection algorithms boasts advantages like high speed, clear structure, and convenient deployment, making them suitable for rapid defect identification in ground-penetrating radar (GPR) images. The YOLOv12 detection framework typically consists of three parts: a backbone network, a neck network, and a detection head. The backbone network extracts image features from shallow to deep layers, the neck network fuses features at different scales, and the detection head outputs target location and category at multiple scales. However, directly applying the original YOLOv12 to GPR crack and loose area detection still has shortcomings: fixed convolutional blocks struggle to simultaneously adapt to both narrow cracks and large loose areas; multi-scale fusion may introduce redundant background responses; and a fixed feature extraction path applies the same processing method to different samples, making it difficult to select appropriate feature extraction methods based on the target scale, reflection direction, and background interference level of the current sample.

[0006] Therefore, a method is needed to improve the deep feature extraction path based on the YOLOv12 framework, so that the network can adaptively select the feature extraction method according to the scale, orientation and background interference features of the input sample, so as to simultaneously adapt to the detection needs of two types of disease targets: cracks and loose areas. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a ground-penetrating radar detection method for geological defects based on dynamic gating hybrid experts. This method enables the network to adaptively select expert branches for feature extraction based on the scale, orientation, and background interference characteristics of the input sample, thereby improving the detection accuracy of both fracture and loose zone defects.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows: A ground-penetrating radar method for detecting geological defects based on dynamic gating hybrid experts includes the following steps: Step 1, Construct a deep learning algorithm model: Based on the YOLO single-stage object detection framework, including the backbone network, neck network, and detection head; replace the A2C2f module used for deep feature extraction in the backbone network with a dynamic gated hybrid expert module. This module contains a dynamic gated module and multiple Inception deep convolutional experts, which are used to automatically select the appropriate expert for feature extraction based on the input sample. Step 2, train the model: train the model using a ground-penetrating radar B-scan image dataset labeled with cracks and loose areas of disease; Step 3, Disease Detection: Input the ground-penetrating radar B-scan image to be detected into the trained model. After processing by the backbone network, neck network and detection head, the detection boxes and category labels of cracks and loose areas are output.

[0009] In the above scheme, the dynamic gated hybrid expert module includes: a convolutional layer, channel partitioning, a dynamic gating module, an expert selection and weighted fusion module, and a channel fusion and residual output module. After the module feature map is input, it is expanded through a convolutional layer, and then split into two parts along the channel dimension. The first part is used as the feature map input for the dynamic gating module, and the second part is used as the expert input feature map input for the expert selection and weighted fusion module. The dynamic gating module outputs an expert activation mask and expert fusion weights to the expert selection and weighted fusion module based on its input feature map. The expert selection and weighted fusion module contains four Inception deep convolutional experts (Expert 1, Expert 2, Expert 3, and Expert 4). It determines which experts to select based on the received expert activation mask, and sends the received expert input features to the selected experts for calculation. Then, it performs weighted fusion on the outputs of the selected experts according to the expert fusion weights to obtain a dynamic expert fusion feature map, which is output to the channel fusion and residual output module. The channel fusion and residual output module completes channel integration and residual connection before outputting the final feature.

[0010] In the above scheme, the dynamic gating module includes: a max pooling layer, an average pooling layer, vectorization processing, a fully connected layer one, and a fully connected layer two. Top-K Expert routing, Mask The process involves constructing an expert activation mask and expert scoring for selection. The dynamic gating module first extracts channel statistics from the feature map input through max pooling and average pooling layers. The sum of these two layers is then vectorized to form a vector with a length equal to the number of input channels. This vector then enters two fully connected branches. The upper branch passes through fully connected layers one and two... LeakyReLU The activation generates the final base scores of four experts, and the next branch passes through the fully connected layer 2. Softplus Activation and noise perturbation normalization generate the final noise perturbation values ​​for the four experts; the final base score is added to the final noise perturbation value to form the gating score for the four experts. Top-K Expert routing selects the two experts with the highest gating scores from among four experts. Mask The expert activation mask is constructed by converting the selected expert index into an expert activation mask. The final base score of the expert scores, after filtering out unselected experts, is then processed. Softmax The expert fusion weight is obtained by calculating the expert weights.

[0011] In the above scheme, the calculation formula for the noise disturbance normalization is: ; In the formula, Indicates the first b Each sample corresponds to four initial values ​​for expert noise perturbation. and These represent the mean and standard deviation of the four initial noise disturbance values ​​for the sample, respectively. Indicates the final value of the noise disturbance; The formula for calculating the gate control score is as follows: ; In the formula, Indicates the first b Gating scores from four experts corresponding to each sample. Indicates the first b The final expert baseline score for each sample.

[0012] In the above scheme, the expert scoring screening retains the final base scores of selected experts, while setting the final base scores of unselected experts to negative infinity; the expert fusion weighting is... Softmax The expert weights are calculated using the following formula: ; In the formula, Indicates the first b The sample is assigned to the first e The expert fusion weight of each expert This represents the final value of the expert baseline score after screening.

[0013] In the above scheme, in the expert selection and weighted fusion module, the outputs of the selected experts are weighted according to their corresponding expert fusion weights, and then the dynamic expert fusion feature map is obtained by summing the features from multiple experts. The calculation formula is as follows: ; In the formula, Indicates the first b Dynamic expert fusion feature map of each sample Indicates the first e One expert, Indicates the first b Expert input features of a sample Indicates the first b The sample is assigned to the first e The expert fusion weight of each expert.

[0014] In the above scheme, among the four Inception deep convolution experts, each expert first divides the expert input features into channels, and then enters the identity branch, the square deep convolution branch, the horizontal strip deep convolution branch, and the vertical strip deep convolution branch respectively. The outputs of each branch are concatenated through channels to form the expert output. The identity branch is used to preserve the original channel information, the square deep convolution branch is used to extract the two-dimensional local structure, the horizontal strip deep convolution branch is used to model the changes in the horizontal reflection structure on the coordinates of the sample image, and the vertical strip deep convolution branch is used to model the changes in the vertical reflection structure on the coordinates of the sample image.

[0015] In the above scheme, the four experts use a combination of receptive fields from small to large. Expert 1 uses a 3×3 square depth convolution kernel combined with a horizontal stripe depth convolution kernel and a vertical stripe depth convolution kernel of length 5. Expert 2 uses a 5×5 square depth convolution kernel combined with a horizontal stripe depth convolution kernel and a vertical stripe depth convolution kernel of length 7. Expert 3 uses a 7×7 square depth convolution kernel combined with a horizontal stripe depth convolution kernel and a vertical stripe depth convolution kernel of length 9. Expert 4 uses a 9×9 square depth convolution kernel combined with a horizontal stripe depth convolution kernel and a vertical stripe depth convolution kernel of length 11.

[0016] In the above scheme, in the channel fusion and residual output module, the dynamic expert fusion feature map first undergoes channel integration through a channel fusion convolutional layer, and is then added to the original input features of the dynamic gated hybrid expert module through channel fusion residual connections to obtain the gated multi-expert output. Subsequently, the gated multi-expert output sequentially passes through BN batch normalization, FFN pointwise convolutional feedforward network, layer scaling, and DropPath random depth regularization, and is then enhanced through feedforward residual connections to obtain the feature map after feature enhancement. The calculation formula is as follows: ; ; In the formula, F represents the original input features of the dynamic gating hybrid expert module, and Y represents the dynamic expert fusion feature map. This indicates a channel-fused convolutional layer. This indicates that the gating is controlled by multiple experts. This represents the feature map after feature enhancement, γ represents the layer scaling parameter, ⊙ represents element-wise multiplication, and DP represents the DropPath random depth regularization operation. This represents the output of the FFN pointwise convolutional feedforward network.

[0017] In the above scheme, the detection head is a three-scale detection head, which receives high, medium, and low-scale fused features output by the neck network respectively; during the model training phase, the detection head participates in the loss calculation at all three scales simultaneously, and the formula for calculating the total loss function is: ; In the formula, Indicates the total training loss. Indicates the bounding box localization loss. Represents the category classification loss. Indicates the distribution focus loss. , and These represent the weights of the corresponding loss terms.

[0018] The above technical solution provides a ground-penetrating radar detection method for geological defects based on dynamic gating hybrid experts, which has the following beneficial effects: First, this invention proposes a Dynamically Gated Hybrid Expert Module (DGME) and places it in the deep layers of the YOLOv12 backbone to improve the original A2C2f-type deep feature aggregation structure. Four Inception deep convolutional experts employ different kernel combinations to cover receptive fields ranging from small to large, while also considering two-dimensional local structures, lateral reflection structure variations in sample image coordinates, and longitudinal reflection structure variations. The dynamic gating module generates expert activation masks and expert fusion weights based on input features. The expert selection and weighted fusion module only calls the two selected experts and performs weighted fusion. This structure enhances the model's adaptability to differences in disease scale, changes in reflection direction, and complex background interference.

[0019] Second, the dynamic gating module simultaneously acquires local strong responses and overall channel energy through max pooling and average pooling, and then utilizes the fully connected layer to generate expert baseline scores and... LeakyReLU The activation process generates the final expert baseline score, which describes the fit between the sample and different experts. The noise perturbation value generated by the fully connected layer 2 is then processed... Softplus After activation becomes positive and noise perturbation is normalized, a final noise perturbation value is formed, which is used to improve the diversity of expert selection during the training phase. The final base score and the final noise perturbation value together form the gating score. Top-K Expert routing, Mask Constructing expert activation masks, expert scoring and screening, and Softmax The expert weight calculation sequentially completes the expert selection, suppression of unselected experts, and normalization of the selected expert weights. Therefore, the expert selection and weighted fusion module only calculates the selected experts and performs weighted summation, enabling the model to improve its representational capabilities while reducing redundant expert calculations, thus enhancing computational efficiency and sample adaptability.

[0020] Third, this invention replaces only the A2C2f module in the deep backbone layer with a dynamically gated hybrid expert module on the YOLOv12 framework, while maintaining the overall connection method of the original backbone network, neck network, and detection head. This design allows the dynamically gated hybrid expert module to be easily integrated into the existing YOLO framework, exhibiting good compatibility and deployability.

[0021] Fourth, based on experimental data, the DGME-YOLO model proposed in this invention outperforms YOLOv12n and YOLOv13n in key metrics such as Precision, mAP50, mAP75, and mAP50-95. Specifically, Precision improved from 0.8555 to 0.9406, mAP50-95 from 0.5939 to 0.6478, and mAP75 improved by 0.0396, indicating that the model maintains good classification ability while exhibiting better boundary localization quality under more stringent matching conditions. Visualized detection results further validate the advantages of the proposed model in reducing false positives and false negatives. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0023] Figure 1 A schematic diagram of the YOLOv12 network structure before improvement; Figure 2 This is a schematic diagram of the DGME-YOLO network structure of the present invention; Figure 3 This is a schematic diagram of the Dynamic Gated Hybrid Expert Module (DGME). Figure 4 This is a schematic diagram of the dynamic gating module structure; Figure 5 A schematic diagram of the expert selection and weighted fusion module structure; Figure 6 A schematic diagram of the Inception deep convolutional expert module structure; Figure 7 This is a schematic diagram of the channel fusion and residual output module structure; Figure 8 A visualization comparison of the model confidence detection results is provided, where (a) is the original label image, (b) is the YOLOv12n detection result, and (c) is the DGME-YOLO detection result of this invention. Figure 9 The visualization comparison of the detection results for model misses is shown in the figure, where (a) is the original label image, (b) is the detection result of YOLOv12n, and (c) is the detection result of DGME-YOLO of this invention. The red arrows indicate the miss locations. Figure 10 The image shows a comparison of the detection results for missed detections and false detections by the model. (a) is the original label image, (b) is the detection result of YOLOv12n, and (c) is the detection result of DGME-YOLO of this invention. The red arrows indicate the locations of missed detections, and the green arrows indicate the locations of false detections. Detailed Implementation

[0024] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0025] I. Overall Structure of DGME-YOLO Network This embodiment provides a ground-penetrating radar detection method for geological defects based on dynamic gating hybrid experts. This method, in situations such as... Figure 1 The DGME-YOLO detection network is built upon the YOLO single-stage object detection framework shown. For example... Figure 2 As shown, DGME-YOLO consists of three parts: a backbone network, a neck network, and a detection head.

[0026] Step 1: Building the Model The backbone network, from shallow to deep, consists of a basic convolutional layer (Conv), a basic local feature aggregation module (DS-C3K2), a depthwise separable downsampling convolutional layer (DSConv), and a dynamically gated hybrid expert module (DGME). The DGME module replaces the A2C2f module used for deep feature extraction in the original YOLO framework. Internally, the DGME module incorporates a dynamic gating module and multiple Inception deep convolutional experts, automatically selecting the appropriate expert for feature extraction based on the input samples. The backbone network outputs features at three resolutions—shallow, intermediate, and deep—to the neck network.

[0027] The neck network achieves multi-scale feature fusion through upsampling, feature concatenation, DS-C3K2, and Conv, combining deep semantic information with shallow spatial information.

[0028] The detection head employs a three-scale detection head (Detect), receiving high-, medium-, and low-scale fused features from the neck network output to locate small cracks, medium-scale defects, and large loose areas. The high-resolution detection branch is more effective at locating small cracks, while the medium- and low-resolution branches are better at identifying larger loose areas and deep structural anomalies. The detection head is the output component of the basic detection framework. Its output bounding box distribution parameters are decoded and transformed to obtain predicted boxes. The class responses are mapped to obtain class probabilities, and finally, confidence filtering and non-maximum suppression are used to obtain the final detection result.

[0029] Step 2: Train the model The model was trained using a dataset of ground-penetrating radar B-scan images labeled with cracks and loose areas of disease.

[0030] Step 3: Disease Detection The ground-penetrating radar B-scan image to be detected is input into the trained model. After processing by the backbone network, neck network and detection head, the model outputs the detection boxes, category labels and detection scores for cracks and loose areas.

[0031] II. Dynamic Gated Hybrid Expert Module like Figure 3 As shown, the input feature map of the Dynamic Gated Hybrid Expert Module (DGME module) is denoted as... This module consists of five parts: convolutional layers, channel partitioning, a dynamic gating module, an expert selection and weighted fusion module, and a channel fusion and residual output module. The DGME dynamic gating hybrid expert module is a deep adaptive feature extraction structure proposed in this invention, designed to address the problems of large differences in disease scale, complex changes in reflection direction, and insufficient adaptability of fixed convolutional paths. Starting with the DGME module feature map input, the 3×3 convolutional layer first completes local context integration and channel expansion, providing more comprehensive feature representation for subsequent gating discrimination and expert computation. Channel partitioning splits the output features of the 3×3 convolutional layer into the dynamic gating module feature map input and expert input features. The former is used to generate expert activation masks and expert fusion weights, while the latter is fed into the selected expert branches for feature extraction. The expert selection and weighted fusion module determines the selection status of experts 1 to 4 based on the expert activation masks and weights the outputs of the selected experts according to the expert fusion weights. After expert fusion, the channel fusion and residual output module sequentially completes channel integration, channel fusion residual connection, feedforward enhancement, and feedforward enhancement residual connection, finally obtaining the DGME module feature map output.

[0032] 1. Convolutional Layers and Channel Division In the DGME dynamically gated hybrid expert module, the input feature is denoted as F. First, a 3×3 convolutional layer expands the number of channels from C to 2C, outputting the projected feature. Its expression is: ; In the formula, This indicates a projection operation on a 3×3 convolutional layer. This represents the projected features; the projection completes local context integration while keeping the spatial dimensions unchanged, and generates two sets of features for subsequent gating branches and expert branches.

[0033] Subsequently, the channel partitioning module will Divide the data into two equal parts along the channel dimension, with each part containing 10 channels. The first one As the input to the dynamic gating module, the second copy As expert input features, these features are fed into the expert selection and weighted fusion module, and their expression is: ; This indicates a bisection along the channel dimension. Input the feature map for the dynamic gating module. Input features for experts. F g This is used to determine which experts are more suitable for the current sample and to generate expert activation masks and expert fusion weights. F e The input is used for feature extraction by the selected Inception deep convolution expert.

[0034] 2. Dynamic Gating Module The dynamic gating module is used to determine the experts to be selected for each sample and their fusion ratio. The dynamic gating module is based on... Generate expert activation masks and expert fusion weights for each sample. For example... Figure 4 As shown, it internally includes: max pooling layer, average pooling layer, vectorization processing, fully connected layer 1, and fully connected layer 2. Top-K Expert routing, Mask Construct an expert-activated mask and use expert scoring for screening.

[0035] The dynamic gating module first extracts channel statistics from the feature map input using max pooling and average pooling layers. The max pooling layer preserves local salient responses such as fracture boundaries and strong reflection points, while the average pooling layer describes the overall energy distribution of each channel. The sum of these two layers is then vectorized to form a length of... C The vector is used as the global basis for expert routing. The upper branch generates the initial base scores of the four experts through the fully connected layer, and then... LeakyReLU The activation yields the final base scores from four experts, representing the degree of matching between the current sample and the receptive fields and directional modeling abilities of different experts. The lower branch generates initial noise perturbation values ​​for the four experts through a second fully connected layer, and then sequentially passes through... Softplus The activation process transforms the noise perturbation into a positive value and normalizes the noise perturbation, resulting in final noise perturbation values ​​for the four experts. This is used during the training phase to improve the diversity of expert selection and prevent long-term concentration of samples in a single expert. Subsequently, the final base scores of the four experts are added to their final noise perturbation values ​​to form the gated scores for the four experts. Top-K The expert routing selects the two experts with the highest gating scores. Mask The expert activation mask is constructed by converting the selected expert index into an expert activation mask, which controls whether each expert participates in the current sample calculation. Expert scoring filtering is used to remove scores from unselected experts, ensuring that only those experts are included in the calculation. Softmax Experts no longer contribute weight in the weight calculation. SoftmaxThe expert weight calculation normalizes the scores of the selected experts into expert fusion weights, which are used to determine the contribution ratio of the selected experts in the subsequent fusion results.

[0036] (1) Max pooling, average pooling and vectorization Simultaneously, the data enters both the max-pooling and average-pooling layers. Max-pooling highlights local salient responses (such as fracture boundaries and strong reflection points), while average-pooling describes the overall energy distribution. The sum of both layers is then vectorized to form a global discrimination vector. : ; In the formula, This indicates max pooling. Indicates average pooling. Indicates vectorization operation, This is the channel statistics matrix for a batch of samples. Max pooling is used to retain strong response information in each channel, while average pooling is used to describe the overall response level of each channel. The combination of the two provides input for calculating the initial values ​​of the base scores from the four experts and the noise perturbation values.

[0037] (2) Basic scoring branch Entering the fully connected layer 1, after linear mapping, it is... LeakyReLU Activation (negative half-axis slope) Generate the final base scores of the four experts. (Regarding the first) (samples) ; In the formula, Based on the final value matrix of the basic score, and For the weights and biases of the fully connected layer 1, The activation function retains small gradients when the values ​​are negative.

[0038] (3) Noise disturbance branch Simultaneously, it enters the second fully connected layer, and after linear mapping, it is... Softplus activation( Generate initial noise perturbation values ​​for four experts. The expression is: ; In the formula, and For the weights and biases of the fully connected layer 2, Softplus This represents a non-negative mapping function that transforms noise disturbances into positive values. NThis represents the positive value matrix of noise perturbation. The noise here is neither the noise collected by the bottom-penetrating radar nor the detection error, but rather a learnable expert selection auxiliary quantity generated by the gating network based on the input features, used to adjust the expert selection process during the training phase.

[0039] To ensure that noise perturbation primarily reflects the relative differences between different experts within the same sample, the dynamic gating module normalizes the positive noise perturbation values ​​along the expert dimension, obtaining the final noise perturbation values ​​for the four experts. For the b-th sample, its standardized expression is: ; In the formula, For the first The initial vector of noise perturbation for each sample. and These are the mean and standard deviation of the four initial values ​​for the sample, respectively. This is the final value vector of the noise perturbation. This processing is used to reduce the overall magnitude difference of the noise perturbation values, making the expert selection focus more on the relative differences between the four experts.

[0040] (4) Gating score and Top-K Expert Router During the expert selection phase, the dynamic gating module will... b The final baseline scores of the four experts for each sample are added together with the final noise perturbation scores of the four experts to form the gating scores of the four experts, which are expressed as follows: ; In the formula, For the first The gating score vector for each sample.

[0041] After the gating score is obtained, Top-K Expert routing selects the two experts with the highest gating scores from four experts. ):

[0042] In the formula, For the two selected expert index sets. TopK express Top-K Expert routing, in this embodiment K =2. This step ensures that only two of the four expert branches are activated for each sample, thus controlling computational overhead while preserving multi-expert expressiveness. Based on the selected set of expert indices, the dynamic gating module further... Mask Construct an expert activation mask to control the selection state of each expert.

[0043] (5) Mask Construction and expert scoring screening according to Construction expert activates mask Experts activate mask m b,e Indicates the first b The nth sample pair e The selection status of an expert. A value of 1 indicates that the expert is selected and participates in the feature extraction of the current sample, while a value of 0 indicates that the expert is not selected and does not participate in the feature extraction of the current sample.

[0044] Once the expert pool is determined, the dynamic gating module performs expert scoring and screening, retaining only the final baseline scores of the selected experts and setting the scores of the unselected experts to negative infinity, in order to... Softmax In the expert weight calculation, the weight of unselected experts is reset to 0. The expression is as follows: ; In the formula, This represents the final score of the expert baseline rating after screening.

[0045] It should be noted that the gating scores of the four experts are used to determine the expert index set. After the expert set is determined, the final expert fusion weight is calculated from the final values ​​of the basic scores of the selected experts. Therefore, noise disturbances affect the expert selection process, but do not directly change the final fusion ratio among the selected experts.

[0046] (6) Softmax Expert weight calculation The scores after screening are then evaluated. Softmax The expert fusion weights are calculated as follows: ; In the formula, For the first The sample is assigned to the first The fusion weight of each expert is calculated as follows: the weight of an unselected expert is 0, and the sum of the weights of the two selected experts is 1.

[0047] The dynamic gating module will activate the expert mask. Integration weight with experts Output to the expert selection and weighted fusion module.

[0048] 3. Expert Selection and Weighted Fusion Module like Figure 5 As shown, this module contains four parallel Inception deep convolutional experts: Expert 1, Expert 2, Expert 3, and Expert 4. The module receives three inputs: expert input features. (From channel division), expert activation mask Integration weight with experts (All from the dynamic gating module).

[0049] The expert selection and weighted fusion module uses the output expert activation mask and expert fusion weights as control information. The expert activation mask represents the selection status of experts 1, 2, 3, and 4; ON indicates an expert is selected, and OFF indicates an expert is not selected. Unselected experts skip the current sample calculation, thus avoiding redundant computation caused by all experts participating simultaneously. The expert fusion weights adjust the contribution of the selected experts' outputs to the fusion result, enabling the network to adaptively allocate expert responses based on the disease scale, reflection direction, and background interference level of the current sample. The outputs of the selected experts are weighted according to their corresponding weights and then summed through multi-expert feature fusion to form a dynamic expert fusion feature map. This process allows DGME to obtain multi-expert expressive power while controlling computational load.

[0050] (1) Internal structure of each Inception deep convolution expert like Figure 6 As shown, each expert first... The channels are divided into four branches: identity branch, square depth convolution branch, horizontal strip depth convolution branch, and vertical strip depth convolution branch.

[0051] Channel partitioning: Let the number of input channels be... The number of channels in each convolutional branch The division method is as follows:

[0052] In the formula, For the identity branch input, Input for the square convolution branch, For the input of the horizontal strip convolution branch, This is the input for the vertical strip convolution branch. This splitting method preserves most of the original channels while extracting the two-dimensional local structure, the lateral reflection structure changes and the vertical reflection structure changes on the sample image coordinates with fewer channels.

[0053] Processing the four branches: Identity Branches: Output directly, preserving the original channel information.

[0054] Square depthwise convolution branch: for implement Depthwise convolution extracts two-dimensional local structures.

[0055] Lateral strip depthwise convolution branch: implement Depth convolution models lateral reflection variations.

[0056] Vertical stripe depthwise convolution branch: for implement Depth convolution models vertical reflection variations.

[0057] The outputs of each branch are concatenated along the channel to obtain the expert output: ; In the formula, Indicates the first One expert, For channel splicing, Represents depthwise convolution. and These represent the size of the square convolution kernel and the length of the strip convolution kernel, respectively.

[0058] Convolutional kernel configurations for four experts (receptive field from smallest to largest): Expert 1 uses a 3×3 square depth convolution kernel, combined with horizontal and vertical stripe depth convolution kernels of length 5, suitable for extracting fine crack boundaries and local reflection perturbations. Expert 2 uses a 5×5 square depth convolution kernel, combined with horizontal and vertical stripe depth convolution kernels of length 7, suitable for extracting medium-scale cracks and the edges of local loose regions. Expert 3 uses a 7×7 square depth convolution kernel, combined with horizontal and vertical stripe depth convolution kernels of length 9, suitable for extracting large-scale reflection structure variations. Expert 4 uses a 9×9 square depth convolution kernel, combined with horizontal and vertical stripe depth convolution kernels of length 11, suitable for extracting large-scale loose regions, strong reflection anomalies, and long stripe reflection structures.

[0059] That is: Expert 1: , Expert 2: , Expert 3: , Expert 4: , .

[0060] The four experts are not simply stacked, but rather form a sequence of receptive fields from small to large. Expert 1 is better suited for extracting crack boundaries and small-scale reflection changes; Expert 2 is better suited for extracting medium-scale cracks and the edges of local loose areas; Expert 3 is better suited for extracting large-scale reflection structure changes; and Expert 4 is better suited for large-scale loose areas, strong reflection anomalies, and long strip reflection structures. Through this set of experts, DGME can simultaneously cover local two-dimensional structures, lateral reflection structure changes on the sample image coordinates, and longitudinal reflection structure changes.

[0061] (2) Expert selection and weighted fusion In the expert selection and weighted fusion module, four experts are selected sequentially according to their numbers. If no sample is assigned to a particular expert in the current batch, that expert remains unselected and the calculation is skipped. If a sample is assigned, that expert is selected, and only the expert input features of the corresponding sample are fed to that expert, while the expert output is accumulated according to the expert fusion weights. This is based on the expert activation mask. Only selected experts will receive [the information]. The calculation is performed, skipping unselected experts. The outputs of the selected experts are weighted and summed to obtain the dynamic expert fusion feature map. : ; In the formula, For the first Dynamic expert fusion feature map of each sample For the first One expert, For the first b Expert input features of a sample Indicates the first b The sample is assigned to the first e The expert fusion weights are calculated for each expert. Since only two experts in each sample have fusion weights greater than 0, the outputs of only the two selected experts are accumulated during the actual fusion process.

[0062] After four experts complete the selection, feature extraction, and weighted summation, a dynamic expert fusion feature map is obtained. This dynamic expert fusion feature map is then input into the channel fusion and residual output module.

[0063] 4. Channel Fusion and Residual Output Module This module is for Post-processing is performed to output the final features. .like Figure 7 As shown, the internal structure includes, in sequence: channel fusion convolutional layer, channel fusion residual connection, batch normalization (BN), pointwise feedforward convolutional network (FFN), layer scaling, droppath random depth regularization, and feedforward enhanced residual connection.

[0064] The channel fusion and residual output module is used for post-processing of the dynamic expert fusion feature map, including channel fusion residual connections and feedforward enhancement residual connections. The input dynamic expert fusion feature map first passes through a channel fusion convolutional layer to integrate channels, and is then added to the module feature input through a channel fusion residual connection to obtain the gated multi-expert output. Subsequently, the gated multi-expert output passes through batch normalization (BN), a pointwise feedforward convolutional network (FFN), layer scaling, and DropPath stochastic depth regularization, and is then enhanced through feedforward enhancement residual connections to obtain the feature map. Specifically, the channel fusion convolutional layer integrates channel information from different expert outputs; the channel fusion residual connection preserves the original features in the DGME module input; batch normalization (BN) stabilizes the batch feature distribution; the pointwise feedforward convolutional network (FFN) enhances the nonlinear expression between channels; layer scaling and DropPath stochastic depth regularization improve the stability of deep training; and the feedforward enhancement residual connection maintains a stable residual propagation path after feedforward enhancement.

[0065] (1) Channel fusion convolution and first residual connection First, pass through the channel fusion convolutional layer (1×1 convolution) restores the number of channels to the same level as the original input. Same, then with Adding them together yields the gated multi-expert output. : ;

[0066] In the formula, For channel fusion convolutional layers, This step involves gated multi-expert outputs. It corresponds to channel fusion residual connections, used to integrate channel information from different expert outputs while preserving the original features from the DGME module input.

[0067] (2) BN, FFN, layer scaling, DropPath and second residual connection After batch normalization (BN), the signal is fed into the FFN pointwise convolutional feedforward network. The FFN consists of a pre-processor pointwise convolution (channel expansion), GELU activation, Dropout, and a post-processor pointwise convolution (channel compression), outputting... Then multiply by the learnable layer scaling parameter. Then, it undergoes DropPath random depth regularization, and finally is compared with the original input. Adding them together gives the final output: ; In the formula, For batch normalization, and These represent the pre-pointwise convolution and the post-pointwise convolution, respectively. For activation function, For random inactivation, For layer scaling parameters, For element-wise multiplication, This is a DropPath operation. This is the final output feature map for the module. Represents the normalized features. express FFN Output of pointwise convolutional feedforward network.

[0068] At this point, the DGME module has completed the process from input... To output The complete calculation.

[0069] III. Three-Scale Detector Head and Loss Function The detection head is a three-scale detection head, receiving high, medium, and low-scale fused features from the neck network. The high-resolution detection head is used to locate small cracks, while the medium- and low-resolution detection heads are used to identify large-scale loose areas and deep structural anomalies. Each detection head simultaneously outputs bounding box distribution parameters and class responses at each spatial location. The dimensional relationship between the detection heads at each scale location and the bounding box distribution regression results and class branch results can be expressed as follows: P box ∈ R 4N_d , P cls ∈ R N_c ; In the formula, P box This represents the bounding box distribution regression output. P cls Indicates the output of the category branch. N d This represents the number of intervals in a discrete distribution along the boundary. N c This indicates the number of detection categories. In this embodiment, N c =2, corresponding to the loose zone and the crack, respectively.

[0070] During the inference phase, the bounding box distribution regression output first passes through... Softmax Normalization is performed, and the continuous boundary distances of the predicted bounding boxes are obtained through expectation decoding. The class branch output is then processed... SigmoidThe function maps to class probabilities. After completing bounding box decoding and class probability calculation, the model further removes low-confidence and duplicate prediction boxes through confidence filtering and non-maximum suppression, ultimately obtaining the detection box location, class label, and detection score of the disease target.

[0071] During the training phase, the detector head participates in loss calculation simultaneously at three scales. The total model loss consists of bounding box localization loss, class classification loss, and distribution focus loss, and its expression is: ;

[0072] In the formula, For total training loss, The bounding box localization loss (constraining the consistency of the predicted box and the ground truth box position) is used. Classification loss for categories (classification of cracks and loose regions). This is for distribution focus loss (improving the regression accuracy of discrete distributions of bounding boxes to continuous distances). , , The corresponding weights are: bounding box localization loss, which constrains the consistency between the predicted and ground truth bounding boxes; category classification loss, which supervises the classification of loose areas and cracks; and distribution focus loss, which improves the regression accuracy of the distance from the discrete distribution of the bounding box to the continuous boundary. Through the above joint optimization, the detection head can simultaneously improve the localization accuracy and classification accuracy of the disease target.

[0073] IV. Model Training and Experimental Validation 1. Dataset A dataset was constructed using a mixture of publicly available measured data and self-tested data. The publicly available data came from high-resolution GPR measured data acquired by Mohammed V University using GSSI ground-penetrating radar, including two types of defects: loose zones and fissures. The self-tested data was collected using a self-developed step-frequency drilling radar hardware system (antenna center frequency 520MHz). The data were uniformly organized into B-scan images and corresponding annotation files, and divided into training, validation, and test sets in a 7:2:1 ratio.

[0074] 2. Training parameters Data augmentation: random flipping (0.5 probability), ±5° rotation, 0.8~1.2x scaling, and color dithering. The optimizer used was AdamW with an initial learning rate of 0.001, cosine annealing scheduling, and weight decay of 0.0005. Batch size was 16, and training lasted 300 epochs. The model weights with the best mAP50 on the validation set were saved.

[0075] 3. Quantitative Experimental Results The results of comparing YOLOv13n, YOLOv12n, and the DGME-YOLO model of this invention on the test set are shown in Table 1: Table 1. Performance comparison of different models in the task of detecting underground defects using bottom-penetrating radar.

[0076] As shown in Table 1, the DGME-YOLO model of this invention outperforms YOLOv13n and YOLOv12n in key metrics such as Precision, mAP50, mAP75, and mAP50-95.

[0077] Compared to YOLOv13n, DGME-YOLO's Precision increased from 0.8481 to 0.9406, Recall from 0.7329 to 0.7493, mAP50 from 0.8193 to 0.8623, mAP75 from 0.6746 to 0.7250, and mAP50-95 from 0.5659 to 0.6478, indicating that the model of this invention has achieved improvements in false detection control, target recall, and multi-threshold localization evaluation.

[0078] Compared to YOLOv12n, DGME-YOLO's Precision improved from 0.8555 to 0.9406, mAP50 from 0.8369 to 0.8623, mAP75 from 0.6854 to 0.7250, and mAP50-95 from 0.5939 to 0.6478. Specifically, the improvements in mAP50-95 (0.0539) and mAP75 (0.0396) indicate that the model not only maintains good target classification capabilities but also exhibits better boundary localization quality under more stringent matching conditions. Although DGME-YOLO's Recall is slightly lower than YOLOv12n by 0.0099, its Precision and various mAP metrics are significantly higher, demonstrating the model's advantages in reducing false detections and improving overall detection accuracy.

[0079] 4. Visualization Results In addition to the quantitative indicators shown in Table 1, this embodiment provides the following to further visually verify the model's detection effect on loose zones and fissures: Figures 8 to 10 The visualization of the detection results is shown. In the figure, the dark blue box represents the detected loose areas, and the light blue box represents the detected cracks.

[0080] like Figure 8 As shown, in this set of samples, both YOLOv12n and the DGME-YOLO of the present invention can detect the target, but the detection box of DGME-YOLO is more consistent with the original label and has the highest target confidence, indicating that the model of the present invention has a stronger discrimination ability and a more stable target response for disease targets.

[0081] Depend on Figure 9It can be seen that YOLOv12n exhibited missed detections during the detection process; the red arrow indicates the location of the undetected disease targets. In contrast, DGME-YOLO was able to completely detect the corresponding targets without this missed detection problem, indicating that the model of this invention has a better ability to identify disease areas with weak reflection, indistinct boundaries, or small scale.

[0082] Depend on Figure 10 It is known that YOLOv12n suffers from both false negatives and false positives, with red arrows indicating false negatives and green arrows indicating false positives. The present invention, DGME-YOLO, can more accurately distinguish between real disease targets and background interference in the same set of samples, without exhibiting the corresponding false negatives or false positives.

[0083] Combining the quantitative results in Table 1 and Figures 8 to 10 The visualization results show that the model of this invention can more comprehensively and accurately complete the detection of underground defects by bottom-penetrating radar.

[0084] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily 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 invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A ground-penetrating radar detection method for geological defects based on dynamic gating hybrid experts, characterized in that, Includes the following steps: Step 1, Construct a deep learning algorithm model: Based on the YOLO single-stage object detection framework, including the backbone network, neck network, and detection head; replace the A2C2f module used for deep feature extraction in the backbone network with a dynamic gated hybrid expert module. This module contains a dynamic gated module and multiple Inception deep convolutional experts, which are used to automatically select the appropriate expert for feature extraction based on the input sample. Step 2, train the model: train the model using a ground-penetrating radar B-scan image dataset labeled with cracks and loose areas of disease; Step 3, Disease Detection: Input the ground-penetrating radar B-scan image to be detected into the trained model. After processing by the backbone network, neck network and detection head, the detection boxes and category labels of cracks and loose areas are output.

2. The method according to claim 1, characterized in that, The dynamic gated hybrid expert module includes: convolutional layers, channel partitioning, a dynamic gating module, an expert selection and weighted fusion module, and a channel fusion and residual output module. After the module feature map is input, it is expanded through convolutional layers, and then split into two parts along the channel dimension. The first part is used as the feature map input for the dynamic gating module, and the second part is used as the expert input feature map input for the expert selection and weighted fusion module. The dynamic gating module outputs an expert activation mask and expert fusion weights based on its input feature map to the expert selection and weighted fusion module. The expert selection and weighted fusion module contains four Inception deep convolutional experts (Expert 1, Expert 2, Expert 3, and Expert 4). It determines which experts to select based on the received expert activation mask, and sends the received expert input features to the selected experts for calculation. Then, it performs weighted fusion of the selected expert outputs according to the expert fusion weights to obtain a dynamic expert fusion feature map, which is output to the channel fusion and residual output module. The channel fusion and residual output module completes channel integration and residual concatenation before outputting the final feature.

3. The method according to claim 2, characterized in that, The dynamic gating module includes: a max pooling layer, an average pooling layer, vectorization processing, a fully connected layer 1, and a fully connected layer 2. Top-K Expert routing, Mask The process involves constructing an expert activation mask and expert scoring for selection. The dynamic gating module first extracts channel statistics from the feature map input through max pooling and average pooling layers. The sum of these two layers is then vectorized to form a vector with a length equal to the number of input channels. This vector then enters two fully connected branches. The upper branch passes through fully connected layers one and two... LeakyReLU The activation generates the final base scores of four experts, and the next branch passes through the fully connected layer 2. Softplus Activation and noise perturbation normalization generate the final noise perturbation values ​​for the four experts; the final base score is added to the final noise perturbation value to form the gating score for the four experts. Top-K Expert routing selects the two experts with the highest gating scores from among four experts. Mask The expert activation mask is constructed by converting the selected expert index into an expert activation mask. The final base score of the expert scores, after filtering out unselected experts, is then processed. Softmax The expert fusion weight is obtained by calculating the expert weights.

4. The method according to claim 3, characterized in that, The formula for calculating the normalization of the noise disturbance is: ； In the formula, Indicates the first b Each sample corresponds to four initial values ​​for expert noise perturbation. and These represent the mean and standard deviation of the four initial noise disturbance values ​​for the sample, respectively. Indicates the final value of the noise disturbance; The formula for calculating the gate control score is as follows: ； In the formula, Indicates the first b Gating scores from four experts corresponding to each sample. Indicates the first b The final expert baseline score for each sample.

5. The method according to claim 3, characterized in that, The expert scoring and screening process retains the final base scores of selected experts and sets the final base scores of unselected experts to negative infinity; the expert fusion weighting is... Softmax The expert weights are calculated using the following formula: ； In the formula, Indicates the first b The sample is assigned to the first e The expert fusion weight of each expert This represents the final value of the expert baseline score after screening.

6. The method according to claim 2, characterized in that, In the expert selection and weighted fusion module, the outputs of the selected experts are weighted according to their corresponding expert fusion weights, and then summed through multi-expert feature fusion to obtain a dynamic expert fusion feature map. The calculation formula is as follows: ； In the formula, Indicates the first b Dynamic expert fusion feature map of each sample Indicates the first e One expert, Indicates the first b Expert input features of a sample Indicates the first b The sample is assigned to the first e The expert fusion weight of each expert.

7. The method according to claim 2, characterized in that, In the four Inception deep convolution experts, each expert first divides the expert input features into channels, and then enters the identity branch, square deep convolution branch, horizontal strip deep convolution branch, and vertical strip deep convolution branch respectively. The outputs of each branch are concatenated through channels to form the expert output. The identity branch is used to preserve the original channel information, the square deep convolution branch is used to extract the two-dimensional local structure, the horizontal strip deep convolution branch is used to model the changes in the horizontal reflection structure on the coordinates of the sample image, and the vertical strip deep convolution branch is used to model the changes in the vertical reflection structure on the coordinates of the sample image.

8. The method according to claim 7, characterized in that, The four experts used a combination of receptive fields ranging from small to large. Expert 1 used a 3×3 square depth convolution kernel combined with a horizontal stripe depth convolution kernel and a vertical stripe depth convolution kernel of length 5. Expert 2 used a 5×5 square depth convolution kernel combined with a horizontal stripe depth convolution kernel and a vertical stripe depth convolution kernel of length 7. Expert 3 used a 7×7 square depth convolution kernel combined with a horizontal stripe depth convolution kernel and a vertical stripe depth convolution kernel of length 9. Expert 4 used a 9×9 square depth convolution kernel combined with a horizontal stripe depth convolution kernel and a vertical stripe depth convolution kernel of length 11.

9. The method according to claim 2, characterized in that, In the channel fusion and residual output module, the dynamic expert fusion feature map first undergoes channel fusion convolutional layer for channel integration, and is then added to the original input features of the dynamic gated hybrid expert module via channel fusion residual connections to obtain the gated multi-expert output. Subsequently, the gated multi-expert output passes through BN batch normalization, FFN pointwise convolutional feedforward network, layer scaling, and DropPath random depth regularization, and is further enhanced by feedforward residual connections to obtain the feature map after feature enhancement. The calculation formula is as follows: ； ； In the formula, F represents the original input features of the dynamic gating hybrid expert module, and Y represents the dynamic expert fusion feature map. This indicates a channel-fused convolutional layer. This indicates that the gating is controlled by multiple experts. This represents the feature map after feature enhancement, γ represents the layer scaling parameter, ⊙ represents element-wise multiplication, and DP represents the DropPath random depth regularization operation. This represents the output of the FFN pointwise convolutional feedforward network.

10. The method according to claim 1, characterized in that, The detection head is a three-scale detection head, which receives high, medium and low scale fused features output by the neck network respectively; During the model training phase, the detector head participates in the loss calculation simultaneously at three scales. The formula for calculating the total loss function is as follows: ； In the formula, Indicates the total training loss. Indicates the bounding box localization loss. Represents the category classification loss. Indicates the distribution focus loss. , and These represent the weights of the corresponding loss terms.

Images