Ground penetrating radar underground disease identification method, device and equipment based on bidirectional mamba YOLOv12 and medium

Abstract

The application discloses a bidirectional Mamba-based YOLOv12 ground penetrating radar underground disease identification method and device, equipment and medium, and relates to the technical field of road disease identification. The method comprises the following steps: constructing a four-channel physical perception tensor containing original signals, instantaneous amplitudes, phases and frequencies by processing ground penetrating radar gray-scale B-Scan data; designing an improved MPS-YOLOv12 network, expanding the input layer to 4 channels, decoupling physical features through four differentiated projection heads, modeling global dependence through bidirectional Mamba scanning, and finally outputting disease detection results through a multi-scale detection head by using the residual fusion mechanism of amplitude gating and phase compensation. By fully mining the physical information of GPR signals and using bidirectional Mamba, the disease identification accuracy is improved, and the false detection rate under complex geological background is effectively reduced.

Description

Technical Field

[0001] This invention relates to the field of road defect identification technology, and in particular to a method, device, equipment and medium for identifying underground defects using YOLOv12 ground penetrating radar based on bidirectional Mamba. Background Technology

[0002] Currently, underground disease detection mostly employs traditional target detection models (such as the YOLO series). These models use ground-penetrating radar (GPR) grayscale B-Scan data as input to a single-channel or three-channel visual image model, relying on convolutional neural networks or Transformer attention mechanisms to extract visual features such as texture and shape, and then output the disease detection results through a detection head. Some improved solutions perform simple preprocessing on the grayscale data (such as denoising and enhancement), but do not fully utilize the physical properties inherent in the GPR signal.

[0003] However, existing underground disease detection methods have limitations: First, they only utilize the brightness information of grayscale data, discarding frequency features sensitive to changes in medium properties and phase features sensitive to structural boundaries, making it difficult to distinguish weak diseases from background noise, resulting in high false detection and false negative rates. Second, feature extraction modules of general-purpose networks (such as traditional convolutional units and region attention units) are unable to capture long-distance geological structural dependencies spanning hundreds of scan channels in GPR data, leading to poor identification of long-span diseases. Third, the models are not adapted to the physical characteristics of GPR data, lacking robustness in complex scenarios such as strong reflection interference and waveform superposition, and some architectures based on state-space models suffer from training instability. Fourth, single-scale detection heads are difficult to adapt to the detection needs of diseases of different depths and sizes, resulting in limited detection accuracy.

[0004] Therefore, how to fully explore the physical attribute information of GPR signals and construct a disease identification model adapted to the characteristics of GPR data to improve the accuracy of underground disease detection has become an urgent problem to be solved. Summary of the Invention

[0005] The main objective of this application is to provide a method, device, equipment, and medium for identifying underground defects using YOLOv12 ground-penetrating radar based on bidirectional Mamba, aiming to solve the technical problem of how to improve the accuracy of underground defect detection.

[0006] To achieve the above objectives, this application proposes a YOLOv12 ground-penetrating radar method for identifying underground defects based on bidirectional Mamba, comprising:

[0007] Acquire grayscale B-Scan data from ground-penetrating radar;

[0008] The grayscale B-Scan data is processed by multiphysics input to obtain a four-channel physical sensing tensor;

[0009] The four-channel physical sensing tensor is input into the improved MPS-YOLOv12 network to obtain the disease detection results. The improved MPS-YOLOv12 network includes convolutional layers, shallow layers, deep layers, and a detection head. The number of input convolutional kernel channels in the convolutional layers is expanded to 4. The shallow layers retain A2C2f units and C3k2 units, and the deep layers are replaced with MPS-Block units. The MPS-Block units adopt a dual-path structure and internally nest multiple MPS-Mamba sub-units. The deep layers are layers with downsampling ratios of 16x and 32x. The detection head includes a high-scale detection head, a medium-scale detection head, and a low-scale detection head.

[0010] In one embodiment, the step of performing multiphysics input construction processing on the grayscale B-Scan data to obtain a four-channel physical sensing tensor includes:

[0011] The grayscale B-Scan data is subtracted from the pixel value reference number along the time axis to perform DC removal processing, resulting in a zero-mean radar signal;

[0012] Applying a Hilbert transform along the time axis to the zero-mean radar signal generates a complex analytic signal;

[0013] The instantaneous amplitude, instantaneous phase, and instantaneous frequency are calculated based on the complex analytic signal.

[0014] The zero-mean radar signal, instantaneous amplitude, instantaneous phase, and instantaneous frequency are respectively subjected to minimum-maximum normalization processing and then stacked along the channel dimension to obtain a four-channel physical sensing tensor.

[0015] In one embodiment, the step of calculating the instantaneous amplitude, instantaneous phase, and instantaneous frequency based on the complex analytic signal includes:

[0016] Calculate the squares and square roots of the real and imaginary parts of the complex analytic signal to obtain the instantaneous amplitude that fits the GPR energy distribution;

[0017] Calculate the arctangent values ​​of the imaginary and real parts of the complex analytic signal to obtain the instantaneous phase that adapts to the phase change characteristics of GPR;

[0018] Perform a phase unwrapping operation on the instantaneous phase to eliminate jumps and obtain a continuous phase function;

[0019] Perform a difference operation on the continuous phase function along the time axis to obtain the difference result;

[0020] The instantaneous frequency that adapts to the GPR dispersion characteristics is calculated based on the difference results and the time sampling interval.

[0021] In one embodiment, the step of inputting the four-channel physical sensing tensor into the improved MPS-YOLOv12 network to obtain the disease detection result includes:

[0022] The four-channel physical perception tensor is input into the input convolutional layer to complete channel adaptation and encode it into an initial feature map.

[0023] The initial feature map is input into the shallow A2C2f and C3k2 units to extract the local texture and edge features of the GPR image, thus obtaining the shallow local features;

[0024] The shallow local features are input into the deep MPS-Block unit, and the residual features and deep global features are obtained by processing them through a dual-path structure.

[0025] The residual features and the depth global features are concatenated by channels and then fused by 1×1 convolution to obtain a fusion result adapted for underground disease detection.

[0026] The fusion results are input into the high-scale detection head, the medium-scale detection head and the low-scale detection head respectively to generate corresponding multi-scale prediction feature maps;

[0027] Each of the multi-scale prediction feature maps is decoded to obtain a prediction result, wherein the prediction result includes bounding box coordinates, class confidence and target presence probability;

[0028] Non-maximum suppression is performed on the prediction results at all scales to obtain filtered prediction results, wherein the non-maximum suppression is performed using a preset intersection-over-union threshold and a preset confidence threshold.

[0029] The filtered prediction results are integrated by category to output disease detection results that include disease location, type, and confidence level.

[0030] In one embodiment, the step of inputting the shallow local features into a deep MPS-Block unit and processing them through a dual-path structure to obtain residual features and deep global features includes:

[0031] The shallow local features are input into the first path of the MPS-Block unit, and after 1×1 convolution, they are directly passed to obtain residual features with the same dimension as the input features.

[0032] The shallow local features are input into the second path of the MPS-Block unit, and the channels are compressed to half of the original number of channels through 1×1 convolution to obtain the compressed features;

[0033] The compressed features are input into the MPS-Mamba sub-unit to decouple the physical features and model global dependencies, thus obtaining the sub-unit output features.

[0034] Based on the element-wise multiplication of the energy features and the output features of the sub-unit, low-confidence features are filtered out to obtain the gated features;

[0035] The gated features are added element by element to the edge features to obtain the deep global features.

[0036] In one embodiment, the step of inputting the compressed features into an MPS-Mamba subunit, decoupling physical features and modeling global dependencies to obtain the subunit output features includes:

[0037] The compressed features are decoupled by four differentiated parallel projection heads to obtain texture features, energy features, dispersion features and edge features, wherein the four differentiated parallel projection heads are the mainstream semantic head, amplitude gating head, frequency head and phase head, respectively.

[0038] The dispersive features are superimposed element by element onto the texture features, and then subjected to layer normalization to obtain the enhanced features;

[0039] Perform a forward state space scan on the enhanced features to obtain forward context features;

[0040] After performing time reversal on the enhanced features, a forward scan is performed and then reversed again to obtain the reverse context features;

[0041] The positive context features are added to the negative context features to obtain bidirectional global features;

[0042] The bidirectional global feature and the energy feature are multiplied element by element to obtain the gated global feature;

[0043] The gating global feature, the texture feature, and the edge feature are added element by element to obtain the sub-unit output feature.

[0044] In one embodiment, the step of decoupling the compressed features through four differential parallel projection heads to obtain texture features, energy features, dispersion features, and edge features includes:

[0045] The compressed features are input into the mainstream semantic head, and the compressed features are dimension-mapped by 1×1 point convolution to obtain texture features.

[0046] The compressed features are input into the amplitude gating head, and the compressed features are extracted with a large receptive field through a 7×7 depth separable convolution. Then, the energy features are obtained by processing the compressed features through the Sigmoid activation function.

[0047] The compressed features are input into the frequency head, and the chromatic information of the compressed features is extracted by a 3×3 dilated convolution with a dilation rate of 2 to obtain the chromatic features.

[0048] The compressed features are input into the phase head, and edge information is extracted from the compressed features through a 3×3 standard convolution to obtain edge features.

[0049] Furthermore, to achieve the above objectives, this application also proposes a YOLOv12 ground-penetrating radar underground defect identification device based on bidirectional Mamba, the YOLOv12 ground-penetrating radar underground defect identification device based on bidirectional Mamba comprising:

[0050] The acquisition module is used to acquire grayscale B-Scan data from ground-penetrating radar;

[0051] The construction module is used to perform multi-physics input construction processing on the grayscale B-Scan data to obtain a four-channel physical sensing tensor;

[0052] The results module is used to input the four-channel physical sensing tensor into the improved MPS-YOLOv12 network to obtain the disease detection results. The improved MPS-YOLOv12 network includes convolutional layers, shallow layers, deep layers, and a detection head. The number of input convolutional kernel channels of the convolutional layers is expanded to 4. The shallow layers retain A2C2f units and C3k2 units. The deep layers are replaced with MPS-Block units. The MPS-Block units adopt a dual-path structure and internally nest multiple MPS-Mamba sub-units. The deep layers are layers with downsampling ratios of 16x and 32x. The detection head includes a high-scale detection head, a medium-scale detection head, and a low-scale detection head.

[0053] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the YOLOv12 ground-penetrating radar underground disease identification method based on bidirectional Mamba as described above.

[0054] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the YOLOv12 ground-penetrating radar underground disease identification method based on bidirectional Mamba as described above.

[0055] This application constructs a four-channel physical sensing tensor containing the original signal, instantaneous amplitude, phase, and frequency from ground-penetrating radar grayscale B-Scan data through processing. An improved MPS-YOLOv12 network is designed, expanding the input layer to four channels. In the shallow layer, A2C2f / C3k2 units are retained to extract local features, while in the deeper layers, dual-path MPS-Block units are used at the 16x and 32x downsampling levels. Physical features are decoupled through four differentiated projection heads, and global dependencies are modeled using bidirectional Mamba scanning. A residual fusion mechanism combining amplitude gating and phase compensation is employed, and finally, the disease detection results are output through a multi-scale detection head. By fully exploiting the physical information of the GPR signal and utilizing bidirectional Mamba, the accuracy of disease identification is improved, effectively reducing the false detection rate in complex geological backgrounds. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 This is a flowchart illustrating the first embodiment of the YOLOv12 ground-penetrating radar method for identifying underground defects based on bidirectional Mamba in this application.

[0058] Figure 2 This is a flowchart illustrating the second embodiment of the YOLOv12 ground-penetrating radar method for identifying underground defects based on bidirectional Mamba in this application.

[0059] Figure 3 This is a schematic diagram of the module structure of the YOLOv12 ground-penetrating radar underground disease identification device based on bidirectional Mamba, as described in this application.

[0060] Figure 4 This is a schematic diagram of the hardware operating environment involved in the YOLOv12 ground-penetrating radar underground disease identification method based on bidirectional Mamba in the embodiments of this application.

[0061] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0062] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0063] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0064] Currently, underground disease detection mostly employs traditional target detection models (such as the YOLO series). These models use ground-penetrating radar (GPR) grayscale B-Scan data as input to a single-channel or three-channel visual image model, relying on convolutional neural networks or Transformer attention mechanisms to extract visual features such as texture and shape, and then outputting the disease detection results through a detection head. Some improved solutions perform simple preprocessing on the grayscale data (such as denoising and enhancement), but do not fully utilize the physical properties inherent in the GPR signal; the network architecture mostly follows the structure of general visual detection models, without customizing the design for the anisotropy, weak texture, and long-distance dependence characteristics of GPR data.

[0065] However, existing underground disease detection methods have limitations: First, they only utilize the brightness information of grayscale data, discarding frequency features sensitive to changes in medium properties and phase features sensitive to structural boundaries, making it difficult to distinguish weak diseases from background noise, resulting in high false detection and false negative rates. Second, feature extraction modules of general-purpose networks (such as traditional convolutional units and region attention units) are unable to capture long-distance geological structural dependencies spanning hundreds of scan channels in GPR data, leading to poor identification of long-span diseases. Third, the models are not adapted to the physical characteristics of GPR data, lacking robustness in complex scenarios such as strong reflection interference and waveform superposition, and some architectures based on state-space models suffer from training instability. Fourth, single-scale detection heads are difficult to adapt to the detection needs of diseases of different depths and sizes, resulting in limited detection accuracy.

[0066] Based on the above, this application also provides a method for identifying underground defects using YOLOv12 ground-penetrating radar based on bidirectional Mamba, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the YOLOv12 ground-penetrating radar method for identifying underground defects based on bidirectional Mamba, as described in this application.

[0067] In this embodiment, the YOLOv12 ground-penetrating radar method for identifying underground defects based on bidirectional Mamba includes steps S10 to S30:

[0068] Step S10: Obtain grayscale B-Scan data from the ground penetrating radar.

[0069] It should be noted that grayscale B-Scan data is time-series data of the response of underground media to electromagnetic waves recorded by ground penetrating radar during detection. It is presented in the form of grayscale images, and the pixel values ​​reflect the intensity of the electromagnetic wave reflection signal.

[0070] Specifically, firstly, raw grayscale B-Scan data is acquired continuously at equal intervals along the detection route using a vehicle-mounted or portable ground-penetrating radar system. During acquisition, the radar transmission frequency is set to 900MHz and the time window to 20ns to ensure data coverage of the underground media information within the target depth range. This setup achieves a balance between detection depth and resolution, adapting to the needs of road structure layer defect detection. Secondly, the acquired raw data is converted into a single-channel two-dimensional matrix format. The number of rows in the matrix is ​​fixed at 1080, corresponding to the depth direction of each time sampling point. The number of columns is determined by the length of the survey line, corresponding to the horizontal spatial position. Each pixel value is stored as an 8-bit unsigned integer in the range of 0-255. This formatting standardizes the data input size, facilitating subsequent batch processing and neural network input adaptation. Then, abnormal samples caused by uneven antenna-to-ground coupling, environmental electromagnetic interference, and other factors during the acquisition process are screened and removed, retaining only valid data with a signal-to-noise ratio that meets the requirements. This is done to ensure the quality of the input data and prevent excessively noisy samples from interfering with model training convergence. Finally, the filtered grayscale B-Scan data is used as the raw input and directly applied to subsequent multiphysics construction without any compression or format conversion, thus preserving the complete dynamic range and physical information of the original data.

[0071] Step S20: Perform multi-physics input construction processing on the grayscale B-Scan data to obtain a four-channel physical sensing tensor.

[0072] It should be noted that the four-channel physical sensing tensor is a three-dimensional data structure formed by stacking four types of physical feature data along the channel dimension. The four types of data are zero-mean radar signal, instantaneous amplitude, instantaneous phase and instantaneous frequency.

[0073] Further, step S20 includes: First, subtracting the pixel value reference number from the grayscale B-Scan data along the time axis to perform DC removal processing to obtain a zero-mean radar signal, specifically using the following formula:

[0074]

[0075] in This represents the original input grayscale B-Scan data. These represent the time sampling points for the horizontal spatial coordinates along the survey line and the time sampling points for the depth direction, respectively. This represents the preprocessed zero-mean radar signal with a pixel reference number of 127.5. This is done to eliminate DC component interference in the signal, meet the mathematical requirement of Hilbert transform for zero-mean input, and avoid introducing low-frequency artifacts after the transform that would cause distortion in physical feature extraction.

[0076] Next, a Hilbert transform is applied to the zero-mean radar signal along the time axis to generate a complex analytic signal. The specific formula is as follows:

[0077]

[0078] in Represents a complex-valued analytic signal. The imaginary unit satisfies , This represents the Hilbert transform function, which produces a 90° phase shift in a real-valued signal, thus forming the imaginary part of the complex signal.

[0079] Then, the instantaneous amplitude, instantaneous phase, and instantaneous frequency are calculated based on the complex analytic signal. Specifically, the squares and square roots of the real and imaginary parts of the complex analytic signal are calculated to obtain the instantaneous amplitude adapted to the GPR energy distribution, using the following formula:

[0080]

[0081] in Indicates instantaneous amplitude.

[0082] Calculate the arctangent values ​​of the imaginary and real parts of the complex analytic signal to obtain the instantaneous phase that adapts to the GPR phase change characteristic. The specific formula is as follows:

[0083]

[0084] in Represents the imaginary part of a complex analytic signal. This represents the real part of a complex analytic signal.

[0085] A phase unwrapping operation is performed on the instantaneous phase to eliminate jumps, resulting in a continuous phase function. A difference operation is then performed on the continuous phase function along the time axis to obtain the difference result. Based on the difference result and the time sampling interval, the instantaneous frequency adapted to the GPR dispersion characteristics is calculated. The specific formula is as follows:

[0086]

[0087] in Represents the instantaneous change in phase. Represents a continuous phase function, eliminating discontinuous transitions The time sampling interval of the radar system. Indicates the length of the radar recording time window. This represents the height of the grayscale B-Scan data. In this embodiment, the time window length is set to 20 ns. =1080.

[0088] Finally, the zero-mean radar signal, instantaneous amplitude, instantaneous phase, and instantaneous frequency are respectively subjected to minimum-maximum normalization and then stacked along the channel dimension to obtain a four-channel physical sensing tensor. Specifically, in order to eliminate the numerical differences caused by different physical dimensions, all four channels ( , , All values ​​are mapped to the [0,1] range using the Min-Max normalization method, and then multiplied by 255 to restore the values ​​to the pixel value range. The specific formula is as follows:

[0089]

[0090] Where X represents the original data matrix of the physical channel, including ( , , ). and These represent the minimum and maximum values ​​within the channel, respectively. To prevent the use of tiny constants with a denominator of zero, this embodiment uses 10^-8. This represents the normalized data, stacked along the channel dimension to form the final four-channel physical sensing tensor. This is done to unify the numerical scale of different physical quantities and eliminate dimensional differences. At the same time, the four-channel parallel input allows the neural network to independently encode each physical attribute in the initial layer, avoiding the information bottleneck problem of traditional single-channel input.

[0091] Step S30: Input the four-channel physical sensing tensor into the improved MPS-YOLOv12 network to obtain the disease detection results.

[0092] It should be noted that the improved MPS-YOLOv12 network is a target detection network optimized based on the YOLOv12 architecture. It is customized for the characteristics of ground-penetrating radar data by expanding the number of channels in convolutional layers, replacing deep feature extraction units, and configuring multi-scale detection heads to adapt to multi-physics data and improve the detection capability of underground defects. The improved MPS-YOLOv12 network includes convolutional layers, shallow layers, deep layers, and a detection head. Specifically, the convolutional layer is the initial feature extraction layer of the improved MPS-YOLOv12 network. The number of channels in the input convolutional kernel of the convolutional layer is expanded to 4, specifically adapted to the input of a four-channel physical perception tensor, enabling preliminary encoding of four types of physical features. The shallow layer is the feature extraction layer near the input end of the network, retaining the original YOLOv12 A2C2f and C3k2 units. The deep layer consists of feature extraction layers with downsampling ratios of 16x and 32x. This deep layer is replaced by MPS-Block units, which employ a dual-path structure with multiple nested MPS-Mamba sub-units. The dual-path structure is the core of the MPS-Block unit, comprising a first path (global residual edge) and a second path (main computation branch). The first path ensures lossless feature transfer, while the second path performs deep feature extraction; both work synergistically to enhance feature representation capabilities. The MPS-Mamba sub-unit is the core computational unit of the MPS-Block unit, integrating differentiated parallel projection heads, bidirectional state-space scanning, and robust residual integration mechanisms. This decouples physical features and models long-distance dependencies, improving feature discriminative power. The deep layer consists of layers with downsampling ratios of 16x and 32x. The detection head includes high-scale, medium-scale, and low-scale detection heads, corresponding to feature maps of different sizes, adaptable to underground diseases of varying sizes, achieving full-scale disease coverage detection.

[0093] Specifically, a four-channel physical sensing tensor is input into a convolutional layer, and initial feature encoding is performed using a 4-channel input convolutional kernel to obtain an initial feature map. Then, the initial feature map is input into shallow A2C2f and C3k2 units to extract local texture and edge features. The shallow output features are then input into MPS-Block units at 16x and 32x downsampling levels. Local and global features are fused through a dual-path structure and MPS-Mamba subunits to obtain multi-scale fused features. Finally, the multi-scale fused features are input into high-scale, medium-scale, and low-scale detection heads. After target box regression, confidence filtering, and non-maximum suppression, the disease detection results are output.

[0094] Furthermore, to verify the effectiveness of the MPS-YOLOv12 model proposed in this embodiment for detecting underground defects in GPR, a simulation dataset was constructed and multiple sets of experiments were designed. This embodiment used gprMax 3.0 electromagnetic simulation software to construct a Hybrid GPR Dataset. The simulation environment was set as a typical highway road structure, including three layers: asphalt layer, base course, and subgrade, and four common defect targets—cavity, void, loose, and water-rich zone—were embedded. The radar transmission frequency was set to 900 MHz, the time window to 20 ns, and the antenna scanned and acquired data at equal intervals along the survey line. After screening and cleaning, the dataset contained 1000 high-quality B-Scan images, with the image size uniformly adjusted to 1080×1080 pixels. All raw single-channel data were processed according to the aforementioned multiphysics construction process, undergoing DC removal and Hilbert transform, and converted into 4-channel TIF format files containing original signal, instantaneous amplitude, instantaneous phase, and instantaneous frequency information. All targets were labeled with pixel-level bounding boxes using the LabelImg tool, and the label files followed the standard YOLO format. The dataset was randomly divided into a training set of 800 images, a validation set of 100 images, and a test set of 100 images in an 8:1:1 ratio. The training process employed a stochastic gradient descent (SGD) optimizer with a momentum factor of 0.937, a weight decay coefficient of 0.0005, an initial learning rate of 0.01 decaying using cosine annealing, a batch size of 8, and 200 training epochs. Mosaic and Mixup data augmentation were enabled during training, while rectangular inference was disabled during validation and testing to avoid image distortion affecting physical features.

[0095] To comprehensively evaluate the performance of MPS-YOLOv12, three representative state-of-the-art (SOTA) models were selected for horizontal comparison: the classic two-stage baseline Faster R-CNN, the real-time Transformer SOTA: RT-DETR, and the latest real-time CNN / hybrid architecture YOLOv11m and YOLOv12m. All models were trained in the same hardware and software environment, with a uniform input size of 256×256. Except for MPS-YOLOv12, which uses 4-channel input, the others used standard RGB input. Detailed performance data for each model on the GPR test set is shown in Table 1.

[0096] Table 1. Detailed performance data of each model on the GPR test set.

[0097]

[0098] As can be seen, compared with YOLOv12m, this embodiment improves mAP50 by 2.6%, accuracy by 3.1%, and the number of parameters decreases from 20.1M to 19.0M (a reduction of approximately 5.5%), indicating that the MPS-Mamba module has higher parameter efficiency than the original A2C2f attention module. Compared with RT-DETR-L, this embodiment achieves a 5.0% improvement in mAP50 with 58% fewer parameters (19.0M vs 32.8M), proving that the Physically Aware SSM architecture is more suitable for GPR data than the general Transformer. Compared with Faster R-CNN, this embodiment improves mAP50 by 6.1%.

[0099] This embodiment processes ground-penetrating radar (GPR) grayscale B-Scan data to construct a four-channel physical sensing tensor containing the original signal, instantaneous amplitude, phase, and frequency. An improved MPS-YOLOv12 network is designed, expanding the input layer to four channels. In the shallow layer, A2C2f / C3k2 units are retained to extract local features, while in the deeper layers, dual-path MPS-Block units are used at the 16x and 32x downsampling levels. Physical features are decoupled using four differentiated projection heads. Global dependencies are modeled using bidirectional Mamba scanning, and a residual fusion mechanism combining amplitude gating and phase compensation is employed. Finally, the disease detection results are output through a multi-scale detection head. By fully exploiting the physical information of the GPR signal and utilizing bidirectional Mamba, the accuracy of disease identification is improved, effectively reducing the false detection rate in complex geological backgrounds.

[0100] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 The YOLOv12 ground-penetrating radar method for identifying underground defects based on bidirectional Mamba, step S30, further includes steps S201 to S208:

[0101] Step S201: Input the four-channel physical perception tensor into the input convolutional layer to complete channel adaptation and encode it into an initial feature map.

[0102] Specifically, a four-channel physical perception tensor is input to the input convolutional layer of the improved MPS-YOLOv12 network. This layer reconstructs the standard YOLOv12's 3-channel input convolutional kernel into a 4-channel one. By adjusting the input dimension of the convolutional kernel weight matrix, it expands from 3×k×k to 4×k×k, where k is the convolutional kernel size. This allows the network front-end to simultaneously receive data from four physical channels: the original signal, instantaneous amplitude, instantaneous phase, and instantaneous frequency. This is done to avoid the traditional 3-channel RGB input being forced to discard a physical channel or perform redundant channel duplication, ensuring that the four physical attributes are independently encoded at the input stage. Then, the input convolutional layer performs a downsampling convolution operation with a stride of 2 on the four-channel tensor. Through a sliding window calculation, the 1080×1080×4 input tensor is mapped to an initial feature map of 540×540×C, where C is the number of output channels. This is done to extract preliminary local feature patterns while reducing spatial dimensions, providing a feature representation of appropriate scale for the subsequent deep processing of shallow networks. Finally, the initial feature map retains the physical information contained in the four-channel input, such as the amplitude energy distribution, phase transition location, and frequency dispersion characteristics. Pixel-level local associations are established through convolution weight learning. This is done because although only standard convolution operations are performed at this stage, the four-channel input enables the network to distinguish the response patterns of different physical quantities at the front end, avoiding the aliasing of physical information in the input layer.

[0103] Step S202: Input the initial feature map into the shallow A2C2f and C3k2 units to extract the local texture and edge features of the GPR image and obtain the shallow local features.

[0104] Specifically, the initial feature map is input into the first-level A2C2f unit of the shallow network. This unit divides the feature map into multiple local regions through a region-partitioning attention mechanism, calculating the dependencies between pixels within each region instead of global computation. This is done to reduce the computational complexity of the shallow high-resolution feature map while maintaining a large receptive field, avoiding the memory explosion problem caused by global attention. Then, the output of the A2C2f unit is downsampled and input into the C3k2 unit. This unit adopts a cross-stage partial connection structure, dividing the feature flow into two branches. One branch directly connects across stages, while the other branch is processed by multiple bottleneck convolutional layers before fusion. This is done to enhance the gradient backpropagation path, alleviate the gradient vanishing problem during deep network training, and reduce the number of parameters through the bottleneck structure. Next, alternately stacked A2C2f and C3k2 units extract local texture and edge features from the GPR image layer by layer. Shallow units focus on capturing subtle disease boundaries, layering changes, and local curvature features of hyperbolic diffraction waves. This is because GPR B-Scan images maintain high spatial resolution in the shallow stage, making them suitable for extracting fine local patterns using standard convolution operations. Diseases often exhibit local features such as phase discontinuities and energy abrupt changes in the image. Finally, after processing by multiple sets of A2C2f and C3k2 units, shallow local features are obtained. This feature map has a resolution of 1 / 4 to 1 / 8 of the input size, with the number of channels gradually increasing. This preserves the precise location information of disease edges and encodes the statistical regularities of local textures. This approach establishes a sufficient foundation of local features before subsequent deep network global context modeling, preventing the loss of crucial detailed boundary information during long-distance dependency modeling.

[0105] Step S203: Input the shallow local features into the deep MPS-Block unit, and process them through a dual-path structure to obtain residual features and deep global features respectively.

[0106] It should be noted that the MPS-Block unit is a core feature extraction unit for improving the deep layers of the MPS-YOLOv12 network. Specifically adapted for 16x and 32x downsampling levels, it employs a dual-path structure and nests multiple MPS-Mamba sub-units, enabling the fusion of local and global features to enhance feature representation capabilities. The dual-path structure is the core architecture of the MPS-Block unit, comprising a first path (global residual edges) and a second path (main computation branch). The first path ensures lossless feature transfer, while the second path performs deep feature extraction; both work synergistically to balance feature integrity and abstraction. Residual features are the output of the first path of the MPS-Block unit, directly preserving the core information of shallow local features and avoiding information loss during deep feature extraction. Deep global features are the output of the second path of the MPS-Block unit, obtained after channel compression and global dependency modeling by MPS-Mamba sub-units. They can capture long-range contextual information in ground-penetrating radar data, reflecting the global distribution characteristics of underground defects.

[0107] Further, step S203 includes: inputting shallow local features into the first path of the MPS-Block unit, passing them directly after a 1×1 convolution to obtain residual features with the same dimension as the input features; inputting shallow local features into the second path of the MPS-Block unit, compressing the channels to half the original number of channels after a 1×1 convolution to obtain compressed features; inputting the compressed features into the MPS-Mamba sub-unit, decoupling physical features and modeling global dependencies to obtain sub-unit output features; multiplying the energy features and sub-unit output features element-wise to filter low-confidence features to obtain gated features; and adding the gated features and edge features element-wise to obtain deep global features.

[0108] It should be noted that energy features are physical features decoupled from the compressed features, reflecting the energy distribution of the ground-penetrating radar signal. They can be used as gating signals to filter high-confidence features and invalid noise information. Sub-cell output features are features obtained by processing the compressed features using MPS-Mamba sub-cells. They contain decoupled physical features and global context information, forming the core foundation for deep global feature construction. Gated features are the result of element-wise multiplication of energy features and sub-cell output features. The gating effect of energy features filters low-confidence features, retaining high-value features related to subsurface defects. Edge features are physical features decoupled from the compressed features, reflecting the structural edge information of the subsurface medium. They can compensate for lost local details during global feature modeling, improving feature integrity.

[0109] Specifically, shallow local features are input in parallel into the dual-path structure of the MPS-Block unit. In the first path, after adjusting the channel dimensions through a 1×1 convolution, they are directly passed to the end of the module, resulting in residual features with the same dimensions as the input features. This is done to establish global residual connections, ensuring that shallow features can be losslessly passed to deeper layers, preventing gradient vanishing, and forming the basic path for the safety failure mechanism. Secondly, the shallow local features are input into the second path, where a 1×1 convolution compresses the number of channels to half of the original number, resulting in compressed features. This is done to reduce the computational complexity of subsequent Mamba scans, and simultaneously, through dimensionality compression, forces the network to learn a more compact feature representation, removing redundant information. Then, the compressed features are input into the MPS-Mamba sub-unit. Texture features, energy features, dispersion features, and edge features are decoupled using four differentiated parallel projection heads. The dispersion features are element-wise superimposed onto the texture features, then normalized by layers, and fed into a bidirectional state-space model for forward and reverse scanning. The outputs from both directions are summed to obtain the bidirectional global features. This is done because GPRB-Scan images exhibit significant spatial non-causality; the hyperbolic features of underground targets are symmetrically distributed in the left-right direction. Bidirectional scanning ensures that each pixel perceives the entire image context, addressing the spatial blind spot problem of unidirectional Mamba. Subsequently, the bidirectional global features are element-wise multiplied with the energy features. The energy features are used as a gating signal to filter low-confidence regions, resulting in gated features. This is done because amplitude energy reflects signal reflection intensity; high-energy regions correspond to real underground structures, while low-energy regions are mostly noise. The gating operation suppresses background noise interference, retaining only the high-confidence global context. Finally, the gated features and edge features are added element by element to obtain the deep global features. This is done to compensate for the smoothing of local details that may be caused by global scanning. The edge features extracted by the phase head are directly injected into the output to sharpen the disease boundaries. At the same time, the residual features are spliced ​​and fused with the deep global features at the end of the module to ensure that the network can still degenerate into a standard CNN even when the Mamba branch fails.

[0110] Furthermore, the steps of inputting the compressed features into the MPS-Mamba sub-unit, decoupling the physical features, and modeling global dependencies to obtain the sub-unit output features include:

[0111] First, the compressed features are decoupled using four differentiated parallel projection heads to obtain texture features, energy features, dispersion features, and edge features. Specifically, the four differentiated parallel projection heads are the mainstream semantic head, amplitude gating head, frequency head, and phase head. Specifically, the compressed features are input to the mainstream semantic head, where a 1×1 point convolution is used to map the dimensions of the compressed features to obtain texture features; the compressed features are input to the amplitude gating head, where a 7×7 depthwise separable convolution is used to extract a large receptive field, followed by processing with a sigmoid activation function to obtain energy features; the compressed features are input to the frequency head, where a 3×3 dilated convolution with a dilation rate of 2 is used to extract dispersion information to obtain dispersion features; and the compressed features are input to the phase head, where a 3×3 standard convolution is used to extract edge information to obtain edge features.

[0112] The dispersion features are element-wise superimposed onto the texture features, and then subjected to layer normalization to obtain the enhanced features. A forward state space scan is performed on the enhanced features to obtain the forward context features. The enhanced features are then time-reversed and subjected to a forward scan followed by reversal to obtain the reverse context features. The forward and reverse context features are added together to obtain the bidirectional global features. The bidirectional global features and energy features are multiplied element-wise to obtain the gated global features. The gated global features, texture features, and edge features are then element-wise added together to obtain the sub-unit output features. Specifically, the dispersion features output from the frequency head are element-wise superimposed onto the texture features output from the mainstream semantic head, and then subjected to layer normalization to obtain the enhanced features. This is done because the frequency dispersion features reflect the propagation characteristics of electromagnetic waves in different media; injecting them into the mainstream features enhances the model's ability to perceive medium properties, while layer normalization stabilizes the data distribution and accelerates the convergence of the subsequent state space model. Secondly, a forward state space scan is performed on the enhanced features. The two-dimensional spatial features are flattened into a one-dimensional sequence in row-major order. The hidden state is updated time-by-time using discretized state equations, and the output is calculated to obtain the forward context features. This is done to establish long-distance dependencies along the survey line direction, enabling each location to perceive the context information of all locations in front of it. Then, the enhanced features are time-reversed, i.e., the sequence is reversed along the sequence dimension, and the same forward scan is performed before reversing back to the original order to obtain the reverse context features. This is done because the spatial correlation of underground targets in GPR B-Scan images is omnidirectional. Unidirectional scanning will lead to spatial blind spots. Reverse scanning ensures that each location can also perceive the context of all locations behind it, solving the problem of spatial non-causality in hyperbolic symmetric structures. Subsequently, the forward and reverse context features are added element-wise to obtain bidirectional global features. This is done to aggregate context information from both directions, enabling each pixel on the feature map to perceive dependencies from the entire image, establishing a global receptive field. Next, the bidirectional global features and energy features are multiplied element-wise to obtain the gated global features. This is done by using amplitude energy as a gating signal; high-energy regions activate long-range context, while low-energy regions suppress noise propagation, achieving feature filtering for physical perception. Finally, the gated global features, texture features, and edge features are added element-wise to obtain the sub-unit output features. This is done because the gated global features provide filtered long-range dependencies, texture features act as an identity path to ensure stable gradient backpropagation, and edge features compensate for local detail sharpness. The fusion of these three features forms a deep feature representation that combines global context, local detail, and training robustness.

[0113] Step S204: The residual features and the deep global features are concatenated by channels and then fused by 1×1 convolution to obtain a fusion result adapted for underground disease detection.

[0114] Specifically, the residual features output from the first path and the deep global features output from the second path are concatenated along the channel dimension to obtain a concatenated feature with double the number of channels. This is done to integrate the information flow of the two paths. The residual features retain the precise location information of shallow local details, while the deep global features contain long-range physical context. Channel concatenation ensures that the two features are encoded independently before subsequent fusion, avoiding premature information aliasing. Then, the concatenated features are input into a 1×1 convolutional layer. The number of channels is restored to the target dimension through pointwise convolution operations. At the same time, the optimal fusion weights of the residual features and the deep global features are learned to obtain a fusion result adapted for underground disease detection. This is because the 1×1 convolution can adaptively weight and combine the two features at each spatial location, achieving a dynamic balance between local details and global semantics. The output dimension matches the input requirements of the subsequent detection head, providing a feature representation with both accurate localization capabilities and physical interpretability for the final disease localization and classification.

[0115] Step S205: Input the fusion results into the high-scale detection head, the medium-scale detection head and the low-scale detection head respectively to generate the corresponding multi-scale prediction feature maps.

[0116] It should be noted that the fusion result is input into a high-scale detection head, which receives high-semantic, low-resolution feature maps output from the deep layers of the network. Through upsampling and convolution operations, it restores the spatial resolution and generates a high-scale predicted feature map. This is done to detect small-sized disease targets in the image. Although deep features have low spatial resolution, they have a high degree of semantic abstraction, enabling the identification of the category attributes of minute diseases. The localization accuracy restored after upsampling is sufficient to define the location of small targets. Next, the fusion result is input into a mid-scale detection head, which directly operates on the mid-resolution intermediate layer features to generate a mid-scale predicted feature map. This is done to detect mid-sized disease targets, balancing localization accuracy and semantic information, and adapting to the detection needs of most conventionally sized diseases. Then, the fusion result is input into a low-scale detection head, which utilizes the high-resolution features retained from the shallow layers to generate a low-scale predicted feature map. This is done to detect large-sized disease targets. High-resolution features retain precise edge location information, accurately defining the boundary range of large targets. Finally, the three detection heads output predicted feature maps of different scales in parallel, forming a feature pyramid structure. This is because the size of disease targets in GPR images varies significantly, ranging from small cracks to large areas of voids. The multi-scale detection heads ensure that disease of all sizes can be detected by the appropriate feature layer, improving detection coverage and overall accuracy.

[0117] Step S206: Decode each multi-scale prediction feature map separately to obtain the prediction result.

[0118] It should be noted that the prediction results include bounding box coordinates, class confidence, and target presence probability. Bounding box regression decoding is performed on each multi-scale predicted feature map. A convolutional layer maps each spatial location of the feature map to the center point coordinates, width, and height offset of the bounding box. The final bounding box coordinates are then calculated using the anchor box parameters corresponding to that location. This is done to transform abstract convolutional features into specific spatial location information of the disease, achieving a conversion from pixel-level localization to target-level localization. Secondly, class classification decoding is performed on each multi-scale predicted feature map. A convolutional layer and a softmax activation function are used to calculate the class confidence for each bounding box corresponding to various diseases (cavities, voids, loose bodies, water-rich bodies). This is done to determine the type of disease the detected target belongs to, providing a classification basis for subsequent engineering decisions. Finally, target presence decoding is performed on each multi-scale predicted feature map. A convolutional layer and a sigmoid activation function are used to calculate the target presence probability for each bounding box. This is done to distinguish foreground targets from background regions, filtering out invalid predicted boxes that do not contain diseases and reducing the false detection rate. Finally, all prediction results output by the three detection heads are aggregated to form an initial detection set. Each prediction result contains three sets of values: bounding box coordinates, class confidence, and target presence probability. This is done to provide complete prediction information for subsequent non-maximum suppression processing and to select the optimal detection box through multi-dimensional scoring.

[0119] Step S207: Perform nonmaximum suppression processing on the prediction results of all scales to obtain the filtered prediction results.

[0120] Specifically, prediction results at all scales are sorted in descending order based on the product of the target's existence probability and class confidence, generating a candidate prediction box list. This prioritizes high-confidence predictions, ensuring the most reliable detection boxes are retained in subsequent screening. Next, the prediction box with the highest confidence is selected as the baseline box from the candidate list. The intersection-union ratio (IU) of this baseline box with all other prediction boxes is calculated. Prediction boxes with an IU greater than a preset threshold (e.g., 0.5) are marked as redundant and removed. This eliminates the possibility of the same disease target being detected repeatedly by multiple overlapping boxes, avoiding redundant detection results and ensuring that each real target corresponds to only one optimal detection box. Non-maximum suppression uses a preset IU threshold and a preset confidence threshold for screening. Then, the baseline box is retained in the screening results, and it, along with all redundant boxes, are removed from the candidate list. This selection and comparison process is repeated until the candidate list is empty. This iteratively processes all prediction results, gradually filtering out non-overlapping high-confidence detection boxes. Finally, the retained detection boxes are filtered a second time according to a preset confidence threshold (e.g., 0.25) to remove low-confidence predictions below the threshold, thus obtaining the filtered prediction results. This is done to further suppress false detections caused by background noise and blurred targets, and only output highly reliable disease detection results to improve engineering practicality.

[0121] Step S208: Integrate the filtered prediction results by category and output the disease detection results including the location, type and confidence level of the disease.

[0122] It should be noted that the filtered prediction results are grouped and integrated according to disease categories (cavities, voids, loose bodies, and water-rich bodies). The detection quantity, spatial distribution, and confidence distribution of each disease category are statistically analyzed. This is done to form structured detection results, facilitating engineers to quickly grasp the development status of various diseases. Secondly, the bounding box coordinates of each detection frame are transformed, mapping the normalized coordinates output by the network back to the pixel coordinates of the original B-Scan image or actual engineering coordinates (station number, depth) to obtain the actual spatial location of the disease. This is done to make the detection results interpretable for engineering purposes, guiding on-site drilling verification or maintenance operations. Then, the category confidence score of each detection frame is extracted as a quantitative indicator of the reliability of disease identification, while the probability of target presence is recorded as an auxiliary reference. This is done to provide a reliable basis for engineering decisions; high-confidence results can be directly used to formulate maintenance plans, while low-confidence results are recommended for manual review. Finally, the location, type, and confidence level information of the defects are formatted and output as an inspection report or a visually labeled image. This is done to form a complete inspection deliverable that supports database storage, report generation, and GIS system integration, thus meeting the digital needs of road maintenance management.

[0123] This embodiment improves the MPS-YOLOv12 network with a four-channel physical sensing tensor input. After encoding by a four-channel input convolutional layer, shallow A2C2f and C3k2 units extract local texture edge features, while deep MPS-Block units generate residual features and depth global features through a dual-path structure. These features are then concatenated and fused before being decoded and predicted by high, medium, and low-scale detection heads. Non-maximum suppression is then used to filter and output the location, type, and confidence level of the disease. By retaining the complete information of the GPR signal through the four-channel physical input, and combining bidirectional Mamba global modeling and a physical gated residual mechanism with the dual-path MPS-Block, the network improves long-distance dependent detection capability while suppressing background noise. Multi-scale detection adapts to diseases of different sizes, achieving high-precision and robust real-time underground disease identification, meeting the rapid detection needs of engineering sites.

[0124] Based on the first embodiment of this application, this application also provides a YOLOv12 ground-penetrating radar underground defect identification device based on bidirectional Mamba. Please refer to... Figure 3 The device includes:

[0125] The acquisition module 10 is used to acquire grayscale B-Scan data from the ground penetrating radar.

[0126] Module 20 is used to perform multi-physics input construction processing on grayscale B-Scan data to obtain a four-channel physical sensing tensor.

[0127] The result module 30 is used to input the four-channel physical sensing tensor into the improved MPS-YOLOv12 network to obtain the disease detection results. The improved MPS-YOLOv12 network includes convolutional layers, shallow layers, deep layers, and a detection head. The number of input convolutional kernel channels in the convolutional layers is expanded to 4. The shallow layers retain A2C2f units and C3k2 units, while the deep layers are replaced with MPS-Block units. The MPS-Block units adopt a dual-path structure and internally nest multiple MPS-Mamba sub-units. The deep layers are layers with downsampling ratios of 16x and 32x. The detection head includes a high-scale detection head, a medium-scale detection head, and a low-scale detection head.

[0128] The YOLOv12 ground-penetrating radar underground disease identification device based on bidirectional Mamba provided in this application adopts the YOLOv12 ground-penetrating radar underground disease identification method based on bidirectional Mamba in the above embodiments, which can solve the technical problem of how to improve the accuracy of underground disease detection. Compared with the prior art, the beneficial effects of the YOLOv12 ground-penetrating radar underground disease identification device based on bidirectional Mamba provided in this application are the same as the beneficial effects of the YOLOv12 ground-penetrating radar underground disease identification method based on bidirectional Mamba provided in the above embodiments, and other technical features in the YOLOv12 ground-penetrating radar underground disease identification device based on bidirectional Mamba are the same as the features disclosed in the methods of the above embodiments, and will not be repeated here.

[0129] This application provides a YOLOv12 ground-penetrating radar underground defect identification device based on bidirectional Mamba. The YOLOv12 ground-penetrating radar underground defect identification device based on bidirectional Mamba includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the YOLOv12 ground-penetrating radar underground defect identification method based on bidirectional Mamba in the first embodiment described above.

[0130] The following is for reference. Figure 4 This document illustrates a structural schematic diagram of a YOLOv12 ground-penetrating radar underground defect identification device based on bidirectional Mamba, suitable for implementing embodiments of this application. The YOLOv12 ground-penetrating radar underground defect identification device based on bidirectional Mamba in these embodiments may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. Figure 4 The YOLOv12 ground-penetrating radar underground disease identification device based on bidirectional Mamba shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0131] like Figure 4As shown, the YOLOv12 ground-penetrating radar underground defect identification device based on bidirectional Mamba may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to programs stored in read-only memory (ROM) 1002 or programs loaded from storage device 1003 into random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the YOLOv12 ground-penetrating radar underground defect identification device based on bidirectional Mamba. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the YOLOv12 ground-penetrating radar underground disease identification device based on two-way Mamba to wirelessly or wiredly communicate with other devices to exchange data. Although various YOLOv12 ground-penetrating radar underground disease identification devices based on two-way Mamba are shown in the figures, it should be understood that implementation or possession of all of them is not required. More or fewer of these devices may be implemented alternatively.

[0132] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0133] The YOLOv12 ground-penetrating radar underground disease identification device based on bidirectional Mamba provided in this application, employing the YOLOv12 ground-penetrating radar underground disease identification method based on bidirectional Mamba in the above embodiments, can solve the technical problem of how to improve the accuracy of underground disease detection. Compared with the prior art, the beneficial effects of the YOLOv12 ground-penetrating radar underground disease identification device based on bidirectional Mamba provided in this application are the same as the beneficial effects of the YOLOv12 ground-penetrating radar underground disease identification method based on bidirectional Mamba provided in the above embodiments, and other technical features in the YOLOv12 ground-penetrating radar underground disease identification device based on bidirectional Mamba are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0134] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0135] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0136] This application provides a computer-readable medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the YOLOv12 ground-penetrating radar underground disease identification method based on bidirectional Mamba in the above embodiments.

[0137] The computer-readable medium provided in this application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor devices, or any combination thereof. More specific examples of computer-readable media may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable medium may be any tangible medium containing or storing a program that can be executed by instructions, used by a device, or used in conjunction with it. The program code contained on the computer-readable medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0138] The aforementioned computer-readable medium may be included in the YOLOv12 ground-penetrating radar underground defect identification device based on bidirectional Mamba; or it may exist independently and not assembled into the YOLOv12 ground-penetrating radar underground defect identification device based on bidirectional Mamba.

[0139] The aforementioned computer-readable medium carries one or more programs that, when executed by the YOLOv12 ground-penetrating radar underground disease identification device based on bidirectional Mamba, enable the device to write computer program code for performing the operations of this application in one or more programming languages ​​or combinations thereof. These programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—and conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0140] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of this application. In this regard, all blocks in the flowcharts or block diagrams may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that all blocks in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using dedicated hardware-based implementations that perform the specified functions or operations, or using a combination of dedicated hardware and computer instructions.

[0141] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0142] The readable medium provided in this application is a computer-readable medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described YOLOv12 ground-penetrating radar underground defect identification method based on bidirectional Mamba, thereby solving the technical problem of how to improve the accuracy of underground defect detection. Compared with the prior art, the beneficial effects of the computer-readable medium provided in this application are the same as those of the YOLOv12 ground-penetrating radar underground defect identification method based on bidirectional Mamba provided in the above embodiments, and will not be repeated here.

[0143] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the YOLOv12 ground-penetrating radar underground disease identification method based on bidirectional Mamba as described above.

[0144] The computer program product provided in this application can solve the technical problem of how to improve the accuracy of underground disease detection. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the underground disease identification method based on bidirectional Mamba YOLOv12 ground penetrating radar provided in the above embodiments, and will not be repeated here.

[0145] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for identifying underground defects using YOLOv12 ground-penetrating radar based on bidirectional Mamba, characterized in that, The method includes: Acquire grayscale B-Scan data from ground-penetrating radar; The grayscale B-Scan data is processed by multiphysics input to obtain a four-channel physical sensing tensor; The four-channel physical sensing tensor is input into the improved MPS-YOLOv12 network to obtain the disease detection results. The improved MPS-YOLOv12 network includes convolutional layers, shallow layers, deep layers, and a detection head. The number of input convolutional kernel channels in the convolutional layers is expanded to 4. The shallow layers retain A2C2f units and C3k2 units, and the deep layers are replaced with MPS-Block units. The MPS-Block units adopt a dual-path structure and internally nest multiple MPS-Mamba sub-units. The deep layers are layers with downsampling ratios of 16x and 32x. The detection head includes a high-scale detection head, a medium-scale detection head, and a low-scale detection head. The step of inputting the four-channel physical sensing tensor into the improved MPS-YOLOv12 network to obtain the disease detection results includes: The four-channel physical perception tensor is input into the input convolutional layer to complete channel adaptation and encode it into an initial feature map. The initial feature map is input into the shallow A2C2f and C3k2 units to extract the local texture and edge features of the GPR image, thus obtaining the shallow local features; The shallow local features are input into the deep MPS-Block unit, and the residual features and deep global features are obtained by processing them through a dual-path structure. The residual features and the depth global features are concatenated by channels and then fused by 1×1 convolution to obtain a fusion result adapted for underground disease detection. The fusion results are input into the high-scale detection head, the medium-scale detection head and the low-scale detection head respectively to generate corresponding multi-scale prediction feature maps; Each of the multi-scale prediction feature maps is decoded to obtain a prediction result, wherein the prediction result includes bounding box coordinates, class confidence and target presence probability; Non-maximum suppression is performed on the prediction results at all scales to obtain filtered prediction results, wherein the non-maximum suppression is performed using a preset intersection-over-union threshold and a preset confidence threshold. The filtered prediction results are integrated by category, and the disease detection results, including disease location, type and confidence level, are output. The step of inputting the shallow local features into the deep MPS-Block unit and processing them through a dual-path structure to obtain residual features and deep global features includes: The shallow local features are input into the first path of the MPS-Block unit, and after 1×1 convolution, they are directly passed to obtain residual features with the same dimension as the input features. The shallow local features are input into the second path of the MPS-Block unit, and the channels are compressed to half of the original number of channels through 1×1 convolution to obtain the compressed features; The compressed features are input into the MPS-Mamba sub-unit to decouple the physical features and model global dependencies, thus obtaining the sub-unit output features. Based on the element-wise multiplication of the energy features and the output features of the sub-unit, low-confidence features are filtered out to obtain the gated features; The gated features are added element-wise to the edge features to obtain the deep global features; The step of inputting the compressed features into the MPS-Mamba sub-unit, decoupling the physical features and modeling global dependencies to obtain the sub-unit output features includes: The compressed features are decoupled by four differentiated parallel projection heads to obtain texture features, energy features, dispersion features and edge features, wherein the four differentiated parallel projection heads are the mainstream semantic head, amplitude gating head, frequency head and phase head, respectively. The dispersive features are superimposed element by element onto the texture features, and then subjected to layer normalization to obtain the enhanced features; Perform a forward state space scan on the enhanced features to obtain forward context features; After performing time reversal on the enhanced features, a forward scan is performed and then reversed again to obtain the reverse context features; The positive context features are added to the negative context features to obtain bidirectional global features; The bidirectional global feature and the energy feature are multiplied element by element to obtain the gated global feature; The gating global feature, the texture feature, and the edge feature are added element by element to obtain the sub-unit output feature.

2. The method as described in claim 1, characterized in that, The step of performing multiphysics input construction processing on the grayscale B-Scan data to obtain a four-channel physical sensing tensor includes: The grayscale B-Scan data is subtracted from the pixel value reference number along the time axis to perform DC removal processing, resulting in a zero-mean radar signal; Applying a Hilbert transform along the time axis to the zero-mean radar signal generates a complex analytic signal; The instantaneous amplitude, instantaneous phase, and instantaneous frequency are calculated based on the complex analytic signal. The zero-mean radar signal, instantaneous amplitude, instantaneous phase, and instantaneous frequency are respectively subjected to minimum-maximum normalization processing and then stacked along the channel dimension to obtain a four-channel physical sensing tensor.

3. The method as described in claim 2, characterized in that, The steps for calculating the instantaneous amplitude, instantaneous phase, and instantaneous frequency based on the complex analytic signal include: Calculate the squares and square roots of the real and imaginary parts of the complex analytic signal to obtain the instantaneous amplitude that fits the GPR energy distribution; Calculate the arctangent values ​​of the imaginary and real parts of the complex analytic signal to obtain the instantaneous phase that adapts to the phase change characteristics of GPR; Perform a phase unwrapping operation on the instantaneous phase to eliminate jumps and obtain a continuous phase function; Perform a difference operation on the continuous phase function along the time axis to obtain the difference result; The instantaneous frequency that adapts to the GPR dispersion characteristics is calculated based on the difference results and the time sampling interval.

4. The method as described in claim 1, characterized in that, The step of decoupling the compressed features using four differentiated parallel projection heads to obtain texture features, energy features, dispersion features, and edge features includes: The compressed features are input into the mainstream semantic head, and the compressed features are dimension-mapped by 1×1 point convolution to obtain texture features. The compressed features are input into the amplitude gating head, and the compressed features are extracted with a large receptive field through a 7×7 depth separable convolution. Then, the energy features are obtained by processing the compressed features through the Sigmoid activation function. The compressed features are input into the frequency head, and the chromatic information of the compressed features is extracted by a 3×3 dilated convolution with a dilation rate of 2 to obtain the chromatic features. The compressed features are input into the phase head, and edge information is extracted from the compressed features through a 3×3 standard convolution to obtain edge features.

5. A YOLOv12 ground-penetrating radar underground defect identification device based on bidirectional Mamba, characterized in that, The device is applied to the YOLOv12 ground-penetrating radar method for identifying underground defects based on bidirectional Mamba as described in any one of claims 1-4, and the device comprises: The acquisition module is used to acquire grayscale B-Scan data from ground-penetrating radar; The construction module is used to perform multi-physics input construction processing on the grayscale B-Scan data to obtain a four-channel physical sensing tensor; The results module is used to input the four-channel physical sensing tensor into the improved MPS-YOLOv12 network to obtain the disease detection results. The improved MPS-YOLOv12 network includes convolutional layers, shallow layers, deep layers, and a detection head. The number of input convolutional kernel channels of the convolutional layers is expanded to 4. The shallow layers retain A2C2f units and C3k2 units. The deep layers are replaced with MPS-Block units. The MPS-Block units adopt a dual-path structure and internally nest multiple MPS-Mamba sub-units. The deep layers are layers with downsampling ratios of 16x and 32x. The detection head includes a high-scale detection head, a medium-scale detection head, and a low-scale detection head.

6. A YOLOv12 ground-penetrating radar underground defect identification device based on bidirectional Mamba, characterized in that, The device includes: a memory, a processor, and a YOLOv12 ground-penetrating radar subsurface defect identification program based on bidirectional Mamba, stored in the memory and running on the processor, the YOLOv12 ground-penetrating radar subsurface defect identification program based on bidirectional Mamba configured to implement the steps of the YOLOv12 ground-penetrating radar subsurface defect identification method based on bidirectional Mamba as described in any one of claims 1-4.

7. A storage medium, characterized in that, The storage medium stores a YOLOv12 ground-penetrating radar underground defect identification program based on bidirectional Mamba. When the YOLOv12 ground-penetrating radar underground defect identification program based on bidirectional Mamba is executed by the processor, it implements the steps of the YOLOv12 ground-penetrating radar underground defect identification method based on bidirectional Mamba as described in any one of claims 1-4.

Images