Two-dimensional ground penetrating radar data real-time underground hidden danger detection method and system based on YOLOv11

Abstract

The application provides a YOLOv11-based real-time underground hidden danger detection method and system for two-dimensional ground penetrating radar data, comprising the following steps: S1, performing signal-to-noise ratio analysis on the collected ground penetrating radar image, directly outputting the result and skipping the subsequent convolution calculation when determining that it is a non-hidden danger area, and sending it to a neural network when determining that it is a suspected target area; S2, replacing the standard feature extraction module in the YOLOv11 backbone network with a dynamic feature extraction module, and adaptively selecting the corresponding calculation path according to the complexity of the input signal; S3, evaluating the importance of the GPR feature channel through sparse training, realizing sparse pruning processing, reconstructing a lightweight network structure and training; S4, converting the optimized model into a format supported by a mobile terminal inference framework, and deploying it on a smartphone. The application can not only dynamically adjust the calculation according to the signal complexity, but also adapt to the physical characteristics of GPR and can be deployed on a smartphone.

Description

Technical Field

[0001] This invention relates to the field of ground-penetrating radar detection and mobile edge artificial intelligence technology, and more particularly to a real-time underground hazard detection method based on YOLOv11 two-dimensional ground-penetrating radar data, and further to a system employing this real-time underground hazard detection method based on YOLOv11 two-dimensional ground-penetrating radar data. Background Technology

[0002] Current ground-penetrating radar (GPR) underground hazard detection technology mainly relies on post-processing, requiring the collected data to be uploaded to a server for offline analysis using standard deep learning models such as YOLO. This existing technical solution has three main drawbacks: First, it suffers from computational redundancy. The general-purpose YOLO model uses a "static computation graph," consuming the same computing power regardless of whether the detection area is a simple, uniform soil layer or a complex area with dense pipelines. Even in areas with no hazards, which constitute a large portion of the GPR data, it still performs full / standard convolution calculations, resulting in a significant waste of power and computing power on mobile devices. Second, it suffers from poor feature adaptability. The standard YOLO model is designed for natural images (RGB), while GPR data (GPR data) depends on waveform amplitude and phase changes. Many feature channels in the model not only do not contribute to the GPR signal but also introduce noise, leading to poor feature adaptability. Third, it suffers from mobile device thermal throttling. High-load AI models running on smartphones for extended periods can cause overheating and frequency reduction. Simple quantization cannot address the inherent computational load of the model structure, resulting in the inability to maintain real-time performance during long-term operation. Therefore, there is an urgent need for a lightweight real-time detection method that can dynamically adjust the calculation based on the signal complexity, adapt to the physical characteristics of GPR, and can be deployed on smartphones. Summary of the Invention

[0003] The technical problem this invention aims to solve is to provide a real-time underground hazard detection method based on YOLOv11-based two-dimensional ground-penetrating radar (GPR) data. The goal is to achieve a lightweight technical solution that can dynamically adjust calculations based on signal complexity, adapt to GPR physical characteristics, and be deployed on smartphones. Furthermore, this invention provides a system employing this YOLOv11-based real-time underground hazard detection method.

[0004] To address this issue, the present invention provides a method for real-time underground hazard detection based on two-dimensional ground-penetrating radar data using YOLOv11, comprising the following steps:

[0005] Step S1: Perform signal-to-noise ratio analysis on the acquired ground-penetrating radar image. If it is determined to be a no-hazard area, output the result directly and skip the subsequent convolution calculation. If it is determined to be a suspected target area, send it into the neural network for processing.

[0006] Step S2: Replace the standard feature extraction module in the YOLOv11 backbone network with a dynamic feature extraction module, and adaptively select the corresponding computation path according to the complexity of the input signal.

[0007] Step S3: Evaluate the importance of GPR feature channels through sparse training, implement sparse pruning, reconstruct the lightweight network structure, and train it.

[0008] Step S4: Convert the optimized model into a format supported by the mobile inference framework and deploy it on a smartphone.

[0009] A further improvement of the present invention is that step S1 includes the following sub-steps:

[0010] Step S101: Perform sliding window processing on the acquired ground-penetrating radar B-Scan image;

[0011] Step S102: Calculate the information entropy and local variance within the window to obtain the signal-to-noise ratio score. ;

[0012] Step S103, score the signal-to-noise ratio. With preset gate threshold If a comparison is made, If the result is positive, the region is determined to be without potential risks, and the result is directly output without potential risks, skipping subsequent convolution calculations; otherwise, it is determined to be a suspected target region, and the image is sent to the neural network for processing.

[0013] A further improvement of the present invention is that, in step S102, the information entropy and local variance are weighted and summed, or the information entropy and local variance are multiplied and normalized to obtain the signal-to-noise ratio score. .

[0014] A further improvement of the present invention is that step S2 includes the following sub-steps:

[0015] Step S201: Insert a lightweight fully connected layer between the feature layers as a routing controller, taking the current feature map as input and outputting a routing probability vector. ;

[0016] Step S202: Set up two computation paths, a fast path and a slow path, where the fast path uses a single-layer convolutional structure and the slow path uses a multi-layer Bottleneck structure.

[0017] Step S203: Based on the complexity of the input signal, the routing probability vector is used. The fast path output and slow path output are dynamically weighted and fused to obtain the final feature extraction result.

[0018] A further improvement of the present invention is that, in step S202, if the current feature map corresponds to a shallow subsurface signal, the routing controller activates the fast path calculation path; if the current feature map corresponds to a deep signal with hyperbolic characteristics, the routing controller activates the slow path calculation path.

[0019] A further improvement of the present invention is that, in step S203, the formula is used... The fast path output and slow path output are dynamically weighted and fused to obtain the final feature extraction result. ,in, Indicates slow path output. This indicates the fast path output.

[0020] A further improvement of the present invention is that step S3 includes the following sub-steps:

[0021] Step S301: Add an L1 regularization term to the loss function of the YOLOv11 training layer, targeting the scaling factor of the BatchNormalization layer. Conduct sparse punishment training;

[0022] Step S302: After training, calculate the absolute value of the scaling factor for each convolutional channel. To evaluate the importance of each feature channel for GPR hyperbolic feature detection;

[0023] Step S303: Set the pruning rate according to the absolute value of the scaling factor. Sort the data from smallest to largest, remove redundant channels whose sorting rate is lower than the pruning rate, and reconstruct a lightweight network structure.

[0024] Step S304: Train the pruned lightweight network to restore detection accuracy.

[0025] A further improvement of the present invention is that, in step S301, the formula is used... Scaling factor for Batch Normalization layer Sparse punishment training is conducted, in which... Indicates the total loss; This represents the original YOLO loss; Indicates the sparsity penalty coefficient; This represents the L1 regularization term.

[0026] A further improvement of the present invention is that step S4 includes the following sub-steps:

[0027] Step S401: Export the pruned lightweight network model as the ONNX Open Neural Network Exchange format.

[0028] Step S402: Use the NCNN mobile neural network inference framework or the TNN mobile inference framework to translate the model;

[0029] Step S403: Enable FP16 half-precision inference and Vulkan hardware acceleration, and deploy on smartphones.

[0030] This invention also provides a real-time underground hazard detection system based on YOLOv11 two-dimensional ground-penetrating radar data, characterized in that it employs the aforementioned real-time underground hazard detection method based on YOLOv11 two-dimensional ground-penetrating radar data and is deployed on a smartphone; the real-time underground hazard detection system based on two-dimensional ground-penetrating radar data includes:

[0031] A dynamic input gating module is constructed to perform signal-to-noise ratio analysis on the acquired ground-penetrating radar images. When the area is determined to be a safe area, the result is directly output and subsequent convolution calculations are skipped. When the area is determined to be a suspected target area, it is fed into the neural network for processing.

[0032] The dynamic routing feature extraction module replaces the standard feature extraction module in the YOLOv11 backbone network with a dynamic feature extraction module, which adaptively selects the corresponding computation path according to the complexity of the input signal.

[0033] The GPR feature channel sparsity pruning module evaluates the importance of GPR feature channels through sparse training, implements sparse pruning, reconstructs a lightweight network structure, and performs training.

[0034] The mobile deployment module converts the optimized model into a format supported by the mobile inference framework and deploys it on smartphones.

[0035] Compared with existing technologies, the advantages of this invention are as follows: First, signal-to-noise ratio (SNR) analysis is performed on the acquired ground-penetrating radar (GPR) images to construct a dynamic input gating system. When a region is determined to be free of hidden dangers, the result is directly output and subsequent convolution calculations are skipped. When a region is determined to be a suspected target region, it is fed into a neural network for processing. This allows for the direct filtering of regions free of hidden dangers, skipping redundant convolution calculations and significantly reducing the power consumption and heat generation of smartphones. Then, through a dynamic routing feature extraction mechanism, the computational path is adaptively selected based on the complexity of the input signal. Simple signals are processed quickly, while complex signals are analyzed in depth, achieving dynamic allocation of computing resources. Finally, sparse training is used to evaluate the GPR characteristics. The importance of the signal channels is emphasized, and sparse pruning is implemented. L1 regularization is used to automatically filter channels sensitive to hyperbolic features and remove redundant channels sensitive to RGB color. Finally, the optimized model is converted into a format supported by the mobile inference framework and deployed on smartphones. This enables smartphones to become intelligent terminals that integrate data reception, real-time AI interpretation, and result visualization and interaction. This solves the technical problems of high cost, inconvenient deployment, and poor real-time performance of existing technologies. It can not only dynamically adjust the calculation according to the signal complexity, but also adapt to the physical characteristics of GPR and can be deployed on smartphones, effectively improving detection efficiency and reducing equipment costs. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the workflow of one embodiment of the present invention. Detailed Implementation

[0037] In the description of this invention, the term "several" means one or more; the term "multiple" means two or more; the terms "greater than," "less than," and "exceeding" are all understood to exclude the stated number; and the terms "above," "below," and "within" are all understood to include the stated number. The terms "first," "second," etc., are understood to be used only to distinguish identical or similar technical feature names, and should not be construed as implying / indicating the relative importance of the technical features, the number of technical features, or the sequential relationship between the technical features.

[0038] Before further describing the preferred embodiments of the present invention, the application fields of the present invention will be introduced first.

[0039] This invention relates to the fields of ground-penetrating radar (GPR) detection and mobile edge computing, specifically to a method and system for real-time, online, and automatic identification of underground hazards during the data acquisition process of two-dimensional GPR based on the YOLOv11 deep learning target detection model. The optimized model of this invention (i.e., a lightweight network model, also known as a lightweight AI model or a lightweight YOLOv11 model) can be directly deployed on mobile terminals such as smartphones, achieving deep integration of industry-level professional detection capabilities with consumer-grade mobile devices. It is suitable for rapid, low-cost, and highly flexible detection of underground anomalies such as cavities, voids, loose bodies, and pipelines in scenarios such as urban roads, pipeline corridors, and construction areas.

[0040] The preferred embodiments of the present invention will now be described in further detail with reference to the accompanying drawings.

[0041] like Figure 1 As shown, this embodiment provides a real-time underground hazard detection method based on YOLOv11 two-dimensional ground penetrating radar data, including the following steps:

[0042] Step S1: Perform signal-to-noise ratio analysis on the acquired ground-penetrating radar image. If it is determined to be a no-hazard area, output the result directly and skip the subsequent convolution calculation. If it is determined to be a suspected target area, send it into the neural network for processing.

[0043] Step S2: Replace the standard feature extraction module (standard C3k2 module) in the YOLOv11 backbone network with the dynamic feature extraction module (Dy-C3k2 module, also known as Dynamic C3k2 module), and adaptively select the corresponding computation path according to the complexity of the input signal;

[0044] Step S3: Evaluate the importance of GPR feature channels through sparse training, implement sparse pruning, reconstruct the lightweight network structure, and train it.

[0045] Step S4: Convert the optimized model into a format supported by the mobile inference framework and deploy it on a smartphone.

[0046] Step S1 in this embodiment is used to construct dynamic input gating (pre-screening) based on signal-to-noise ratio entropy (SNR-Entropy). The aim is to determine whether the current echo is worth depth calculation by a low-computing dynamic input gating module before the image enters the YOLO backbone network.

[0047] Preferably, step S1 in this embodiment includes the following sub-steps:

[0048] Step S101: Perform sliding window processing on the acquired ground-penetrating radar B-Scan image;

[0049] Step S102: Calculate the information entropy and local variance within the window to obtain the signal-to-noise ratio score. ;

[0050] Step S103, score the signal-to-noise ratio. With preset gate threshold If a comparison is made, If the result is positive, the region is determined to be a region without potential risks, i.e., a uniform medium / no signal. The result of no potential risks is directly output and all subsequent convolution calculations are skipped. Otherwise, the region is determined to be a suspected target region and the image is sent to the neural network for processing.

[0051] More specifically, in step S102 of the embodiment, it is preferable to use a formula Calculate local variance .in, , which refers to local variance, is used to represent the local activity score within a window. These are the width and height of the sliding window, respectively (e.g.) (pixels). The pixel grayscale value at that location. This is the arithmetic mean of all pixels within the current window.

[0052] Preferred formula Calculate information entropy .in, Information entropy is used to reflect the disorder and structure of a signal through fluctuation entropy scores. The simplified Sobel gradient operator is calculated as follows: . For statistical functions, the gradient value within the window is equal to... The number of pixels. The gradient quantization level (usually 0 to 15 to reduce the histogram dimension). This represents the total number of pixels within the window. These are adaptive weighting coefficients; this is a constant array, suitable for high-frequency, small fluctuations. Values ​​are assigned high weights, for =0 (smooth region) is assigned zero weight.

[0053] Preferred formula Calculate the signal-to-noise ratio score , This represents the final judgment score for entering the YOLO outpost, abbreviated as... . This indicates the environmental sensitivity factor; the amplification factor preset based on the antenna frequency of the GPR device. This is a system bias term; used to offset the system thermal noise substrate generated by the radar hardware.

[0054] In step S103 of the embodiment, a preset gate threshold is used. The calculation preferably consists of two parts: "benchmark mean" and "tolerance offset", which can be expressed by the formula: .

[0055] in, This represents the background mean score, indicating the area known to be free of hazards (background area) under the current detection environment. Average score. This represents the background standard deviation, used to reflect the intensity of fluctuations in the background signal; if the strata are chaotic, the fluctuations are large; if the strata are flat, the fluctuations are small. This is the sensitivity coefficient, which is an artificially adjustable constant, typically ranging from [2,5]. The smaller the value, the more sensitive the system is and the less likely it is to miss detections, but the YOLO module will be woken up frequently (increasing power consumption). The larger the value, the more conservative the system becomes. YOLO will only be triggered by very obvious hidden dangers, making it suitable for scenarios with low battery or simple geological conditions.

[0056] Preferably, in step S102 of this embodiment, the information entropy and local variance can be weighted and summed, or the information entropy and local variance can be multiplied and normalized to obtain the signal-to-noise ratio score. .

[0057] This embodiment, through the dynamic input gating in step S1, can filter out approximately 60% of invalid background calculations, significantly reducing smartphone heat generation and improving FPS (Frames Per Second).

[0058] Step S2 in this embodiment is used for the dynamic routing mechanism based on the YOLOv11 C3k2 module.

[0059] Preferably, step S2 in this embodiment includes the following sub-steps:

[0060] Step S201: Insert a lightweight fully connected layer as a router between the feature layers, taking the current feature map as input and outputting a routing probability vector. Lightweight fully connected layers can use a variant of SE-Block; SE-Block stands for Squeeze-and-Excitation Block, which is a channel attention mechanism in deep learning.

[0061] Step S202: Set up two computation paths, a fast path and a slow path, where the fast path uses a single-layer convolutional structure and the slow path uses a multi-layer Bottleneck structure.

[0062] Step S203: Based on the complexity of the input signal, the routing probability vector is used. The fast path output and slow path output are dynamically weighted and fused to obtain the final feature extraction result.

[0063] Specifically, in step S202 of this embodiment, if the current feature map corresponds to a shallow underground signal with simple texture, the routing controller activates the fast path calculation path; if the current feature map corresponds to a complex deep signal with hyperbolic features, the routing controller activates the slow path calculation path.

[0064] In the improved architecture of YOLOv11, the Bottleneck structure is the core component for achieving high-performance feature extraction. The multi-layer Bottleneck structure includes: 1. A dimensionality reduction layer (1×1 convolution), used to compress the high-dimensional number of input channels (e.g., 512) to a low dimension (e.g., 128); 2. A feature extraction layer (3×3 convolution), used to extract spatial features in the low-dimensional space; 3. A dimensionality increase layer (1×1 convolution), used to restore the number of channels to the original dimension (or higher).

[0065] For YOLOv11 real-time inference on mobile devices, performing 3×3 convolution after 1×1 convolution reduces the computational cost significantly less than performing two 3×3 convolutions directly.

[0066] In step S203 of this embodiment, it is preferable to use the formula... The fast path output and slow path output are dynamically weighted and fused to obtain the final feature extraction result. End-to-end training is achieved using the differentiable sampling technique Gumbel-Softmax. Indicates slow path output. This indicates the fast path output.

[0067] Step S3 in this embodiment is used to implement channel pruning of GPR feature channels. The reason for this is that GPR is single-channel (grayscale) or pseudo-color data, and a large number of convolutional kernels responsible for color texture in natural image models are redundant.

[0068] Preferably, step S3 in this embodiment includes the following sub-steps:

[0069] Step S301 is used to implement sparse training by adding an L1 regularization term to the loss function of YOLOv11 training, which is a scaling factor for the Batch Normalization layer (also known as the BN layer). Conduct sparse punishment training;

[0070] Step S302 is used to evaluate channel importance. After training, the absolute value of the scaling factor of each convolutional channel is calculated. The importance of each feature channel to GPR hyperbolic feature detection was evaluated; for GPR data, the weight of high-frequency texture channels is usually very low.

[0071] Step S303 is used to implement physical pruning, setting the pruning rate according to the absolute value of the scaling factor. Sort the data from smallest to largest and remove redundant channels whose ranking is lower than the pruning rate, then reconstruct a lightweight network structure. The pruning rate can be customized and adjusted according to actual conditions and needs, such as 40%.

[0072] Step S304 is used to perform fine-tuning, retraining the pruned lightweight network to restore detection accuracy.

[0073] In step S301 of this embodiment, it is preferable to use the formula... Scaling factor for BatchNormalization layer Sparse punishment training is conducted, in which... Indicates the total loss; This represents the original YOLO loss; Represents the sparsity penalty coefficient. The larger the value, the stronger the sparsity. Represents the scaling factor for all Batch Normalization (BN) layers in the network. A set; This represents the L1 regularization term, which represents the absolute value of the scaling factors of all BN layers. The sum of these is added to the total loss as a penalty. .

[0074] Step S4 in this embodiment is used to achieve heterogeneous deployment on mobile devices.

[0075] Preferably, step S4 in this embodiment includes the following sub-steps:

[0076] Step S401: Export the pruned lightweight network model as the ONNX Open Neural Network Exchange format.

[0077] Step S402: Use the NCNN mobile neural network inference framework or the TNN mobile inference framework to translate the model;

[0078] Step S403: Enable FP16 half-precision inference and Vulkan hardware acceleration, and deploy on smartphones.

[0079] Preferably, step S401 is used to export ONNX, exporting the trained YOLOv11 model into the general ONNX format, i.e., the ONNX Open Neural Network Exchange format. Step S402 is used to translate into a framework format. Specifically, in the model translation of the NCNN mobile neural network inference framework, the onnx2ncnn tool is used to convert the ONNX Open Neural Network Exchange format into .param (network structure) and .bin (weight parameters) files. In the model translation of the TNN mobile inference framework, the onnx2tnn tool is used to convert it into .tnnproto and .tnnmodel files.

[0080] Furthermore, an optimization and pruning fusion step can be optionally included, by executing ncnnoptimize (a model optimization tool for the NCNN framework) or an optimization tool for the TNN mobile inference framework. This preferred step performs operator fusion and thoroughly simplifies the computation graph based on the physical pruning results of step S303.

[0081] Finally, in step S403, FP16 half-precision inference is enabled by setting `opt.use_fp16_packed = true` and `opt.use_fp16_storage = true` in `ncnn::Option`. Vulkan is a cross-platform, low-overhead graphics API used to check if the phone hardware supports Vulkan, and then enable it.

[0082] This embodiment also provides a real-time underground hazard detection system based on YOLOv11 two-dimensional ground-penetrating radar data. Its characteristic is that it employs the aforementioned real-time underground hazard detection method based on YOLOv11 two-dimensional ground-penetrating radar data and deploys it on a smartphone. The real-time underground hazard detection system based on YOLOv11 two-dimensional ground-penetrating radar data includes:

[0083] A dynamic input gating module is constructed to perform signal-to-noise ratio analysis on the acquired ground-penetrating radar images. When the area is determined to be a safe area, the result is directly output and subsequent convolution calculations are skipped. When the area is determined to be a suspected target area, it is fed into the neural network for processing.

[0084] The dynamic routing feature extraction module replaces the standard feature extraction module in the YOLOv11 backbone network with a dynamic feature extraction module, which adaptively selects the corresponding computation path according to the complexity of the input signal.

[0085] The GPR feature channel sparsity pruning module evaluates the importance of GPR feature channels through sparse training, implements sparse pruning, reconstructs a lightweight network structure, and performs training.

[0086] The mobile deployment module converts the optimized model into a format supported by the mobile inference framework and deploys it on smartphones.

[0087] In summary, this embodiment first performs signal-to-noise ratio analysis on the acquired ground-penetrating radar images and constructs a dynamic input gating system. When an area is determined to be free of potential hazards, the result is directly output and subsequent convolution calculations are skipped. When an area is determined to be a suspected target area, it is fed into the neural network for processing. This allows for the direct filtering of areas free of potential hazards, skipping redundant convolution calculations and significantly reducing the power consumption and heat generation of the smartphone. Then, through a dynamic routing feature extraction mechanism, the computation path is adaptively selected based on the complexity of the input signal. Simple signals are processed quickly, while complex signals are analyzed in depth, achieving dynamic allocation of computing resources. Finally, sparse training is used to evaluate the weight of the GPR feature channels. Essentially, this approach implements sparse pruning, automatically filtering channels sensitive to hyperbolic features using L1 regularization and removing redundant channels sensitive to RGB color. Finally, the optimized model is converted into a format supported by the mobile inference framework and deployed on smartphones. This enables smartphones to become intelligent terminals integrating data reception, real-time AI interpretation, and result visualization and interaction, solving the technical problems of high cost, inconvenient deployment, and poor real-time performance of existing technologies. It can not only dynamically adjust calculations according to signal complexity but also adapt to GPR physical characteristics and can be deployed on smartphones, effectively improving detection efficiency and reducing equipment costs.

[0088] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for real-time underground hazard detection based on YOLOv11 two-dimensional ground-penetrating radar data, characterized in that, Includes the following steps: Step S1: Perform signal-to-noise ratio analysis on the acquired ground-penetrating radar image. If it is determined to be a no-hazard area, output the result directly and skip the subsequent convolution calculation. If it is determined to be a suspected target area, send it into the neural network for processing. Step S2: Replace the standard feature extraction module in the YOLOv11 backbone network with a dynamic feature extraction module, and adaptively select the corresponding computation path according to the complexity of the input signal. Step S3: Evaluate the importance of GPR feature channels through sparse training, implement sparse pruning, reconstruct the lightweight network structure, and train it. Step S4: Convert the optimized model into a format supported by the mobile inference framework and deploy it on a smartphone.

2. The method for real-time underground hazard detection based on YOLOv11 two-dimensional ground-penetrating radar data according to claim 1, characterized in that, Step S1 includes the following sub-steps: Step S101: Perform sliding window processing on the acquired ground-penetrating radar B-Scan image; Step S102: Calculate the information entropy and local variance within the window to obtain the signal-to-noise ratio score. ; Step S103, score the signal-to-noise ratio. With preset gate threshold If a comparison is made, If the result is positive, the region is determined to be without potential risks, and the result is directly output without potential risks, skipping subsequent convolution calculations; otherwise, it is determined to be a suspected target region, and the image is sent to the neural network for processing.

3. The method for real-time underground hazard detection based on YOLOv11 two-dimensional ground-penetrating radar data according to claim 2, characterized in that, In step S102, the information entropy and local variance are weighted and summed, or the information entropy and local variance are multiplied and normalized to obtain the signal-to-noise ratio score. .

4. The method for real-time underground hazard detection based on YOLOv11 two-dimensional ground-penetrating radar data according to any one of claims 1 to 3, characterized in that, Step S2 includes the following sub-steps: Step S201: Insert a lightweight fully connected layer between the feature layers as a routing controller, taking the current feature map as input and outputting a routing probability vector. ; Step S202: Set up two computation paths, a fast path and a slow path, where the fast path uses a single-layer convolutional structure and the slow path uses a multi-layer Bottleneck structure. Step S203: Based on the complexity of the input signal, the routing probability vector is used. The fast path output and slow path output are dynamically weighted and fused to obtain the final feature extraction result.

5. The method for real-time underground hazard detection based on YOLOv11 two-dimensional ground penetrating radar data according to claim 4, characterized in that, In step S202, if the current feature map corresponds to a shallow subsurface signal, the routing controller activates the fast path calculation path; if the current feature map corresponds to a deep signal with hyperbolic characteristics, the routing controller activates the slow path calculation path.

6. The method for real-time underground hazard detection based on YOLOv11 two-dimensional ground-penetrating radar data according to claim 4, characterized in that, In step S203, the formula is used. The fast path output and slow path output are dynamically weighted and fused to obtain the final feature extraction result. ,in, Indicates slow path output. This indicates the fast path output.

7. The method for real-time underground hazard detection based on YOLOv11 two-dimensional ground-penetrating radar data according to any one of claims 1 to 3, characterized in that, Step S3 includes the following sub-steps: Step S301: Add an L1 regularization term to the loss function of the YOLOv11 training layer, which is the scaling factor for the Batch Normalization layer. Conduct sparse punishment training; Step S302: After training, calculate the absolute value of the scaling factor for each convolutional channel. To evaluate the importance of each feature channel for GPR hyperbolic feature detection; Step S303: Set the pruning rate according to the absolute value of the scaling factor. Sort the data from smallest to largest, remove redundant channels whose sorting rate is lower than the pruning rate, and reconstruct a lightweight network structure. Step S304: Train the pruned lightweight network to restore detection accuracy.

8. The method for real-time underground hazard detection based on YOLOv11 two-dimensional ground-penetrating radar data according to claim 7, characterized in that, In step S301, the formula is used. Scaling factor for Batch Normalization layer Sparse punishment training is conducted, in which... Indicates the total loss; This represents the original YOLO loss; Indicates the sparsity penalty coefficient; This represents the L1 regularization term.

9. The method for real-time underground hazard detection based on YOLOv11 two-dimensional ground-penetrating radar data according to any one of claims 1 to 3, characterized in that, Step S4 includes the following sub-steps: Step S401: Export the pruned lightweight network model as the ONNX Open Neural Network Exchange format. Step S402: Use the NCNN mobile neural network inference framework or the TNN mobile inference framework to translate the model; Step S403: Enable FP16 half-precision inference and Vulkan hardware acceleration, and deploy on smartphones.

10. A real-time underground hazard detection system based on YOLOv11 two-dimensional ground-penetrating radar data, characterized in that, The method for real-time underground hazard detection based on YOLOv11 two-dimensional ground penetrating radar data, as described in any one of claims 1 to 9, was adopted and deployed on a smartphone. The real-time underground hazard detection system based on two-dimensional ground-penetrating radar data includes: A dynamic input gating module is constructed to perform signal-to-noise ratio analysis on the acquired ground-penetrating radar images. When the area is determined to be a safe area, the result is directly output and subsequent convolution calculations are skipped. When the area is determined to be a suspected target area, it is fed into the neural network for processing. The dynamic routing feature extraction module replaces the standard feature extraction module in the YOLOv11 backbone network with a dynamic feature extraction module, which adaptively selects the corresponding computation path according to the complexity of the input signal. The GPR feature channel sparsity pruning module evaluates the importance of GPR feature channels through sparse training, implements sparse pruning, reconstructs a lightweight network structure, and performs training. The mobile deployment module converts the optimized model into a format supported by the mobile inference framework and deploys it on smartphones.

Images