Buried pipeline identification and positioning method based on target detection

Abstract

The invention discloses a buried pipeline identification and positioning method based on target detection, and belongs to the technical field of pipeline detection, and the method comprises the following steps: S1, inputting a marked PE pipeline welding defect image; s2, introducing an ADown module to improve a Conv module: through down-sampling operation, the ADown module effectively reduces the spatial dimension of a feature map, and captures high-level features; the calculation amount is reduced by reducing redundant feature information and optimizing the calculation efficiency; the GPR signal data of the underground pipeline is acquired through the signal acquisition module and is preprocessed, including signal denoising and feature enhancement, so that the data is adaptive to the input requirement of the deep learning model, and an ADown module in YOLOv9 is introduced to reduce the spatial dimension of a feature map, reduce the calculation amount and memory use, and improve the accuracy of the GPR signal data. And the calculation efficiency of the model is remarkably improved on the premise of keeping key feature information.

Description

Technical Field

[0001] The present invention belongs to the technical field of pipeline detection, and in particular relates to a method for identifying and locating buried pipelines based on target detection. Background Art

[0002] Buried pipeline detection is an important part of modern urban infrastructure management, especially in the safe operation of pipeline systems such as natural gas and oil. Although traditional pipeline detection methods such as manual detection, acoustic detection and traditional ground penetrating radar (GPR) technology are widely used, their accuracy and efficiency are often limited in complex environments. With the rapid development of deep learning technology, target detection methods based on deep learning have gradually become an emerging technology for GPR data processing, which can significantly improve the automation and accuracy of pipeline detection. In traditional GPR signal processing methods, manual feature extraction and signal analysis are usually relied on, which is not only labor-intensive and prone to human errors, but also difficult to handle complex background noise and targets of different scales. In recent years, deep learning, especially convolutional neural networks (CNNs), has made significant progress in the field of image recognition and has become an effective tool for solving GPR signal recognition problems. As a real-time target detection framework, the YOLO series of algorithms has become an important representative of deep learning in the field of target detection with its efficient feature extraction and target positioning capabilities.

[0003] With the development of deep learning, object detection algorithms have made significant progress in the field of computer vision. In particular, the YOLO series has been widely used for its efficient real-time detection capabilities. YOLOv8 optimizes the network structure and feature extraction module to improve detection accuracy while maintaining real-time performance and speed. Although traditional YOLO has limitations in processing complex scenes and multi-scale objects, YOLOv8 significantly enhances its performance in complex tasks by improving the convolution module and multi-scale feature fusion strategy. Although YOLOv8 performs well in many visual tasks, it still faces the following problems in the processing and recognition of buried pipeline GPR signals: GPR images usually contain a lot of background noise and complex geological interference, which will affect the accuracy of target detection. Therefore, a buried pipeline recognition and positioning method based on target detection is needed to solve the above problems. Summary of the invention

[0004] The purpose of the present invention is to provide a method for identifying and locating buried pipelines based on target detection to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: a method for identifying and locating buried pipelines based on target detection, comprising the following steps: S1. Input annotated PE pipeline welding defect images: collect a large number of GPR data sets covering a variety of buried pipeline scenarios, and pre-process them through signal processing technology, including denoising and feature enhancement. Then, use data annotation tools to accurately annotate specific targets in the signal (such as the buried depth of the pipeline), and divide the data set into a training set and a validation set. S2. Introduce the ADown module to improve the Conv module: Through downsampling operations, the ADown module effectively reduces the spatial dimension of the feature map, which not only reduces the computational complexity, but also can capture the key feature information in target detection while ensuring performance; This module improves the detection capability of the model in complex backgrounds by reducing redundant feature information and optimizing computational efficiency. The introduction of the ADown module not only improves the real-time performance of the model, but also maintains high detection performance in resource-limited industrial environments. It is particularly suitable for target recognition tasks in complex scenarios. S3. Introducing the HS-FPN (High-Level Screening Fusion Pyramid Network) structure in the MFDS-DETR model: Optimizing the Neck part of the original YOLOv8 model to achieve efficient fusion of multi-scale features, thereby enhancing the model's feature expression capabilities in complex target recognition tasks; The HS-FPN structure can filter, fuse and enhance feature information from different scales, making the model more superior in processing diverse and multi-scale targets; By effectively fusing multi-scale features, the model can more accurately capture the details of the target, especially when identifying small-scale targets and blurred edge areas in complex backgrounds, the performance is significantly improved.

[0006] S4. Improve C2f with DualCorv: The DualConv technology is introduced to lightweight the C2f module of the original YOLOv8 model, and the 3×3 and 1×1 convolution kernels are improved to process the same input feature channels, which significantly improves the feature extraction efficiency. The convolution filters are optimally arranged through group convolution technology, which reduces the computational cost and the number of parameters, thereby improving the model's operating efficiency under resource-constrained conditions; DualConv technology not only retains the detection accuracy of the model, but also further makes the model lightweight, ensuring its excellent performance in high-efficiency, low-latency tasks, meeting the performance and real-time requirements of industrial-grade detection tasks; S5, model testing, wherein the model testing includes performance evaluation and calculation of mPA; The model parameters were carefully adjusted according to specific inspection requirements, and the improved model was fully tested under multiple experimental conditions to verify its accuracy and inspection efficiency in identifying defect signals in GPR datasets. Through experimental testing and performance evaluation, the detection capability of the model was further optimized to ensure that it can maintain efficient and stable performance in different data environments and actual applications, providing strong technical support for pipeline detection and positioning; S6. Ablation experiment.

[0007] In the detection process of this method, the preprocessed GPR signal is transmitted through an interface to a platform deployed with an improved algorithm, and the depth of the buried pipeline is identified and located in real time. The detection results output by the model are combined with the original GPR signal. The buried depth position and spatial distribution of the pipeline can be obtained by calculating the data on the obtained detection result map. Through this optimized integration method, the full process coverage from GPR signal acquisition, algorithm processing to result output is achieved, which provides an integrated and intelligent solution for the efficient detection of buried PE pipelines, and is suitable for a variety of practical scenarios such as pipeline safety monitoring and depth measurement.

[0008] As a preferred solution, in "S1", a large number of GPR data sets are obtained, B-Scan images under different conditions are selected from the data sets, and the pipeline burial depth positions are annotated by labellmg. The annotated GPR data sets are divided into a validation set and a training set by a python script.

[0009] As a preferred solution, in "S2", the improved ADown module reduces the number of channels of the input feature map to half of the original through convolution operation, and then downsamples through the average pooling layer to reduce the spatial dimension of the feature map. This not only reduces the size of the feature map, but also retains the important feature information of the image, so that the model can capture the key features of the image at a higher level. The downsampling operation reduces the amount of calculation and memory usage, and improves the operation efficiency; As a preferred solution, in “S3”, HS-FPN combines low-level detail features and high-level semantic features to generate a more representative comprehensive feature map; The HS-FPN structure pays special attention to the combination of low-level detail features and high-level semantic features. Low-level features usually contain fine information of the image, while high-level features represent the overall semantic information of the image. By fusing these two types of features, HS-FPN can generate a more comprehensive and representative comprehensive feature map. The multi-scale feature fusion method significantly enhances the performance of the model in complex scenes, making it more accurate and reliable when processing a variety of detection tasks. Whether it is an image with rich details or a complex scene, HS-FPN can improve the accuracy and robustness of the model and ensure more efficient target detection.

[0010] As a preferred solution, in "S4", group convolution technology is used to efficiently arrange convolution filters, which significantly reduces the computational cost and the number of parameters; The improved C2f can more effectively capture feature information of different scales, making the model more flexible and efficient in handling complex tasks; The 3×3 convolution kernel captures local features, while the 1×1 convolution kernel can integrate more contextual information; The introduction of group convolution technology realizes the efficient arrangement of convolution filters, significantly reducing the computational cost and the number of parameters. While maintaining high performance, this technology effectively reduces the computational burden and memory requirements of the model.

[0011] As a preferred solution, in "S5", set the parameters, set the optimal combination of iteration rounds, batches, optimizers and learning rate momentum, and create a py file for training. Then, quote the best model weight obtained from the training for testing. Use the confusion matrix in the obtained data results to calculate the precision and recall and compare them with the original YOLOv8n model to measure the classification accuracy of the model. Compare the calculated precision and recall values ​​with the original YOLOv8n model to obtain the mAP value to evaluate the overall effect of the improved model. As a preferred solution, in “S6”, the three added improved models are verified by ablation experiments to be superior to the single improved model in terms of accuracy and robustness, further proving the effectiveness of the improvement.

[0012] Compared with the prior art, the present invention has the following beneficial effects: The present invention obtains the GPR signal data of the underground pipeline through the signal acquisition module and preprocesses it, including signal denoising and feature enhancement, so that the data adapts to the input requirements of the deep learning model; The present invention introduces the ADown module in YOLOv9 to reduce the spatial dimension of the feature map, while reducing the amount of calculation and memory usage, and significantly improves the computational efficiency of the model while retaining key feature information; The present invention realizes efficient fusion of multi-scale features by integrating the HS-FPN structure in MFDS-DETR, significantly enhancing the feature expression capability of the model in complex scenarios, especially showing higher detection accuracy and robustness when dealing with noise interference and diversified features in GPR signals. The present invention adopts the C2f module improved by DualConv, combines 3×3 and 1×1 convolution kernels, and uses group convolution technology to significantly reduce the computational cost and the number of parameters, while optimizing the information processing and feature extraction process; The present invention achieves a good balance between the number of model parameters, detection accuracy and computational efficiency, significantly improves the recognition accuracy and real-time performance of GPR signals of buried pipelines, lays a solid foundation for the practical application of ground penetrating radar technology and the deployment of terminal equipment, and provides reliable technical guarantee for improving the monitoring efficiency and safety of underground pipeline systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 is a flow chart of the present invention; Figure 2 B-Scan diagrams under different conditions of the present invention; Figure 3 It is a structural diagram of the ADwon network of the present invention; Figure 4 It is a schematic diagram of the structure of the HS-FPN network of the present invention; Figure 5 It is the working flow chart of the SFF module of the present invention; Figure 6 It is a visualization diagram of the DualConv structure of the present invention; Figure 7 It is the network structure diagram of ADC-yolov8 of the present invention; Figure 8 This is a comparison chart of the detection results of different modules introduced by YOLOv8n of the present invention. DETAILED DESCRIPTION

[0014] The present invention will be further described below in conjunction with the embodiments.

[0015] The following examples are used to illustrate the present invention, but they cannot be used to limit the scope of protection of the present invention. The conditions in the examples can be further adjusted according to specific conditions. Simple improvements to the method of the present invention under the premise of the concept of the present invention belong to the scope of protection claimed in the present invention.

[0016] See also Figure 1-8 The present invention provides a method for identifying and locating a buried pipeline based on target detection, comprising the following steps: S1. Input the labeled PE pipeline welding defect image: obtain a large number of GPR data sets; select B-Scan images under different conditions in the data set, and annotate the pipeline burial depth position through labellmg, and then divide the annotated GPR data set into a validation set and a training set through a python script; S2. Introduce the ADown module to improve the Conv module: add ADown to the Conv of feature extraction in the backbone of the original yaml file for use; The ADown module in YOLOv9 is an innovative convolutional block design, which is mainly used for downsampling operations in target detection tasks; Downsampling is a common technique in deep learning models. It helps the model capture higher-level features and reduces computational burden by reducing the spatial dimensions of feature maps. The ADown module is designed to balance efficiency and performance, aiming to achieve efficient downsampling operations at a low computational cost; The core features of the ADown module include: Lightweight design: This module significantly reduces the complexity of the model and the computing resource requirements by reducing the number of parameters in the convolutional layer, which is especially suitable for resource-constrained environments. Information preservation: Although the ADown module reduces the resolution of the feature map, it retains key information, ensuring that the model does not lose important features during the downsampling process, thereby improving the accuracy of object detection; Learning capability: The ADown module has adaptive capabilities and can dynamically adjust its parameters according to different data scenarios to optimize model performance; Improved accuracy: Using the ADown module not only reduces the model size, but also improves the detection accuracy, which provides a new idea for performance optimization of target detection tasks; Flexibility: The ADown module can be integrated into the backbone and head parts of YOLOv9, providing multiple configuration options to adapt to different network structures and task requirements; The application of the ADown module in YOLOv9 is mainly reflected in replacing traditional downsampling operations, such as the Conv module. By introducing ADown in the backbone, the model can perform efficient downsampling between different feature layers. In the head part, ADown helps to further refine the feature map, laying the foundation for more accurate target detection. The steps to implement the ADown module include: Convolution operation: extract key information in the feature map through the convolution layer; Stride adjustment: adjust the stride of the convolutional layer to reduce the spatial dimension of the feature map; Parameter optimization: Reduce the computational complexity of the model by reducing the number of parameters in the convolutional layer; In practical applications, the introduction of the ADown module in YOLOv9 has achieved a significant reduction in the number of parameters while maintaining or even improving the accuracy of target detection; for example, when improving YOLOv8, ADown was added to the backbone and head parts, providing a variety of configuration options to help achieve higher performance.

[0017] S3. Introduce the HS-FPN structure in the MFDS-DETR model: add the HS-FPN structure to the head of the original yaml file and replace the head of the original yaml file for use; HS-FPN (High-level Screening Fusion Pyramid Networks) was originally designed for white blood cell detection, but its excellent performance in processing multi-scale targets and complex backgrounds can also help ground penetrating radar pipeline detection. For the detection of pipelines of different depths and scales, HS-FPN can integrate multi-resolution feature maps to enhance the recognition ability of small-sized or shallowly buried pipeline locations. At the same time, it can effectively suppress background noise interference by screening relevant high-level features. Its optimized feature expression mechanism and adaptability to complex scenes help improve the accuracy and signal-to-noise ratio of GPR pipeline detection, and provide reliable support for pipeline depth positioning detection. Its core includes the following two key parts: Feature selection module: This module uses channel attention (CA) and dimension matching (DM) mechanisms to effectively screen feature maps of different scales. Through pooling operations such as global average pooling and global maximum pooling, as well as weight calculation, this module can efficiently extract important information in each channel. This method of selectively focusing on important features helps to better capture relevant features at each scale, as described below: Input and initial processing of the CA module: The CA module first processes the input feature map , whose shape is , where C represents the number of channels, H represents the height, and W represents the width; The feature map is processed by two pooling layers: global average pooling and global maximum pooling; Weight calculation: The pooled feature maps are combined together, and then the Sigmoid activation function is used to calculate the weight value of each channel. The resulting weight matrix is ; The purpose of pooling operation, the main functions of pooling include: downsampling and dimensionality reduction, reducing the spatial dimension of feature maps; eliminating redundant data, compressing features and reducing the number of parameters; achieving translation, rotation and scale invariance; The CA module uses global average pooling and global maximum pooling to extract the average and maximum value of each channel respectively. Maximum pooling extracts the most relevant data, while average pooling averages all data. Information extraction and feature map filtering: By combining the maximum pooling and average pooling methods, the CA module is able to extract the most representative information from each channel while minimizing information loss. The filtered feature map is generated by multiplying the weight information with the original feature map at the corresponding scale; Dimension matching (DM) module: Before feature fusion, feature maps of different scales need to be dimensionally matched because they have different numbers of channels. The DM module uses 1×1 convolution to reduce the number of feature map channels of each scale to 256 to achieve dimension matching; Feature fusion module: In the feature fusion module, the filtered features are integrated through the selective feature fusion (SFF) mechanism. The module first expands the high-level features, then resizes them through bilinear interpolation or transposed convolution, and finally fuses them with low-level features to enhance the model's ability to express image features. Through this fusion, high-level and low-level features are organically combined, enabling the network to better identify and detect targets at different scales. The specific contents are as follows: Characteristics of multi-scale feature maps: The multi-scale feature maps generated by the backbone network are very rich in semantic information, but relatively coarse in object positioning. On the contrary, low-level features provide precise object locations but limited semantic information; Limitations of traditional solutions: To resolve this contradiction, the usual solution is to directly sum the upsampled high-level features and low-level features pixel by pixel to enrich the semantic information of each layer. However, this technique does not perform feature selection, but simply sums the pixel values ​​of multiple feature layers; Introduction of SFF module: To overcome this limitation, the present invention introduces the SFF module, which uses high-level features as weights to filter important semantic information embedded in low-level features, thereby achieving effective feature fusion; The workflow of the SFF module: Given an input high-level feature and an input low-level feature First, we use the transposed convolution (T-Conv) with a stride of 2 and a convolution kernel of 3×3 to expand the high-level features to obtain the feature size ; Unify the dimensions of high-level and low-level features: In order to unify the dimensions of high-level features and low-level features, bilinear interpolation is used to upsample or downsample high-level features to obtain ; Application of attention mechanism: The CA module is used to convert high-level features into corresponding attention weights to filter low-level features. After obtaining features of consistent dimensions, the filtered low-level features are finally fused with high-level features to enhance the feature representation ability of the model. ; Formula describes the feature selection process: Equations (1) and (2) show the fusion process of feature selection: (1) (2); Advantages of combining transposed convolution and bilinear interpolation: In the image sampling process, transposed convolution and bilinear interpolation are used to restore the scale of high-level features. The bilinear interpolation method is simple and fast, and can directly operate on pixels to achieve image scaling; Advantages of transposed convolution: By adapting to the data through learnable parameters, the output can not only expand the feature map, but also implement the convolution operation by filling zeros after the feature map is expanded through the convolution kernel, thereby reconstructing the input; by sampling different areas of the input image at different positions of the output image, the problem of non-uniform sampling is handled; Improve C2f with DualCorv: A lightweight C2f is proposed, and C2f_Dual is added to the head part of the modified yaml file to replace the original C2f. like Figure 6 As shown in the figure, the basic principle of DualConv is as follows: Combining 3×3 and 1×1 convolution kernels: The design concept of combining 3×3 and 1×1 convolution kernels in the DualConv structure is to integrate the advantages of these two convolution kernels: 3×3 convolution kernels can capture more spatial information when performing feature extraction, while 1×1 convolution kernels can interact and integrate information between feature channels without increasing too many parameters and computational complexity. It shows how to combine 3×3 and 1×1 convolution kernels: In DualConv, 3×3 convolution kernels are used to extract spatial features of feature maps, while 1×1 convolution kernels are used to integrate these features and reduce model parameters: The convolution kernels in each group process a part of the input channel separately, and then the output is merged, thereby achieving efficient flow and integration of information between different feature map channels. This structural design not only maintains the network depth and representation ability, but also reduces the computational complexity and model size, making it suitable for resource-constrained environments; Using group convolution technology: DualConv uses group convolution technology, which is an effective strategy to reduce parameters and computation. In group convolution, the input and output feature maps are divided into multiple groups, and the convolution filter of each group only processes part of the corresponding input feature map, which reduces the complexity of the model. DualConv uses this technology to further reduce the computational cost because it allows different convolution kernels in the group (such as 3×3 and 1×1) to process the same group of input channels in parallel, optimizing the information flow and feature extraction efficiency while maintaining the representation ability of the network;

[0018] DualConv's structural layout: 3×3 and 1×1 convolution kernels are arranged in parallel on the input feature map channel. Specifically, this layout uses the group convolution technique to group the convolution kernels and use convolution kernels of different sizes in parallel within the same group. This design helps to simultaneously utilize the spatial feature extraction capabilities of large-sized convolution kernels and the computational efficiency of small-sized convolution kernels, thereby reducing the number of model parameters and computational costs while maintaining accuracy.

[0019] like Figure 7 As shown, model testing includes performance evaluation and calculation of mP: setting parameters, setting the optimal combination of iteration rounds, batches, optimizers, learning rate momentum, etc. and creating a py file for training, and then referencing the best model weight obtained from training for testing; The confusion matrix in the obtained data results is used to calculate precision and recall and compared with the original Yolov8n model, and then mAP is calculated using precision and recall; S6. Ablation experiment: Through ablation experiments, it is verified that the three added improved models have higher accuracy and robustness than a single improved model.

[0020] Working principle and use process of the present invention: The method is experimented on Windows 11 operating system, runs based on PyTorch 2.0.0, and uses an NVIDIA RTX 1650 12GB GPU for calculation. All experimental models are executed in the same experimental environment to ensure the comparability of the results. The image input size is set to 640×640, and the training is performed for a total of 200 epochs, with each batch size of 4; In order to test the detection performance of the improved model proposed in this paper, this paper uses precision, recall, mAP@0.5, model parameter scale, model calculation amount (floating point operations, FLOPs), model training time and detection speed (unit: frame / s) as evaluation indicators; the following parameters are used in the formula of the above evaluation indicators: TP (predicted as a positive sample, which is actually a positive sample), FP (predicted as a positive sample, although it is actually a negative sample) and FN (predicted as a negative sample, although it is actually a positive sample); Precision is the ratio of the number of positive samples predicted by the model to the number of all detected samples. The calculation formula is as follows: , Recall is the ratio of the number of positive samples correctly predicted by the model to the number of positive samples that actually appeared. The calculation formula is as follows: , The average precision (AP) is the area under the precision and recall curves. The mean average precision (mAP@0.5) is the weighted average of the AP values ​​of all sample categories, which is used to evaluate the detection performance of the model in all categories. The threshold of the intersection ratio of the predicted box and the true box is set to 0.5. The formula is as follows: , In order to verify the detection performance after adding the ADown module, the HS-FPN structure in the MFDS-DETR model, and the DualConv module to the YOLOv8 model, the present invention conducted an ablation experiment to compare the YOLOv8 model with the ADown module, the model with the CA module in the HS-FPN structure in the MFDS-DETR model, and the model with the DualConv. All experiments were conducted under the same settings, and the average precision (mAP@0.5) and single image inference time (FPS) were used to evaluate the detection accuracy and speed of the model. In the experimental comparison on the acquired GPR dataset, this paper conducted experimental analysis on the YOLOv8 model with the ADown module, the YOLOv8 model with the HS-FPN structure introduced in the MFDS-DETR model, and the YOLOv8 model with the DualConv. The experimental results were compared with the original YOLOv8 model, the CA module in the HS-FPN structure introduced in the MFDS-DETR model, and the ADC-YOLOv8 model with the DualConv. The experimental results are as follows Figure 8 As shown, the detection results of YOLOv8n introduced different modules for comparison; Through in-depth analysis of the experimental results, the following conclusions can be drawn: Significant performance improvement: Experiments on the ground penetrating radar (GPR) pipeline dataset using the YOLOv8n model (ADC-YOLOv8n) with ADown, HS-FPN's CA module, and DualConv module show that ADC-YOLOv8n improves Precision and mAP@50 by 3.5% and 1.5% respectively, while reducing FLOPs by 24% and the number of parameters by 50% compared to the original YOLOv8n model. These improvements significantly improve the detection performance and computational efficiency of the model, and verify the effectiveness and reliability of each module in the target detection task. Among them, the CA module effectively enhances the model's perception of pipeline targets and improves detection accuracy by strengthening the importance of feature channels. Reasonable computational cost control: The ADC-YOLOv8n model introduces a variety of improved modules, which increases the computational complexity to a certain extent, but the FLOPs is only 5.8G, which is about 24% less than the 8.2G FLOPs of the baseline YOLOv8n model. This optimization not only improves the detection efficiency of the model, but also reduces resource consumption, making it more applicable in practical applications, especially in environments with limited computing resources. This is due to the lightweight design of the ADown and DualConv modules, which can effectively utilize feature information and perform fine feature fusion, thereby improving detection performance without significantly increasing computational costs, and comprehensively improving precision and recall:

[0021] The ADC-YOLOv8n model not only performs well in accuracy, reaching 0.971, but also achieves a significant improvement in recall, reaching 0.815. This means that the introduced module combination can effectively improve the model's ability to detect and identify targets, and is more robust in complex scenarios. Compared with the model with ADown, CA or DualConv modules added separately, the comprehensive performance advantage of ADC-YOLOv8n is more obvious, further proving the synergistic effect of these modules when used together. In summary, the experimental comparison and result analysis of the YOLOv8n model with ADown, HS-FPN's CA module and DualConv module fully demonstrates the effectiveness and practicality of these improved methods; These improvements not only achieved significant breakthroughs in detection accuracy, but also enhanced the practical application value of the model while maintaining a low computational cost, providing important reference and guidance for further optimizing the target detection model, especially the ADC-YOLOv8n model, which will be more competitive in practical applications and suitable for tasks with high requirements for accuracy and efficiency.

[0022] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can refer to each other. For the system module disclosed in the embodiment, since it corresponds to the method part, its description is relatively simple, and the relevant details can be referred to the description of the method part. The above description of the embodiment enables professionals and technicians in this field to implement or use the present invention. Various modifications and changes to these embodiments will be obvious to professionals and technicians. The general principles defined in this article can be applied to other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown in this article, but should include the broadest range consistent with the principles and novel features disclosed herein.

[0023] In the embodiments provided in the present application, it should be understood that the disclosed apparatus / terminal equipment and method can be implemented in other ways, and the above embodiments are only illustrative examples, wherein the division of the modules or units only represents a logical function allocation. In actual implementation, these modules or units can be divided in different ways, for example, multiple units or components can be combined, or integrated into other systems, and certain features can also be ignored or not executed. It should be pointed out that the description of the above embodiments is only used to illustrate the technical solution of the present application, and is not intended to limit it. Although the above embodiments have been described in detail, those skilled in the art should understand that these technical solutions can still be modified, or some technical features can be replaced by equivalents. These modifications or replacements will not cause the essence of the technical solution to deviate from the technical concept and protection scope of the present application, and should all be included in the protection scope of the present application.

Claims

1. A method for identifying and locating buried pipelines based on target detection, characterized in that: The following steps are involved: S1. Input the labeled PE pipe welding defect image; S2. Introduce the ADown module to improve the Conv module: Through downsampling operations, the ADown module effectively reduces the spatial dimension of the feature map and captures high-level features; Reduce the amount of calculation by reducing redundant feature information and optimizing calculation efficiency; S3. Introduce the HS-FPN structure in the MFDS-DETR model: optimize the Neck part of the original YOLOv8 model to achieve efficient fusion of multi-scale features and improve the feature expression ability of the model; S4. Using DualCorv to improve C2f: The DualConv technology is introduced to lightweight improve the C2f module of the original YOLOv8 model, combining 3×3 and 1×1 convolution kernels to process the same input feature channel, significantly improving the feature extraction efficiency; S5, model testing, wherein the model testing includes performance evaluation and calculation of mPA; The model parameters were carefully adjusted according to specific inspection requirements, and the improved model was fully tested under multiple experimental conditions to verify its accuracy and inspection efficiency in GPR dataset defect signal identification. Through experimental testing and performance evaluation, the detection capability of the model was further optimized to ensure that it can maintain efficient and stable performance in different data environments and actual applications, providing strong technical support for pipeline detection and positioning; S6. Ablation experiment.

2. The method for identifying and locating buried pipelines based on target detection according to claim 1, characterized in that: In "S1", a large number of GPR data sets are obtained, B-Scan images under different conditions are selected from the data sets, and the pipeline burial depth positions are annotated by labellmg. The annotated GPR data sets are divided into a validation set and a training set by a python script.

3. The method for identifying and locating buried pipelines based on target detection according to claim 2, characterized in that: In "S2", the improved ADown module reduces the number of channels of the input feature map to half of the original through convolution operation, and then downsamples through the average pooling layer to reduce the spatial dimension of the feature map, which not only reduces the size of the feature map but also retains the important feature information of the image, enabling the model to capture the key features of the image at a higher level.

4. The method for identifying and locating buried pipelines based on target detection according to claim 3 is characterized in that: In "S3", HS-FPN combines low-level detail features and high-level semantic features to generate more representative comprehensive feature maps.

5. The method for identifying and locating buried pipelines based on target detection according to claim 4, characterized in that: In "S4", group convolution technology is used to efficiently arrange convolution filters, significantly reducing the computational cost and the number of parameters.

6. The method for identifying and locating buried pipelines based on target detection according to claim 5, characterized in that: In "S5", set the parameters, set the optimal combination of iteration rounds, batches, optimizers, and learning rate momentum, and create a py file for training. Then, reference the best model weights obtained from training for testing. Use the confusion matrix in the obtained data results to calculate precision and recall and compare them with the original Yolov8n model. Then, calculate mAP through precision and recall.

7. The method for identifying and locating buried pipelines based on target detection according to claim 1, characterized in that: In "S6", ablation experiments are performed to verify that the three added improved models have higher accuracy and robustness than a single improved model.

Images