A method for measuring the size of defects in oil pipelines based on HMCNet

Abstract

The application discloses a kind of HMCNet-based oil pipeline defect size measurement method, comprising: S1, initialization oil pipeline shooting equipment parameter, oil pipeline is photographed at multiple angles, obtains the multiple groups of image information of different parts of oil pipeline;S2, the image information obtained is input to the deep learning network HMCNet that is trained using a large number of known defect samples in advance;S3, the input image features are divided into multiple regions by HMCNet, the local attention to target image is enhanced using the region perception module, and the defect information is identified and located;S4, obtain the processing result of HMCNet, output the position, type and defect size of oil pipeline defect.The application realizes the rapid, accurate detection of the position, type and size of oil pipeline defect, greatly improves the detection efficiency and anti-interference ability, and is more convenient to operate.

Description

Technical Field

[0001] This invention relates to the field of oil pipeline inspection technology, specifically to a method for measuring the size of oil pipeline defects based on HMCNet. Background Technology

[0002] As a critical infrastructure for oil transportation, the safety of oil pipelines is of paramount importance. Pipeline defects leading to oil leaks not only cause enormous economic losses but also result in severe consequences such as environmental pollution and safety incidents. Accurately measuring the dimensions of oil pipeline defects is crucial for timely assessment of pipeline safety and the development of appropriate maintenance plans. Therefore, the development of a high-precision and efficient method for measuring the dimensions of oil pipeline defects is urgently needed.

[0003] Traditional techniques for measuring the dimensions of defects in oil pipelines mainly include ultrasonic testing, magnetic particle testing, and radiographic testing. Ultrasonic testing utilizes the propagation characteristics of ultrasound waves in different media to detect defects, but it has certain errors in judging the shape and location of defects and is difficult to detect defects in complex structures. Magnetic particle testing is only applicable to ferromagnetic materials, which has significant limitations. Radiographic testing, while offering high accuracy, poses radiation hazards, is complex to operate, and is costly. These traditional techniques often require specialized personnel to operate, have low testing efficiency, and cannot meet the demands of large-scale, rapid inspection of modern oil pipelines.

[0004] With the development of computer technology and image processing technology, image processing-based defect detection technology for oil pipelines has gradually emerged. Some existing deep learning-based detection methods have improved the accuracy and efficiency of defect detection to a certain extent. However, these methods still suffer from insufficient detection accuracy when dealing with minute defects in complex backgrounds. Moreover, most existing technologies can only detect the presence of defects, making it difficult to accurately measure their size. Furthermore, existing technologies are lacking in multimodal information fusion and local feature extraction, failing to fully utilize the various information in pipeline images, leading to inaccurate judgments of defect type and size.

[0005] Therefore, neither traditional detection techniques nor existing deep learning-based detection methods can fully meet the high-precision, high-efficiency, and intelligent requirements for measuring the size of defects in oil pipelines. Therefore, this application proposes an HMCNet-based method for measuring the size of defects in oil pipelines, overcoming the shortcomings of existing technologies and providing a more reliable guarantee for the safe operation of oil pipelines. Summary of the Invention

[0006] Based on the above, a method for measuring the defect size of oil pipelines based on HMCNet specifically includes the following steps:

[0007] S1. Initialize the parameters of the oil pipeline imaging equipment, take multi-angle shots of the oil pipeline, and obtain multiple sets of image information of different parts of the oil pipeline.

[0008] S2. Input the acquired image information into HMCNet, a deep learning network that has been trained in advance using a large number of known defect samples;

[0009] S3. The input image features are divided into multiple regions using HMCNet, and the region awareness module is used to enhance local attention to the target image, identify and locate defect information.

[0010] S4. Obtain the processing results from HMCNet and output the location, type, and size of the defects in the oil pipeline.

[0011] Preferably, in step S1, an intelligent gimbal handheld device integrating an inertial measurement unit and a laser rangefinder is used. With the central axis of the pipe as a reference, the laser rangefinder measures the distance in real time and the integrated inertial measurement unit monitors the attitude. The device moves around the pipe according to a spiral trajectory to capture images from multiple perspectives.

[0012] Preferably, the multi-view image information acquired in S1 includes an overall appearance image of the pipe and a close-up image of a part of the pipe; the overall appearance image of the pipe includes images of the pipe connection parts and the connection area of ​​the support structure; the close-up image of the pipe includes images of the weld, the surface texture of the pipe, and small pits or protrusions.

[0013] Preferably, the deep learning network HMCNet in S2 includes a backbone network, a neck network, and a head network; the acquired image information is transmitted to the backbone network to extract multi-scale features of the image, the multi-scale features extracted by the backbone network are fused through the neck network to extract semantic information, and the head network is used to perform target detection based on the semantic information provided by the neck network to generate detection results.

[0014] Preferably, the backbone network adopts a multi-branch heterogeneous fusion architecture, which includes a lightweight branch and a deep convolutional branch. The lightweight branch reduces computation through grouped convolution and pointwise convolution, and extracts the basic features of the pipeline image. The deep convolutional branch dynamically adjusts the weights of different convolutional kernels according to the input image features, enhancing the ability to capture complex defect features. The feature maps of the lightweight branch and the deep convolutional branch are fused through a cross-channel interaction module, and the attention mechanism is used to redistribute features in the channel dimension to obtain multi-scale features of the image.

[0015] Preferably, the neck network adopts a cross-scale feature interaction architecture. By constructing a multi-scale feature fusion tree, the feature maps of different scales output by the backbone network are used as nodes of the tree. Each node is connected to other nodes at different levels through skip connections to perform cross-scale feature interaction.

[0016] Preferably, the head network adopts a multimodal fusion architecture to detect defects in oil pipelines; the head network is composed of a target classification subnetwork, a bounding box regression subnetwork, and a key point detection subnetwork in parallel; the target classification subnetwork introduces a meta-learning mechanism, which pre-trains a set of general classifier weights and adjusts the weights when facing different types of pipeline defect images, and uses multimodal features to classify defect types;

[0017] The bounding box regression subnetwork uses an adaptive anchor point generation strategy to dynamically adjust the scale and proportion of anchor points based on the feature map output by the neck network, adapting to defects of different sizes and shapes.

[0018] The key point detection subnetwork adopts a graph convolutional network structure, which treats the key points of the defect as nodes in the graph and learns the spatial distribution characteristics of the key points through the connection relationship between the nodes.

[0019] The outputs of the target classification subnetwork, bounding box regression subnetwork, and key point detection subnetwork are integrated through a dynamic fusion mechanism. Based on the uncertainty of the input features and the confidence of each subnetwork, the weights in the output results are adjusted to output the defect classification, bounding box, and key point information of the image.

[0020] Preferably, in step S3, the input image features are divided into multiple regions using HMCNet, and the local attention to the target image is enhanced by a region-aware module, as shown in the formula: in, R(·) represents the features after region perception, and R(·) represents the region perception module. HMCNET divides the image features into multiple regions and then uses the region perception module to distinguish the defect types.

[0021] Preferably, in step S3, the defect location information is captured using the Transformer's self-attention mechanism to locate the defect; the image features and target query information are processed by the Transformer's encoder and decoder respectively; the encoder process is as follows: Z enc =TransformerEnc(F r Z pos ), where Z enc These are features processed by the encoder. It is a feature derived from region perception, Z pos It is positional encoding; the decoder process is: Z dec=TransformerDec(Z) enc Z query ), where Z query It is the target query vector, Z dec It is the result output by the decoder; the image information is captured and the defect location is output through the self-attention mechanism of the Transformer.

[0022] Preferably, in step S4, after obtaining the processing results of HMCNet, the location, type, and size of the oil pipeline defect are output to the intelligent early warning platform to obtain the defect information in real time and issue an early warning. The defect information is then notified to relevant personnel for handling through different early warning channels, including SMS, email, or communication software.

[0023] Compared with the prior art, the technical solution of this application has the following technical effects:

[0024] This invention utilizes a smart gimbal handheld device integrating an inertial measurement unit and a laser rangefinder to capture multi-view images of the pipeline by moving along a spiral trajectory around the pipeline, using the pipeline's central axis as a reference. This solves the technical problem that traditional shooting methods struggle to comprehensively capture images of different parts of the oil pipeline, leading to the easy omission of some defects. The invention achieves the technical effect of acquiring multi-view image information, including the overall appearance of the pipeline (such as connection parts and supporting structure connection areas) as well as close-up details (such as welds, surface textures, and small pits or protrusions), providing a rich and comprehensive data foundation for subsequent accurate detection and measurement of defects.

[0025] This invention employs the deep learning network HMCNet, whose backbone network adopts a multi-branch heterogeneous fusion architecture, the neck network adopts a cross-scale feature interaction architecture, and the head network adopts a multi-modal fusion architecture. This technical solution solves the technical problems of insufficient detection accuracy of existing deep learning-based detection methods in handling small defects in complex backgrounds and the difficulty in accurately measuring defect size. It achieves the technical effect of effectively capturing complex defect features, fusing multi-scale features and multi-modal information, thereby accurately identifying and locating oil pipeline defect information, accurately measuring defect size, and accurately determining defect type.

[0026] This invention employs a region-aware module to enhance local attention to target images, processes image features, and utilizes the Transformer's self-attention mechanism to capture image information and locate defects. This technical solution solves the technical problem of difficulty in focusing on minute defects and accurately locating their positions during detection. It achieves the technical effect of enhancing attention to specific areas of the target image, more accurately capturing defect information in the image, and thus realizing precise location of defects in oil pipelines, providing accurate location data for subsequent defect assessment and repair.

[0027] This invention provides a technical solution that outputs the processing results of HMCNet to an intelligent early warning platform and notifies relevant personnel through different early warning channels such as SMS, email, or communication software. This solves the technical problem that the failure to transmit information in a timely manner after a defect is detected leads to the inability to handle the defect in a timely manner and poses a safety hazard. The invention achieves the technical effect of being able to obtain defect information in real time and notify relevant personnel in a timely manner, which greatly improves the timeliness and safety of oil pipeline maintenance and reduces pipeline operation risks.

[0028] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0029] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0030] 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, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0031] Figure 1 This is a flowchart of a method for measuring the defect size of an oil pipeline based on HMCNet according to the present invention.

[0032] Figure 2 This is a flowchart of the deep learning network HMCNet, which is used in the oil pipeline defect size measurement method based on HMCNet according to the present invention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0034] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0035] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0036] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0037] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0038] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0039] Example 1

[0040] This embodiment mainly describes a method for measuring the defect size of oil pipelines based on HMCNet, such as... Figure 1 As shown, it includes the following steps:

[0041] S1. Initialize the parameters of the oil pipeline imaging equipment, take multi-angle shots of the oil pipeline, and obtain multiple sets of image information of different parts of the oil pipeline.

[0042] S2. Input the acquired image information into HMCNet, a deep learning network that has been trained in advance using a large number of known defect samples;

[0043] S3. The input image features are divided into multiple regions using HMCNet, and the region awareness module is used to enhance local attention to the target image, identify and locate defect information.

[0044] S4. Obtain the processing results from HMCNet and output the location, type, and size of the defects in the oil pipeline.

[0045] Furthermore, in S1, an intelligent gimbal handheld device integrating an inertial measurement unit and a laser rangefinder is used. With the central axis of the pipeline as a reference, the laser rangefinder measures the distance in real time and the integrated inertial measurement unit monitors the attitude. The device moves around the pipeline according to a spiral trajectory to capture images from multiple perspectives.

[0046] The multi-angle shooting utilizes a handheld device equipped with a smart gimbal. The smart gimbal integrates a high-precision inertial measurement unit (IMU) and environmental sensing sensors, such as a laser rangefinder and a vision sensor. The laser rangefinder measures the distance to the oil pipeline in real time, and combined with the pipeline contour information obtained by the vision sensor, the smart gimbal automatically plans the shooting path based on the pipeline's central axis according to a preset shooting strategy. During the shooting process, the IMU monitors the handheld device's posture changes in real time to ensure that the angle deviation of each shot is controlled within a very small range. The handheld device moves around the pipeline in a spiral upward or downward trajectory, taking pictures at different axial positions and circumferential angles. The axial spacing and circumferential angle interval of each shot are dynamically adjusted according to the pipeline's diameter and preset accuracy requirements, thereby acquiring high-resolution image information covering various parts of the oil pipeline from different perspectives.

[0047] Furthermore, the multi-view image information acquired in S1 includes an overall appearance image of the pipeline and a close-up image of a part of the pipeline; the overall appearance image of the pipeline includes images of the pipeline connection parts and the connection area of ​​the support structure; the close-up image of the pipeline includes images of the weld seam, the surface texture of the pipeline, and small pits or protrusions.

[0048] Furthermore, the deep learning network HMCNet in S2 includes a backbone network, a neck network, and a head network. The acquired image information is transmitted to the backbone network to extract multi-scale features of the image. The neck network fuses the multi-scale features extracted by the backbone network to extract semantic information. The head network then uses the semantic information provided by the neck network to perform target detection and generate detection results.

[0049] Furthermore, the backbone network adopts a multi-branch heterogeneous fusion architecture, which includes a lightweight branch and a deep convolutional branch. The lightweight branch reduces computation through grouped convolution and pointwise convolution, extracting the basic features of the pipeline image. The deep convolutional branch dynamically adjusts the weights of different convolutional kernels according to the input image features, enhancing the ability to capture complex defect features. The feature maps of the lightweight branch and the deep convolutional branch are fused through a cross-channel interaction module, and the attention mechanism is used to redistribute features in the channel dimension to obtain multi-scale features of the image.

[0050] Furthermore, the neck network adopts a cross-scale feature interaction architecture. By constructing a multi-scale feature fusion tree, the feature maps of different scales output by the backbone network are used as nodes of the tree. Each node is connected to other nodes at different levels through skip connections to perform cross-scale feature interaction.

[0051] Furthermore, the head network adopts a multimodal fusion architecture to detect defects in oil pipelines. The head network consists of a target classification subnetwork, a bounding box regression subnetwork, and a key point detection subnetwork that are constructed in parallel. The target classification subnetwork introduces a meta-learning mechanism, which adjusts the weights when faced with different types of pipeline defect images by pre-training a set of general classifier weights and using multimodal features to classify defect types.

[0052] The bounding box regression subnetwork dynamically adjusts the scale and proportion of anchor points based on the feature map output by the neck network through an adaptive anchor point generation strategy, adapting to defects of different sizes and shapes.

[0053] The key point detection subnetwork adopts a graph convolutional network structure, treating the key points of the defect as nodes in the graph, and learning the spatial distribution characteristics of the key points through the connection relationship between the nodes;

[0054] The outputs of the target classification subnetwork, bounding box regression subnetwork, and key point detection subnetwork are integrated through a dynamic fusion mechanism. Based on the uncertainty of the input features and the confidence of each subnetwork, the weights in the output results are adjusted to output the defect classification, bounding box, and key point information of the image.

[0055] Furthermore, in S3, the input image features are divided into multiple regions using HMCNet, and the region-aware module enhances local attention to the target image. The formula is as follows: in, R(·) represents the features after region perception, and R(·) represents the region perception module. HMCNET divides the image features into multiple regions and then uses the region perception module to distinguish the defect types.

[0056] Furthermore, in S3, defect location information is captured using the Transformer's self-attention mechanism to locate defects; the Transformer's encoder and decoder process image features and target query information respectively; the encoder process is as follows: Z enc =TransformerEnc(Fr,Z) pos ), where Z enc These are features processed by the encoder. It is a feature derived from region perception, Z pos It is position encoding; the decoder process is: Z dec =TransformerDec(Z) enc Z query ), where Z query It is the target query vector, Z dec It is the result output by the decoder; the image information is captured and the defect location is output through the self-attention mechanism of the Transformer.

[0057] Furthermore, in S4, after obtaining the processing results from HMCNet, the location, type, and size of the oil pipeline defects are output to the intelligent early warning platform. The defect information is acquired in real time, and early warnings are issued. The defect information is then notified to relevant personnel through different early warning channels, including SMS, email, or communication software, for handling.

[0058] This embodiment details how this application utilizes an intelligent gimbal handheld device integrating an inertial measurement unit and a laser rangefinder to acquire multi-view image information of oil pipelines. This information is then input into a deep learning network HMCNet, combined with a region perception module and a Transformer self-attention mechanism for defect identification and localization. Finally, the processing results are output to an intelligent early warning platform to notify relevant personnel. This solves the problems of low accuracy, low efficiency, and difficulty in detecting defects in complex parts by traditional detection technologies, as well as the shortcomings of existing deep learning detection methods in detecting minute defects, accurately measuring dimensions, and providing timely early warnings.

[0059] Based on Example 1, this example describes how the deep learning network HMCNet processes oil pipeline images and outputs defect information, such as... Figure 2 As shown, specifically:

[0060] Image information of oil pipelines is acquired and transmitted to the HMCNet backbone network. The backbone network adopts a multi-branch heterogeneous fusion architecture, including a lightweight branch and a deep convolution branch. The lightweight branch reduces computational load through grouped convolution and pointwise convolution, thereby extracting basic features of the pipeline image. This reduces computational burden and improves processing efficiency while ensuring basic feature extraction. The deep convolution branch dynamically adjusts the weights of different convolution kernels according to the input image features, enhancing the ability to capture complex defect features and effectively identifying various complex defect situations. The feature maps generated by the lightweight branch and the deep convolution branch are fused through a cross-channel interaction module, and the features are redistributed along the channel dimension using an attention mechanism, thereby obtaining multi-scale features of the image and providing rich feature information for subsequent processing.

[0061] The multi-scale features processed by the backbone network enter the neck network. The neck network adopts a cross-scale feature interaction architecture. By constructing a multi-scale feature fusion tree, the feature maps of different scales output by the backbone network are used as nodes of the tree. Each node is connected to other nodes at different levels through skip connections to perform cross-scale feature interaction, so that features of different scales can complement and fuse each other, further extracting more representative semantic information and improving the ability to understand image content.

[0062] The semantic information output from the neck network is fed into the head network, which employs a multimodal fusion architecture consisting of a target classification subnetwork, a bounding box regression subnetwork, and a keypoint detection subnetwork operating in parallel. The target classification subnetwork introduces a meta-learning mechanism, pre-training a set of general classifier weights to adjust the weights when faced with different types of pipeline defect images. Utilizing multimodal features, it classifies defect types accurately. The bounding box regression subnetwork uses an adaptive anchor point generation strategy, dynamically adjusting the scale and proportion of anchor points based on the feature map output from the neck network to adapt to defects of different sizes and shapes, thus more accurately determining the defect boundaries. The keypoint detection subnetwork uses a graph convolutional network structure, treating the defect's keypoints as nodes in the graph and learning the spatial distribution features of keypoints through the connections between nodes. The outputs of the three subnetworks are integrated through a dynamic fusion mechanism, adjusting the weights in the output result based on the uncertainty of the input features and the confidence of each subnetwork, outputting the defect classification, bounding box, and keypoint information of the image.

[0063] HMCNet enhances local attention to target images through a region-aware module. This module divides the input image features into multiple regions, allowing the network to focus more on defect areas and distinguish defect types. Furthermore, it utilizes the Transformer's self-attention mechanism to capture image information for defect localization. The encoder and decoder process image features and target query information respectively, ultimately accurately outputting the location information of oil pipeline defects. Through this complex yet orderly processing flow, HMCNet can accurately process oil pipeline images and output key information such as the location, type, and size of defects.

[0064] For the region-aware module, when processing oil pipeline images, HMCNet first divides the input image features into multiple regions. The region-aware module then functions by focusing on specific local areas of the target image, enhancing the extraction and analysis of features in these areas, thereby better distinguishing different types of defects. For example, when processing images containing minor pits or weld defects, the region-aware module allows the network to focus on the defective locality, ignoring irrelevant background information. This effectively improves the sensitivity and recognition ability of defect features, providing strong support for subsequent accurate identification and location of oil pipeline defects and accurate determination of defect types, ensuring a more precise and efficient detection process.

[0065] This embodiment describes in detail the deep learning network HMCNet, which efficiently processes oil pipeline images through a multi-branch heterogeneous fusion backbone network, a neck network with cross-scale feature interaction, a multimodal fusion head network, a region perception module, and a Transformer self-attention mechanism. It accurately captures multi-scale features of tiny defects in complex backgrounds from multi-view images, accurately identifies and locates defects through multimodal information fusion, accurately determines the defect type, and accurately measures the defect size.

[0066] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. A method for measuring the size of defects in oil pipelines based on HMCNet, characterized in that, Includes the following steps: S1. Initialize the parameters of the oil pipeline imaging equipment, take multi-angle shots of the oil pipeline, and obtain multiple sets of image information of different parts of the oil pipeline. S2. Input the acquired image information into HMCNet, a deep learning network that has been trained in advance using a large number of known defect samples; S3. The input image features are divided into multiple regions using HMCNet, and the region awareness module is used to enhance local attention to the target image, identify and locate defect information. S4. Obtain the processing results from HMCNet and output the location, type, and size of the defects in the oil pipeline. The deep learning network HMCNet in S2 includes a backbone network, a neck network, and a head network. The acquired image information is transmitted to the backbone network to extract multi-scale features of the image. The neck network fuses the multi-scale features of the image extracted by the backbone network to extract semantic information. The head network uses the semantic information provided by the neck network to perform target detection and generate detection results. The backbone network adopts a multi-branch heterogeneous fusion architecture, which includes lightweight branches and deep convolutional branches. The lightweight branch reduces computation through grouped convolution and pointwise convolution, extracting basic features of the pipeline image; the deep convolution branch dynamically adjusts the weights of different convolution kernels based on the input image features, enhancing the ability to capture complex defect features; the feature maps of the lightweight branch and the deep convolution branch are fused through a cross-channel interaction module, and the attention mechanism is used to redistribute features in the channel dimension to obtain multi-scale features of the image.

2. The method for measuring the defect size of oil pipelines based on HMCNet according to claim 1, characterized in that, In step S1, a smart gimbal handheld device integrating an inertial measurement unit and a laser rangefinder is used. With the central axis of the pipeline as a reference, the laser rangefinder measures the distance in real time and the integrated inertial measurement unit monitors the attitude. The device moves around the pipeline according to a spiral trajectory to capture images from multiple perspectives.

3. The method for measuring the defect size of oil pipelines based on HMCNet according to claim 2, characterized in that, The multi-view image information acquired in S1 includes an overall appearance image of the pipeline and a close-up image of a part of the pipeline; the overall appearance image of the pipeline includes images of the pipeline connection parts and the connection area of ​​the support structure; the close-up image of the pipeline includes images of the weld seam, the surface texture of the pipeline, and small pits or protrusions.

4. The method for measuring the defect size of oil pipelines based on HMCNet according to claim 1, characterized in that, The neck network adopts a cross-scale feature interaction architecture. By constructing a multi-scale feature fusion tree, the feature maps of different scales output by the backbone network are used as nodes of the tree. Each node is connected to other nodes at different levels through skip connections to perform cross-scale feature interaction.

5. The method for measuring the defect size of oil pipelines based on HMCNet according to claim 1, characterized in that, The head network adopts a multimodal fusion architecture to detect defects in oil pipelines. The head network consists of a target classification subnetwork, a bounding box regression subnetwork, and a key point detection subnetwork in parallel. The target classification subnetwork introduces a meta-learning mechanism, which pre-trains a set of general classifier weights and adjusts the weights when faced with different types of pipeline defect images, using multimodal features to classify defect types. The bounding box regression subnetwork uses an adaptive anchor point generation strategy to dynamically adjust the scale and proportion of anchor points based on the feature map output by the neck network, adapting to defects of different sizes and shapes. The key point detection subnetwork adopts a graph convolutional network structure, which treats the key points of the defect as nodes in the graph and learns the spatial distribution characteristics of the key points through the connection relationship between the nodes. The outputs of the target classification subnetwork, bounding box regression subnetwork, and key point detection subnetwork are integrated through a dynamic fusion mechanism. Based on the uncertainty of the input features and the confidence of each subnetwork, the weights in the output results are adjusted to output the defect classification, bounding box, and key point information of the image.

6. The method for measuring the defect size of an oil pipeline based on HMCNet according to claim 1, characterized in that, In step S3, the input image features are divided into multiple regions using HMCNet, and the region-aware module enhances local attention to the target image. The formula is as follows: ,in, These are features derived from region perception. It is a region-aware module; HMCNET divides image features into multiple regions and then uses the region-aware module to distinguish defect types.

7. A method for measuring the defect size of an oil pipeline based on HMCNet according to claim 1 or 6, characterized in that, In step S3, defect location information is captured using the Transformer's self-attention mechanism to locate defects; the encoder and decoder of the Transformer process image features and target query information respectively; the encoder process is as follows: ,in, These are features processed by the encoder. These are features derived from region perception. It is position encoding; the decoder process is as follows: ,in, It is the target query vector. It is the result output by the decoder; the image information is captured and the defect location is output through the self-attention mechanism of the Transformer.

8. The method for measuring the defect size of oil pipelines based on HMCNet according to claim 1, characterized in that, In step S4, after obtaining the processing results of HMCNet, the location, type, and size of the oil pipeline defect are output to the intelligent early warning platform. The defect information is acquired in real time, and an early warning is issued. The defect information is then notified to relevant personnel through different early warning channels, including SMS, email, or communication software, for handling.

Images