An electric power inspection unmanned aerial vehicle image acquisition and identification device and system

Abstract

The application relates to the technical field of image acquisition and analysis, and particularly discloses an electric power inspection unmanned aerial vehicle image acquisition and recognition device and system, which comprises an adaptive holder and a multispectral polarization imaging module. The system solves the image quality deterioration problem under strong wind and complex illumination through a data acquisition and dynamic stabilization module and a multispectral polarization imaging and fusion module; through an edge enhancement and feature extraction module, a three-dimensional reconstruction and accurate positioning module and a path planning and autonomous obstacle avoidance module, accurate spatial perception and autonomous flight planning of a target under a complex background are realized; through an intelligent diagnosis module integrated with a hot texture double-flow recognition model and an edge-cloud collaborative architecture, non-device attachment interference is effectively filtered out, and low-power-consumption and high-precision real-time defect diagnosis is realized. The application significantly improves the full-autonomous inspection capability, operation efficiency and diagnosis reliability of the unmanned aerial vehicle under a severe natural environment.

Description

Technical Field

[0001] This invention relates to the field of image acquisition and analysis technology, specifically to an image acquisition and recognition device and system for power line inspection drones. Background Technology

[0002] With the continuous development of smart grids and the ongoing expansion of power grid scale, the natural environment in which transmission lines are located is becoming increasingly complex. Traditional manual inspection methods are no longer sufficient to meet the demands for efficient and safe operation and maintenance. Against this backdrop, unmanned aerial vehicle (UAV) power line inspection technology, due to its flexibility, efficiency, and accessibility, has become an important means of monitoring the condition and identifying defects in transmission lines. Current technological development focuses on improving the autonomy and intelligence of UAV inspections, involving the research and development of dedicated UAV platforms, the optimization of airborne sensing and imaging systems, automatic defect identification based on computer vision and artificial intelligence, and autonomous path planning combined with three-dimensional spatial information. The aim is to build a closed-loop inspection system integrating environmental perception, data acquisition, intelligent analysis, and autonomous operation.

[0003] For example, application number "CN202411602423.0" discloses an autonomous flight path planning and image acquisition method and system for unmanned aerial vehicles (UAVs). Firstly, intelligent 3D mesh generation and dynamic flight path planning significantly improve the coverage and efficiency of feature area monitoring. Secondly, the integration of real-time meteorological data and historical feature data enhances the targeting and adaptability of the acquisition, helping to obtain high-quality images in complex and ever-changing environments. Thirdly, the combination of real-time processing and ground-based deep learning analysis greatly improves the accuracy and efficiency of feature extraction. Based on an adaptive supplementary sampling mechanism for key areas of interest, rapid response and precise positioning of suspected feature areas are achieved, providing reliable data support for subsequent detailed analysis and decision-making. However, existing UAV power line inspection technology still faces a series of key challenges in practical applications. Firstly, in complex and ever-changing external environments (such as strong winds in mountainous areas and strong light reflection from snow), the UAV platform is prone to severe shaking, resulting in serious motion blur and geometric distortion in the acquired images. Furthermore, the lack of intelligent anti-interference shooting strategies for the gimbal makes it impossible to guarantee the quality of the original images, directly affecting the reliability of all subsequent processing stages. Secondly, even with relatively clear images, existing feature extraction and matching algorithms still fall short in identifying and locating small targets (such as insulator strings and equipotential rings) on power transmission lines under complex backgrounds and varying lighting conditions. This makes it difficult to support high-precision 3D scene reconstruction, thus limiting the fully autonomous and adaptive trajectory planning capabilities based on 3D models. Existing path planning methods also have limited efficiency and adaptability when facing dynamic obstacles and complex spatial constraints. Finally, in the intelligent defect diagnosis stage, existing recognition models based on visible light images are susceptible to interference from non-target attachments such as bird nests and vines, leading to increased false alarm rates. Furthermore, considering the strict limitations on computing power and power consumption of UAV onboard platforms, how to deploy a lightweight intelligent diagnostic model that balances high precision and low power consumption on embedded devices to achieve real-time and accurate defect identification remains a critical technical bottleneck that urgently needs to be overcome. Summary of the Invention

[0004] (a) Technical problems to be solved This invention provides an image acquisition and recognition device and system for power line inspection drones, which solves the problems mentioned in the background art above.

[0005] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: an image acquisition and recognition device for power line inspection drones, comprising: An adaptive gimbal bracket is fixedly connected to the front end of the UAV. It integrates an embedded meteorological sensor unit, which includes a miniature anemometer, a humidity sensor, and a light intensity sensor. The adaptive gimbal bracket also integrates a flight control and data preprocessing sub-board. A multispectral polarization imaging module includes a camera that integrates visible light, infrared thermal imaging, and ultraviolet light. The multispectral polarization imaging module is fixedly installed at the bottom of the adaptive gimbal bracket. The multispectral polarization imaging module is internally configured with an electrically controlled rotating polarizer, which is driven by a micro stepper motor. The flight control and data preprocessing sub-board built into the adaptive gimbal bracket is electrically connected to the multispectral polarization imaging module via a cable, and is used to collect sensor data and control the movement of the electronically controlled rotating polarizer.

[0006] A power line inspection drone image acquisition and recognition system includes: The data acquisition and dynamic stabilization module is used to acquire the UAV's flight attitude, environmental meteorological data and raw multispectral image sequences in real time, and perform dynamic image correction based on visual inertial odometry and Kalman filtering to eliminate motion blur and geometric distortion caused by disturbances such as strong winds, and obtain the stabilized image sequence and attitude stability score (SSI). The multispectral polarization imaging and fusion module is used to receive the stabilized image sequence, control the electrically controlled rotating polarizer to adjust the polarization direction according to the real-time illumination angle and the material characteristics of the target surface, acquire multi-band images under different polarization states, and fuse multispectral information and polarization information to suppress specular reflection and scattered light interference. The edge enhancement and feature extraction module is used to process the image after fusion by the multispectral polarization imaging and fusion module based on the multispectral edge enhancement network (MSEEN), enhance the edge features of the power transmission equipment target, and adaptively adjust the processing parameters in combination with the attitude stability score (SSI) to output the target image with enhanced edge features and the initial localization box. The 3D reconstruction and precise positioning module is used to perform a visual SLAM process based on the improved ORB feature operator and Local Relevance Point Set Matching (LSCC) algorithm using the target image sequence with enhanced edge features, and combined with power Beidou positioning data to perform real-time 3D spatial reconstruction of the inspection scene and centimeter-level precise positioning of inspection targets such as poles and insulators. The path planning and autonomous obstacle avoidance module is used to calculate the best shooting waypoint position in real time based on the positioning results of the 3D reconstruction and precise positioning module, the preset 3D model of the tower and the inspection requirements, and the improved ant colony algorithm, and generate the global optimal flight path. At the same time, it integrates airborne sensor data to perform real-time local obstacle avoidance. The intelligent diagnosis and defect recognition module is used to perform preliminary target detection on the located target image using a lightweight improved YOLOv5-GhostNet model deployed on an embedded platform, and further combine infrared thermal imaging data to distinguish the equipment body from the attachments through a thermal texture dual-stream recognition model, and perform intelligent diagnosis of defects such as insulator damage and hardware corrosion. The collaborative processing module is used to execute front-end lightweight algorithms on the flight control and data preprocessing sub-board and the onboard DSP / GPU heterogeneous computing unit, including image stabilization, edge enhancement and preliminary detection. At the same time, it synchronizes high-dimensional data and processing results to the cloud through the 5G power grid, and the cloud performs high-precision 3D modeling, massive data deep analysis and model iterative training. The secure communication and task management module is used to ensure the security of all data transmission and control commands through the power grid intranet communication protocol, and to manage the entire process from task issuance, autonomous route execution, data acquisition, real-time processing to report generation.

[0007] Furthermore, the specific workflow of the data acquisition and dynamic stabilization module is as follows: First, three-dimensional acceleration, three-dimensional angular velocity, wind speed, and illumination data are acquired in real time through an embedded meteorological sensor unit and the UAV flight control system. Second, the original multispectral image sequence is timestamped and aligned with the inertial measurement unit data. Then, a tightly coupled visual-inertial odometry calculation method is adopted, which estimates the precise pose change of the UAV under high-speed motion through joint optimization of feature point tracking and IMU pre-integration. Next, a Kalman filter-based prediction and correction framework is used to perform reverse motion compensation on each frame of the image according to the pose change to eliminate blur and jitter. Finally, the sharpness and feature point stability of the compensated image sequence are calculated, and a quantitative attitude stability score (SSI) is output to indicate the reliability of image quality and the benchmark for parameter adjustment of subsequent modules.

[0008] Furthermore, the specific workflow of the multispectral polarization imaging and fusion module is as follows: First, based on the real-time solar azimuth and elevation angle information collected by the embedded meteorological sensor unit, and the preliminary target surface normal direction obtained from the 3D reconstruction and precise positioning module, the theoretically strongest specular reflection direction is calculated. Second, the electrically controlled rotating polarizer is rotated to an angle perpendicular to the polarization direction of the strongest reflection direction to suppress glare to the greatest extent. Visible and infrared band images are then acquired sequentially at four typical polarization directions: 0 degrees, 45 degrees, 90 degrees, and 135 degrees. Then, the Stokes vector method is used to calculate the polarization degree and polarization angle of the target images for different polarization directions within the same band. Finally, the visible light intensity image, infrared thermal radiation image, and polarization feature image are adaptively weighted and fused at the pixel level. The weights are dynamically adjusted based on the attitude stability score (SSI) and regional contrast to generate an enhanced multispectral fusion image that combines intensity information, thermal information, and polarization-suppressed reflection information for subsequent processing.

[0009] Furthermore, the specific workflow of the edge enhancement and feature extraction module is as follows: First, the enhanced multispectral fusion image is input into a multispectral edge enhancement network (MSEEN), which includes a shared feature encoder and multiple parallel decoding branches targeting visible light, infrared, and polarization features. These branches interact with each other through an attention mechanism. Next, the network outputs a comprehensive edge saliency map, which exhibits high response at small structures such as insulator string contours and broken conductor strands. Then, the feature map of the network's decoding layer is gated using the attitude stability score (SSI). When the SSI is low, the extraction of low-frequency contour features is enhanced; when the SSI is high, the extraction of high-frequency detail features is emphasized. Next, an improved connected component analysis algorithm is run on the generated edge saliency map. Combined with prior size constraints, potential target regions are initially determined, and initial bounding boxes are generated. Finally, this edge saliency map is concatenated with the original fusion image along the channel dimension to form a feature image rich in edge information, which is then passed to the subsequent recognition and localization modules.

[0010] Furthermore, the specific workflow of the 3D reconstruction and precise positioning module is as follows: First, robust feature points are extracted from the feature image rich in edge information using an improved FAST corner detection and ORB descriptor algorithm. Second, the Local Correlation-Based Set Matching (LSCC) algorithm is used to match feature points between consecutive frames, effectively eliminating mismatched points caused by repetitive textures or similar structures. Then, based on the RANSAC and Radial Consistency (RAC) algorithms, the precise relative pose between frames is calculated using matched point pairs and UAV IMU data, and a local image area network is constructed. Next, combined with the absolute position information provided by the power grid Beidou positioning system, the local area network is incorporated into the global coordinate system, and a dense point cloud is generated using the SGM algorithm to achieve real-time 3D mesh model reconstruction of the power transmission corridor scene. Finally, in the reconstructed 3D model, the initial positioning box is back-projected into 3D space, and the precise coordinates and poses of each insulator, equipotential ring, and hanging hardware in 3D space are accurately located by performing iterative nearest point (ICP) registration with a preset tower 3D model library.

[0011] Furthermore, the specific workflow of the path planning and autonomous obstacle avoidance module is as follows: First, based on the precise 3D coordinates of the inspection target, the target type, and the preset shooting specifications, the optimal observation waypoint for each target is automatically calculated. The shooting specifications include shooting distance, azimuth angle, and gimbal pitch angle. Second, a 3D path planning map is constructed using all waypoints to be photographed as nodes. The cost between nodes integrates flight distance, time, energy consumption, and the smoothness of shooting angle switching. Then, an ant colony algorithm is used for global path optimization. In the algorithm, the initial pheromone distribution is set according to the importance of waypoints, the heuristic factor considers the angle between the current attitude of the UAV and the direction of the next waypoint, and the pheromone update strategy integrates the globally optimal path and iterative optimization information. Next, during flight along the planned path, data from the airborne millimeter-wave radar or visual depth sensor is read in real time. If an unmodeled obstacle is detected on the planned path, local replanning is triggered, and a local obstacle avoidance trajectory is generated using the Dynamic Window Method (DWA) combined with real-time 3D point cloud. Finally, the final optimized 3D flight path and the shooting parameters for each waypoint are sent to the UAV flight control system and gimbal control system to achieve fully autonomous flight inspection.

[0012] Furthermore, the specific workflow of the intelligent diagnosis and defect identification module is as follows: First, on the embedded side, a lightweight, improved YOLOv5-GhostNet model is used to quickly detect and classify the target image within the initial positioning box, identifying equipment types such as insulators, vibration dampers, and wire clamps, and obtaining their two-dimensional bounding boxes. Second, for the identified equipment targets, their corresponding infrared thermal imaging areas are cropped, and the average temperature, maximum temperature, and temperature distribution variance of the area are calculated. Then, a thermal texture dual-stream recognition network is constructed, where the texture stream takes a visible light cropped image as input, and the thermal stream takes a normalized temperature distribution map as input, and the dual-stream features are stitched together in a fusion layer. Next, the network is trained to distinguish between normal equipment, defective equipment, and non-equipment attachments, where the non-equipment attachment category specifically refers to interference objects such as bird nests and vines that have similar textures to equipment but different thermal radiation characteristics. Finally, for targets identified as defective equipment, a more refined classifier is called to determine the defect type, such as insulator spontaneous explosion, hardware overheating, and conductor strand breakage, and the diagnostic results, location coordinates, and evidence images are packaged into structured data.

[0013] Furthermore, the specific workflow of the collaborative processing module is as follows: First, on the UAV side, the flight control and data preprocessing sub-board and its onboard DSP / GPU heterogeneous computing unit are responsible for running all the algorithms of the data acquisition and dynamic stabilization module, the multispectral polarization imaging and fusion module, the edge enhancement and feature extraction module, and the lightweight YOLOv5-GhostNet detection steps in the intelligent diagnosis and defect recognition module, performing real-time front-end processing and preliminary diagnosis. Second, through the integrated 5G intelligent processing unit, the intermediate data and results generated by the processing, including the stabilized image sequence, 3D point cloud fragments, target localization and diagnostic results, are encrypted and transmitted to the cloud secure data pool through a dedicated power APN access point. Then, on the high-performance computing cluster in the cloud, high-computational-load tasks are executed, including globally consistent 3D scene fine reconstruction, comparative analysis and trend prediction of massive historical inspection data, and incremental training and optimization of complex AI models such as the thermal texture dual-stream recognition model. Finally, the cloud sends the updated model parameters and optimized flight path strategies to the front-line UAV to complete the iteration of algorithms and knowledge.

[0014] Furthermore, the specific workflow of the secure communication and task management module is as follows: First, the task management subsystem receives inspection work orders from the power production management system. These work orders include the line segment, tower number, inspection type, and level. Second, based on the work order, it automatically retrieves the corresponding line's detailed 3D model, historical inspection data, and optimized flight strategy from the cloud to generate an executable inspection task plan, which is then distributed to the designated drone. Next, during task execution, all control commands, status information, and collected data streams between the drone and the ground station, and between the ground station and the cloud, are encrypted and verified end-to-end using a power intranet communication protocol based on digital certificates and national cryptographic algorithms, ensuring the security of operational data. Then, the system monitors the task execution status, equipment health status, and data processing progress in real time, providing a visual monitoring interface. Finally, after the task is completed, the system automatically summarizes the processing results from each module, generating a standard inspection report including defect reports, 3D model updates, and statistical analysis charts, which is then sent back to the production management system.

[0015] (III) Beneficial Effects This invention provides an image acquisition and recognition device and system for power line inspection drones. It has the following beneficial effects: (I) This power line inspection UAV image acquisition and recognition device and system, by integrating a data acquisition and dynamic stabilization module, a multispectral polarization imaging and fusion module, and coordinating with an adaptive gimbal support and a multispectral polarization imaging device, solves the fundamental problem of unstable image quality under harsh environments such as strong winds and complex lighting conditions. Specifically, the system uses a visual inertial odometry and Kalman filtering to perform real-time dynamic compensation for images under high-speed motion, effectively suppressing motion blur and jitter caused by strong winds, ensuring the clarity and usability of the original data. On this basis, it further introduces an electronically controlled polarizer array and multispectral (visible light, infrared, ultraviolet) imaging technology. It can adaptively adjust the polarization direction according to the real-time ambient light and target surface characteristics to suppress specular reflection and glare interference on snow, water, or wet insulator surfaces to the greatest extent. It generates enhanced images through multi-band information fusion, which not only directly improves the quality of a single image, but also provides a reliable quality assessment benchmark for all subsequent processing stages by outputting a quantified attitude stability score (SSI). This fundamentally ensures the quality of input data for subsequent feature extraction, matching, and 3D reconstruction algorithms, overcoming the pain point of traditional UAV inspections where images are unusable under adverse weather conditions and repeated inspections are required. It significantly improves the first-time success rate and economy of inspection operations.

[0016] (II) This power line inspection UAV image acquisition and recognition device and system, through deep fusion of edge enhancement and feature extraction modules, 3D reconstruction and precise positioning modules, and path planning and autonomous obstacle avoidance modules, constructs a closed loop from 2D image understanding to 3D spatial perception and decision-making. This achieves fully autonomous, high-precision intelligent inspection in complex natural environments. The system employs a multispectral edge enhancement network (MSEEN) and a local correlation-based point set matching (LSCC) algorithm to achieve robust edge enhancement and high-precision feature matching for small targets such as insulators and hardware under complex backgrounds and varying lighting conditions, providing robust edge enhancement and high-precision feature matching for 3D reconstruction. The system utilizes high-quality data, combined with improved visual SLAM technology and high-precision BeiDou positioning for power transmission corridors, to achieve real-time centimeter-level 3D reconstruction and precise spatial positioning of inspection targets. Crucially, based on this accurate 3D spatial information, the system employs real-time 3D path planning technology based on an improved ant colony algorithm to automatically calculate the optimal shooting waypoint and plan the globally optimal flight path. It also integrates real-time obstacle avoidance capabilities. This series of technologies enables full-process automation, greatly reducing the degree of human intervention and solving the core problems of high human involvement, poor route reusability, and low efficiency of autonomous inspection in existing technologies.

[0017] (III) This power inspection UAV image acquisition and recognition device and system, by deploying a lightweight intelligent diagnostic model on the front-end embedded platform and introducing a thermal texture dual-stream recognition mechanism, constructs an edge-cloud collaborative intelligent diagnostic system, significantly improving the accuracy and practicality of defect recognition while ensuring low power consumption and high performance. Addressing the computing power and power consumption limitations of the UAV platform, the system employs a compressed and optimized lightweight improved YOLOv5-GhostNet model for rapid front-end target detection. Furthermore, to overcome the false alarm problem caused by bird nests, vines, and other attachments, a thermal texture dual-stream recognition model is proposed, fully utilizing the infrared thermal radiation characteristics of power equipment and biological attachments. By leveraging the differences in image quality, the system effectively filters out interference and accurately diagnoses real defects (such as insulator damage and overheating of fittings). Simultaneously, the system utilizes a 5G power intranet to form an edge-cloud collaborative architecture. The front end performs real-time processing and preliminary diagnosis, while the cloud performs complex model training and big data analysis, enabling algorithm iteration. This not only raises the average detection accuracy (mAP) of defects such as insulators, broken strings, and damage to a higher level, but also significantly reduces false alarms and missed detections caused by image quality or interference from attachments in practical applications. This allows the massive amounts of image data collected by drones to be efficiently and accurately transformed into defect reports and decision-making suggestions that can directly guide operation and maintenance, truly realizing the intelligent and lean management of transmission line inspection. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall structure of the present invention; Figure 2This is a schematic diagram of the system flow of the present invention; Figure 3 This is a schematic diagram of the technical route for constructing a local area network for local images according to the present invention; Figure 4 This is a schematic diagram of the real-time three-dimensional path planning technology for UAVs according to the present invention; Figure 5 This is a schematic diagram comparing the LSCC matching algorithm of this invention with existing algorithms; Figure 6 This is a schematic diagram comparing the YOLOv5-GhostNet detection algorithm of this invention with mainstream algorithms; Figure 7 This is a schematic diagram of the improved YOLOv5-GhostNet network structure of this invention.

[0019] In the figure: 1. Adaptive gimbal support; 2. Multispectral polarization imaging module. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] First embodiment: as follows Figures 1 to 7 As shown, the present invention provides a technical solution: an image acquisition and recognition device for power line inspection drones, comprising: Adaptive gimbal bracket 1 is fixedly connected to the front end of the UAV. It integrates an embedded meteorological sensor unit, which includes a miniature anemometer, a humidity sensor and a light intensity sensor. The adaptive gimbal bracket 1 also integrates a flight control and data preprocessing sub-board. The multispectral polarization imaging module 2 includes a camera that integrates visible light, infrared thermal imaging and ultraviolet light. The multispectral polarization imaging module 2 is fixedly installed at the bottom of the adaptive gimbal bracket 1. The multispectral polarization imaging module 2 is independently configured with an electrically controlled rotating polarizer, which is driven by a micro stepper motor. The adaptive gimbal bracket 1 has a built-in flight control and data preprocessing sub-board that is electrically connected to the multispectral polarization imaging module 2 via a cable. This sub-board is used to collect sensor data and control the movement of the electrically controlled rotating polarizer.

[0022] A power line inspection drone image acquisition and recognition system includes: The data acquisition and dynamic stabilization module is used to acquire the UAV's flight attitude, environmental meteorological data and raw multispectral image sequences in real time, and perform dynamic image correction based on visual inertial odometry and Kalman filtering to eliminate motion blur and geometric distortion caused by disturbances such as strong winds, and obtain the stabilized image sequence and attitude stability score (SSI). The multispectral polarization imaging and fusion module is used to receive the stabilized image sequence, control the electrically controlled rotating polarizer to adjust the polarization direction according to the real-time illumination angle and the material characteristics of the target surface, acquire multi-band images under different polarization states, and fuse multispectral information and polarization information to suppress specular reflection and scattered light interference. The edge enhancement and feature extraction module is used to process the image after fusion by the multispectral polarization imaging and fusion module based on the multispectral edge enhancement network (MSEEN), enhance the edge features of the power transmission equipment target, and adaptively adjust the processing parameters in combination with attitude stability score (SSI) to output the target image with enhanced edge features and the initial localization box. The 3D reconstruction and precise positioning module is used to perform a visual SLAM process based on the improved ORB feature operator and the Local Relevance Point Set Matching (LSCC) algorithm using target image sequences with enhanced edge features. Combined with power grid Beidou positioning data, it performs real-time 3D spatial reconstruction of inspection scenarios and centimeter-level precise positioning of inspection targets such as poles and insulators. The path planning and autonomous obstacle avoidance module is used to calculate the best shooting waypoint position in real time based on the positioning results of the 3D reconstruction and precise positioning module, the preset 3D model of the tower and the inspection requirements, and the improved ant colony algorithm, and generate the global optimal flight path. At the same time, it integrates airborne sensor data to perform real-time local obstacle avoidance. The intelligent diagnosis and defect recognition module is used to perform preliminary target detection on the located target image using a lightweight improved YOLOv5-GhostNet model deployed on an embedded platform, and further combine infrared thermal imaging data to distinguish the equipment body from the attachments through a thermal texture dual-stream recognition model, and perform intelligent diagnosis of defects such as insulator damage and hardware corrosion. The collaborative processing module is used to execute front-end lightweight algorithms on the flight control and data preprocessing sub-board and the onboard DSP / GPU heterogeneous computing unit, including image stabilization, edge enhancement and preliminary detection. At the same time, it synchronizes high-dimensional data and processing results to the cloud through the 5G power grid, and the cloud performs high-precision 3D modeling, massive data deep analysis and model iterative training. The secure communication and task management module is used to ensure the security of all data transmission and control commands through the power grid intranet communication protocol, and to manage the entire process from task issuance, autonomous route execution, data acquisition, real-time processing to report generation.

[0023] The specific workflow of the data acquisition and dynamic stabilization module is as follows: First, three-dimensional acceleration, three-dimensional angular velocity, wind speed, and illumination data are acquired in real time through an embedded meteorological sensor unit and the UAV flight control system. Second, the original multispectral image sequence is timestamped and aligned with the inertial measurement unit data. Then, a tightly coupled visual-inertial odometry calculation method is adopted, which estimates the precise pose changes of the UAV under high-speed motion through joint optimization of feature point tracking and IMU pre-integration. Next, a Kalman filter-based prediction and correction framework is used to perform reverse motion compensation on each frame of the image according to the pose changes to eliminate blur and jitter. Finally, the sharpness and feature point stability of the compensated image sequence are calculated, and a quantitative attitude stability score (SSI) is output to indicate the reliability of image quality and the benchmark for parameter adjustment of subsequent modules.

[0024] Specifically, when a drone equipped with an image acquisition device is inspecting high-altitude, windy areas, its data acquisition and dynamic stabilization module begins operation. It uses a miniature anemometer, gyroscope, and accelerometer built into the gimbal to collect real-time raw data on environmental wind speed and the drone's attitude. Simultaneously, it acquires raw image sequences from visible light, infrared, and ultraviolet cameras. Next, a tightly coupled visual-inertial odometry (VIO) method is employed to synchronize and jointly optimize image feature points with inertial measurement unit (IMU) data, accurately estimating the high-frequency attitude changes of the drone caused by strong wind disturbances. Then, using a Kalman filter-based predictive correction framework, pixel-level motion compensation and geometric correction are performed on each frame of the image based on the estimated pose changes, effectively eliminating motion blur and jitter, outputting a stable and clear image sequence, and calculating a quantified attitude stability score (SSI). Subsequently, the multispectral polarization imaging and fusion module is activated. Based on the real-time solar azimuth information obtained from meteorological sensors and the preliminary identification of target surface geometry, it calculates the directions that are prone to strong specular reflection (such as snow, ice, or wet insulator surfaces) and drives the electrically controlled rotating polarizer to rotate to the optimal angle to suppress glare. It then sequentially acquires multi-band images under multiple polarization directions. Finally, it uses the Stokes vector method to solve the polarization features and adaptively weights and fuses the intensity image, thermal radiation image, and polarization feature image to generate an enhanced multispectral fusion image that can effectively resist strong light interference. The beneficial effect of this embodiment is that by combining dynamic image stabilization and adaptive polarization imaging technology, the quality of images acquired under complex meteorological conditions is fundamentally guaranteed, providing a high-precision and high-reliability data foundation for subsequent feature extraction and 3D reconstruction.

[0025] Second embodiment: as follows Figures 1 to 7 As shown, the specific workflow of the multispectral polarization imaging and fusion module is as follows: First, based on the real-time solar azimuth and elevation angle information collected by the embedded meteorological sensor unit, and the preliminary target surface normal direction obtained from the 3D reconstruction and precise positioning module, the theoretically strongest direction of specular reflection is calculated. Second, the electrically controlled rotating polarizer is rotated to an angle perpendicular to the polarization direction of the strongest reflection direction to suppress glare to the greatest extent. Visible and infrared band images are then collected sequentially at four typical polarization directions: 0 degrees, 45 degrees, 90 degrees, and 135 degrees. Then, the Stokes vector method is used to calculate the polarization degree and polarization angle of the target images for images with different polarization directions in the same band. Finally, the visible light intensity image, infrared thermal radiation image, and polarization feature image are adaptively weighted and fused at the pixel level. The weights are dynamically adjusted according to the attitude stability score (SSI) and regional contrast to generate an enhanced multispectral fusion image that combines intensity information, thermal information, and polarization suppression reflection information for subsequent processing.

[0026] The specific workflow of the edge enhancement and feature extraction module is as follows: First, the enhanced multispectral fusion image is input into a multispectral edge enhancement network (MSEEN), which includes a shared feature encoder and multiple parallel decoding branches targeting visible light, infrared, and polarization features. These branches interact with each other through an attention mechanism. Second, the network outputs a comprehensive edge saliency map, which exhibits high response at small structures such as insulator string contours and broken conductor strands. Then, the feature map of the network's decoding layer is gated using a pose stability score (SSI). A lower SSI enhances the extraction of low-frequency contour features, while a higher SSI emphasizes the extraction of high-frequency detail features. Next, an improved connected component analysis algorithm is run on the generated edge saliency map, combined with prior size constraints, to initially determine potential target regions and generate initial bounding boxes. Finally, this edge saliency map is concatenated with the original fusion image along the channel dimension to form a feature image rich in edge information, which is then passed to the subsequent recognition and localization modules.

[0027] The specific workflow of the 3D reconstruction and precise positioning module is as follows: First, robust feature points are extracted from feature images rich in edge information using an improved FAST corner detection and ORB descriptor algorithm. Second, the Local Correlation-Based Set Matching (LSCC) algorithm is used to match feature points between consecutive frames, effectively eliminating mismatched points caused by repetitive textures or similar structures. Then, based on the RANSAC and Radial Consistency (RAC) algorithms, the precise relative pose between frames is calculated using matched point pairs and UAV IMU data, and a local image area network is constructed. Next, combined with the absolute position information provided by the power grid BeiDou positioning system, the local area network is incorporated into the global coordinate system, and the SGM algorithm is used to generate a dense point cloud, realizing the real-time 3D mesh model reconstruction of the power transmission corridor scene. Finally, in the reconstructed 3D model, the initial positioning box is back-projected into 3D space, and the precise coordinates and poses of each insulator, equipotential ring, and hanging hardware in 3D space are accurately located by performing iterative nearest point (ICP) registration with a preset tower 3D model library.

[0028] The specific workflow of the path planning and autonomous obstacle avoidance module is as follows: First, based on the precise 3D coordinates of the inspection target, the target type, and the preset shooting specifications, the optimal observation waypoint for each target is automatically calculated. The shooting specifications include shooting distance, azimuth angle, and gimbal pitch angle. Second, a 3D path planning map is constructed using all waypoints to be photographed as nodes. The cost between nodes integrates flight distance, time, energy consumption, and the smoothness of shooting angle switching. Then, an ant colony algorithm is used for global path optimization. In the algorithm, the initial pheromone distribution is set according to the importance of waypoints, the heuristic factor considers the angle between the current attitude of the UAV and the direction of the next waypoint, and the pheromone update strategy integrates the globally optimal path and iterative optimization information. Next, during flight along the planned path, data from the airborne millimeter-wave radar or visual depth sensor is read in real time. If an unmodeled obstacle is detected on the planned path, local replanning is triggered, and a local obstacle avoidance trajectory is generated using the dynamic window method (DWA) combined with real-time 3D point cloud. Finally, the final optimized 3D flight path and the shooting parameters for each waypoint are sent to the UAV flight control system and gimbal control system to achieve fully autonomous flight inspection.

[0029] Specifically, the edge enhancement and feature extraction module first receives the enhanced multispectral fusion image and inputs it into a specially designed multispectral edge enhancement network (MSEEN). This network, through a shared encoder and multi-branch decoder structure, fuses visible light texture, infrared thermal profile, and polarization features to generate a comprehensive edge saliency map. This map particularly enhances small and critical target features such as insulator string edges and wires. Referring to the attitude stability score (SSI) output in Example 1, the network's focus is dynamically adjusted. When the SSI is low (large sway), the emphasis is on extracting robust overall contours. When the SSI is high, the focus is on fine details. Then, based on this saliency map, connected component analysis and preliminary localization are performed to generate the initial 2D bounding box of the target. Next, the 3D reconstruction and precise localization module begins its work. It extracts feature points from the feature image rich in edge information using FAST corner detection and ORB descriptors, and utilizes the Local Correlation-Based Set Matching (LSCC) algorithm for efficient and accurate cross-frame feature matching, effectively overcoming the interference of repetitive textures in the transmission line scene. Subsequently, combining the matched point pairs with IMU data, the precise localization is calculated using an algorithm based on Radial Consistency (RAC). The system determines the pose and integrates power grid BeiDou positioning information to construct a local image area network. Then, the SGM algorithm is used to generate a real-time dense 3D point cloud of the inspection area, achieving digital reconstruction of the scene. Finally, the preliminary 2D positioning frame is back-projected into this 3D space and registered with a pre-set refined 3D model of the tower. This enables centimeter-level precise positioning and attitude measurement of each insulator, equipotential ring, and other inspection target. Based on this, the path planning and autonomous obstacle avoidance module is activated. It automatically calculates the pose of each target based on the precisely located 3D coordinates, target type, and preset shooting specifications. The optimal observation waypoint for the target is identified, and then, using all waypoints as nodes, a real-time 3D path planning technology based on an improved ant colony algorithm is employed. By setting pheromones, heuristic factors, and dynamic update strategies, a globally optimal flight path is planned in 3D space. During flight, local obstacle avoidance is performed by combining real-time perception data. The beneficial effect of this embodiment is that, through a coherent technology chain from image enhancement and feature matching to 3D reconstruction and path planning, accurate spatial perception of inspection targets and intelligent planning of autonomous operation paths are achieved in complex environments, providing core technical support for efficient and safe unmanned inspection.

[0030] Third embodiment: as follows Figures 1 to 7 As shown, the specific workflow of the intelligent diagnosis and defect identification module is as follows: First, on the embedded side, a lightweight, improved YOLOv5-GhostNet model is used to quickly detect and classify target images within the initial localization box, identifying equipment types such as insulators, vibration dampers, and wire clamps, and obtaining their two-dimensional bounding boxes. Second, for the identified equipment targets, their corresponding infrared thermal imaging regions are cropped, and the average temperature, maximum temperature, and temperature distribution variance of the region are calculated. Then, a thermal texture dual-stream recognition network is constructed, where the texture stream takes a visible light cropped image as input, and the thermal stream takes a normalized temperature distribution map as input, with the dual-stream features stitched together in a fusion layer. Next, this network is trained to distinguish between normal equipment, defective equipment, and non-equipment attachments, where the non-equipment attachment category specifically refers to interference objects such as bird nests and vines that have similar textures to equipment but different thermal radiation characteristics. Finally, for targets identified as defective equipment, a more refined classifier is called to determine the defect type, such as insulator spontaneous explosion, hardware overheating, and conductor strand breakage, and the diagnostic results, location coordinates, and evidence images are packaged into structured data.

[0031] The specific workflow of the collaborative processing module is as follows: First, on the drone side, the flight control and data preprocessing sub-board and its onboard DSP / GPU heterogeneous computing unit are responsible for running all algorithms of the data acquisition and dynamic stabilization module, multispectral polarization imaging and fusion module, edge enhancement and feature extraction module, as well as the lightweight YOLOv5-GhostNet detection steps in the intelligent diagnosis and defect recognition module, performing front-end real-time processing and preliminary diagnosis. Second, through the integrated 5G intelligent processing unit, the intermediate data and results generated by the processing, including stabilized image sequences, 3D point cloud fragments, target localization and diagnostic results, are encrypted and transmitted to the cloud secure data pool through a dedicated power APN access point. Then, on the high-performance computing cluster in the cloud, high-computational-load tasks are executed, including globally consistent 3D scene fine reconstruction, comparative analysis and trend prediction of massive historical inspection data, and incremental training and optimization of complex AI models such as thermal texture dual-stream recognition models. Finally, the cloud sends the updated model parameters and optimized flight path strategies to the front-line drones to complete the iteration of algorithms and knowledge.

[0032] The specific workflow of the secure communication and task management module is as follows: First, the task management subsystem receives inspection work orders from the power production management system. These work orders include the line segment, tower number, inspection type, and level. Second, based on the work order, it automatically retrieves the corresponding line's detailed 3D model, historical inspection data, and optimized flight strategy from the cloud to generate an executable inspection task plan, which is then distributed to the designated drone. Next, during task execution, all control commands, status information, and collected data streams between the drone and the ground station, and between the ground station and the cloud, are encrypted and verified end-to-end using a power intranet communication protocol based on digital certificates and national cryptographic algorithms, ensuring the security of operational data. Then, the system monitors the task execution status, equipment health status, and data processing progress in real time, providing a visual monitoring interface. Finally, after the task is completed, the system automatically summarizes the processing results from each module, generating a standard inspection report including defect reports, 3D model updates, and statistical analysis charts, which is then sent back to the production management system.

[0033] Specifically, the intelligent diagnosis and defect identification module is activated first. Its embedded front-end uses a lightweight, improved YOLOv5-GhostNet model to quickly identify and classify the target area image after localization, initially selecting devices such as insulators and vibration dampers. Then, the module captures infrared thermal imaging data of the corresponding area and calculates its temperature distribution statistical characteristics. Next, the core thermal texture dual-flow recognition model starts working. The texture flow of this model takes a visible light cropped image as input and learns the shape and texture features of the device; the heat flow takes a normalized temperature distribution map as input and learns the thermal radiation patterns of the device in normal and defective states. The dual-flow features interact in the fusion layer, especially for bird nests (textured but with temperatures close to ambient temperature) and vines (textured and with a different texture than metal). The system models the temperature difference and texture differences between common attachments (such as heat capacity) and real power equipment, thereby effectively distinguishing normal equipment, defective equipment, and non-equipment attachments. Finally, for equipment identified as defective, such as insulators, the system further analyzes cracks and damaged textures in visible light images and abnormal local overheating or undercooling areas in infrared images to accurately diagnose specific defect types such as spontaneous explosion, contamination, and zero values. All diagnostic results, spatial coordinates, and evidence images are associated and packaged. The beneficial effect of this embodiment is that by introducing a thermal texture dual-flow analysis mechanism, it utilizes the essential differences in thermal radiation characteristics between the equipment body and biological attachments, significantly reducing false alarms caused by interference from bird nests, vines, and other objects, and improving the accuracy and practicality of intelligent defect diagnosis.

[0034] It should 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, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0035] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An image acquisition and recognition device for power line inspection drones, characterized in that: include: An adaptive gimbal bracket (1) is fixedly connected to the front end of the UAV. An embedded meteorological sensor unit is integrated inside the bracket. The embedded meteorological sensor unit includes a miniature anemometer, a humidity sensor and a light intensity sensor. The adaptive gimbal bracket (1) also integrates a flight control and data preprocessing sub-board. The multispectral polarization imaging module (2) includes a camera that integrates visible light, infrared thermal imaging and ultraviolet light. The multispectral polarization imaging module (2) is fixedly installed at the bottom of the adaptive gimbal bracket (1). The multispectral polarization imaging module (2) is independently configured with an electrically controlled rotating polarizer, which is driven by a micro stepper motor. The flight control and data preprocessing sub-board built into the adaptive gimbal bracket (1) is electrically connected to the multispectral polarization imaging module (2) via a cable, and is used to collect sensor data and control the movement of the electronically controlled rotating polarizer.

2. A power line inspection drone image acquisition and recognition system, comprising the power line inspection drone image acquisition and recognition device as described in claim 1, characterized in that: include: The data acquisition and dynamic stabilization module is used to acquire the UAV's flight attitude, environmental meteorological data and raw multispectral image sequences in real time, and perform dynamic image correction based on visual inertial odometry and Kalman filtering to obtain image sequences and attitude stability score (SSI). The multispectral polarization imaging and fusion module is used to receive the stabilized image sequence, control the electrically controlled rotating polarizer to adjust the polarization direction according to the real-time illumination angle and the material characteristics of the target surface, acquire multi-band images under different polarization states, and fuse multispectral information and polarization information. The edge enhancement and feature extraction module is used to process the image after fusion by the multispectral polarization imaging and fusion module based on the multispectral edge enhancement network (MSEEN), and adaptively adjust the processing parameters in combination with the attitude stability score (SSI) to output a target image with enhanced edge features and an initial localization box. The 3D reconstruction and precise positioning module is used to perform a visual SLAM process based on the improved ORB feature operator and Local Correlation Set Matching (LSCC) algorithm using the target image sequence with enhanced edge features, and combined with power Beidou positioning data to perform real-time 3D spatial reconstruction of the inspection scene and positioning of inspection targets such as poles and insulators. The path planning and autonomous obstacle avoidance module is used to calculate the best shooting waypoint position in real time based on the positioning results of the 3D reconstruction and precise positioning module, the preset 3D model of the tower and the inspection requirements, and the improved ant colony algorithm, and generate the global optimal flight path. At the same time, it integrates airborne sensor data to perform real-time local obstacle avoidance. The intelligent diagnosis and defect recognition module is used to perform preliminary target detection on the located target image using a lightweight improved YOLOv5-GhostNet model deployed on an embedded platform, and further combine infrared thermal imaging data to distinguish the equipment body from the attachments through a thermal texture dual-stream recognition model, and perform intelligent diagnosis of defects such as insulator damage and hardware corrosion. The collaborative processing module is used to execute front-end lightweight algorithms on the flight control and data preprocessing sub-board and the onboard DSP / GPU heterogeneous computing unit, including image stabilization, edge enhancement and preliminary detection. At the same time, it synchronizes high-dimensional data and processing results to the cloud through the 5G power grid, and the cloud performs high-precision 3D modeling, massive data deep analysis and model iterative training. The secure communication and task management module is used to ensure the security of all data transmission and control commands through the power grid intranet communication protocol, and to manage the entire process from task issuance, autonomous route execution, data acquisition, real-time processing to report generation.

3. The image acquisition and recognition device and system for power line inspection drones according to claim 2, characterized in that: The specific workflow of the data acquisition and dynamic stabilization module is as follows: First, three-dimensional acceleration, three-dimensional angular velocity, wind speed, and illumination data are acquired in real time through an embedded meteorological sensor unit and the UAV flight control system. Second, the original multispectral image sequence is timestamped and aligned with the inertial measurement unit data. Then, a tightly coupled visual-inertial odometry method is used, through joint optimization of feature point tracking and IMU pre-integration, to estimate the precise pose change of the UAV under high-speed motion. Next, a Kalman filter-based prediction and correction framework is used to perform reverse motion compensation on each frame of the image based on the pose change. Finally, the sharpness and feature point stability of the compensated image sequence are calculated, and a quantitative attitude stability score (SSI) is output.

4. The image acquisition and recognition device and system for power line inspection drones according to claim 3, characterized in that: The specific workflow of the multispectral polarization imaging and fusion module is as follows: First, based on the real-time solar azimuth and elevation angle information collected by the embedded meteorological sensor unit, and the preliminary target surface normal direction obtained from the 3D reconstruction and precise positioning module, the theoretically strongest specular reflection direction is calculated. Second, the electrically controlled rotating polarizer is rotated to an angle perpendicular to the polarization direction of the strongest reflection direction to suppress glare to the greatest extent. Visible and infrared band images are then acquired sequentially at four typical polarization directions: 0 degrees, 45 degrees, 90 degrees, and 135 degrees. Then, the Stokes vector method is used to calculate the polarization degree and polarization angle of the target image from images of different polarization directions within the same band. Finally, the visible light intensity image, infrared thermal radiation image, and polarization feature image are adaptively weighted and fused at the pixel level. The weights are dynamically adjusted based on the attitude stability score (SSI) and regional contrast to generate an enhanced multispectral fusion image that combines intensity information, thermal information, and polarization-suppressed reflection information.

5. The image acquisition and recognition device and system for power line inspection drones according to claim 4, characterized in that: The specific workflow of the edge enhancement and feature extraction module is as follows: First, the enhanced multispectral fused image is input into a multispectral edge enhancement network (MSEEN), which includes a shared feature encoder and multiple parallel decoding branches targeting visible light, infrared, and polarization features. These branches interact with each other through an attention mechanism. Next, the network outputs a comprehensive edge saliency map, which exhibits high response at small structures such as insulator string contours and broken conductor strands. Then, the feature map of the network's decoding layer is gated using the attitude stability score (SSI). When the SSI is low, the extraction of low-frequency contour features is enhanced; when the SSI is high, the extraction of high-frequency detail features is emphasized. Next, an improved connected component analysis algorithm is run on the generated edge saliency map. Combined with prior size constraints, potential target regions are initially determined, and initial bounding boxes are generated. Finally, this edge saliency map is concatenated with the original fused image along the channel dimension to form a feature image rich in edge information.

6. The image acquisition and recognition device and system for power line inspection drones according to claim 5, characterized in that: The specific workflow of the 3D reconstruction and precise positioning module is as follows: First, robust feature points are extracted from the feature image rich in edge information using an improved FAST corner detection and ORB descriptor algorithm. Second, the Local Correlation-Based Set Matching (LSCC) algorithm is used to match feature points between consecutive frames, effectively eliminating mismatched points caused by repetitive textures or similar structures. Then, based on the RANSAC and Radial Consistency (RAC) algorithms, the precise relative pose between frames is calculated using matched point pairs and UAV IMU data, and a local image area network is constructed. Next, combined with the absolute position information provided by the power grid Beidou positioning system, the local area network is incorporated into the global coordinate system, and a dense point cloud is generated using the SGM algorithm to achieve real-time 3D mesh model reconstruction of the power transmission corridor scene. Finally, in the reconstructed 3D model, the initial positioning box is back-projected into 3D space, and the precise coordinates and poses of each insulator, equipotential ring, and hanging hardware in 3D space are accurately located by performing iterative nearest point (ICP) registration with a preset tower 3D model library.

7. The image acquisition and recognition device and system for power line inspection drones according to claim 6, characterized in that: The specific workflow of the path planning and autonomous obstacle avoidance module is as follows: First, based on the precise 3D coordinates of the inspection target, the target type, and the preset shooting specifications, the optimal observation waypoint for each target is automatically calculated. The shooting specifications include shooting distance, azimuth angle, and gimbal pitch angle. Second, a 3D path planning map is constructed using all waypoints to be photographed as nodes. The cost between nodes integrates flight distance, time, energy consumption, and the smoothness of shooting angle switching. Then, an ant colony algorithm is used for global path optimization. In the algorithm, the initial pheromone distribution is set according to the importance of waypoints, the heuristic factor considers the angle between the current attitude of the UAV and the direction of the next waypoint, and the pheromone update strategy integrates the globally optimal path and iterative optimization information. Next, during flight along the planned path, data from the airborne millimeter-wave radar or visual depth sensor is read in real time. If an unmodeled obstacle is detected on the planned path, local replanning is triggered, and a local obstacle avoidance trajectory is generated using the Dynamic Window Method (DWA) combined with real-time 3D point cloud. Finally, the final optimized 3D flight path and the shooting parameters for each waypoint are sent to the UAV flight control system and gimbal control system to achieve fully autonomous flight inspection.

8. The image acquisition and recognition device and system for power line inspection drones according to claim 7, characterized in that: The specific workflow of the intelligent diagnosis and defect identification module is as follows: First, on the embedded side, a lightweight, improved YOLOv5-GhostNet model is used to quickly detect and classify the target image within the initial positioning box, identifying equipment types such as insulators, vibration dampers, and wire clamps, and obtaining their two-dimensional bounding boxes. Second, for the identified equipment targets, their corresponding infrared thermal imaging regions are cropped, and the average temperature, maximum temperature, and temperature distribution variance of the region are calculated. Then, a thermal texture dual-stream recognition network is constructed, where the texture stream takes a visible light cropped image as input, and the thermal stream takes a normalized temperature distribution map as input, and the dual-stream features are stitched together in a fusion layer. Next, this network is trained to distinguish between normal equipment, defective equipment, and non-equipment attachments. Finally, for targets identified as defective equipment, a more refined classifier is called to determine the defect type, and the diagnostic results, location coordinates, and evidence images are packaged into structured data.

9. The image acquisition and recognition device and system for power line inspection drones according to claim 8, characterized in that: The specific workflow of the collaborative processing module is as follows: Firstly, on the drone side, the flight control and data preprocessing sub-board and its onboard DSP / GPU heterogeneous computing unit are responsible for running all the algorithms of the data acquisition and dynamic stabilization module, the multispectral polarization imaging and fusion module, the edge enhancement and feature extraction module, as well as the lightweight YOLOv5-GhostNet detection steps in the intelligent diagnosis and defect recognition module, to perform front-end real-time processing and preliminary diagnosis. Secondly, through the integrated 5G intelligent processing unit, the intermediate data and results generated by the processing, including the stabilized image sequence, 3D point cloud fragments, target localization and diagnostic results, are encrypted and transmitted to the cloud secure data pool through a dedicated power APN access point. Then, on the high-performance computing cluster in the cloud, high-computational-load tasks are executed, including fine reconstruction of globally consistent 3D scenes, comparative analysis and trend prediction of massive historical inspection data, and incremental training and optimization of complex AI models such as the thermal texture dual-stream recognition model; finally, the cloud sends the updated model parameters and optimized flight path strategies to the front-line UAVs to complete the iteration of algorithms and knowledge.

10. The image acquisition and recognition device and system for power line inspection drones according to claim 9, characterized in that: The specific workflow of the secure communication and task management module is as follows: First, the task management subsystem receives inspection work orders from the power production management system. The work orders include line sections, tower numbers, inspection types, and levels. Second, based on the work orders, it automatically retrieves detailed 3D models of the corresponding lines, historical inspection data, and optimized flight strategies from the cloud to generate an executable inspection task plan, which is then sent to the designated drone. Then, during the mission execution, all control commands, status information, and collected data streams between the UAV and the ground station, and between the ground station and the cloud are encrypted and verified end-to-end through the power intranet communication protocol based on digital certificates and national cryptographic algorithms. Next, the mission execution status, equipment health status, and data processing progress are monitored in real time, providing a visual monitoring interface. Finally, after the mission is completed, the processing results of each module are automatically summarized, a standard inspection report including defect reports, 3D model updates, and statistical analysis charts is generated, and then sent back to the production management system.

Images