Unmanned aerial vehicle visual inspection and defect real-time identification method based on edge computing

Abstract

This invention discloses a UAV visual inspection and real-time defect identification method based on edge computing, relating to the field of power line inspection technology. The invention obtains power line images by deploying an airborne edge computing unit and adaptively adjusting image acquisition parameters; it then constructs a power line spatial model based on the power line images to obtain a masked image; it performs defect image template matching on the masked image and generates suspected defect regions based on similarity; it performs continuous multi-frame temporal confirmation on the suspected defect regions to obtain actual defects, and executes graded alarms based on the defect type of the confirmed actual defects; a ground server performs cluster analysis on the actual defects, dynamically updates the defect image templates, and marks priorities for subsequent inspections. This invention achieves efficient adaptive detection and dynamic optimization of power line defects.

Description

Technical Field

[0001] This invention relates to the field of power line inspection technology, specifically a method for real-time defect identification and inspection of unmanned aerial vehicles (UAVs) based on edge computing. Background Technology

[0002] Power lines are widely distributed and in complex environments. Long-term exposure can easily lead to defects such as insulator damage, broken conductor strands, and loose hardware, which directly affect the safety of the power grid. Traditional manual inspection has problems such as high labor intensity, low efficiency, and high safety risks. With the maturity of drone technology, aerial inspection has become the mainstream, which has greatly improved the coverage. However, the current mainstream drone inspection mode is mostly "collect first, process later": only images are stored during flight, and the data is transmitted back to the ground data center or cloud after the mission is completed, where defects are identified by humans or back-end models. This mode has obvious shortcomings: First, high-definition image data is large in volume, and transmission requires high bandwidth and long time, making it impossible to provide real-time warnings during flight, and critical defects often miss the best time to be dealt with; Second, in remote mountainous areas or areas with strong electromagnetic interference, the communication link between the drone and the ground station is unstable, and data transmission is prone to interruption, resulting in the loss of inspection data or mission failure. Therefore, how to achieve high real-time, low-latency, and high-precision identification of power line defects under the condition of limited airborne computing power is a problem we need to solve. To this end, we now provide a UAV visual inspection and real-time defect identification method based on edge computing. Summary of the Invention

[0003] The purpose of this invention is to provide a method for real-time visual inspection and defect identification of unmanned aerial vehicles based on edge computing.

[0004] The objective of this invention can be achieved through the following technical solution: a method for real-time visual inspection and defect identification of unmanned aerial vehicles based on edge computing, comprising the following steps: Step S1: Deploy the airborne edge computing unit to adaptively adjust the image acquisition parameters and obtain power line images; Step S2: Construct a spatial model of the power lines based on the power line image, and then obtain the masked image; Step S3: Perform defect image template matching on the masked image to generate suspected defect areas; Step S4: Perform time-series confirmation on suspected defect areas to obtain actual defects, and issue graded alarms based on the confirmed actual defect types; Step S5: The ground server performs cluster analysis and updates the defect image template.

[0005] Furthermore, the process of deploying airborne edge computing units includes: An airborne edge computing unit is deployed on a drone, and the airborne edge computing unit is electrically connected to the drone's visible light camera, infrared thermal imaging camera and communication module; The design parameters of the transmission lines in the inspection section are pre-entered into the airborne edge computing unit. The design parameters include the position coordinates of each tower, the conductor arrangement of the power lines between adjacent towers, the conductor design sag, and the conductor type. Defect image templates of various typical defects are pre-collected, each defect image template is scaled to a preset pixel size, and stored in the airborne edge computing unit; The airborne edge computing unit acquires the flight status parameters of the UAV in real time. These parameters include ground flight speed, relative altitude, real-time differential positioning coordinates, UAV attitude angles, and ambient illuminance values.

[0006] Furthermore, the process by which the airborne edge computing unit adaptively adjusts image acquisition parameters to obtain power line images includes: The image acquisition parameters include resolution, frame rate, and the spectral mode of acquisition; The airborne edge computing unit dynamically adjusts the image acquisition parameters according to the acquired relative altitude, ground flight speed, and ambient illuminance values, respectively, following resolution adaptive acquisition rules, frame rate adaptive acquisition rules, and spectral mode adaptive acquisition rules. Each captured power line image, along with its complete metadata, is stored in the onboard edge computing unit.

[0007] Furthermore, the process of constructing the electric field space model includes: Obtain the real-time differential positioning coordinates and attitude angles of the UAV, and obtain the corresponding span and height difference based on the position coordinates of adjacent towers; Establish a local coordinate system and obtain the catenary parameters based on the span, conductor type, and conductor design sag. Based on the catenary parameters, the catenary equation of the power line is constructed, and the local coordinate system equation is transformed into a parametric curve in the geographic coordinate system through local coordinate transformation, thus completing the construction of the power line spatial model.

[0008] Furthermore, the process of obtaining the masked image based on the electric field line space model includes: The electric field line spatial model is projected onto the image plane to obtain the electric field line trajectory; The relative distance between the UAV and the power line is obtained based on the real-time differential positioning coordinates of the UAV and the parameterized curve of the power line. Based on the relative distance, the mask width is generated, and then the initial mask image is obtained. In the initial mask image, pixels on the electric field line trajectory are assigned to the foreground, and the remaining areas are the background. The mask edges of the initial mask image are smoothed to obtain a refined mask image; The refined mask image is compared pixel by pixel with the current frame power line image to obtain the masked image.

[0009] Furthermore, the process of obtaining suspected defect regions by performing defect image template matching based on the masked image includes: During flight, the airborne edge computing unit evenly divides the masked image into several image blocks. For each image block, if the proportion of foreground pixels in the image block is lower than a preset ratio threshold, the image block is skipped and no matching is performed. If the proportion of foreground pixels within the image block reaches the preset proportion threshold, then the absolute value of pixel difference is summed and matched with all defective image templates in sequence to obtain the sum of absolute pixel differences. The similarity between the defective image template and the corresponding image block is obtained based on the sum of absolute pixel differences. If the similarity is greater than the preset similarity threshold, the image block is determined to be a suspected defect region, and the position coordinates of the image block in the masked image and the type of the matched defect image template are recorded.

[0010] Furthermore, the process of temporal confirmation of suspected defective areas includes: For each suspected defect area in the output, obtain the masked image of the current frame and the next 3 consecutive frames, and denote the current frame as k; In the (k+1)th frame, the defect image template is matched again within the preset search window, with the center point of the suspected defect area in the kth frame as the center. If a defect image template of the same type as the k-th frame is found within the search window, then the match is recorded as successful. Repeat the above defect image template matching operation for the (k+2)th and (k+3)th frames; If at least two out of three consecutive frames match successfully, the suspected defect area is confirmed as a real defect; otherwise, it is discarded.

[0011] Furthermore, the process of issuing graded alerts based on the confirmed actual defect type includes: The defect level is determined according to the defect image template type corresponding to the actual defect and a preset mapping relationship. When the defect level is determined to be a critical defect, an audible and visual alarm is triggered, and the real-time differential positioning coordinates of the actual defect and the defect type are sent to the ground control center through the communication module. When the defect level is severe or general, no audible and visual alarms are triggered. Instead, the real-time differential positioning coordinates of the actual defect, the defect type, and the corresponding image blocks are stored in the airborne edge computing unit. After the inspection mission is completed, the data will be uniformly transmitted back to the ground server for subsequent analysis.

[0012] Furthermore, the process of cluster analysis and defect image template updating by the ground server includes: After each inspection mission is completed, the ground server receives all the real defect data transmitted back. Each piece of real defect data includes the defect type, real-time differential positioning coordinates and detection timestamp. Cluster analysis was performed on all real defect data. After traversing all real defect data, several clusters were obtained, and each cluster represented an independent defect location point. For each cluster, the arithmetic mean of all coordinate points within the cluster is taken as the representative coordinate of that location, and the frequency of occurrence of each defect type within the cluster is counted. If the number of occurrences of a certain defect type in the cluster is greater than the preset repeated confirmation threshold, then it is determined that there is a repeated defect of that type at that location point; If the same type of defect is repeatedly detected at the same location point in multiple inspection tasks, the ground server will divide all historical defect images corresponding to that location point into blocks and perform pixel-level weighted averaging with the currently stored template to generate an updated defect image template. The weights of the weighted average are allocated according to the detection time, with newer images having higher weights. The updated defect image template replaces the original defect image template and is marked as a high-priority template. Before the next inspection mission begins, the drone automatically downloads the updated defect image template via the ground base station. When the drone flies to the coordinates again, the onboard edge computing unit prioritizes the high-priority template corresponding to that location for matching, instead of traversing all templates, thereby improving the matching speed.

[0013] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention obtains power line images by deploying an airborne edge computing unit and adaptively adjusting image acquisition parameters. The airborne edge computing unit can process image data in real time during flight, avoiding the need for large-scale raw data backhaul. The adaptive adjustment of image acquisition parameters can dynamically optimize image quality based on environmental conditions such as lighting and weather, ensuring clear and effective data acquisition even under complex conditions. Furthermore, by constructing a spatial model of the power line based on the power line image, and then obtaining a masked image, the location of the power line can be accurately determined. The masking process effectively removes background interference, highlighting the main area of ​​the power line, and significantly reducing the false alarm and false negative rates of subsequent defect detection. This invention improves detection accuracy by performing defect image template matching on the masked image and generating suspected defect areas based on similarity, thus achieving preliminary defect screening. Suspected defect areas are then confirmed sequentially across multiple frames to obtain actual defects. A tiered alarm system is implemented based on the confirmed defect type to ensure reliable defect identification. The tiered alarm strategy allows for differentiated response measures based on defect severity, enabling rational resource allocation. The ground server performs cluster analysis on actual defects, dynamically updates defect image templates, and marks priorities for subsequent inspections, achieving continuous iterative improvement in inspection efficiency. This invention realizes efficient adaptive detection and dynamic optimization of power line defects. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0015] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0016] like Figure 1 As shown, the edge computing-based UAV visual inspection and real-time defect identification method includes the following steps: Step S1: Deploy the airborne edge computing unit to adaptively adjust the image acquisition parameters and obtain power line images; Step S2: Construct a spatial model of the power lines based on the power line image, and then obtain the masked image; Step S3: Perform defect image template matching on the masked image to generate suspected defect areas; Step S4: Perform time-series confirmation on suspected defect areas to obtain actual defects, and issue graded alarms based on the confirmed actual defect types; Step S5: The ground server performs cluster analysis and updates the defect image template.

[0017] It should be further explained that, in the specific implementation process, the deployment of airborne edge computing units includes: An airborne edge computing unit is deployed on a drone, and the airborne edge computing unit is electrically connected to the drone's visible light camera, infrared thermal imaging camera and communication module; The design parameters of the transmission lines in the inspection section are pre-entered into the airborne edge computing unit. The design parameters include the position coordinates of each tower, the conductor arrangement of the power lines between adjacent towers, the conductor design sag, and the conductor type. Defect image templates of various typical defects are collected in advance, including insulator damage, broken conductor strands, loose hardware, bird nests, and tree obstacles. Each defect image template is scaled to 64×64 pixels and stored in the airborne edge computing unit. The airborne edge computing unit acquires the flight status parameters of the UAV in real time. These parameters include ground flight speed, relative altitude, real-time differential positioning coordinates, UAV attitude angles, and ambient illuminance values.

[0018] It should be further explained that, in the specific implementation process, the airborne edge computing unit adaptively adjusts the image acquisition parameters to obtain power line images, including the following steps: The image acquisition parameters include resolution, frame rate, and the spectral mode of acquisition; The airborne edge computing unit dynamically adjusts the image acquisition resolution according to the acquired relative altitude and the resolution adaptive acquisition rules. Specifically: When the relative height is greater than 20 meters, the pixel area occupied by the power line and insulator string in the image plane is small. At this time, the low resolution mode is used for image acquisition. The image resolution of the low resolution mode is 640×480 pixels. When the relative height is greater than 5 meters and less than or equal to 20 meters, image acquisition is performed using a medium resolution mode, where the image resolution of the medium resolution mode is 1280×720 pixels. When the relative height is less than or equal to 5 meters, the power line and insulator string occupy a large pixel area in the image plane. At this time, the high resolution mode is used for image acquisition. The image resolution of the high resolution mode is 1920×1080 pixels. The airborne edge computing unit dynamically adjusts the image acquisition frame rate according to the acquired ground flight speed and the frame rate adaptive acquisition rule; Specifically: When the speed of flight to the ground is less than or equal to At that time, the frame rate was set to 15 frames per second; When the speed of flight to the ground is greater than and less than or equal to At that time, the frame rate was set to 25 frames per second; When the speed of flight to the ground is greater than At that time, the frame rate was set to 30 frames per second; The airborne edge computing unit dynamically adjusts the spectral mode of image acquisition according to the acquired ambient illuminance value and the adaptive acquisition rules of spectral mode. Specifically: When the ambient illuminance value is greater than or equal to At that time, only the visible light camera is turned on for image acquisition; When the ambient illuminance value is greater than or equal to and less than At the same time, the visible light camera and the infrared thermal imaging camera are turned on simultaneously for synchronous data acquisition; When the ambient illuminance value is less than At that time, only the infrared thermal imaging camera is turned on for image acquisition; Each captured power line image, along with its complete metadata, is stored in the airborne edge computing unit. The metadata includes image acquisition timestamp, real-time differential positioning coordinates, UAV attitude angle, current frame resolution, and acquisition frame rate.

[0019] It should be further explained that, in the specific implementation process, the process of constructing the power line space model includes: Obtain the real-time differential positioning coordinates and attitude angles of the UAV, denoted as follows: , ; The coordinates of the adjacent tower positions are denoted as follows: , The span and height difference are obtained based on the coordinates of adjacent towers, and are denoted as follows: and ,Right now: , ; The span refers to the horizontal distance between the suspension points of the power transmission lines of two adjacent towers, and the height difference refers to the vertical height difference between two adjacent towers. With towers The projection point on the geographic plane is the origin. , in order to point to The horizontal direction is Establish a local coordinate system along the positive axis; The local coordinate system is a one-dimensional curvilinear coordinate system, with coordinate parameters... Indicates starting from the origin The distance traveled along the curve in the direction of the gear length. The range of values ​​is ; The parameters of the catenary are calculated based on the span, conductor type, and designed sag of the conductor, and denoted as follows: ,Right now: ; in, Design a sag for the conductor; The catenary equation for the electric power line is constructed based on the catenary parameters, denoted as: ,Right now: ; in, This represents the power line height coordinate at a point with coordinate parameter s along the span direction in the local coordinate system, where s is the position parameter along the span direction and C is a constant term. The catenary equation in the local coordinate system is transformed into a parametric curve in the geographic coordinate system through local coordinate transformation, denoted as... ,Right now: ; Where t is the curve parameter corresponding to s; Complete the construction of the electric field line spatial model.

[0020] It should be further explained that, in the specific implementation process, the process of obtaining the masked image based on the electric field line space model includes: The electric field line spatial model is projected onto the image plane to obtain the electric field line trajectory; The relative distance between the UAV and the power line is obtained based on the UAV's real-time differential positioning coordinates and the parametric curve of the power line, denoted as . ,Right now: ; The mask width is dynamically calculated based on the relative distance to obtain an initial mask image. The mask width refers to the pixel width of the power line mask in the power line image. ; Where W is the mask width, The preset base width is given by k, the distance coefficient is given by d, and the relative distance between the drone and the power line is given by d. The initial mask image is a binary image of the same size as the current frame's power line image. Pixels on the power line trajectory are assigned the foreground value, and the remaining areas are the background. Gaussian blurring is applied to smooth the mask edges of the initial mask image, and the mask boundary is further refined by combining image gradient direction constraints to obtain a refined mask image. The refined mask image is compared pixel by pixel with the current frame power line image. The image content corresponding to the foreground area in the refined mask image is retained, and the background area is removed to obtain the masked image.

[0021] It should be further explained that, in the specific implementation process, the process of obtaining suspected defect areas by performing defect image template matching based on the masked image includes: During flight, the airborne edge computing unit evenly divides the masked image into 64×64 pixel image blocks. For each image block, if the proportion of foreground pixels in the image block is lower than a preset ratio threshold, the image block is skipped and no matching is performed. If the proportion of foreground pixels within the image block reaches the preset proportion threshold, then the absolute value of the pixel difference is summed and matched with all defective image templates in sequence, and the sum of the absolute pixel differences is denoted as... ,Right now: ; in, Indicates the location of the defect image template The pixel grayscale value at that location, Indicates the position of the corresponding image block. The pixel grayscale value at that location; Calculate the similarity between the defective image template and the corresponding image block, denoted as . ,Right now: ; like If the image block is identified as a suspected defect area, the position coordinates of the image block in the masked image and the type of the matched defect image template are recorded.

[0022] It should be further explained that, in the specific implementation process, the process of confirming suspected defective areas in a timely manner includes: For each suspected defect area in the output, obtain the masked image of the current frame and the next 3 consecutive frames, and denote the current frame as k; In the (k+1)th frame, defect image template matching is performed again within a 10×10 pixel search window, centered on the center point of the suspected defect area in the kth frame. If a defect image template of the same type as the k-th frame is found within the search window, then the match is recorded as successful. Repeat the above defect image template matching operation for the (k+2)th and (k+3)th frames; If at least two out of three consecutive frames match successfully, the suspected defect area is confirmed as a real defect; otherwise, the suspected defect area is discarded as a false alarm.

[0023] It should be further explained that, in the specific implementation process, the process of issuing graded alerts based on the confirmed actual defect type includes: The defect level is determined according to the defect image template type corresponding to the actual defect and a preset mapping relationship. Specifically: The template for broken conductor strands corresponds to critical defects, the template for damaged insulators corresponds to serious defects, and the templates for loose fittings, bird nests, and tree obstacles correspond to general defects. When the defect level is determined to be a critical defect, the airborne edge computing unit triggers the audible and visual alarm device installed on the UAV body, controls the audible and visual alarm device to emit a buzzing sound and a red flashing light, and at the same time sends the real-time differential positioning coordinates of the actual defect and the defect type to the ground control center through the communication module. When the defect level is severe or general, no audible and visual alarms are triggered. Instead, the real-time differential positioning coordinates of the actual defect, the defect type, and the corresponding image blocks are stored in the airborne edge computing unit. After the inspection mission is completed, the data will be uniformly transmitted back to the ground server for subsequent analysis.

[0024] It should be further explained that, in the specific implementation process, the process of cluster analysis and defect image template updating by the ground server includes: After each inspection mission is completed, the ground server receives all the real defect data transmitted back. Each piece of real defect data includes the defect type, real-time differential positioning coordinates and detection timestamp. Let the real-time differential positioning coordinates of the p-th real defect data be denoted as . ,in , This represents the total number of actual defects detected in this mission. Initialize the cluster set and set the coordinates of the first real defect data. The initial center of the first cluster is denoted as . ; For the coordinates of each subsequent real defect data Calculate the spatial distance between the r-th cluster and the center point of all currently established clusters, and mark the center point of the r-th cluster as follows: ; Calculate the spatial distance between the real-time differential location coordinates of the p-th real defect data and the coordinates of the center point of the r-th cluster, denoted as . ,Right now: ; If a certain cluster exists If the spatial distance is less than a preset clustering distance threshold, then... Assign it to the cluster and update the coordinates of the center point of the cluster to the arithmetic mean of all coordinate points within the cluster; If the spatial distance for all existing clusters is greater than or equal to the clustering distance threshold... Then create a new cluster, with As the initial center point of the newly created cluster; After traversing all the real defect data, we obtain There are 3 clusters, each cluster representing an independent defect location point; For each cluster, the arithmetic mean of all coordinate points within the cluster is taken as the representative coordinate of that location, and the frequency of occurrence of each defect type within the cluster is counted. If the number of occurrences of a certain defect type in the cluster is greater than the preset repeated confirmation threshold, then it is determined that there is a repeated defect of that type at that location point; If the same type of defect is repeatedly detected at the same location point in multiple inspection tasks, the ground server will divide all historical defect images corresponding to that location point into blocks and perform pixel-level weighted averaging with the currently stored template to generate an updated defect image template. The weights of the weighted average are allocated according to the detection time, with newer images having higher weights. The updated defect image template replaces the original defect image template and is marked as a high-priority template. Before the next inspection mission begins, the drone automatically downloads the updated defect image template via the ground base station. When the drone flies to the coordinates again, the onboard edge computing unit prioritizes the high-priority template corresponding to that location for matching, instead of traversing all templates, thereby improving the matching speed.

[0025] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications or equivalent substitutions made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for real-time visual inspection and defect identification of unmanned aerial vehicles (UAVs) based on edge computing, characterized in that, Includes the following steps: Step S1: Deploy the airborne edge computing unit to adaptively adjust the image acquisition parameters and obtain power line images; Step S2: Construct a spatial model of the power lines based on the power line image, and then obtain the masked image; Step S3: Perform defect image template matching on the masked image to generate suspected defect areas; Step S4: Perform time-series confirmation on suspected defect areas to obtain actual defects, and issue graded alarms based on the confirmed actual defect types; Step S5: The ground server performs cluster analysis and updates the defect image template.

2. The UAV visual inspection and real-time defect identification method based on edge computing according to claim 1, characterized in that, The process of deploying airborne edge computing units includes: An airborne edge computing unit is deployed on a drone, and the airborne edge computing unit is electrically connected to the drone's visible light camera, infrared thermal imaging camera and communication module; The design parameters of the transmission lines in the inspection section are pre-entered into the airborne edge computing unit. The design parameters include the position coordinates of each tower, the conductor arrangement of the power lines between adjacent towers, the conductor design sag, and the conductor type. Defect image templates of various typical defects are pre-collected, each defect image template is scaled to a preset pixel size, and stored in the airborne edge computing unit; The airborne edge computing unit acquires the flight status parameters of the UAV in real time. These parameters include ground flight speed, relative altitude, real-time differential positioning coordinates, UAV attitude angles, and ambient illuminance values.

3. The UAV visual inspection and real-time defect identification method based on edge computing according to claim 2, characterized in that, The process by which the airborne edge computing unit adaptively adjusts image acquisition parameters to obtain power line images includes: The image acquisition parameters include resolution, frame rate, and the spectral mode of acquisition; The airborne edge computing unit dynamically adjusts the image acquisition parameters according to the acquired relative altitude, ground flight speed, and ambient illuminance values, respectively, following resolution adaptive acquisition rules, frame rate adaptive acquisition rules, and spectral mode adaptive acquisition rules. Each captured power line image, along with its complete metadata, is stored in the onboard edge computing unit.

4. The UAV visual inspection and real-time defect identification method based on edge computing according to claim 3, characterized in that, The process of constructing a space model of electric power lines includes: Obtain the real-time differential positioning coordinates and attitude angles of the UAV, and obtain the corresponding span and height difference based on the position coordinates of adjacent towers; Establish a local coordinate system and obtain the catenary parameters based on the span, conductor type, and conductor design sag. Based on the catenary parameters, the catenary equation of the power line is constructed, and the local coordinate system equation is transformed into a parametric curve in the geographic coordinate system through local coordinate transformation, thus completing the construction of the power line spatial model.

5. The UAV visual inspection and real-time defect identification method based on edge computing according to claim 4, characterized in that, The process of obtaining the masked image based on the electric field line space model includes: The electric field line spatial model is projected onto the image plane to obtain the electric field line trajectory; The relative distance between the UAV and the power line is obtained based on the real-time differential positioning coordinates of the UAV and the parameterized curve of the power line. Based on the relative distance, the mask width is generated, and then the initial mask image is obtained. In the initial mask image, pixels on the electric field line trajectory are assigned to the foreground, and the remaining areas are the background. The mask edges of the initial mask image are smoothed to obtain a refined mask image; The refined mask image is compared pixel by pixel with the current frame power line image to obtain the masked image.

6. The UAV visual inspection and real-time defect identification method based on edge computing according to claim 5, characterized in that, The process of obtaining suspected defect regions by performing defect image template matching based on the masked image includes: During flight, the airborne edge computing unit evenly divides the masked image into several image blocks. For each image block, if the proportion of foreground pixels in the image block is lower than a preset ratio threshold, the image block is skipped and no matching is performed. If the proportion of foreground pixels within the image block reaches the preset proportion threshold, then the absolute value of pixel difference is summed and matched with all defective image templates in sequence to obtain the sum of absolute pixel differences. The similarity between the defective image template and the corresponding image block is obtained based on the sum of absolute pixel differences. If the similarity is greater than the preset similarity threshold, the image block is determined to be a suspected defect region, and the position coordinates of the image block in the masked image and the type of the matched defect image template are recorded.

7. The UAV visual inspection and real-time defect identification method based on edge computing according to claim 6, characterized in that, The process of time-series confirmation of suspected defective areas includes: For each suspected defect area in the output, obtain the masked image of the current frame and the next 3 consecutive frames, and denote the current frame as k; In the (k+1)th frame, the defect image template is matched again within the preset search window, with the center point of the suspected defect area in the kth frame as the center. If a defect image template of the same type as the k-th frame is found within the search window, then the match is recorded as successful. Repeat the above defect image template matching operation for the (k+2)th and (k+3)th frames; If at least two out of three consecutive frames match successfully, the suspected defect area is confirmed as a real defect; otherwise, it is discarded.

8. The UAV visual inspection and real-time defect identification method based on edge computing according to claim 7, characterized in that, The process of issuing graded alerts based on the confirmed actual defect type includes: The defect level is determined according to the defect image template type corresponding to the actual defect and a preset mapping relationship. When the defect level is determined to be a critical defect, an audible and visual alarm is triggered, and the real-time differential positioning coordinates of the actual defect and the defect type are sent to the ground control center through the communication module. When the defect level is severe or general, no audible and visual alarms are triggered. Instead, the real-time differential positioning coordinates of the actual defect, the defect type, and the corresponding image blocks are stored in the airborne edge computing unit. After the inspection mission is completed, the data will be uniformly transmitted back to the ground server for subsequent analysis.

9. The UAV visual inspection and real-time defect identification method based on edge computing according to claim 8, characterized in that, The process of cluster analysis and defect image template updating by the ground server includes: After each inspection mission is completed, the ground server receives all the real defect data transmitted back. Each piece of real defect data includes the defect type, real-time differential positioning coordinates and detection timestamp. Cluster analysis was performed on all real defect data. After traversing all real defect data, several clusters were obtained, and each cluster represented an independent defect location point. For each cluster, the arithmetic mean of all coordinate points within the cluster is taken as the representative coordinate of that location, and the frequency of occurrence of each defect type within the cluster is counted. If the number of occurrences of a certain defect type in the cluster is greater than the preset repeated confirmation threshold, then it is determined that there is a repeated defect of that type at that location point; If the same type of defect is repeatedly detected at the same location point in multiple inspection tasks, the ground server will divide all historical defect images corresponding to that location point into blocks and perform pixel-level weighted averaging with the currently stored template to generate an updated defect image template. The weights of the weighted average are allocated according to the detection time, with newer images having higher weights. The updated defect image template replaces the original defect image template and is marked as a high-priority template. Before the next inspection mission begins, the drone automatically downloads the updated defect image template via the ground base station. When the drone flies to the coordinates again, the onboard edge computing unit prioritizes the high-priority template corresponding to that location for matching, instead of traversing all templates, thereby improving the matching speed.

Images