Abnormal site swing shooting and collecting method of sample library

Abstract

The invention provides an abnormal scene swing shooting and collecting method for a normal sample library, and relates to the technical field of image detection and generation. According to the scheme, the method comprises the steps that S1, an original image sequence is collected, and a normal sample library is constructed; s2, preprocessing the images in the normal sample library; s3, screening out abnormal samples from the normal sample library, and sending the abnormal samples into the abnormal samples; s4, classifying the abnormal samples according to the types of the anomalies; s5, constructing a new abnormal sample according to the screened abnormal samples; s6, carrying out the preprocessing operation in the step S2 on a new abnormal sample, and then carrying out the screening operation in the step S3; and S7, separating an abnormal site, fusing the abnormal site with the background image to obtain a new abnormal sample, and executing the step S6 on the new abnormal sample. The method has the beneficial effects that the problem of few samples is effectively solved, the problem of difficulty in swing shooting of some real scenes is overcome, the number of abnormal samples can be increased in batches, and the cost and labor are reduced.

Description

Technical Field

[0001] This invention relates to the field of image detection and generation technology, specifically to a method for staging and collecting abnormal scene photographs of a sample library. Background Technology

[0002] Overhead power lines are mostly located in mountainous areas, facing environmental challenges such as rapid vegetation growth, kites on the conductors, dust nets, and tangled film, as well as inherent faults in the lines themselves, such as abnormalities in the drain lines. Failure to detect these faults in a timely manner can have two consequences: rapid vegetation growth may lead to insufficient safety distances, resulting in single-phase grounding faults, tripping, and potentially causing wildfires; tangled film that is not cleared in time may cause phase-to-phase short circuits during rainfall, leading to tripping; secondly, abnormal drain lines and tilted towers may result in insufficient safety distances, tower collapse, discharge, or tripping, and even wildfires, causing significant economic losses and social impact; thirdly, dry winter weather increases the risk of wildfires, which can easily damage overhead power line corridors and cause line tripping.

[0003] However, the environment in which overhead lines are located makes manual inspections difficult and costly. Therefore, the trend in anomaly detection for overhead lines is to use anomaly detection and identification algorithms to investigate and issue warnings for the aforementioned anomalies. Anomaly detection and identification algorithms utilize the abnormal features of the abnormal site to identify and detect anomalies, but these algorithms suffer from the problem of insufficient sample size. Summary of the Invention

[0004] To address the problems mentioned above, this invention provides a method for staging and collecting abnormal scenes in a sample library, which can solve the problem of scarce abnormal samples and improve the efficiency and accuracy of model training.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for staging and collecting abnormal scenes in a normal sample library, comprising the following steps: S1. Collect raw image sequences and construct a normal sample library; S2. Preprocess the images in the normal sample library; S3. Perform abnormal sample screening operation, filter out abnormal samples from the normal sample library, and send the abnormal samples into the abnormal sample library. S4. Classify abnormal samples according to the type of abnormality; S5. Construct new abnormal samples based on the selected abnormal samples; S6. Perform the preprocessing operation in step S2 on the newly constructed abnormal samples, and then perform the filtering operation in step S3 to obtain usable abnormal samples and send them into the abnormal sample library. S7. Separate the abnormal scene from the abnormal sample and fuse it with the background image in the independent background library to obtain a new abnormal sample, and perform step S6 on the new abnormal sample.

[0006] Furthermore, the detailed process of step S1 is as follows: Using drones equipped with high-resolution cameras, images were collected from sections of overhead power lines. During the collection process, the drone's flight altitude and angle were adjusted to capture raw images and build a normal sample library. The drone is also equipped with temperature and humidity sensors and light sensors to record environmental data during the collection process.

[0007] Furthermore, the detailed process of step S2 is as follows: S201. Perform Gaussian filtering on the image to remove noise using a Gaussian kernel. The Gaussian kernel function is: (1); in, Standard deviation x and y These are the horizontal and vertical coordinates, respectively; S202. Sharpen the filtered image to highlight the texture and contours in the image.

[0008] Furthermore, the detailed process of step S3 is as follows: S301. Set the grayscale mutation threshold to identify provisional outliers; Grayscale mutation threshold The threshold will be dynamically adjusted according to the collected light intensity data. Specifically, when the light intensity is 500-3000 lux, the grayscale change threshold is set to 55, and when the light intensity is 3000-10000 lux, the grayscale change threshold is set to 50. When a pixel meets the judgment condition, the pixel is recorded as a temporary outlier. The judgment condition is as follows: (2); in, L This refers to the grayscale pixel value of a certain pixel. This is the average pixel grayscale value of the points adjacent to this point; S302. Form a connected region P from consecutive provisional outliers. Filter by area feature, retaining regions that meet the outlier screening criteria as suspected outlier regions. The screening criteria are: (3); in, Let P be the area. , These are the minimum and maximum area thresholds, respectively; S303. Perform preliminary screening of abnormal samples: If there are no suspected abnormal areas in the sample, the image is directly determined as a normal sample and sent back to the normal sample library; if there are suspected abnormal areas in the sample, proceed to step S304. S304. Manually review samples with suspected abnormal areas: If the image in the sample is abnormal, send it to the abnormal sample database; otherwise, send it back to the normal sample database.

[0009] Furthermore, the types of anomalies include: faults in the overhead line itself, foreign objects entangled in the transmission line, smoke and fire in the line and surrounding environment, and excessively rapid growth of vegetation around the overhead line; Faults in overhead power lines include: broken strands, abnormal lead wires, and tilted towers.

[0010] Furthermore, the detailed process of step S5 is as follows: S501. Use an image feature extraction algorithm to extract abnormal features from abnormal samples; S502. Based on the extracted abnormal features, perform a staged photography operation to obtain new abnormal samples; Staged photography includes both staged photography in real-world settings and modeling. Furthermore, the process of posing for the photos is as follows: Construct staged shooting scenes, create different environments, add auxiliary elements, add abnormal features of various abnormal types to the staged shooting scenes, determine the shooting angle and distance, and collect images of each abnormal state in different environments in close-up, medium-range and long-range according to the shooting angle and distance; The shooting angle is related to the type of anomaly; During the data acquisition process, keep the camera's focal length, aperture, ISO, and other parameters consistent.

[0011] Furthermore, the detailed process of step S7 is as follows: S701. Use an image separation algorithm to separate abnormal samples in the abnormal sample library to obtain the abnormal scene and background image; S702: Generate more background images, and all background images are combined into a background image set; S703. After randomly arranging and combining the abnormal scene with the background images in the background image set, the images are fused to obtain new abnormal samples. The fusion process is as follows: (4); in, The merged pixel values , These are the pixel values ​​of the abnormal scene and the background image, respectively. Weights for abnormal scenarios; S704. Use the GAN repair network model to eliminate splicing traces in the fused abnormal samples and output usable abnormal samples. S705. Perform step S6 on the abnormal samples after processing by the GAN repair network model.

[0012] Furthermore, in step S702, AI is used to generate background images in batches.

[0013] Furthermore, the manual review process is as follows: For any sample, professional inspectors examine suspected abnormal areas one by one to determine whether they are genuine abnormalities or interference. If there is at least one genuine abnormality in a sample, the sample is determined to be an abnormal sample; otherwise, it is determined to be a normal sample.

[0014] Beneficial effects: (1) Using staged photography can effectively solve the problem of small sample size; (2) The combination of real-world posing and modeling in posing method can overcome the difficulty of posing in certain real-world scenarios while ensuring the authenticity of the samples. (3) Integrating abnormal sites with the background database can increase the number of abnormal samples in batches, reducing costs and manpower. Attached Figure Description

[0015] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the method steps of the present invention. Detailed Implementation

[0016] like Figure 1 As shown, a method for staging and collecting abnormal scenes from a normal sample library is used to provide training samples needed for overhead line anomaly detection. The steps are as follows: S1. Acquire raw image sequences and construct a normal sample library. The detailed process is as follows: A drone equipped with a 20-megapixel high-definition camera was used to collect images of a 35kV overhead power line section. During the collection process, the drone's flight altitude and angle were adjusted to acquire images from horizontal, vertical, and 45° tilted perspectives. The collection distances were 3m, 15m, and 30m. Simultaneously, environmental data was recorded using temperature and humidity sensors and light sensors on the drone: temperature and humidity ranged from 10℃ to 35℃, humidity from 40% to 70%, and light intensity ranged from 1000 to 8000 lux. A total of 10,000 original images were collected to construct a normal sample library.

[0017] S2. Preprocess the images in the normal sample library to improve image quality. The detailed process is as follows: S201. Perform Gaussian filtering on the image using a 3×3 Gaussian kernel. The Gaussian kernel function is: (1); in, Standard deviation x and y These are the horizontal and vertical coordinates, respectively; this processing removes Gaussian noise from the image while preserving key image features. S202. Use the Laplacian operator to sharpen the filtered image, highlighting the texture contours in the image, such as wire edges, foreign object shapes, plant outlines, etc., to improve the recognition of abnormal features.

[0018] S3. Perform anomaly sample screening. Based on preset grayscale mutation thresholds and area thresholds, anomaly samples are selected from the normal sample library. The detailed process is as follows: S301. Set the grayscale mutation threshold to identify provisional outliers; Grayscale mutation threshold The threshold will be dynamically adjusted according to the collected light intensity data. Specifically, when the light intensity is 500-3000 lux, the grayscale change threshold is set to 55, and when the light intensity is 3000-10000 lux, the grayscale change threshold is set to 50. When a pixel meets the judgment condition, the pixel is recorded as a temporary outlier. The judgment condition is as follows: (2); in, L This refers to the grayscale pixel value of a certain pixel. This is the average pixel grayscale value of the points adjacent to this point; S302. Form a connected region P from consecutive provisional outliers. Filter by area feature, retaining regions that meet the outlier screening criteria as suspected outlier regions. The screening criteria are: (3); in, Let P be the area. , These are the minimum and maximum area thresholds, respectively; S303. Perform preliminary screening of abnormal samples: If there are no suspected abnormal areas in the sample, the image is directly determined as a normal sample and sent back to the normal sample library; if there are suspected abnormal areas in the sample, proceed to step S304. S304. Manually review samples with suspected abnormal areas: If the image in the sample is abnormal, send it to the abnormal sample database; otherwise, send it back to the normal sample database. The manual review process is as follows: For any sample, professional inspectors will check the suspected abnormal areas one by one to determine whether they are real abnormalities or interference. If there is at least one real abnormality in the sample, the sample will be judged as an abnormal sample; otherwise, it will be judged as a normal sample. The types of anomalies include: faults in the overhead line itself, foreign objects entangled in the transmission line, smoke and fire in the line and surrounding environment, and excessively rapid growth of vegetation around the overhead line; Faults in overhead power lines include: broken strands, abnormal lead wires, and tilted towers.

[0019] S4. Classify abnormal samples according to the type of abnormality.

[0020] S5. Construct new abnormal samples based on the selected abnormal samples. The detailed process is as follows: S501. Use an image feature extraction algorithm to extract abnormal features from abnormal samples; abnormal features include: shape parameters, texture features, gray-level distribution features, and relative positional relationship between the abnormal region and the background; The shape parameters of the abnormal region include: area, perimeter, and aspect ratio; Texture features include: such as gray-level co-occurrence matrix and local binary pattern; Gray-level distribution characteristics include: gray-level mean, variance, and histogram distribution; The relative positional relationship between the abnormal area and the background refers to the position of the foreign object relative to the overhead line, such as the position of the foreign object wrapped around the conductor, the distance between the plants and the conductor, etc. Commonly used image feature extraction algorithms include SIFT algorithm, HOG algorithm, etc. S502. Based on the extracted abnormal features, perform a staged photography operation to obtain new abnormal samples; Staged photography includes both staged photography in real-world settings and modeling. The process of posing for photos in real-world scenarios is as follows: Constructing a staged shooting scene: Select an open space and build simulated poles, conductors, insulators and other core components according to the scale, ensuring that the size and relative position of the components are consistent with the actual overhead lines; Construct different environments: Adjust temperature, humidity, and light intensity using devices such as humidifiers, heaters, awnings, and supplemental lighting to simulate different weather conditions over time, such as sunny evenings and rainy mornings; the temperature range is 10℃-35℃, the humidity range is 40%-70%, and the light intensity range is 1000-8000 lux. Add auxiliary elements: Add natural disturbance elements to the scene, such as weeds on the ground, mountain models or real scenery in the distance, and cloud scenery hanging in the sky, to restore the outdoor environmental characteristics of the overhead line. Add abnormal features for each type of abnormality to the above staged shooting scenarios; For broken strands, conductor samples with the same material and diameter as those in actual overhead lines are selected, and the faults are treated according to three levels: 1 / 3 broken strand, 1 / 2 broken strand, and complete broken strand. The fracture surface is simulated with irregular shapes in actual faults, such as some conductor strands sticking out and oxidation marks at the fracture surface. The treated conductor samples are fixed to the insulator strings of the simulated tower to ensure that the conductor tension is consistent with that of the actual overhead line. Oxidation marks at the fracture surface can be simulated by applying a small amount of brown pigment. For abnormalities in the drainage line, a flexible conductor with the same specifications as the actual conductor is selected to simulate three abnormal states: slack drooping, i.e., the drooping amount is 1.5-3 times that of the normal state; excessive stretching, i.e. the stretching length is 1.1-1.3 times that of the normal state; and offset collision, i.e. the distance between the drainage line and the tower component is less than 0.5m. Fixtures are used to fix both ends of the drainage line, and the angle is adjusted to simulate different installation deviations. For tower tilting, a 1:50 scale tower model is used, made of ABS engineering plastic. The tilt angle is adjusted from 3 to 15 degrees by adjusting the base support device, with tilt directions including along and across the power line. Simulated terrain, including mountains and flat areas, is arranged around the tower, and auxiliary anomalies such as loose guy wires or foundation settlement are added. For transmission lines entangled with foreign objects, kites, dustproof nets, or films can be used. Kite: Select kites of different materials with an area of ​​0.5-2㎡, simulate kite line winding, including two states: single strand winding 2-5 times and kite body covering. Fix the kite on the simulated guide wire and adjust the kite's posture, including: horizontal, vertical, and tilted at an angle of 30-60 degrees.

[0021] Dustproof netting and film: Use polyethylene dustproof netting and plastic film with a thickness of 0.02-0.1mm, cut into sheets of 1-3㎡, to simulate loose wrapping, i.e., 2-4 layers of wrapping, and tight coverage, i.e., the coverage area accounts for 1 / 4-1 / 2 of the conductor span. The wrapping parts include the middle section of the conductor, near the insulator, and at the conductor joint. For smoke and fire hazards along the power lines and in the surrounding environment, smoke generators and flame simulators are used to simulate three smoke and fire intensities: weak smoke (visibility 5-10m), medium smoke (visibility 2-5m), and strong smoke (visibility less than 2m). The flame height is set to 0.5-3m, and the flame color gradually changes from orange-red to blue. The orange-red temperature is approximately 600-800 degrees Celsius, and the blue temperature is greater than 1000 degrees Celsius. The placement locations include: the bottom of the tower, 10-30m below the conductor, and vegetation areas within the power line corridor. Background lighting at different times of day needs to be added. For cases where vegetation around overhead power lines grows too quickly, common tree species such as pine, poplar, and shrubs, which are consistent with the line corridor, are selected. A model is built at a scale of 1:20 to simulate the distance between the plants and the conductors. When the distance is less than 3m, it is considered a danger distance; when it is greater than or equal to 3m but less than 5m, it is considered a warning distance; and when it is greater than or equal to 5m but less than 8m, it is considered an approach distance. Plant morphology includes upright growth, growth towards the line at an angle of 10-45 degrees, and branches entwined around the conductors. Design the angle for posed photos; posed angles include: basic angles and specific angles. The basic angles cover three major perspectives: horizontal, vertical, and 45-degree tilt. Each basic perspective is further subdivided into sub-angles every 15 degrees to achieve full-angle coverage. The horizontal perspective refers to the perspective parallel to the conductor axis; the vertical perspective refers to the perspective perpendicular to the conductor axis, including both downward and upward views; and the 45-degree tilt perspective refers to the perspective at a 45-degree angle to the horizontal direction. A specific angle is determined based on different anomaly characteristics, specifically as follows: For broken strands, specific angles are: perpendicular to the direction of the break, the side of the break and parallel to the direction of the break, and taking pictures of the overall state of the conductor from a distance. For abnormalities in the drain wire, specific angles are: side view in relaxed / stretched state, front view of offset collision, and close-up view of the connection between the drain wire and the insulator. For tower tilt, the specific angles are: the frontal view along the line direction, the frontal view across the line direction, and the top-down view of the tower offset. For transmission lines entangled with foreign objects, specific angles include: close-up view of the entangled part, view of the overall distribution of the foreign object, and view of the details of the contact point between the foreign object and the conductor. For fireworks and fires affecting the line and its surrounding environment, specific angles include: a close-up view of the source of the fireworks, a panoramic view of the spread of the fireworks, and a view of the fireworks obscuring the line. For fireworks related to the line and its surrounding environment, specific angles are: a top-down view of the distance between the top of the plant and the conductor, a side view of the plant's tilt direction, and a close-up view of the contact point between the plant branches and the conductor. The shooting distances for the specific shooting angles mentioned above are divided into three distance levels: close-up, medium shot, and long shot. Close-up is 0.5-3m, used to highlight abnormal details; medium shot is 3-15m, used to show the relationship between the abnormality and surrounding components; long shot is 15-50m, used to show the position of the abnormality in the overall line scene. Using a high-definition camera with at least 20 megapixels, images of each abnormal state in different environments were collected at preset shooting angles and distances, including close-up, medium-range, and distant views. Since each distance level corresponds to a range, several different distances could be randomly selected within that range for shooting. During the collection process, the camera's focal length, aperture, ISO, and other parameters were kept consistent to avoid affecting sample consistency due to changes in equipment parameters.

[0022] The modeling process follows the same logic as on-site scene photography, only implemented in modeling software, specifically as follows: Constructing a staged photography scene model: Create a new SolidWorks part file, draw the main body of the tower, crossarm, and connectors according to the drawing parameters, and complete the 1:1 scale 3D model using the "Extrude, Array, Assemble" function, retaining details such as guy wire fixing points and insulator installation positions; Create the wire component, using a steel core and aluminum stranded wire structure, setting the number of strands to 24 and the diameter of each strand to 2.2mm. Use the "spiral line + scan" function to generate the stranding pattern, reserving a cutting surface for strand breakage faults. Auxiliary component modeling: Complete the modeling of insulators and hardware in sequence, ensuring that the component interface dimensions match; the insulator material is: porcelain / composite material, including metal joints; the hardware includes: wire clamps and hanging rings; Assemble all components into a complete "overhead line core model" and export it as a STEP or STL file for easy import by subsequent tools; the STL file is a backup file. To build the scene environment in Unity3D, first import the STEP file of the "overhead line core model". Then, use the "scale and pan" function to place the model at the scene origin to ensure that the model is not deformed and the size is accurate. Add environmental and auxiliary elements, specifically: import DEM terrain data of the actual route corridor, generate 1:1 scale terrain using the "Terrain Tool", set terrain texture, adjust the terrain according to the actual route direction to make the base of the tower conform to the terrain; add "Directional Light" to simulate sunlight, adjust parameters according to time of day, and enable shadow casting; add "Skybox", set fog effect, and enable atmospheric scattering to make the scene lighting and shadows more natural; use the "Detail Drawing Tool" to add weeds, trees, etc. in batches on the surface of the foreground terrain to avoid dense stacking; To add anomaly characteristics in Blender, specifically: Broken strand fault: Select the wire model, cut the wire in "Edit Mode", delete the corresponding strands according to 1 / 3, 1 / 2, and complete breakage, manually adjust the curling of the broken wire strands, and add an oxidation texture map; Kite entanglement: Create a new flexible mesh, select plastic film material, draw the kite outline using "curve modeling", add wind field simulation to create folds, wrap the kite line around the guide wire, and export it as FBX format.

[0023] Dustproof netting entanglement: Create a grid-like model, simulate the entanglement shape using the "stretch and bend" function, cover wire length 0.3-1m, and export as FBX format.

[0024] Fireworks effects: Create a new particle system, set flame particles and smoke particles, position them at the bottom of the tower or below the conductor, and export them as FBX format.

[0025] Plant growth: Use Blender's "Curves + Array" function to create shrub / tree models, set the trunk height to 5-15m and the branch tilt angle to 10-45 degrees, make the branches close to or wrap around the guide wires, and export as FBX format; Add multiple virtual cameras in Unity3D, set the resolution to 3840×2160 to ensure image clarity; set the shooting angle to be the same as the real scene, and set the camera to automatically switch according to the preset angle and distance; enable batch rendering, name the captured images according to "abnormal type-angle-distance", such as "broken leg-0 degrees-close view-3m.jpg", and save them to the specified folder; The output images are inspected to ensure that abnormal features are clear, the scene is free of distortion, and the lighting is natural. Unqualified images are retaken.

[0026] S6. Perform the preprocessing operation in step S2 on the newly constructed abnormal samples, and then perform the filtering operation in step S3 to obtain usable abnormal samples and send them into the abnormal sample library.

[0027] S7. Separate the abnormal scene from the abnormal sample and fuse it with the background image in the independent background library to obtain a new abnormal sample. Then, perform step S6 on the new abnormal sample. The detailed process is as follows: S701. Use an image separation algorithm to separate abnormal samples in the abnormal sample library to obtain the abnormal scene and background image; S702: Use AI to generate more background images in batches, and all background images are combined into a background image set; S703. After randomly arranging and combining the abnormal scene with the background images in the background image set, the images are fused to obtain new abnormal samples. The fusion process is as follows: (4); in, The merged pixel values , These are the pixel values ​​of the abnormal scene and the background image, respectively. Weights for abnormal scenarios; S704. Use the GAN repair network model to eliminate splicing traces in the fused anomalous samples and output usable anomalous samples; after training, the GAN repair network model can identify and remove splicing traces, and then output seamless images for filling. S705. Perform step S6 on the abnormal samples after processing by the GAN repair network model.

[0028] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A method for staging and collecting abnormal on-site images from a normal sample library, used to provide training samples needed for anomaly detection in 35kV overhead lines, characterized in that, Includes the following steps: S1. Collect raw image sequences and construct a normal sample library; S2. Preprocess the images in the normal sample library; S3. Perform abnormal sample screening operation, filter out abnormal samples from the normal sample library, and send the abnormal samples into the abnormal sample library. S4. Classify abnormal samples according to the type of abnormality; S5. Construct new abnormal samples based on the selected abnormal samples; S6. Perform the preprocessing operation in step S2 on the newly constructed abnormal samples, and then perform the filtering operation in step S3 to obtain usable abnormal samples and send them into the abnormal sample library. S7. Separate the abnormal scene from the abnormal sample and fuse it with the background image in the independent background library to obtain a new abnormal sample, and perform step S6 on the new abnormal sample.

2. The method for staging and collecting abnormal scenes in a normal sample library according to claim 1, characterized in that, The detailed process of step S1 is as follows: Using drones equipped with high-resolution cameras, images were collected from sections of overhead power lines. During the collection process, the drone's flight altitude and angle were adjusted to capture raw images and build a normal sample library. The drone is also equipped with temperature and humidity sensors and light sensors to record environmental data during the collection process.

3. The method for staging and collecting abnormal scenes in a normal sample library according to claim 1, characterized in that, The detailed process of step S2 is as follows: S201. Perform Gaussian filtering on the image to remove noise using a Gaussian kernel. The Gaussian kernel function is: （1）； in, Standard deviation x and y These are the horizontal and vertical coordinates, respectively; S202. Sharpen the filtered image to highlight the texture and contours in the image.

4. The method for staging and collecting abnormal scenes in a normal sample library according to claim 1, characterized in that, The detailed process of step S3 is as follows: S301. Set the grayscale mutation threshold to identify provisional outliers; Grayscale mutation threshold The threshold will be dynamically adjusted according to the collected light intensity data. Specifically, when the light intensity is 500-3000 lux, the grayscale change threshold is set to 55, and when the light intensity is 3000-10000 lux, the grayscale change threshold is set to 50. When a pixel meets the judgment condition, the pixel is recorded as a temporary outlier. The judgment condition is as follows: （2）； in, L This refers to the grayscale pixel value of a certain pixel. This is the average pixel grayscale value of the points adjacent to this point; S302. Form a connected region P from consecutive provisional outliers. Filter by area feature, retaining regions that meet the outlier screening criteria as suspected outlier regions. The screening criteria are: （3）； in, Let P be the area. , These are the minimum and maximum area thresholds, respectively; S303. Perform preliminary screening of abnormal samples: If there are no suspected abnormal areas in the sample, the image is directly determined as a normal sample and sent back to the normal sample library; if there are suspected abnormal areas in the sample, proceed to step S304. S304. Manually review samples with suspected abnormal areas: If the image in the sample is abnormal, send it to the abnormal sample database; otherwise, send it back to the normal sample database.

5. The method for staging and collecting abnormal scenes in a normal sample library according to claim 4, characterized in that, The types of anomalies include: faults in the overhead line itself, foreign objects entangled in the transmission line, smoke and fire in the line and surrounding environment, and excessively rapid growth of vegetation around the overhead line; Faults in overhead power lines include: broken strands, abnormal lead wires, and tilted towers.

6. The method for staging and collecting abnormal scenes in a normal sample library according to claim 1, characterized in that, The detailed process of step S5 is as follows: S501. Use an image feature extraction algorithm to extract abnormal features from abnormal samples; S502. Based on the extracted abnormal features, perform a staged photography operation to obtain new abnormal samples; Staged photography includes both staged photography in real-world settings and modeling.

7. A method for staging and collecting abnormal scenes in a normal sample library according to claim 5 or 6, characterized in that, The process of posing for the photos is as follows: Construct staged shooting scenes, create different environments, add auxiliary elements, add abnormal features of various abnormal types to the staged shooting scenes, determine the shooting angle and distance, and collect images of each abnormal state in different environments in close-up, medium-range and long-range according to the shooting angle and distance; The shooting angle is related to the type of anomaly; During the data acquisition process, keep the camera's focal length, aperture, ISO, and other parameters consistent.

8. The method for staging and collecting abnormal scenes in a normal sample library according to claim 7, characterized in that, The detailed process of step S7 is as follows: S701. Use an image separation algorithm to separate abnormal samples in the abnormal sample library to obtain the abnormal scene and background image; S702: Generate more background images, and all background images are combined into a background image set; S703. After randomly arranging and combining the abnormal scene with the background images in the background image set, the images are fused to obtain new abnormal samples. The fusion process is as follows: （4）； in, The merged pixel values , These are the pixel values ​​of the abnormal scene and the background image, respectively. Weights for abnormal scenarios; S704. Use the GAN repair network model to eliminate splicing traces in the fused abnormal samples and output usable abnormal samples. S705. Perform step S6 on the abnormal samples after processing by the GAN repair network model.

9. The method for staging and collecting abnormal scenes in a normal sample library according to claim 8, characterized in that, In step S702, AI is used to generate background images in batches.

10. A method for staging and collecting abnormal scenes in a normal sample library according to claim 4, characterized in that, The manual review process is as follows: For any sample, professional inspectors will check the suspected abnormal areas one by one to determine whether they are real abnormalities or interference. If there is at least one real abnormality in the sample, the sample will be judged as an abnormal sample; otherwise, it will be judged as a normal sample.

Images