Line equipment defect marking method based on unmanned aerial vehicle electric power line image

Abstract

The invention discloses a line equipment defect labeling method based on an unmanned aerial vehicle power line image, and relates to the technical field of defect labeling, and the method comprises the following steps: obtaining a historical picture of a power transmission line, and carrying out the preprocessing of the historical picture; performing feature extraction on the preprocessed historical picture by using an image segmentation technology to obtain a line image feature value; manually marking defect pictures in the historical pictures to obtain the defect pictures, and marking the line image feature values of the defect pictures as defect image feature values; and taking the defect image feature value and the line image feature value as a training set, and performing model training by using a convolutional neural network to obtain a power line defect identification model. According to the defect identification algorithm, various defects of the power line equipment can be accurately detected and labeled, and the detection rate and accuracy of equipment faults are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of defect annotation, and more specifically, to a method for annotating line equipment defects based on unmanned aerial vehicle power line images. Background Art

[0002] With the continuous advancement of science and technology, drone technology is gradually showing its strong application potential in various industries, especially in the field of equipment inspection and maintenance in the power industry. Traditional power equipment inspection usually relies on manual inspections and periodic inspections, which are inefficient, costly, and prone to human errors. In order to overcome these challenges, a method for marking equipment defects in power line images based on drones has emerged.

[0003] In recent years, the rapid development of drone technology has brought new possibilities for solving power equipment monitoring. Drones are fast, flexible and efficient, and can quickly cover a large area in the air to obtain high-resolution images of power lines. This method not only significantly reduces the manpower and time costs of inspections, but also enables comprehensive monitoring of complex terrain and inaccessible areas, providing strong technical support for the safe operation of power equipment. Drones are equipped with high-resolution cameras and advanced sensors, which can efficiently and quickly collect images of power lines and their equipment. These images not only cover perspectives from high altitude to various angles, but also can obtain data under different weather and lighting conditions. Next, computer vision and image processing algorithms such as object detection, image segmentation and feature extraction are used to analyze these images to identify possible equipment defects on power lines, such as tower tilt and insulator damage. In addition, the application of machine learning and deep learning technologies enables the system to automatically learn and optimize based on a large amount of data, further improving the accuracy and efficiency of detection. The equipment defect labeling method based on drones can not only improve the efficiency of power equipment inspections, but also achieve real-time monitoring and early warning. Once the system detects a problem with the equipment, it will immediately send out an alarm to notify maintenance personnel so that they can take timely repair measures and reduce the risk of power system failure. Overall, this technology not only improves the efficiency and reliability of equipment management in the power industry, but also greatly reduces the safety risks of personnel participating in the inspection process.

[0004] The above disclosed technical solutions have at least the following technical problems: Power line equipment may have many types of defects, such as fracture, corrosion, looseness, etc., and these defects appear in complex and diverse forms in images. The annotation algorithm needs to be able to identify and accurately annotate these different types of defects, which poses a challenge to the robustness and generalization ability of the algorithm.

[0005] Secondly, the amount of power line image data taken by drones is huge, and it is necessary to be able to quickly process and analyze this data in order to promptly discover and deal with defects in line equipment. Real-time processing requires algorithms to have efficient computing and data processing capabilities.

[0006] In view of the above problems, the present invention proposes a solution. Summary of the invention

[0007] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a method for marking line equipment defects based on UAV power line images, which solves the problems raised in the above-mentioned background technology through a convolutional neural network.

[0008] To achieve the above object, the present invention provides the following technical solutions: A method for marking line equipment defects based on UAV power line images, comprising the following steps: Obtain historical pictures of the transmission line and perform preprocessing; use image segmentation technology to extract features from the preprocessed historical pictures to obtain line image feature values; manually annotate defective pictures in the historical pictures to obtain defective pictures, and annotate the line image feature values ​​of the defective pictures as defective image feature values; use the defective image feature values ​​and the line image feature values ​​as training sets and test sets and use a convolutional neural network to perform model training to obtain a power line defect recognition model; input the real-time transmission line picture into the power line defect recognition model, perform defect recognition on the real-time transmission line picture, and annotate it.

[0009] In a preferred embodiment, the preprocessing includes image screening, image cropping and compression, denoising, adjusting brightness and contrast, and image enhancement.

[0010] In a preferred embodiment, the image segmentation technology is specifically as follows: weighted averaging the values ​​of the red, green and blue channels of each pixel of the historical image to obtain a grayscale image; comparing a preset threshold with the grayscale value of the pixel in the grayscale image; dividing the pixels into two categories: marking pixels with grayscale values ​​greater than the preset threshold as targets, and marking pixels with grayscale values ​​less than the preset threshold as backgrounds; combining the pixels of the two categories to generate a binary image; the binary image includes a target part and a background part, the target part is represented by white, and the background part is represented by black.

[0011] In a preferred embodiment, the specific method for obtaining the line image feature value is as follows: use an edge detection algorithm to extract the line edge in the binary image; calculate the straight line features of the line edge based on the Hough transform straight line detection algorithm; perform feature extraction on the straight line features to obtain a first feature; and perform feature post-processing on the first feature to remove unnecessary features and merge related features to obtain the line image feature value; the feature extraction operation includes calculating the length, angle, position information of the straight line, and extracting the endpoint coordinates of the straight line.

[0012] In a preferred embodiment, all the different types of defects include tower foundation defects, grounding wire defects, damper defects, three-plate defects, insulator defects, grading ring defects, contact terminal defects and foreign body defects.

[0013] In a preferred embodiment, the specific method for obtaining the power line defect recognition model is as follows: using defect image feature values ​​and line image feature values ​​as training sets; extracting convolution features from the training sets based on a convolutional neural network to obtain a number of feature maps, and establishing an initial training model; alternately training the initial training model according to a number of feature maps and loss functions, and performing parameter optimization and structural adjustment on the power line defect recognition model to obtain a final power line defect recognition model.

[0014] The technical effects and advantages of the method for marking defects of power line equipment based on drone power line images provided by the present invention are as follows: 1. The present invention can accurately detect and mark various defects of power line equipment, such as breakage, corrosion, looseness, etc., by using defect recognition algorithms based on image processing and computer vision technology. These algorithms are optimized through a large amount of labeled data during the training process, and can achieve higher recognition accuracy, thereby improving the detection rate and accuracy of equipment faults.

[0015] 2. The present invention can quickly and efficiently obtain image data of power lines by using drones, which greatly saves time and labor costs compared to traditional manual inspections or ground vehicle inspections. This automated inspection method can frequently monitor line equipment, reduce the impact of human factors, and improve inspection efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 The present invention is a structural schematic diagram of a method for marking defects of line equipment based on UAV power line images. DETAILED DESCRIPTION

[0017] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] Embodiment 1, Figure 1 The present invention provides a method for marking defects of power line equipment based on UAV power line images, comprising the following steps: Step S1, obtaining a picture around the power transmission line and performing preprocessing; Preprocessing includes image screening, image cropping and compression, denoising, adjusting brightness and contrast, and image enhancement.

[0019] Inspection data generated by drones during the inspection of power facilities and transmission line corridors of transmission lines. Since the sample size of defect data is much smaller than that of normal data, the data of this invention does not use a single inspection line or a single engineering data as the basic data, but collects drone inspection data over the years. This also avoids the problem of low recognition rate in experiments and applications due to a single scene, and further improves the training model capabilities. In the process of collecting raw data, the original images of drone inspections are first classified according to the transmission inspection specifications by manual interpretation. Secondly, the defect data is selected, and eight types of tower base defects, grounding wire defects, damper defects, three-plate defects, insulator defects, equalizing ring defects, contact terminal defects and foreign body defects are selected as the research objects of this invention. Finally, 5,100 images are selected as experimental data.

[0020] The pictures taken by drone inspections are affected by the natural environment, the shooting equipment carried, the jitter of the drone, the technical level of the pilot, etc. The range of the pictures taken is large. In addition, the definition of the shooting camera used by drone inspections is generally good, and the picture quality is very high. The size of a single image file ranges from several megabytes to tens of megabytes, which also increases the workload of later image recognition. For this reason, the experimental pictures were checked one by one, and the images that did not involve the defect were cropped, retaining the main part of the defect and the necessary background. In addition, in order to further reduce the space occupied by the image, the image compression tool is used to further perform lossless compression on the cropped image, and the compressed image is half the size of the image before cropping.

[0021] Image enhancement is to increase the recognition of defect data from drone inspections, while weakening information irrelevant to the defect data, thereby improving the defect recognition accuracy and efficiency of subsequent experiments.

[0022] Step S2 uses image segmentation technology to extract features from the preprocessed historical images to obtain route image feature values; Image segmentation technology is to divide the image into multiple regions with semantic information, each region corresponds to a category, which can be used to distinguish different types of equipment in power line images, such as insulators, conductors, poles, etc. Each equipment type can be assigned a specific category label to achieve mask generation of the region.

[0023] The specific implementation of the image segmentation technology is: According to the preprocessed historical image, threshold segmentation in the image segmentation technology is selected; the values ​​of the three channels of red, green and blue of each pixel of the historical image are weighted averaged to obtain a grayscale image; a preset threshold is applied to the grayscale image, and the grayscale value of each pixel of the grayscale image is compared with the preset threshold, and the pixels are divided into two categories: pixels greater than the preset threshold are marked as targets, and pixels less than the preset threshold are marked as backgrounds; according to the application result of the preset threshold, a binary image is generated; in the binary image, the target part is represented by white, and the background part is represented by black.

[0024] The specific steps of feature extraction are as follows: Using an edge detection algorithm to extract line edges in the binary image; for the image after edge detection, using a Hough transform line detection algorithm to detect straight line features in the image; based on the detected straight line features, performing feature extraction operations and storing the extracted features in the form of numerical values; performing post-processing and analysis of the features, removing unnecessary features, and merging related features to obtain line image feature values; The feature extraction operation includes calculating the length, angle, and position information of the straight line, and extracting the coordinates of the endpoints of the straight line.

[0025] The advantages of using image segmentation technology and feature extraction are as follows: Improved accuracy: Image segmentation technology can accurately separate different device or object areas in an image, avoiding confusion and errors that may occur when performing global feature extraction on the entire image. This ensures that feature extraction in each device area is targeted and accurate.

[0026] Local feature analysis: Feature extraction of each device region can focus on the local characteristics and details of the device without interference from other regions. This helps to gain a deeper understanding of the status, problems or specific attributes of each device.

[0027] Enhanced detection capabilities: By extracting features from each device area, various defects or anomalies can be detected and identified more effectively.

[0028] Personalized processing: Personalized feature extraction and analysis of each device area can be performed according to the functions and usage conditions of different devices. This can optimize maintenance plans, improve fault detection rates, and reduce efficiency reduction caused by unified processing.

[0029] Visualization and interpretability: The segmented equipment areas and their feature extraction results can be more intuitively presented and interpreted, allowing operators or analysts to more easily understand and evaluate the status of the equipment. This visualization can support decision-making processes and problem solving.

[0030] Step S3, manually annotating the defective images in the historical images, and performing quality control and verification to obtain defective images, and annotating the line image feature values ​​of the defective images as defective image feature values; Manual labeling of defects refers to the detailed labeling and classification of defects detected by computer vision and machine learning algorithms by professional operators or domain experts according to pre-set standards and criteria. This process not only includes determining the type, location and size of the defect, but may also involve other specific attributes, such as the severity of the defect or the assessment of its impact on production.

[0031] The specific implementation of manual annotation is as follows: Select a professional annotation tool based on the number of defect images in the historical images; formulate clear annotation rules and guidelines, which include the definition of defect types, detailed annotation steps, and the degree of standardization of annotation; select a set of images that need to be annotated from the historical images, and the image set includes all different types of defects; load the images into the annotation tool, and the annotator uses the tool to annotate the defects in the images, including the defect type and location information corresponding to each image; and verify and review the annotation results.

[0032] The quality control and verification methods are as follows: Determine clear annotation specifications and standards, including the definition of defect categories, annotation formats, and how to use annotation tools. All personnel involved in annotation need to strictly follow these specifications to ensure the consistency and comparability of annotation results. Use a double annotation method, that is, two or more independent annotators annotate the same batch of data. Then compare their annotation results to check for inconsistencies or errors. Regularly check and review the consistency of annotation results. This includes comparing the consistency between annotators, finding possible mis-annotations or misinterpretations, and making necessary corrections. For particularly complex or ambiguous situations, domain experts can be introduced for review. Experts can provide authoritative judgments on the annotation results to ensure that the annotations meet the actual situation and standard requirements. Conduct systematic training for annotators to ensure that they understand the annotation specifications and standards. At the same time, establish a feedback mechanism so that annotators can understand their annotation quality in a timely manner and adjust and improve their annotation skills in a timely manner. Introduce metrics and indicators to evaluate annotation quality, such as accuracy, recall, F1 score, etc. These metrics can help identify problems in the annotation process and provide guidance for improving the annotation process. Continuously improve the annotation process and annotation specifications based on feedback from the quality control and verification process. This includes revising annotation specifications, updating annotation tools, strengthening annotation personnel training and other measures to improve overall annotation quality and efficiency.

[0033] The advantages of quality control and verification of annotation results are as follows: Improve data quality and accuracy: The quality control and verification process can effectively improve the quality and accuracy of the labeled data. By verifying the consistency between different annotators and the matching degree between the annotation results and the actual situation, possible errors and deviations can be discovered and corrected to ensure the accuracy of the annotation results.

[0034] Enhance algorithm performance and stability: The quality of labeled data directly affects the training and performance of machine learning algorithms. High-quality labeled data can provide more reliable training samples, help improve the generalization ability and stability of the algorithm, and make it more stable and reliable in practical applications.

[0035] Reducing AI system errors: Through quality control and verification, AI system errors caused by poor quality of labeled data can be effectively reduced. These errors may have a negative impact on the decision-making and application of the system. By optimizing the quality of labeled data, this risk can be significantly reduced.

[0036] Improve production efficiency and cost-effectiveness: High-quality labeled data means more accurate defect detection and classification results, which can help companies reduce defective product rates, improve product quality, and reduce production costs and resource waste. This is especially important for large-scale production and automated production lines.

[0037] Support continuous improvement and optimization: Quality control and verification processes are not just one-time tasks, they also provide a basis for continuous improvement and optimization. Through feedback mechanisms and data analysis, potential problems in the annotation process can be identified, and adjustments and improvements can be made in a timely manner to further improve the quality and efficiency of the annotation results.

[0038] Step S4, using the defect image feature values ​​and the line image feature values ​​as training sets and using a convolutional neural network to perform model training to obtain a power line defect recognition model; The specific steps of power line defect recognition model training are as follows: The defect image feature values ​​and the line image feature values ​​are divided into a training set and a test set; the convolution features of the training set are extracted based on a convolutional neural network to obtain a number of mappings, and an initial training model is established; the initial training model is alternately trained according to a number of feature mappings and loss functions to obtain an initial power line defect recognition model; the test set is input into the initial power line defect recognition model to obtain an output defect type, and the output defect type is compared with the defect type in the test set; the power line defect recognition model is again parameter optimized and structurally adjusted using different image data to obtain a final power line defect recognition model.

[0039] The feature extraction based on the convolutional neural network is specifically as follows: Use multiple convolution kernels, each of which performs a convolution operation on the input image and activates it through the ReLU function to obtain a feature map; Pool each feature map to obtain the pooled features, and add the convolution kernel again to iterate until the number of iterations reaches the preset number of pooling times; after the iteration is completed, output several feature maps.

[0040] The output calculation formula of the feature map is:

[0041] In the formula, is the i-th feature map of the p-th pooling layer, pool is the pooling operation, is the activation function, is the weight of the feature to be trained, is the feature to be trained, is the bias term.

[0042] The specific calculation formula of the activation function is:

[0043] The labeled data is input into the convolutional neural network to train the model. The training uses convolution and pooling to update the model parameters, iterates for 200,000 steps, and finally completes the model training.

[0044] The trained convolutional neural network model is tested using the test data set to obtain its accuracy. Based on the accuracy, it is determined whether to add data and change training parameters to continue training the model.

[0045] For each type of power transmission equipment, samples of normal power transmission equipment and defective power transmission equipment are collected, and each image is labeled with the defect type. 9 / 10 of the total number of images are used as the training data set, and the remaining 1 / 10 of the images are used as the test data set.

[0046] The classifier is trained for each type of power transmission equipment using the training sample data prepared above.

[0047] For each type of power transmission equipment defect classifier, the accuracy is tested using the test data. Based on the accuracy, it is determined whether it is necessary to add samples or modify training parameters to continue training.

[0048] The above-mentioned trained model and the defect classifier for each type of power transmission equipment are combined for use. The photos of the power transmission line equipment obtained by each drone inspection are first sent to the convolutional neural network to extract the areas containing the power transmission equipment, and determine the type of power transmission equipment contained in each area. Then, according to the type, it is sent to the defect classifier of the equipment to determine whether it has defects and the type of defects. If there are defects, the location of the defective equipment and the type of defect are marked in the photos of the power transmission line equipment.

[0049] We extracted some typical images from the existing machine recognition results and recreated new training samples. We continuously updated the model parameters of the convolutional neural network and defect classifier based on the original model, so that the model is continuously optimized and the accuracy is continuously improved.

[0050] The basic network is an indispensable part of computer vision tasks. It extracts image features for subsequent tasks. The present invention adopts the residual learning idea to solve the problem of feature loss to obtain accuracy. The candidate area is mainly to mine candidate objects through neural network learning, and finally generate feature vectors that are passed to the category decision layer and position regression layer through sliding windows; the category decision layer is mainly used to determine whether the object belongs to the foreground or the background; the position regression layer is used to roughly locate the object.

[0051] The model training steps are: The pre-trained image network is used to initialize the region candidate network, and the region candidate network is fine-tuned to obtain the region decision network; the decision network is initialized using the image network, and the region candidate results are used as the input of the refinement network; finally, the fully connected layer in the refinement network is fine-tuned so that the two networks can share the convolutional layer.

[0052] The number of convolution kernel channels in the classification layer is 18, and that in the regression layer is 36. The position information of each anchor point is described by the coordinates x, y, and the length and width values ​​w, h.

[0053] The loss function can be expressed as:

[0054] Where i is the index of the anchor point; is the predicted probability; If the anchor is positive, i is 1, and if the anchor is negative, i is 0; is the location information of the candidate area; is the position information of the positive anchor point; is the number of classification samples; is the number of regression samples; is the classification loss function, is the position regression loss function.

[0055] is the classification loss function, is the position regression loss function, which can be expressed as:

[0056]

[0057] Where R is the robust loss function

[0058] The robust loss function is designed to reduce the impact of outliers or noise on model performance.

[0059] The probability of the pooling layer area is P = (p0, p1,..., p10). For class u, its loss function can be expressed as

[0060] The output of the regression layer is an array of size 4*k. , , , ) 、 = ( , , , ) represent the true and predicted values ​​of position recognition respectively, and the loss function is:

[0061] The loss function in the training refinement network consists of a decision loss function and a position regression loss function, which can be expressed as:

[0062] in, , They are position regression and decision loss functions respectively; is a hyperparameter, =1.

[0063] Step S5: input the real-time transmission line image into the power line defect recognition model, perform defect recognition on the real-time transmission line image, and mark it.

[0064] The advantages of applying the model to actual circuit line images are as follows: Improve detection accuracy and efficiency: The trained model can accurately identify various defects in power line images through convolutional neural network technology, such as broken insulators, tilted towers, and fallen wires. Compared with manual inspection, the model can process more images in a shorter time and generally provide more accurate detection results.

[0065] Reduce labor and time costs: Automated defect detection systems can greatly reduce human resource investment, and no longer require a large number of people for line inspection and defect identification. The time and cost saved can be used for other more creative and high-value tasks.

[0066] Improve safety and reliability: Timely detection and repair of potential problems in power lines can effectively reduce the risk of line failures and accidents, and improve the overall safety and reliability of the power system. Especially after major weather changes or natural disasters, the model can quickly detect potentially damaged line components, which helps to repair and restore power supply in a timely manner.

[0067] Real-time monitoring and early warning functions: Models deployed in actual environments can achieve real-time monitoring and early warning functions. Once a defect is detected, the system can immediately issue an alarm to notify relevant personnel for further inspection and repair, thereby handling the problem more promptly and reducing potential losses and impacts.

[0068] Data-driven decision support: Collecting and analyzing data output by the model can help power companies make more accurate maintenance plans and decisions. Through long-term data accumulation, trend analysis and prediction can be performed to optimize maintenance cycles and resource allocation, thereby improving the overall efficiency and operating costs of the power system.

[0069] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0070] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination. When implemented by software, the above embodiments may be implemented in whole or in part in the form of a computer program product.

[0071] Those of ordinary skill in the art will appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0072] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0073] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0074] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for marking line equipment defects based on UAV power line images, characterized in that: The following steps are involved: Obtain historical images of transmission lines and perform preprocessing; The pre-processed historical pictures are subjected to image segmentation technology to extract features from the historical pictures and obtain route image feature values; Manually annotate defective images in historical images to obtain defective images, and annotate line image feature values ​​of the defective images as defective image feature values; The defect image feature values ​​and line image feature values ​​are used as training sets and a convolutional neural network is used for model training to obtain a power line defect recognition model. The real-time transmission line image is input into the power line defect recognition model, and the defects of the real-time transmission line image are recognized and annotated.

2. The method for marking line equipment defects based on UAV power line images according to claim 1 is characterized in that: The preprocessing includes image screening, image cropping and compression, denoising, adjusting brightness and contrast, and image enhancement.

3. The method for marking line equipment defects based on UAV power line images according to claim 2 is characterized in that: The image segmentation technology is specifically: The values ​​of the red, green and blue channels of each pixel of the historical image are weighted averaged to obtain a grayscale image; Comparing a preset threshold with the grayscale value of a pixel in the grayscale image; The pixels are divided into two categories: the pixels with gray values ​​greater than a preset threshold are marked as targets, and the pixels with gray values ​​less than the preset threshold are marked as background; Combine pixels of two categories to generate a binary image; The binary image includes a target part and a background part, the target part is represented by white, and the background part is represented by black.

4. The method for marking line equipment defects based on UAV power line images according to claim 3 is characterized in that: The specific method for obtaining the line image feature value is as follows: Extracting line edges in the binary image using an edge detection algorithm; Calculating the straight line features of the line edge based on the Hough transform straight line detection algorithm; Performing feature extraction on the straight line feature to obtain a first feature; and performing feature post-processing on the first feature, removing unnecessary features, and merging relevant features to obtain a line image feature value; The feature extraction operation includes calculating the length, angle, and position information of the straight line, and extracting the coordinates of the endpoints of the straight line.

5. The method for marking line equipment defects based on UAV power line images according to claim 4 is characterized in that: The defect images include tower base defects, grounding wire defects, damper defects, three-plate defects, insulator defects, equalizing ring defects, contact terminal defects and foreign body defects.

6. The method for marking line equipment defects based on UAV power line images according to claim 5 is characterized in that: The specific method for obtaining the power line defect recognition model is as follows: The defect image feature values ​​and the line image feature values ​​are used as training sets; Extracting convolution features from the training set based on a convolutional neural network to obtain a number of feature maps, and establishing an initial training model; The initial training model is trained alternately according to a number of feature maps and loss functions, and the parameters of the power line defect recognition model are optimized and the structure is adjusted to obtain the final power line defect recognition model.

7. The method for marking line equipment defects based on UAV power line images according to claim 6 is characterized in that: The calculation formula of the feature map is: In the formula, is the i-th feature map of the p-th pooling layer, pool is the pooling operation, is the activation function, is the weight of the feature to be trained, is the feature to be trained, is the bias term.

8. The method for marking line equipment defects based on UAV power line images according to claim 7 is characterized in that: The calculation formula of the loss function is: Where i is the index of the anchor point; is the predicted probability; If the anchor is positive, i is 1, and if the anchor is negative, i is 0; is the location information of the candidate area; is the position information of the positive anchor point; is the number of classification samples; is the number of regression samples; is the classification loss function, is the position regression loss function.

9. The method for marking line equipment defects based on UAV power line images according to claim 8 is characterized in that: The classification loss function And the regression loss function The calculation formula is: Where R is the robust function.

10. The method for marking line equipment defects based on UAV power line images according to claim 9 is characterized in that: The loss function in the training refinement network is: in, , They are position regression and decision loss functions respectively; is a hyperparameter, =1.

Images