Power emergency unmanned aerial vehicle power transmission line defect detection method

Abstract

The invention discloses a power emergency unmanned aerial vehicle power transmission line defect detection method, which comprises the following steps: receiving a power transmission line image acquired by a power emergency unmanned aerial vehicle, and preprocessing the image to accurately identify and distinguish defects and backgrounds; carrying out feature extraction on the image, and obtaining statistical property features through texture analysis, a gray-level co-occurrence matrix, a gray-level difference method, discrete wavelet transform and fast Fourier transform methods; fusing and selecting the extracted features to form an optimized feature set; using the feature set to train a super-pixel segmentation model based on deep learning for power transmission line defect detection; and adopting an average threshold method to carry out confidence analysis on the deep learning model after super-pixel segmentation. Through multi-domain feature extraction optimization, high-precision defect detection and robust confidence analysis, the accuracy and efficiency of power transmission line defect detection are significantly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of electric power emergency unmanned aerial vehicle inspection, and in particular to a method for detecting defects in electric power emergency unmanned aerial vehicle transmission lines. Background Art

[0002] As an indispensable and key part of the power system, transmission lines undertake the important task of electric energy transmission. However, their operating environment is complex and changeable, and they are often threatened by natural factors (such as wind, rain, lightning, ice and snow, etc.) and human factors (such as construction damage, theft, etc.), resulting in various defects in transmission lines. These defects not only affect the stability of power supply, but may also cause serious safety accidents. Therefore, timely and accurate defect detection of transmission lines is of great significance.

[0003] Transmission line defects can be roughly divided into two categories, namely, intrinsic defects and external defects, based on their causes and manifestations. Intrinsic defects mainly involve the defects or displacement of key components such as the foundation towers, ground wires, insulators, hardware, and grounding devices of the line body, which directly affect the mechanical and electrical performance of the line. External defects refer to potential threats caused by changes in the external environment, such as illegal buildings, construction machinery, and tall plants in the line protection area, which may cause damage to the line due to being too close or accidentally contacted. These defects are further subdivided into general defects, major defects, and emergency defects according to their severity and urgency, in order to facilitate hierarchical management and rapid response.

[0004] Traditional methods of power transmission line defect detection mainly include manual inspection (human inspection) and mechanical inspection (machine inspection). Human inspection relies on professionals to inspect on the ground or on towers, supplemented by professional tools such as ground resistance testers and infrared testers. Although some defects can be found, they are inefficient, labor-intensive and difficult to cover all areas. Machine inspection, especially helicopter inspection, can overcome geographical barriers and achieve rapid line inspection, but its high cost and complex operation limit its popularization and application.

[0005] In recent years, with the rapid development of power emergency drones and video surveillance technology, transmission line inspections have ushered in new changes. Power emergency drones have been widely used in transmission line inspections due to their simple operation, safety, reliability, and beyond-visual-range flight, significantly improving inspection efficiency and quality. However, although a large number of experts and scholars have conducted in-depth research on the application of power emergency drones in transmission line defect detection and proposed a variety of detection methods, the existing technology still faces many challenges. For example, there is too much redundant information in the original image data, resulting in the extracted image features unable to fully reflect the actual situation of the defects; the accuracy of the deep learning model in complex scenarios needs to be improved, and it is difficult to accurately identify various defects; and the algorithm is highly complex and real-time performance is difficult to guarantee.

[0006] In summary, although power emergency drones have made significant progress in the field of transmission line defect detection, there is still a need to further optimize and improve existing technologies to improve the accuracy, real-time and efficiency of detection, so as to better ensure the safe and stable operation of the power grid. Summary of the invention

[0007] In view of this, the present invention proposes a power emergency UAV power transmission line defect detection method to solve the problems existing in the prior art of inaccurate image processing and the need to improve model accuracy.

[0008] The specific technical solutions of the present invention are as follows:

[0009] A method for detecting power transmission line defects using an electric power emergency drone, comprising:

[0010] Step 1: Receive the transmission line image collected by the power emergency drone, and pre-process the image by grayscale scaling, segmentation, variance screening, and pseudo target elimination to accurately identify and distinguish defects from the background;

[0011] Step 2, extracting features in the time domain, space domain and frequency domain of the preprocessed image, obtaining statistical property features through texture analysis, gray-level co-occurrence matrix, gray-level difference method, discrete wavelet transform and fast Fourier transform; fusing and selecting the extracted features to form an optimized feature set;

[0012] Step 3, using the feature set to train a deep learning-based superpixel segmentation model for power transmission line defect detection, the model adopts a deep learning architecture and is implemented through an encoder-decoder two-layer architecture;

[0013] Step 4: Use the average threshold method to perform confidence analysis on the deep learning model after superpixel segmentation.

[0014] Furthermore, the variance screening in step 1 includes: based on variance sorting, removing image blocks with variances lower than a set threshold.

[0015] Furthermore, the pseudo target elimination in step 1 includes: performing mean calculation and pixel traversal on the remaining image blocks after variance screening, and preliminarily marking suspected defect points; and further eliminating pseudo target points through morphological operations to complete image preprocessing.

[0016] Furthermore, the image texture analysis method in step 2 is: converting the subjective evaluation of smoothness and uniformity in the image into a numerical measurement, wherein the numerical measurement is obtained by calculating the intensity variation difference of the image pixels in spatial distribution.

[0017] Furthermore, the gray level co-occurrence matrix method in step 2 is: under the preset fixed direction and distance conditions, texture highlights are effectively realized by calculating and identifying the occurrence frequency of different pixel combinations on the analyzed image to show the various possible combinations of gray levels inside the image based on specific directions and distances.

[0018] Furthermore, the gray level co-occurrence matrix method involves calculating the gray level co-occurrence matrix at four different distances in the vertical, upward and downward directions, and performing a measurement based on the median value obtained in each analysis direction, while using a grid system to identify and emphasize different texture features in the image in the said direction for a specific distance value.

[0019] Furthermore, the grayscale difference method in step 2 includes: calculating the required density difference between the highest importance area and the lowest importance area in the power line image by contrast modification, and calculating statistical features using four spatial regions formed by all samples in the image and adjacent pixels and reference points, and performing optimization of the grayscale difference method distance calculation for each input image and its corresponding output image.

[0020] Furthermore, in step 3, during the training process, the cross-validation method is used to optimize the model hyperparameters, and a monitoring mechanism is introduced to prevent overfitting.

[0021] Furthermore, step 3 includes: the encoder processes the feature-extracted image through a convolutional layer to generate a high-level feature map, the decoder gradually upsamples these feature maps for prediction, and learns the mapping relationship by training a deep neural network, divides the image into grid cells as initial superpixel seeds in a soft-associative mapping manner, and limits the search to the adjacent grid cells of each pixel for allocation, thereby improving computational efficiency and finally obtaining a trained deep learning superpixel segmentation model.

[0022] Furthermore, step 4 includes: setting the overall mean of the texture feature values ​​extracted from the superpixels corresponding to the defect area in the image as a reference, constructing a confidence interval by calculating the sample mean and sample standard deviation of the defect texture feature values, and determining the threshold with the confidence interval, thereby realizing the confidence assessment of the deep learning model.

[0023] The beneficial effects of the present invention are:

[0024] (1) Multi-domain feature extraction optimization: The present invention has achieved a major breakthrough in image feature extraction. By adopting multiple feature extraction methods, it spans the time domain, spatial domain and frequency domain, greatly improving the effect of image feature extraction. This method not only effectively reduces the redundant information in the original image of the transmission line, such as background noise and irrelevant details, but also maximizes the retention of key features of the image, such as conductor morphology, insulator status, etc. This comprehensive and accurate feature extraction method provides a high-quality data foundation for subsequent defect detection.

[0025] (2) High-precision defect detection: In order to improve the accuracy and efficiency of transmission line defect detection, the present invention introduces a superpixel segmentation method based on deep learning. This method can automatically divide an image into multiple superpixel regions with similar characteristics, and accurately depict the region of interest in the image with the help of a convolutional neural network that is good at capturing complex patterns and features. Compared with traditional segmentation methods, this method not only improves the accuracy of defect identification, but also significantly improves the processing speed, achieving a dual improvement in efficiency and accuracy.

[0026] (3) Robust confidence analysis and decision support: In order to further enhance the reliability of the deep learning model, the present invention also introduces a statistical method to perform confidence analysis on the model. By performing statistical analysis on the probability distribution of the model output, the prediction confidence of the model for different categories can be evaluated. Based on these analysis results, the present invention establishes a series of robust criteria for classifying superpixels as the results of power line defect detection. These criteria not only take into account the prediction confidence of the model, but also combine the actual features and contextual information of the image, thereby ensuring the rationality of the algorithm's decision. In this way, the present invention effectively reduces the risk of misjudgment and missed judgment, and improves the performance and reliability of the entire system. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0028] Figure 1 It is a schematic diagram of the flow of the power emergency UAV power transmission line defect detection method of the present invention. DETAILED DESCRIPTION

[0029] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0030] The present invention proposes a method for power emergency UAV transmission line defect detection, which integrates image preprocessing, feature extraction in the time domain, spatial domain and frequency domain, superpixel segmentation technology training based on deep learning, and final confidence analysis. First, the transmission line images collected by the UAV are preprocessed by denoising, graying, etc. to improve the image quality. Subsequently, multi-dimensional features in the time domain, spatial domain and frequency domain are extracted from the preprocessed image. Next, a deep learning model is used to perform superpixel segmentation training on the feature image to automatically learn and identify potential defects. Finally, the reliability of the model prediction results is evaluated through confidence analysis to ensure the accuracy of the detection results. This set of methods effectively improves the efficiency and accuracy of transmission line inspections and provides a solid guarantee for the stable operation of the power system. The specific steps of this method can be as follows:

[0031] Step 1: Refined image preprocessing and preliminary defect identification.

[0032] During the power emergency inspection, the present invention first collects high-definition input images of the transmission line through drones, and then performs a series of carefully designed preprocessing steps on these images to ensure the accuracy and efficiency of subsequent defect detection. The preprocessing process starts with the key link of grayscale scaling, which aims to convert the input color image into a grayscale image. This not only simplifies the complexity of the image data, but also retains key information sufficient to identify defects.

[0033] Next, the image is evenly segmented into multiple small blocks. This step is critical because it breaks down the huge image data into more manageable units. When segmenting, special attention is paid to the fact that each small block can be composed entirely of background or a mixture of background and defects, but a small block consisting entirely of defects is never allowed to ensure the comprehensiveness and accuracy of the analysis.

[0034] Subsequently, the system calculates the variance value of each small block. Variance, as a statistic to measure the brightness change of pixels within an image block, plays an important role in distinguishing the background from the defect area. The variance value is used to arrange all image blocks in an ordered sequence from small to large. On this basis, the system intelligently removes the first 60% of image blocks with low variance values ​​in the sequence. These blocks mainly represent the background area, and their information contributes little to defect detection, so they are removed to reduce the computational burden of subsequent processing.

[0035] For the remaining 40% high-variance image blocks, the system further performs in-depth analysis. First, the mean of each image block is calculated as a benchmark for distinguishing between background and defects. Then, the system traverses each picture element (pixel) of each image block from left to right and from top to bottom, and preliminarily classifies the pixel as a background point or a defect point by comparing the brightness value of the pixel with the mean of the image block. If the pixel is greater than the mean of the image block, it is marked as a background point, otherwise it is marked as a defect point. Although this step is fast, it may cause misjudgment due to factors such as noise.

[0036] In order to eliminate these false target points that are misjudged, the system adopts a verification method based on a local window. Specifically, for each pixel that is initially marked as a defect, the system builds a 3×3 window with it as the core and counts the number of defective pixels in the window. If at least half of the pixels in the window are marked as defects, the core pixel is confirmed to be a real defect point; otherwise, it is re-marked as a background point. This step significantly improves the accuracy of defect detection and ensures the reliability of the preprocessing results.

[0037] Through the above series of carefully designed preprocessing steps, the present invention finally obtains the strictly screened and optimized power emergency drone transmission line images. These images not only remove most of the noise and redundant information, but also preliminarily distinguish the background and defect areas, laying a solid foundation for subsequent feature extraction and defect detection.

[0038] Step 2: Multi-dimensional feature extraction and in-depth processing.

[0039] After successfully completing the image preprocessing, the present invention further performs comprehensive feature extraction in the time domain, spatial domain and frequency domain on the preprocessed image. This process is used to obtain statistical features from the resized samples to perform image feature extraction, aiming to more effectively process a large amount of data during the model learning process by simplifying the features, thereby improving training efficiency and accuracy.

[0040] First, the present invention performs texture feature analysis on the image. By deeply analyzing the texture content of the image, the transmission line image area is divided into parts with different characteristics. This step not only considers the intuitive quality of the image, such as smoothness and uniformity, but also converts these quality features into the intensity changes of image pixels in space through quantitative means. This conversion makes the texture features of the image more intuitive and easy to process.

[0041] Next, the grayscale co-occurrence matrix technology is applied to image feature extraction. Given a fixed direction and distance, the grayscale co-occurrence matrix T can identify the frequency of occurrence of different pixel combinations on the image, thereby effectively revealing the texture characteristics of the image. In order to more comprehensively capture the texture information of the image, the present invention calculates the grayscale co-occurrence matrix in four directions (horizontal, vertical and two diagonal directions), and performs a comprehensive measurement based on the intermediate values ​​obtained in each direction. In addition, by introducing a grid system, the present invention is able to identify and highlight different texture features on the image in the vertical and oblique upward and downward directions for specific distance values. These features are displayed through the co-occurrence matrix M, revealing various possible combinations of gray levels within the image. Elements p and q are separated by a distance p with a specific direction k, and the matrix is ​​used and deployed in four directions. The specific formula can be expressed as Among them, δ p =∑ p，q p*M(p,q),δ q =∑ p，q q*M(p,q).

[0042] Grayscale difference method is another feature extraction method. It realizes image segmentation by determining the difference between the thickness of grayscale images by statistical grayscale. In the application of the present invention, the method pays special attention to the density difference between the most important area and the least important area in the power transmission line image. To this end, the present invention defines four spatial regions (upper, lower, left, and right) and calculates the statistical characteristics of the grayscale differences of pixels in these regions. These statistical characteristics not only reflect the grayscale changes of the image in different directions, but also provide important reference information for subsequent defect detection. Four spatial regions with a distance of h=8 between all image samples and adjacent pixels (x, y) and reference points in the power transmission line image are used to calculate statistical characteristics. For each input image u and the corresponding output image v, the grayscale difference method distance calculation is performed. The specific formula can be expressed as v(x, y)=|u(x, y)-v(x, y+h)|.

[0043] In order to further explore the detailed information in the image, the present invention also adopts discrete wavelet transform technology. Discrete wavelet transform represents a given discrete time series as wavelet coefficients, which reveal the characteristics of the image at different scales. In the application of the present invention, discrete wavelet transform is used to decompose the image into wavelet coefficients in symmetrical or biorthogonal regions. The form of wavelet representation is clearly defined by a recursive method, and the high-pass and low-pass elements are derived from the channel by applying the input time arrangement. This process not only simplifies the calculation process of wavelet transform, but also improves the efficiency and accuracy of feature extraction. Wavelet can be considered as two N-coefficient channels of finite impulse response, one used as a low-pass channel x j To cut off the check duplication, and the other constitutes a high-pass channel y jThe present invention adopts a recursive method to clearly define this form of wavelet representation, and applies the input time arrangement to initially derive high-pass g1(k) and low-pass g2(k) elements from the channel. The specific formula can be expressed as

[0044] Finally, the present invention uses the fast Fourier transform method to calculate and derive the statistical characteristics of all resized image samples. The specific formula can be expressed as Wherein, F represents the F coefficient channel; d(j) refers to the coefficient of j from 0 to F-1; W refers to the point on the complex plane; j refers to a natural number from 0 to F-1; and m refers to a complex number. As an efficient spectrum analysis tool, the fast Fourier transform can reveal the characteristics of the image in the frequency domain. By applying the fast Fourier transform formula, the present invention conducts an in-depth analysis of the spectral information of the image and extracts features that play a key role in defect detection. These features not only reflect the distribution of the image in the frequency domain, but also establish a close connection with the original information of the image.

[0045] After the above series of feature extraction techniques are used, the present invention finally obtains images after feature extraction. These images contain rich time domain, spatial domain and frequency domain feature information, which can improve the image feature extraction effect, greatly reduce the redundant information of the original image of the transmission line, and retain the features of the original image of the transmission line to the maximum extent, providing a solid data foundation for subsequent defect detection or classification tasks.

[0046] Step 3: Deep learning driven soft-correlation superpixel segmentation and model training.

[0047] In step 3 of the present invention, an innovative superpixel segmentation method based on deep learning is used for further processing of the image that has been feature extracted in step 2. First, we divide the image into uniform grid cells, each grid cell serves as the "seed" of the initial superpixel, and these seeds provide the basic framework for subsequent superpixel generation.

[0048] However, traditional hard assignment methods directly assign each pixel to the nearest seed, which is computationally complex and may lead to inaccurate segmentation results. To overcome this challenge, we train a deep neural network to learn the complex mapping relationship between pixels and superpixel seeds. This learning process is achieved through a soft-associative mapping mechanism, which allows pixels to be assigned to multiple adjacent superpixel seeds with a certain probability instead of a single hard assignment. This method not only improves computational efficiency, but also makes the segmentation results smoother and more accurate because it can better capture the complex structure and edge information in the image.

[0049] In the design of the deep learning model, an encoder-decoder two-layer architecture is adopted. The encoder part is responsible for processing the image after feature extraction and extracting high-level feature maps of the image through multiple convolutional layers. These feature maps contain key information of the image, such as edges, textures, and shapes, which are crucial for subsequent superpixel segmentation. The decoder part gradually upsamples these high-level feature maps to restore the resolution of the image and generate a segmentation map of the same size as the input image. During the decoding process, the decoder not only uses its own convolutional layers and upsampling layers to generate new feature maps, but also merges the features from the corresponding layers of the encoder through jump connections to retain more image detail information. The specific formula can be expressed as s * = argmax s m s (n), where m s (n) represents the probability that pixel n is assigned to each superpixel s; argmax s is the variable x that makes s reach its maximum value.

[0050] Finally, the trained deep learning model can perform superpixel segmentation tasks efficiently and accurately. This model can not only effectively depict the region of interest in the image, but also help accurately identify and locate transmission line defects. Compared with traditional segmentation methods, the superpixel segmentation method based on deep learning adopted in the present invention has higher accuracy and efficiency because it can make full use of the advantages of convolutional neural networks in capturing complex patterns and features of images. This innovative method has brought new breakthroughs and possibilities to the field of transmission line defect detection.

[0051] Step 4: Deep learning model evaluation and optimization based on statistical confidence.

[0052] In step 4 of the present invention, a method based on average threshold is used to conduct an in-depth confidence analysis on the deep learning model trained by superpixel segmentation. crack represents the overall mean of the texture feature values ​​extracted from the superpixels corresponding to the defect area in the image. The threshold Threshold can be determined by calculating the confidence interval of the overall mean of the defect texture features. The confidence interval is constructed using the sample mean x and sample standard deviation s of the defect texture feature values. The specific formula can be expressed as where z is the z-score corresponding to the desired confidence level and n is the number of samples.

[0053] The final result is a deep learning model that can meet the requirements after confidence analysis.

[0054] The beneficial effect of this step is that by using statistical methods to analyze the confidence of the deep learning model, a robust criterion for classifying superpixels as transmission line defect detection is established. This criterion not only ensures that the algorithm's decisions are based on reasonable statistical inferences, but also significantly improves the accuracy and reliability of the deep learning model in defect detection tasks.

[0055] The beneficial effects of the present invention are:

[0056] (1) Multi-domain feature extraction optimization: The present invention has achieved a major breakthrough in image feature extraction. By adopting multiple feature extraction methods, it spans the time domain, spatial domain and frequency domain, greatly improving the effect of image feature extraction. This method not only effectively reduces the redundant information in the original image of the transmission line, such as background noise and irrelevant details, but also maximizes the retention of key features of the image, such as conductor morphology, insulator status, etc. This comprehensive and accurate feature extraction method provides a high-quality data foundation for subsequent defect detection.

[0057] (2) High-precision defect detection: In order to improve the accuracy and efficiency of transmission line defect detection, the present invention introduces a superpixel segmentation method based on deep learning. This method can automatically divide an image into multiple superpixel regions with similar characteristics, and accurately depict the region of interest in the image with the help of a convolutional neural network that is good at capturing complex patterns and features. Compared with traditional segmentation methods, this method not only improves the accuracy of defect identification, but also significantly improves the processing speed, achieving a dual improvement in efficiency and accuracy.

[0058] (3) Robust confidence analysis and decision support: In order to further enhance the reliability of the deep learning model, the present invention also introduces a statistical method to perform confidence analysis on the model. By performing statistical analysis on the probability distribution of the model output, the prediction confidence of the model for different categories can be evaluated. Based on these analysis results, the present invention establishes a series of robust criteria for classifying superpixels as the results of power line defect detection. These criteria not only take into account the prediction confidence of the model, but also combine the actual features and contextual information of the image, thereby ensuring the rationality of the algorithm's decision. In this way, the present invention effectively reduces the risk of misjudgment and missed judgment, and improves the performance and reliability of the entire system.

[0059] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements 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 detecting power transmission line defects using an emergency UAV, characterized in that: include: Step 1: Receive the transmission line image collected by the power emergency drone, and pre-process the image by grayscale scaling, segmentation, variance screening, and pseudo target elimination to accurately identify and distinguish defects from the background; Step 2, extracting features in the time domain, space domain and frequency domain of the preprocessed image, and obtaining statistical property features through texture analysis, gray-level co-occurrence matrix, gray-level difference method, discrete wavelet transform and fast Fourier transform; The extracted features are merged and selected to form an optimized feature set; Step 3, using the feature set to train a deep learning-based superpixel segmentation model for power transmission line defect detection, the model adopts a deep learning architecture and is implemented through an encoder-decoder two-layer architecture; Step 4: Use the average threshold method to perform confidence analysis on the deep learning model after superpixel segmentation.

2. The power emergency UAV transmission line defect detection method according to claim 1 is characterized in that: The variance screening in step 1 includes: based on variance sorting, removing image blocks with variances lower than a set threshold.

3. The power emergency UAV transmission line defect detection method according to claim 1 is characterized in that: The pseudo target elimination in step 1 includes: performing mean calculation and pixel traversal on the remaining image blocks after variance screening, and preliminarily marking suspected defect points; and further eliminating pseudo target points through morphological operations to complete image preprocessing.

4. The power emergency UAV transmission line defect detection method according to claim 1 is characterized in that: The image texture analysis method in step 2 is: converting the subjective evaluation of smoothness and uniformity in the image into a numerical measurement, and the numerical measurement is obtained by calculating the intensity change difference of the image pixels in the spatial distribution.

5. The power emergency UAV transmission line defect detection method according to claim 1, characterized in that: The gray level co-occurrence matrix method in step 2 is: under the preset fixed direction and distance conditions, texture highlights are effectively realized by calculating and identifying the occurrence frequency of different pixel combinations on the analyzed image to show the various possible combinations of gray levels inside the image separated by specific directions and distances.

6. The power emergency UAV transmission line defect detection method according to claim 5 is characterized in that: The gray level co-occurrence matrix method involves calculating the gray level co-occurrence matrix in the vertical, diagonal upward and diagonal downward directions at four different distances, and performing a measurement based on the median value obtained in each analysis direction, while using a grid system to identify and emphasize different texture features in the image in the said direction for a specific distance value.

7. The power emergency UAV transmission line defect detection method according to claim 1 is characterized in that: The grayscale difference method in step 2 includes: calculating the required density difference between the most important area and the least important area in the power line image by contrast modification, and calculating statistical features using four spatial regions formed by all samples in the image and adjacent pixels and reference points, and performing optimization of the grayscale difference method distance calculation for each input image and its corresponding output image.

8. The power emergency UAV transmission line defect detection method according to claim 1, characterized in that: In step 3, during the training process, a cross-validation method is used to optimize the model hyperparameters, and a monitoring mechanism is introduced to prevent overfitting.

9. The power emergency UAV transmission line defect detection method according to claim 1, characterized in that: The step 3 includes: the encoder processes the feature-extracted image through a convolutional layer to generate a high-level feature map, the decoder gradually upsamples these feature maps for prediction, and learns the mapping relationship by training a deep neural network, divides the image into grid units as initial superpixel seeds in a soft-associative mapping manner, and limits the search to the adjacent grid units of each pixel for allocation, thereby improving computational efficiency and finally obtaining a trained deep learning superpixel segmentation model.

10. The power emergency UAV transmission line defect detection method according to claim 1, characterized in that: The step 4 includes: setting the overall mean of the texture feature values ​​extracted from the superpixels corresponding to the defect area in the image as a reference, constructing a confidence interval by calculating the sample mean and sample standard deviation of the defect texture feature values, and determining the threshold with the confidence interval, thereby realizing the confidence assessment of the deep learning model.

Images