Electric power inspection key target identification and tracking method

Abstract

The invention relates to an electric power inspection key target identification and tracking method comprising the following steps: S1, collecting an unmanned aerial vehicle inspection image of a distribution line, using SimCLR to carry out self-supervised feature learning, forming a basic feature platform, using a semi-supervised learning method to generate a pseudo tag, then carrying out artificial fine tuning, and carrying out tracking on the pseudo tag; s2, carrying out preprocessing on the labeled image data, wherein the preprocessing comprises image enhancement and denoising; s3, constructing a target recognition model based on Swin Transform, training based on the preprocessed image data, introducing multi-task learning, and optimizing target detection and mask branching at the same time; and S4, carrying the trained target detection and identification model on the unmanned aerial vehicle, detecting the target, and tracking the detected and identified key target based on the target tracking model. According to the invention, the processing capability of complex target identification and tracking tasks in electric power inspection is effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of image detection, and in particular to a method for identifying and tracking key targets in power inspections. Background Art

[0002] In the maintenance and management of modern power systems, efficient and accurate inspections are crucial for ensuring the safe operation of the grid. Traditional manual inspections are time-consuming and labor-intensive, and their limitations are particularly pronounced when dealing with widely distributed power facilities in complex terrain. Therefore, the use of drones for power inspections has become an emerging technology, not only enabling access to hard-to-reach areas but also reducing human resources while increasing the frequency and accuracy of data collection.

[0003] Target recognition and tracking technology plays a crucial role in drone-based power inspections. The core mission of this technology is to automatically identify power equipment (such as utility poles, transformers, and insulators) from video or image data captured by drones and continuously track them during the inspection process. However, existing target recognition and tracking technologies still face many challenges, such as high real-time requirements, variable environments (such as weather and terrain), and the diversity of power grid equipment. Summary of the Invention

[0004] In order to solve the above problems, the purpose of the present invention is to provide a method for identifying and tracking key targets in power inspections, which effectively improves the processing capabilities of complex target identification and tracking tasks in power inspections.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for identifying and tracking key targets in power inspections includes the following steps:

[0007] S1: Collect drone inspection images of distribution lines, organize the image data into a distribution line drone inspection image library according to the scene, and use SimCLR to perform self-supervised feature learning to form a basic feature platform. Use semi-supervised learning methods to generate pseudo labels and then perform manual fine-tuning to form a sample library of distribution network scenarios;

[0008] S2: Preprocess the labeled image data, including image enhancement and denoising; image enhancement is performed by adjusting contrast and brightness and using adaptive histogram equalization, and denoising is performed using non-local mean;

[0009] S3: Build an object recognition model based on Swin Transformer, train it on preprocessed image data, introduce multi-task learning, and optimize the object detection and mask branches simultaneously;

[0010] The target recognition model includes a backbone network Swin Transformer and two sequence branches, wherein the first sequence branch is used for target detection and the second sequence branch is used for target instance segmentation.

[0011] Extract features through Swin Transformer, which uses a hierarchical network structure to perform multi-scale feature extraction in layers. It uses a moving window mechanism to perform local self-attention calculations to reduce computational costs and capture global features.

[0012] Feature extraction is performed through a shared backbone network. After the backbone Swin Transformer extracts features, these features are sent to two branches. The object detection branch proposes candidate regions through the region proposal network RPN, and the object instance segmentation branch is based on instance segmentation of Mask R-CNN. Multi-task loss optimization is also performed.

[0013] Training is performed based on preprocessed image data, using mixed precision and distributed strategies for acceleration;

[0014] S4: The trained target detection and recognition model is mounted on the UAV to detect the target and track the key targets detected and identified based on the target tracking model.

[0015] Furthermore, S1 is specifically:

[0016] Use drones to collect images of a large number of distribution lines, covering different scenes, weather conditions and angles;

[0017] Organize image data by scene and build a structured image library;

[0018] Use SimCLR for self-supervised feature learning, obtain sample features, and build a basic feature platform H for semi-supervised learning;

[0019] Based on the basic feature platform H, the semi-supervised learning method Label Propagation algorithm is used to perform label propagation, and the confidence level is used to filter out high-confidence pseudo-label data;

[0020] Use the basic feature platform H to perform initial annotation to generate some labeled data Ylabeled;

[0021] Using unlabeled data Yunlabeled and combining it with the basic feature platform H

[0022] Construct a similarity matrix S to describe the similarity between data.

[0023] S ij =exp(-γ||h i -h j|| 2 );

[0024] Among them, γ is the adjustment parameter of the similarity measure; h i , h j is the feature representation;

[0025] Label propagation formula:

[0026] Y (t+1) =αSY (t) +(1-α)Y (0) ;

[0027] Among them, Y (0) is the initial label matrix; Y (t) is the label matrix at the current moment; α is a hyperparameter for adjusting the strength of label propagation;

[0028] Reliable pseudo-labels are screened based on confidence levels and manually corrected or refined to form a final high-quality sample library.

[0029] Furthermore, we use SimCLR for self-supervised feature learning to obtain sample features and build a basic feature platform H for semi-supervised learning, as follows:

[0030] Perform augmentation transformation on each image x, including rotation, cropping and color perturbation, to produce two views x i and x j ;

[0031] Through the neural network f θ Extract feature representation:

[0032] h i =f θ (x i ), h j =f θ (x j );

[0033] Where θ represents the parameters of the neural network;

[0034] Using the projection head φ Map features to vector space:

[0035] z i =g φ (h i ), z j =g φ (h j );

[0036] Where, φ represents the parameters of the projection head;

[0037] Optimize the model using contrastive learning loss:

[0038]

[0039] Among them, z i ,z j is the eigenvector obtained after projection; τ is the temperature parameter;

[0040] Obtain the feature output of all samples and construct the feature platform H for semi-supervised learning.

[0041] Furthermore, S2 is specifically:

[0042] Obtain the annotated image data and adjust the image contrast through linear transformation. Let the original image be I and the enhanced image be I′. The contrast adjustment is expressed as:

[0043] I′=a·I+b;

[0044] Where a is the gain factor used to adjust the contrast, and b is the brightness offset;

[0045] Apply CLAHE to the image for local contrast enhancement;

[0046] Use the NLM algorithm to remove image noise. For each pixel, use the entire image to perform denoising. The formula is as follows:

[0047]

[0048] in, is the pixel value after denoising; I′(y) is the pixel value of the input image at position y, Ω is the entire image domain; w(x,y) is the weight, defined as:

[0049]

[0050] Among them, N x and N y are the adjacent windows of points x and y respectively, Z(x) is the normalization constant, and h is the attenuation coefficient that controls the smoothness; Indicates window N x and N y The sum of the squared differences of the paired pixel intensities.

[0051] Furthermore, CLAHE is applied to the image for local contrast enhancement as follows:

[0052] Divide the image into several non-overlapping grids;

[0053] Calculate the local histogram for each grid; set the clip limit, and clip the level whose frequency in the histogram exceeds this limit. The calculation formula is:

[0054] H clipped (i) = min(H(i), clip limit);

[0055] Where H(i) is the gray level frequency of the i-th histogram;

[0056] Evenly distribute the excess to the entire histogram to avoid noise amplification;

[0057] The equalized cumulative distribution function (CDF) is calculated using the cropped histogram, and the original pixel values ​​are mapped to the new equalized values ​​using the CDF:

[0058]

[0059] Among them, L is the maximum value of gray level (usually 255), CDF min and CDF max are the minimum and maximum values ​​of CDF respectively;

[0060] For each pixel, the final equalized result is obtained from the adjacent grids through bilinear interpolation to avoid image blocking effects, and the final image is composed of all interpolated grids.

[0061] Furthermore, the target detection branch proposes candidate regions through the region proposal network RPN, and the target instance segmentation branch is based on instance segmentation of Mask R-CNN; and multi-task loss optimization is performed as follows:

[0062] The target detection branch includes the region proposal part and the detection part. The region proposal network RPN generates candidate boxes for further detection, generates anchor boxes on the feature map through the sliding network, and outputs the appropriate target area.

[0063] For the input feature F, use 3x3 convolution to generate intermediate features;

[0064] Use 1x1 convolution to give 2k classification scores and 4k bounding box regression coordinate corrections, where k is the number of anchor boxes at each position;

[0065] The irregular proposal area is uniformly pooled into a feature map of the same size, and a fully connected network is used to classify the target category and correct the bounding box prediction for each RoI. The loss function is:

[0066]

[0067] Where p is the target classification probability, y is the label, t* and t are the true and predicted bounding box parameters respectively; L cls is the classification loss function; L reg is the regression loss function;

[0068] Mask R-CNN uses a convolutional network to predict the mask map on the RoI. After the original RoI pooling, a convolutional network is added to generate pixel-level target masks. The mask loss is:

[0069]

[0070] Among them, M ij is the true mask pixel value, 0 represents background and 1 represents foreground; is the predicted mask pixel probability; m×m is the resolution of the mask.

[0071] Comprehensive multi-task loss:

[0072]

[0073] Among them, λ cls ,λ reg ,λ mask is the weight coefficient of the corresponding loss.

[0074] Furthermore, S4 is specifically:

[0075] Process each frame of camera input according to the trained target recognition model to identify and obtain the location information and confidence level of each target;

[0076] Use Deep SORT method for continuous tracking;

[0077] The PID control algorithm is used to adjust the flight path of the UAV so that the UAV can continue to follow the target.

[0078] Furthermore, the Deep SORT method is used for continuous tracking, as follows:

[0079] Match the detection results of the current frame with the previous trajectory and use the Hungarian algorithm for optimal allocation:

[0080] a. Use IoU to measure the overlap between each detection result and the existing track. The element C(i,j) of the cost matrix is ​​the matching cost, which is 1-IoU.

[0081] b. Use the Hungarian algorithm on the cost matrix to find the matching with the minimum total cost;

[0082] c. Determine which detections match which existing trajectories using the output of the Hungarian algorithm;

[0083] d. Unmatched detection results may be initialized as new trajectories, and unmatched trajectories may be marked as lost;

[0084] Initialize and maintain a Kalman filter for each tracked target to perform state prediction and update;

[0085] Use the target detection branch to obtain the visual features of the target and update them to the target trajectory;

[0086] For unmatched detection results, new trajectories and Kalman filters are initialized. Existing trajectories that have not been matched are destroyed after exceeding the set number of frames, and long-term lost trajectories are terminated.

[0087] The present invention has the following beneficial effects:

[0088] 1. The present invention effectively improves the processing capability of complex target identification and tracking tasks in power inspection;

[0089] 2. This invention combines CLAHE and NLM algorithms for preprocessing to improve image quality, making images more robust in machine learning tasks and enhancing the model's ability to understand complex scenarios. It can effectively process image data from distribution line inspections and provide clearer and more informative input for subsequent recognition models.

[0090] 3. The target recognition model constructed based on Swin Transformer in this invention effectively provides excellent performance in complex target recognition tasks. Through the combination of Hungarian algorithm matching, Kalman filter prediction update and trajectory management strategy, the UAV can perform robust target tracking in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] Figure 1 Flow chart of the method of the present invention;

[0092] Figure 2 FIG. 1 is a schematic diagram of tracking in one embodiment of the present invention. DETAILED DESCRIPTION

[0093] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0094] A method for identifying and tracking key targets in power inspections includes the following steps:

[0095] S1: Collect drone inspection images of distribution lines, organize the image data into a distribution line drone inspection image library according to the scene, and use SimCLR to perform self-supervised feature learning to form a basic feature platform. Use semi-supervised learning methods to generate pseudo labels and then perform manual fine-tuning to form a sample library of distribution network scenarios;

[0096] S2: Preprocess the labeled image data, including image enhancement and denoising; image enhancement is performed by adjusting contrast and brightness and using adaptive histogram equalization, and denoising is performed using non-local mean;

[0097] S3: Build an object recognition model based on Swin Transformer, train it on preprocessed image data, introduce multi-task learning, and optimize the object detection and mask branches simultaneously;

[0098] The target recognition model includes a backbone network Swin Transformer and two sequence branches, wherein the first sequence branch is used for target detection and the second sequence branch is used for target instance segmentation.

[0099] Extract features through Swin Transformer, which uses a hierarchical network structure to perform multi-scale feature extraction in layers. It uses a moving window mechanism to perform local self-attention calculations to reduce computational costs and capture global features.

[0100] Feature extraction is performed through a shared backbone network. After the backbone Swin Transformer extracts features, these features are sent to two branches. The object detection branch proposes candidate regions through the region proposal network RPN, and the object instance segmentation branch is based on instance segmentation of Mask R-CNN. Multi-task loss optimization is also performed.

[0101] Training is performed based on preprocessed image data, using mixed precision and distributed strategies for acceleration;

[0102] S4: The trained target detection and recognition model is mounted on the UAV to detect the target and track the key targets detected and identified based on the target tracking model.

[0103] In this embodiment, S1 is specifically:

[0104] Use drones to collect images of a large number of distribution lines, covering different scenes, weather conditions and angles;

[0105] Organize image data by scene (such as urban, rural, and wild) and establish a structured image library;

[0106] Use SimCLR for self-supervised feature learning, obtain sample features, and build a basic feature platform H for semi-supervised learning;

[0107] Based on the basic feature platform H, the semi-supervised learning method Label Propagation algorithm is used to perform label propagation, and the confidence level is used to filter out high-confidence pseudo-label data;

[0108] Use the basic feature platform H to perform initial annotation to generate some labeled data Ylabeled;

[0109] Using unlabeled data Yunlabeled and combining it with the basic feature platform H

[0110] Construct a similarity matrix S to describe the similarity between data.

[0111] S ij =exp(-γ||h i -h j || 2 );

[0112] Among them, γ is the adjustment parameter of the similarity measure; h i , h j is the feature representation;

[0113] Label propagation formula:

[0114] Y (t+1) =αSY (t) +(1-α)Y (0) ;

[0115] Among them, Y (0) is the initial label matrix; Y (t) is the label matrix at the current moment; α is a hyperparameter for adjusting the strength of label propagation;

[0116] Reliable pseudo-labels are screened based on confidence levels and manually corrected or refined to form a final high-quality sample library.

[0117] In this embodiment, SimCLR is used for self-supervised feature learning to obtain sample features and build a basic feature platform H for semi-supervised learning, as follows:

[0118] Perform augmentation transformation on each image x, including rotation, cropping and color perturbation, to produce two views x i and x j ;

[0119] Through the neural network f θ Extract feature representation:

[0120] h i =f θ (x i ), h j =f θ (x j );

[0121] Where θ represents the parameters of the neural network; in this embodiment, a convolutional neural network (CNN) is used;

[0122] Using the projection head φ Map features to vector space:

[0123] z i =g φ (hi ), z j =g φ (h j );

[0124] Where, φ represents the parameters of the projection head;

[0125] Optimize the model using contrastive learning loss:

[0126]

[0127] Among them, z i ,z j is the eigenvector obtained after projection; τ is the temperature parameter;

[0128] Obtain the feature output of all samples and construct the feature platform H for semi-supervised learning.

[0129] In this embodiment, S2 is specifically:

[0130] Obtain the annotated image data and adjust the image contrast through linear transformation. Let the original image be I and the enhanced image be I′. The contrast adjustment is expressed as:

[0131] I′=a·I+b;

[0132] Where a is the gain factor used to adjust the contrast, and b is the brightness offset;

[0133] Apply CLAHE to the image for local contrast enhancement;

[0134] Use the NLM algorithm to remove image noise to reduce the noise impact caused by the enhancement process. For each pixel, use the entire image to perform denoising. The formula is as follows:

[0135]

[0136] in, is the pixel value after denoising; I′(y) is the pixel value of the input image at position y, Ω is the entire image domain; w(x,y) is the weight, defined as:

[0137]

[0138] Among them, N x and N y are the adjacent windows of points x and y respectively, Z(x) is the normalization constant, and h is the attenuation coefficient that controls the smoothness; Indicates window N x and N y The sum of the squared differences of the paired pixel intensities.

[0139] In this embodiment, CLAHE is applied to the image to perform local contrast enhancement, specifically as follows:

[0140] Divide the image into several non-overlapping grids;

[0141] Calculate the local histogram for each grid; set the clip limit, and clip the level whose frequency in the histogram exceeds this limit. The calculation formula is:

[0142] H clipped (i) = min(H(i), clip limit);

[0143] Where H(i) is the gray level frequency of the i-th histogram;

[0144] Evenly distribute the excess to the entire histogram to avoid noise amplification;

[0145] The equalized cumulative distribution function (CDF) is calculated using the cropped histogram, and the original pixel values ​​are mapped to the new equalized values ​​using the CDF:

[0146]

[0147] Among them, L is the maximum value of gray level (usually 255), CDF min and CDF max are the minimum and maximum values ​​of CDF respectively;

[0148] For each pixel, the final equalized result is obtained from the adjacent grids through bilinear interpolation to avoid image blocking effects, and the final image is composed of all interpolated grids.

[0149] In this embodiment, the object detection branch proposes candidate regions through the region proposal network RPN, and the object instance segmentation branch is based on instance segmentation of Mask R-CNN. Multi-task loss optimization is performed as follows:

[0150] The target detection branch includes the region proposal part and the detection part. The region proposal network RPN generates candidate boxes for further detection, generates anchor boxes on the feature map through the sliding network, and outputs the appropriate target area.

[0151] For the input feature F, use 3x3 convolution (shared convolution) to generate intermediate features;

[0152] Use 1x1 convolution to give 2k classification scores and 4k bounding box regression coordinate corrections, where k is the number of anchor boxes at each position;

[0153] The irregular proposal area is uniformly pooled into a feature map of the same size, and a fully connected network is used to classify the target category and correct the bounding box prediction for each RoI. The loss function is:

[0154]

[0155] Where p is the target classification probability, y is the label, t* and t are the true and predicted bounding box parameters respectively; L cls is the classification loss function; L reg is the regression loss function;

[0156] Mask R-CNN uses a convolutional network to predict the mask map on the RoI. After the original RoI pooling, a convolutional network is added to generate pixel-level target masks. The mask loss is:

[0157]

[0158] Among them, M ij is the true mask pixel value, 0 represents background and 1 represents foreground; is the predicted mask pixel probability; m×m is the resolution of the mask.

[0159] Comprehensive multi-task loss:

[0160]

[0161] Among them, λ cls ,λ reg ,λ mask is the weight coefficient of the corresponding loss.

[0162] In this embodiment, S4 is specifically:

[0163] Process each frame of camera input according to the trained target recognition model to identify and obtain the location information and confidence level of each target;

[0164] Use Deep SORT method for continuous tracking;

[0165] The PID control algorithm is used to adjust the flight path of the UAV so that the UAV can continue to follow the target.

[0166] In this embodiment, the Deep SORT method is used for continuous tracking, as follows:

[0167] Match the detection results of the current frame with the previous trajectory and use the Hungarian algorithm for optimal allocation:

[0168] a. Use IoU (Intersection over Union) to measure the overlap between each detection result and the existing track. The element C(i, j) of the cost matrix is ​​the matching cost, which is 1-IoU.

[0169] b. Use the Hungarian algorithm on the cost matrix to find the matching with the minimum total cost;

[0170] c. Determine which detections match which existing trajectories using the output of the Hungarian algorithm;

[0171] d. Unmatched detection results may be initialized as new trajectories, and unmatched trajectories may be marked as lost;

[0172] Initialize and maintain a Kalman filter for each tracked target to perform state prediction and update;

[0173] Use the target detection branch to obtain the visual features of the target and update them to the target trajectory;

[0174] For unmatched detection results, new trajectories and Kalman filters are initialized. Existing trajectories that have not been matched are destroyed after exceeding the set number of frames, and long-term lost trajectories are terminated.

[0175] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0176] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0177] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0178] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0179] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other manner. Any person skilled in the art may utilize the above-disclosed technical content to modify or modify the present invention into equivalent embodiments. However, any simple modifications, equivalent variations, and modifications to the above embodiments that do not depart from the technical content of the present invention and are based on the technical essence of the present invention remain within the scope of protection of the present invention.

Claims

1. A method for identifying and tracking key targets in power inspection, characterized in that: The following steps are involved: S1: Collect UAV inspection images of distribution lines, organize the image data into distribution line UAV inspection image libraries according to the scenes, and use SimCLR for self-supervised feature learning to form a basic feature platform. Use semi-supervised learning methods to generate pseudo labels and then perform manual fine tuning to form a sample library of distribution network scenes; S2: Preprocess the labeled image data, including image enhancement and denoising; image enhancement is performed by adjusting contrast and brightness and using adaptive histogram equalization, and denoising is performed using non-local mean; S3: Build a target recognition model based on Swin Transformer, train based on preprocessed image data, introduce multi-task learning, and optimize target detection and mask branches simultaneously; The target recognition model includes a backbone network Swin Transformer and two sequence branches, wherein the first sequence branch is used for target detection and the second sequence branch is used for target instance segmentation; Extract features through Swin Transformer, which uses a hierarchical network structure to perform multi-scale feature extraction in layers; use a moving window mechanism to perform self-attention calculation locally to reduce computational costs and capture global features; Feature extraction is performed through a shared backbone network. After the backbone Swin Transformer extracts features, these features are sent to two branches. The target detection branch proposes candidate regions through the region proposal network RPN, and the target instance segmentation branch is based on instance segmentation of Mask R-CNN. Multi-task loss optimization is also performed. Based on the preprocessed image data, training is performed using mixed precision and distributed strategies for acceleration; S4: The trained target detection and recognition model is mounted on the drone to detect the target, and the key targets detected and identified are tracked based on the target tracking model.

2. A method for identifying and tracking key targets of power inspection according to claim 1, characterized in that: The S1 is specifically: Use drones to collect images of a large number of distribution lines, covering different scenes, weather conditions and angles; Organize image data by scene and build a structured image library; Use SimCLR for self-supervised feature learning, obtain sample features, and build a basic feature platform H for semi-supervised learning; Based on the basic feature platform H, the semi-supervised learning method LabelPropagation algorithm is used to perform label propagation, and the confidence level is used to filter out pseudo-label data with high confidence. Use the basic feature platform H to perform initial annotation to generate some labeled data Ylabeled; Using unlabeled data Yunlabeled and combining it with the basic feature platform H Construct a similarity matrix S to describe the similarity between data. S ij =exp(-γ||h i -h j || 2 ); Among them, γ is the adjustment parameter of the similarity measure; h i ,h j is a feature representation; Label propagation formula: Y (t+1) =αSY (t) +(1-α)Y (0) ; Among them, Y (0) is the initial labeling matrix; Y (t) is the label matrix at the current moment; α is a hyperparameter for adjusting the strength of label propagation; Reliable pseudo-labels are selected based on confidence levels and manually corrected or refined to form a final high-quality sample library.

3. A method for identifying and tracking key targets of power inspection according to claim 2, characterized in that: The method of using SimCLR to perform self-supervised feature learning, obtain sample features, and construct a basic feature platform H for semi-supervised learning is as follows: Perform an augmentation transformation on each image x, including rotation, cropping, and color perturbation, to produce two views x i and x j ; Through the neural network f θ Extract feature representation: h i =f θ (x i ),h j =f θ (x j ); Where θ represents the parameters of the neural network; Using the projection head φ Map features to vector space: z i =g φ (h i ),z j =g φ (h j ); Among them, φ represents the parameters of the projection head; Optimize the model using contrastive learning loss: Among them, z i ,z j is the eigenvector obtained after projection; τ is the temperature parameter; Get the feature output of all samples and build the feature platform H for semi-supervised learning.

4. A method for identifying and tracking key targets of power inspection according to claim 1, characterized in that: The S2 is specifically: Get the annotated image data, adjust the image contrast through linear transformation, set the original image as I, the enhanced image as I′, and the contrast adjustment is expressed as: I′=a·I+b; Among them, a is the gain coefficient used to adjust the contrast, and b is the brightness offset; Apply CLAHE to the image for local contrast enhancement; Use the NLM algorithm to remove image noise. For each pixel, use the entire image to denoise it. The formula is as follows: in, is the pixel value after denoising; I′(y) is the pixel value of the input image at position y, Ω is the entire image domain; w(x,y) is the weight, defined as: Among them, N x and N y are the adjacent windows of points x and y respectively, Z(x) is the normalization constant, and h is the attenuation coefficient, which controls the smoothness; Indicates window N x and N y The sum of the squared differences of paired pixel intensities.

5. A method for identifying and tracking key targets of power inspection according to claim 4, characterized in that: The application of CLAHE to the image for local contrast enhancement is as follows: Divide the image into several non-overlapping grids; Calculate the local histogram for each grid; set the clip limit, and clip the level whose frequency in the histogram exceeds this limit. The calculation formula is: H clipped (i)=min(H(i),clip limit); Among them, H(i) is the gray level frequency of the i-th histogram; Evenly distribute the excess to the entire histogram to avoid noise amplification; The cropped histogram is used to calculate the equalized cumulative distribution function CDF, and the original pixel value is mapped to the new equalized value through CDF: Where L is the maximum value of the gray level (usually 255), CDF min and CDF max are the minimum and maximum values ​​of CDF respectively; For each pixel, the final equalized result is obtained from the adjacent grids through bilinear interpolation to avoid image blocking effects, and the final image is composed of all interpolated grids.

6. A method for identifying and tracking key targets in power inspection according to claim 1, characterized in that: The target detection branch proposes candidate regions through the region proposal network RPN, and the target instance segmentation branch is based on instance segmentation of Mask R-CNN; and multi-task loss optimization is performed as follows: The target detection branch includes the region proposal part and the detection part. The region proposal network RPN generates candidate boxes for further detection, generates anchor boxes on the feature map through the sliding network, and outputs the appropriate target area. For the input feature F, use 3x3 convolution to generate intermediate features; Use 1x1 convolution to give 2k classification scores and 4k bounding box regression coordinate corrections, where k is the number of anchor boxes at each position; The irregular proposal area is evenly pooled into feature maps of the same size, and a fully connected network is used to classify the target category and correct the bounding box prediction for each RoI. The loss function is: Where p is the target classification probability, y is the label, t* and t are the true and predicted bounding box parameters respectively; L cls is the classification loss function; L reg is the regression loss function; Mask R-CNN uses a convolutional network to predict the mask map on the RoI. After the original RoI pooling, a convolutional network is added to generate pixel-level target masks. The mask loss is: Among them, M ij is the true mask pixel value, 0 represents background and 1 represents foreground; is the predicted mask pixel probability; m×m is the resolution of the mask. Comprehensive Multi-Task Loss: Among them, λ cls ,λ reg ,λ mask is the weight coefficient corresponding to the loss.

7. A method for identifying and tracking key targets of power inspection according to claim 6, characterized in that: The S4 is specifically: Process each frame of the camera input according to the trained target recognition model to identify and obtain the location information and confidence of each target; Use Deep SORT method for continuous tracking; The PID control algorithm is used to adjust the flight path of the drone so that it can continue to follow the target.

8. A method for identifying and tracking key targets of power inspection according to claim 7, characterized in that: The Deep SORT method is used for continuous tracking, as follows: Match the detection results of the current frame with the previous trajectory and use the Hungarian algorithm for optimal allocation: a. Use IoU to measure the overlap between each detection result and the existing track. The element C(i,j) of the cost matrix is ​​the matching cost, which is 1-IoU. b. Use the Hungarian algorithm on the cost matrix to find the match with the minimum total cost; c. Determine which detections match which existing trajectories through the output of the Hungarian algorithm; d. Unmatched detection results may be initialized as new trajectories, and unmatched trajectories may be marked as lost; Initialize and maintain a Kalman filter for each tracked target to perform state prediction and update; Use the target detection branch to obtain the visual features of the target and update them to the target trajectory; For unmatched detection results, new trajectories and Kalman filters are initialized. Existing trajectories that have not been matched are destroyed after exceeding the set number of frames, and long-term lost trajectories are terminated.

Images