Improved ST-YOLO model-based power distribution uninterrupted operation site safety detection method
By using the improved ST-YOLO model, combined with the Swing Transformer and YOLOv5 models, the multi-dimensional detection problem of uninterrupted power distribution operations in complex environments was solved, achieving high-precision and rapid safety detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUANGSHI POWER SUPPLY CO
- Filing Date
- 2026-01-13
- Publication Date
- 2026-04-21
AI Technical Summary
Existing technologies lack the reliability for real-time detection of risks such as the status of the insulated boom of the insulated bucket truck and tool displacement during power distribution uninterrupted operation in complex field environments. They also lack multi-dimensional risk detection, and the YOLOv8n algorithm is difficult to achieve efficient detection of various types of targets.
An improved ST-YOLO model, combined with the Swin Transformer and YOLOv5 models, is used to achieve safety detection of live-line power distribution operations through data augmentation, improved loss function, and spatial relation attention mechanism.
It improves the model's robustness and generalization ability in complex scenarios, enhances the accuracy of detecting objects with spatial connections, and ensures efficient detection of multiple types of targets simultaneously.
Smart Images
Figure CN121904697A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of live-line work in power distribution and image vision technology, and in particular to a method for on-site safety inspection of live-line work in power distribution based on an improved ST-YOLO model. Background Technology
[0002] Live-line work in power distribution is a crucial link in the safe operation and maintenance of power distribution networks. It is an important technical means to ensure continuous power supply to users and stable operation of the power grid. The safety of its operation process is directly related to the life safety of workers and the reliable operation of the power grid. Personal injury and power grid accidents caused by improper wearing of safety protective equipment or hidden dangers in the work equipment result in serious social impact and economic losses. Therefore, implementing systematic and routine target safety inspections on key work equipment such as insulated safety helmets, insulated clothing, insulated gloves, insulated bucket trucks, and shielding covers is an important cornerstone for building a safety defense line for live-line work in power distribution. In-depth research on the condition monitoring and risk prevention and control technologies of work equipment has significant safety value and long-term social significance for improving the safety of live-line work in power distribution networks, reducing power outage time, and adapting to the operational requirements of high-quality development of power distribution networks.
[0003] Currently, research on safety target detection for live-line work in power distribution networks still faces numerous practical challenges and technical bottlenecks. Especially in complex field environments, such as under conditions of strong light, rain, and fog, the reliability of real-time detection of risks such as the status of the insulated boom of insulated bucket trucks and tool displacement needs improvement. Furthermore, there is a lack of comprehensive detection of risks across multiple dimensions of "human-equipment-environment." Overcoming the limitations of existing detection technologies and constructing a high-precision, highly robust, and intelligently interconnected safety perception solution has become an urgent need to improve the safety level of live-line work in power distribution networks.
[0004] In recent years, research on the use of deep learning models such as YOLO for the detection of safety tools and equipment has been continuously developing. However, because convolutional neural networks struggle to focus on global features, researchers have begun to introduce the Transformer architecture into object detection tasks. Against this backdrop, the Swin Transformer model has achieved leading levels in multiple benchmark tests in the field of object detection; however, the Swin Transformer model has a large number of parameters, making real-time monitoring difficult to implement in engineering.
[0005] CN120673339A discloses a method for safety monitoring of live-line work in distribution networks based on an improved YOLOv8n algorithm. The method includes: collecting samples from the first-person perspective of live-line work in the distribution network using a wearable device, and extracting key frames from the video using a segmented sampling method; replacing the C2f module in the backbone network with a C2f_CAS module and the C2f module in the neck network with a C2f_SHViTBlock module, and adding an EFC module to the neck network, based on the existing YOLOv8n structure; training the improved YOLOv8n model using training and validation sets to obtain a target detection model; inputting the image to be tested into the target detection model for detection, and outputting the safety monitoring results of live-line work in the distribution network. However, the YOLOv8n algorithm used in this method is relatively complex and cannot achieve efficient detection of multiple types of targets simultaneously.
[0006] Therefore, researching a power distribution uninterrupted operation safety detection model that balances high detection performance and fast detection speed, enabling simultaneous detection of multiple types of targets, and ensuring detection robustness and generalization performance under various complex working conditions is an urgent problem to be solved by researchers in this field. Summary of the Invention
[0007] In view of this, it is necessary to provide a safety detection method for live-line power distribution operations based on an improved ST-YOLO model, in order to solve the safety detection problem at the live-line power distribution operation site and better ensure the safe conduct of the operation.
[0008] This invention provides a method for on-site safety inspection of power distribution uninterrupted operation based on an improved ST-YOLO model, comprising:
[0009] (1) Collect on-site videos or images of power distribution operations through video surveillance or high-definition cameras, and compile them into a dataset of on-site safety images of power distribution operations;
[0010] (2) Use image annotation software to annotate the images in the dataset and divide them into training set and test set according to a certain ratio;
[0011] (3) Data augmentation is performed on the image data in the training set of safety images of power distribution operation sites to enhance the robustness of subsequent models;
[0012] (4) An improved ST-YOLO model based on the Swing Transformer model and the YOLOv5 model is proposed;
[0013] (5) Input the training set into the improved ST-YOLO model for training to obtain the trained model;
[0014] (6) Input the test set data into the trained model to realize real-time safety monitoring of the power distribution operation site without power interruption.
[0015] Furthermore, the work site safety image dataset includes: low-resolution video images obtained through video surveillance and high-resolution images captured by high-definition cameras. First, video frames are extracted using FFmpeg. Then, the high-resolution images captured by the high-definition cameras are scaled proportionally to a size slightly larger than the low-resolution video images. Finally, the images are precisely cropped from the center to obtain a camera image equal to the low-resolution image.
[0016] Furthermore, the ratio of the training set to the test set is 8:2.
[0017] Furthermore, the data augmentation method includes:
[0018] 1) Illumination-adaptive blending enhancement: First, the average brightness ratio of two images is randomly calculated, and then the area ratio of the blending region is dynamically adjusted as a guide to simulate the complex alternating light and dark effects caused by the position of the light source, equipment reflection, etc. in the real working environment.
[0019] 2) Enhanced geometric consistency across multiple perspectives: A dynamic illumination gradient field associated with the deformation perspective is introduced to simulate the continuous brightness changes caused by the change in the angle of the object surface when a single light source illuminates from a specific direction. The visual effects produced when a camera shoots the power distribution operation site from different angles are simulated by combining perspective transformation, rotation and elastic deformation.
[0020] Furthermore, the improvements to the improved ST-YOLO model include:
[0021] 1) Swin Transformer-based skeleton network: The more balanced global and local feature extraction capabilities of the Swin Transformer (ST) are applied to replace part of the Convolution-BatchNorm-LeakyReLU (CBL) module in the skeleton network of the YOLOv5 model.
[0022] 2) Patch Splitting Module: A patch splitting module is added at the beginning of the YOLOv5 model to split the image into N non-overlapping sizes. P These facets facilitate feature aggregation using the Swing Transformer model later.
[0023] 3) Spatial Relationship Attention Mechanism: Since the detection objects such as safety helmets, workers, and insulating gloves have relatively fixed intrinsic relationships in space, a spatial relationship attention mechanism is proposed to aggregate the spatial location features of different detection categories and improve detection accuracy.
[0024] 4) Improved Loss Function: The original YOLOv5 loss function consists of three parts: bounding box loss, confidence loss, and classification loss. The bounding box loss uses the CIoU loss function, while the confidence loss and classification loss both use the cross-entropy loss function. Based on this, the EIoU loss function replaces the CIoU loss function to accelerate convergence, and the Focal loss function replaces the cross-entropy loss function to improve the detection accuracy for small objects such as occlusions and insulating gloves. The improved formulas for calculating the bounding box loss, confidence loss, and classification loss are as follows:
[0025]
[0026]
[0027]
[0028]
[0029] In the formula, IoU is the crossover union ratio; B p The predicted bounding box, including center point coordinates and width and height; B t For the true bounding box; L box EIoU For the improved rectangular frame loss; ρ 2 This represents the squared distance; c is the length of the diagonal of the smallest bounding rectangle; c w and c h These are the width and height of the diagonal of the smallest bounding rectangle, respectively, b p and b t To predict the center point coordinates of the bounding box, w and h represent the width and height of the bounding box, respectively; N is the total number of samples; y i p is the actual label of the i-th anchor box; i y represents the confidence probability; C represents the number of categories; ij p is the true label of the i-th sample in the j-th category. ij The probability of the j-th category; α o γ o α c and γ c These are the balance factor for confidence loss and classification loss, and the simple sample suppression factor, respectively.
[0030] Furthermore, the objects for real-time safety monitoring at the power distribution uninterrupted operation site include: safety helmets, insulated clothing, workers, insulated gloves, wire shielding covers, insulated bucket trucks, etc.
[0031] Furthermore, the illumination adaptive blending enhancement includes:
[0032] 1) Calculate the brightness ratio between the source image and the target image:
[0033]
[0034] In the formula, β is the brightness ratio, μ is the mean function, V is the brightness channel of the image in the HSV color space, the subscripts s and t represent the source image and the target image respectively, and ε is a very small number to prevent the denominator from being 0;
[0035] 2) Dynamically adjust the mixing coefficient λ based on the brightness ratio. eff :
[0036]
[0037] In the formula, α is the basic mixing strength;
[0038] 3) Determine the dimensions of the cutting area:
[0039]
[0040] In the formula, W and H are the width and height of the image, respectively. c and h c Define the width and height of the cropping area;
[0041] 4) Cropping and pasting at random positions (x, y) to generate enhanced image I. mix :
[0042]
[0043] Furthermore, the multi-view geometric consistency enhancement includes:
[0044] 1) First, apply a perspective transformation T to image I. pespective The deformed image I is obtained. geom and bounding box B geom :
[0045]
[0046] 2) Generate a two-dimensional gradient field G simulating a directional light source. For each pixel position (i, j) in the image, we have:
[0047]
[0048] In the formula, x′ and y′ are normalized coordinate grids, and θ is a randomly generated light source direction angle.
[0049] 3) Then normalize G and map it to the brightness adjustment range [γ]. low , γ high ]:
[0050]
[0051] 4) Apply a random lighting field to the image after perspective transformation:
[0052]
[0053] In the formula, V geom isI geom The luminance channel, the clip function is used to ensure that the pixel value is within the range of 0-255, V final It is an enhanced luminance channel;
[0054] 5) Finally, V final Merging the hue and saturation channels and converting them back to RGB channels yields the enhanced image I. final .
[0055] Furthermore, the spatial relationship attention mechanism includes: after the input feature map is processed by two self-attention mechanisms, global max pooling and global mean pooling are performed respectively, then concatenated along the channel dimension, and finally passed through a convolutional layer and activation function to output the spatial relationship attention weights, which are then multiplied by the broadcast of the input feature map to obtain the output feature map.
[0056] Compared with the prior art, the beneficial effects of the present invention include:
[0057] 1. The proposed on-site safety detection method for live-line power distribution operations based on the improved ST-YOLO model adopts two data augmentation techniques: adaptive illumination hybrid enhancement and multi-view geometric consistency enhancement. This improves the robustness and generalization ability of the model in various complex scenarios and under different shooting angles, thereby enhancing the model's detection performance.
[0058] 2. The proposed on-site safety detection method for live-line power distribution operations based on the improved ST-YOLO model organically integrates the SwinTransformer model and the YOLOv5 model. This improves the overall feature extraction performance and detection accuracy of the model while ensuring the detection speed during model inference. A spatial relationship attention mechanism is proposed to enhance the detection accuracy for objects with spatial relationships. Furthermore, the EIoU loss function and Focal loss function are adopted to enhance the detection performance for small number of small object categories. Attached Figure Description
[0059] Figure 1 The flowchart of the on-site safety detection method for power distribution uninterrupted operation based on the improved ST-YOLO model provided by this invention is shown in the present invention.
[0060] Figure 2 The improved ST-YOLO model framework diagram provided by this invention.
[0061] Figure 3 The Swing Transformer network diagram provided by this invention.
[0062] Figure 4 A framework diagram of the spatial relationship attention mechanism provided for this invention. Detailed Implementation
[0063] This invention provides a method for on-site safety inspection of live-line power distribution operations based on an improved ST-YOLO model, combined with... Figure 1 Let's take a look. Figure 1 The flowchart of the on-site safety detection method for live-line power distribution operations based on the improved ST-YOLO model provided by this invention includes steps S1 to S6, wherein:
[0064] This invention provides a method for on-site safety inspection of power distribution uninterrupted operation based on an improved ST-YOLO model, comprising:
[0065] In step S1, video or images of the power distribution live-line operation site are collected by video surveillance or high-definition cameras and organized into a power distribution operation site safety image dataset.
[0066] In step S2, the images in the dataset are labeled using image annotation software and divided into training set and test set according to a certain ratio;
[0067] In step S3, the image data in the power distribution operation site safety image training set is augmented to enhance the robustness of the subsequent model;
[0068] In step S4, an improved ST-YOLO model based on the Swing Transformer model and the YOLOv5 model is proposed;
[0069] In step S5, the training set is input into the improved ST-YOLO model for training to obtain the trained model;
[0070] In step S6, the test set data is input into the trained model to achieve real-time safety monitoring of power distribution uninterrupted operation sites.
[0071] Preferably, the work site safety image dataset includes: low-resolution video images obtained through video surveillance and high-resolution images captured by a high-definition camera. First, video frames are extracted using FFmpeg. Second, the high-resolution images captured by the high-definition camera are scaled proportionally to a size slightly larger than the low-resolution video images. Then, the images are precisely cropped from the center to obtain a camera image equal to the low-resolution image. As a specific embodiment, the resolution is finally uniformly set to 512×512 in this embodiment of the invention.
[0072] Preferably, the ratio of the training set to the test set is 8:2.
[0073] Preferably, the data augmentation method includes:
[0074] 1) Illumination-adaptive blending enhancement: First, the average brightness ratio of two images is randomly calculated, and then the area ratio of the blending region is dynamically adjusted as a guide to simulate the complex alternating light and dark effects caused by the position of the light source, equipment reflection, etc. in the real working environment.
[0075] 2) Enhanced geometric consistency across multiple perspectives: A dynamic illumination gradient field associated with the deformation perspective is introduced to simulate the continuous brightness changes caused by the change in the angle of the object surface when a single light source illuminates from a specific direction. The visual effects produced when a camera shoots the power distribution operation site from different angles are simulated by combining perspective transformation, rotation and elastic deformation.
[0076] Preferably, combined with Figures 2-4 Let's take a look. Figure 2 This is a framework diagram of the improved ST-YOLO model provided by the present invention. Figure 3 This is the Swing Transformer network diagram provided by the present invention. Figure 4 The spatial relationship attention mechanism framework diagram provided by this invention includes the following improvements to the improved ST-YOLO model:
[0077] 1) The skeleton network based on the Win Transformer leverages the Win Transformer (ST)'s more balanced global and local feature extraction capabilities, such as... Figure 2 As shown, the partial convolution-batchnorm-leakyReLU (CBL) module in the YOLOv5 model skeleton network is replaced, as follows: Figure 3 As shown, the Swin Transformer model consists of layer normalization, a window-based multi-head self-attention mechanism, a shift-window-based multi-head self-attention mechanism, and a multi-layer perception mechanism.
[0078] 2) Patch splitting module: such as Figure 2 As shown, a patch splitting module is added at the beginning of the YOLOv5 model to split the image into N non-overlapping sizes. P These facets facilitate feature aggregation using the Swing Transformer model later.
[0079] 3) Spatial Relationship Attention Mechanism: Since the objects being inspected, such as safety helmets, workers, and insulating gloves, have relatively fixed intrinsic relationships in space, a spatial relationship attention mechanism is proposed, such as... Figure 4As shown, spatial location features of different detection categories are aggregated to improve detection accuracy;
[0080] 4) Improved Loss Function: The original YOLOv5 loss function consists of three parts: bounding box loss, confidence loss, and classification loss. The bounding box loss uses the CIoU loss function, while the confidence loss and classification loss both use the cross-entropy loss function. Based on this, the EIoU loss function replaces the CIoU loss function to accelerate convergence, and the Focal loss function replaces the cross-entropy loss function to improve the detection accuracy for small objects such as occlusions and insulating gloves. The improved formulas for calculating the bounding box loss, confidence loss, and classification loss are as follows:
[0081]
[0082]
[0083]
[0084]
[0085] In the formula, IoU is the crossover union ratio; B p The predicted bounding box, including center point coordinates and width and height; B t For the true bounding box; L box EIoU For the improved rectangular frame loss; ρ 2 This represents the squared distance; c is the length of the diagonal of the smallest bounding rectangle; c w and c h These are the width and height of the diagonal of the smallest bounding rectangle, respectively, b p and b t To predict the center point coordinates of the bounding box, w and h represent the width and height of the bounding box, respectively; N is the total number of samples; y i p is the actual label of the i-th anchor box; i y represents the confidence probability; C represents the number of categories. In a specific embodiment of this invention, the number of categories is 6; ij p is the true label of the i-th sample in the j-th category. ij The probability of the j-th category; α o γ o α c and γ c These are the balance factor for confidence loss and classification loss, and the simple sample suppression factor, respectively. In a specific embodiment of this invention, α... o γ o α c and γ c The values were set to 0.25, 2.0, 0.2, and 1.5 respectively.
[0086] Preferably, the objects for real-time safety monitoring at the power distribution uninterrupted operation site include: safety helmets, insulated clothing, workers, insulated gloves, wire shields, insulated bucket trucks, etc.
[0087] Preferably, the illumination adaptive hybrid enhancement includes:
[0088] 1) Calculate the brightness ratio between the source image and the target image:
[0089]
[0090] In the formula, β is the luminance ratio, μ is the mean function, V is the luminance channel of the image in the HSV color space, the subscripts s and t represent the source image and the target image, respectively, and ε is a very small number to prevent the denominator from being 0. In a specific embodiment of the present invention, ε is taken as 1e. -8 ;
[0091] 2) Dynamically adjust the mixing coefficient λ based on the brightness ratio. eff :
[0092]
[0093] In the formula, α is the basic mixing strength. In this specific embodiment, α is taken as 0.8.
[0094] 3) Determine the dimensions of the cutting area:
[0095]
[0096] In the formula, W and H are the width and height of the image, respectively. c and h c Define the width and height of the cropping area;
[0097] 4) Cropping and pasting at random positions (x, y) to generate enhanced image I. mix :
[0098]
[0099] Preferably, the multi-view geometric consistency enhancement includes:
[0100] 1) First, apply a perspective transformation T to image I. pespective The deformed image I is obtained. geom and bounding box B geom :
[0101]
[0102] 2) Generate a two-dimensional gradient field G simulating a directional light source. For each pixel position (i, j) in the image, we have:
[0103]
[0104] In the formula, x′ and y′ are normalized coordinate grids, and θ is a randomly generated light source direction angle.
[0105] 3) Then normalize G and map it to the brightness adjustment range [γ]. low , γ high ]:
[0106]
[0107] 4) Apply a random lighting field to the image after perspective transformation:
[0108]
[0109] In the formula, V geom isI geom The luminance channel, the clip function is used to ensure that the pixel value is within the range of 0-255, V final It is an enhanced luminance channel;
[0110] 5) Finally, V final Merging the hue and saturation channels and converting them back to RGB channels yields the enhanced image I. final .
[0111] Preferably, such as Figure 4 As shown, the spatial relationship attention mechanism includes: after the input feature map is processed by two self-attention mechanisms, global max pooling and global mean pooling are performed respectively, then concatenated along the channel dimension, and finally passed through a convolutional layer and activation function to output the spatial relationship attention weights, which are then multiplied with the broadcast of the input feature map to obtain the output feature map.
[0112] This invention employs two data augmentation techniques: adaptive illumination hybrid enhancement and multi-view geometric consistency enhancement. These techniques improve the model's robustness and generalization ability in various complex scenes and under different shooting angles, thereby enhancing its detection performance. The Swin Transformer model is organically integrated with the YOLOv5 model, improving overall feature extraction performance and detection accuracy while ensuring detection speed during inference. A spatial relationship attention mechanism is proposed to enhance detection accuracy for spatially related objects. Furthermore, EIoU and Focal loss functions are used to strengthen detection performance for small numbers and small object categories.
[0113] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for on-site safety inspection of live-line power distribution operations based on an improved ST-YOLO model, characterized in that, include: (1) Collect on-site videos or images of power distribution operations without power interruption through video surveillance or high-definition cameras, and compile them into a dataset of on-site safety images of power distribution operations; (2) Use image annotation software to annotate the images in the dataset and divide them into training set and test set according to a certain ratio; (3) Data augmentation is performed on the image data in the training set of safety images of power distribution operation sites to enhance the robustness of subsequent models; (4) An improved ST-YOLO model based on the Swing Transformer model and the YOLOv5 model is proposed; (5) Input the training set into the improved ST-YOLO model for training to obtain the trained model; (6) Input the test set data into the trained model to realize real-time safety monitoring of the power distribution operation site without power interruption.
2. The method for on-site safety inspection of live-line power distribution operations based on the improved ST-YOLO model according to claim 1, characterized in that, The power distribution operation site safety image dataset includes: low-resolution video images obtained through video surveillance and high-resolution images captured by high-definition cameras. First, the video frames are extracted using FFmpeg. Then, the high-resolution images captured by the high-definition cameras are scaled proportionally to a size larger than the low-resolution video images. Finally, the images are precisely cropped from the center to obtain a camera image equal to the low-resolution image.
3. The method for on-site safety inspection of live-line power distribution operations based on the improved ST-YOLO model according to claim 1, characterized in that, The ratio of training set to test set is 8:
2.
4. The method for on-site safety inspection of live-line power distribution operations based on the improved ST-YOLO model according to claim 1, characterized in that, Data augmentation methods include: Illumination-adaptive blending enhancement: First, the average brightness ratio of two images is randomly calculated, and then the area ratio of the blending region is dynamically adjusted as a guide to simulate the complex alternating light and dark effects caused by the position of the light source and equipment reflection in the real working environment. Enhanced geometric consistency across multiple perspectives: A dynamic illumination gradient field associated with the deformation perspective is introduced to simulate the continuous brightness changes caused by the change in the surface angle of an object when a single light source illuminates from a specific direction. By combining perspective transformation, rotation, and elastic deformation, the visual effects produced when a camera captures a power distribution operation site from different angles are simulated.
5. The method for on-site safety inspection of live-line power distribution operations based on the improved ST-YOLO model according to claim 1, characterized in that, The improvements to the ST-YOLO model include: 1) Swing Transformer-based skeleton network: The Swing Transformer's more balanced global and local feature extraction capabilities are used to replace part of the convolution-normalization-activation function module in the YOLOv5 model skeleton network. 2) Patch Splitting Module: A patch splitting module is added at the beginning of the YOLOv5 model to split the image into N non-overlapping sizes. P These facets facilitate feature aggregation using the Swing Transformer model later. 3) Spatial Relationship Attention Mechanism: Since the safety helmet, workers, and insulating gloves being detected have a fixed intrinsic relationship in space, a spatial relationship attention mechanism is proposed to aggregate the spatial location features of different detection categories and improve detection accuracy. 4) Improved loss function: The original YOLOv5 loss function consists of three parts: bounding box loss, confidence loss, and classification loss. The bounding box loss uses the CIoU loss function, while the confidence loss and classification loss both use the cross-entropy loss function. Based on this, the EIoU loss function is used to replace the CIoU loss function to speed up the convergence speed, and the Focal loss function is used to replace the cross-entropy loss function to improve the detection accuracy of small objects such as masking masks and insulating gloves.
6. The method for on-site safety inspection of live-line power distribution operations based on the improved ST-YOLO model according to claim 1, characterized in that, The objects of real-time safety monitoring at the power distribution live-line operation site include: safety helmets, insulated clothing, workers, insulated gloves, wire shielding covers, and insulated bucket trucks.
7. The method for on-site safety inspection of live-line power distribution operations based on the improved ST-YOLO model according to claim 4, characterized in that, Lighting-adaptive blending enhancements include: 1) Calculate the brightness ratio between the source image and the target image: In the formula, β is the brightness ratio, μ is the mean function, V is the brightness channel of the image in the HSV color space, the subscripts s and t represent the source image and the target image respectively, and ε is a very small number to prevent the denominator from being 0; 2) Dynamically adjust the mixing coefficient λ based on the brightness ratio. eff : In the formula, α is the basic mixing strength; 3) Determine the dimensions of the cutting area: In the formula, W and H are the width and height of the image, respectively. c and h c Define the width and height of the cropping area; 4) Cropping and pasting at random positions (x, y) to generate enhanced image I. mix : .
8. The method for on-site safety inspection of live-line power distribution operations based on the improved ST-YOLO model according to claim 4, characterized in that, Multi-view geometric consistency enhancement includes: 1) First, apply a perspective transformation T to image I. pespective The deformed image I is obtained. geom and bounding box B geom : ; 2) Generate a two-dimensional gradient field G simulating a directional light source. For each pixel position (i, j) in the image, we have: In the formula, x′ and y′ are normalized coordinate grids, and θ is a randomly generated light source direction angle; 3) Then normalize G and map it to the brightness adjustment range [γ]. low , γ high ]: , 4) Apply a random lighting field to the image after perspective transformation: In the formula, V geom isI geom The luminance channel, the clip function is used to ensure that the pixel value is within the range of 0-255, V final It is an enhanced luminance channel; 5) Finally, V final Merging the hue and saturation channels and converting them back to RGB channels yields the enhanced image I. final .
9. The method for on-site safety inspection of live-line power distribution operations based on the improved ST-YOLO model according to claim 5, characterized in that, The spatial relation attention mechanism includes: after the input feature map is passed through two self-attention mechanisms, global max pooling and global mean pooling are performed respectively, then concatenated along the channel dimension, and finally passed through a convolutional layer and activation function to output the spatial relation attention weights, which are then multiplied with the input feature map broadcast to obtain the output feature map.