A power transmission line channel floating object detection method for satellite remote sensing images

Abstract

The application discloses a kind of power transmission line channel floating object detection methods for satellite remote sensing image, the method includes: obtaining the satellite remote sensing image of power transmission line channel, it is preprocessed, constructs the rotating target frame data set of power transmission line channel floating object;YOLOv8-OBB network is improved, its improvement includes: introducing MGRC module in main network, introducing GLCA module in feature fusion network, to further construct floating object detection model, and using training set to train;The trained floating object detection model is used to detect the floating object of power transmission line channel, and the target class, the rotating boundary frame coordinate and the corresponding rotation angle are output, to realize the automatic detection and accurate positioning of power transmission line channel floating object target.The application can improve the detection precision and rotating target positioning ability of floating object target under complex background condition, and provide technical support for power transmission line safe operation monitoring.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image detection technology, specifically to a method for detecting floating objects in power transmission line channels based on satellite remote sensing images. Background Technology

[0002] Transmission lines play a crucial role in transmitting electricity over long distances, and their operational status directly impacts the safety and reliability of the power grid. During transmission line operation, the surrounding environment is susceptible to weather conditions and human activities, creating various external safety hazards. Lightweight objects such as plastic greenhouses, agricultural mulch film, dust nets, and corrugated steel sheets are prone to displacement or drifting under strong winds, potentially coming into contact with conductors, insulators, and other live equipment, leading to line tripping, short-circuit faults, or even equipment damage. As the coverage area of ​​transmission lines continues to expand, the hazard of floating objects has become a significant factor affecting the safe operation of transmission lines.

[0003] Currently, the investigation of potential hazards related to floating objects along power transmission line corridors mainly relies on manual inspections and drone patrols. Manual inspections are limited by terrain conditions, inspection cycles, and human resources, making it difficult to achieve efficient monitoring of large-scale corridors. While drone patrols can acquire high-precision on-site images, they suffer from limited flight time, weather-dependent operational efficiency, and high maintenance costs. Neither of these methods can meet the practical needs for rapid detection and continuous large-scale monitoring of potential hazards along power transmission line corridors.

[0004] In recent years, with the continuous improvement of satellite remote sensing technology and remote sensing image acquisition capabilities, environmental monitoring of power transmission line corridors using high-resolution remote sensing imagery has gradually become a research hotspot. Satellite remote sensing imagery enables periodic observation of large areas, providing data support for the investigation of floating object hazards. However, floating objects in remote sensing imagery generally suffer from problems such as large scale differences, indistinct features of small targets, arbitrary target orientation, and complex background interference. Existing target detection methods based on horizontal bounding boxes are difficult to accurately detect floating objects with obvious directional features, easily leading to positioning errors and false positives and false negatives.

[0005] Therefore, there is an urgent need for a new method that can overcome the above challenges and achieve high-precision, automated, and intelligent detection of floating objects in power transmission line corridors in satellite remote sensing images, so as to improve the early warning capability and risk management level of power grid operation and maintenance. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by providing a method for detecting floating objects in power transmission line corridors based on satellite remote sensing imagery. This invention can effectively improve the detection results of floating objects in power transmission line corridors.

[0007] The objective of this invention is achieved through the following technical solution: The first aspect of this invention provides a method for detecting floating objects in power transmission line corridors based on satellite remote sensing imagery, comprising the following steps: S1. Acquire satellite remote sensing images of the power transmission line channel, preprocess them to construct a dataset of rotating target boxes of floating objects in the power transmission line channel, and divide them into training and test sets according to the proportions. S2. Improve the YOLOv8-OBB network to build a floating object detection model, and use the training set to train the floating object detection model to obtain a trained floating object detection model. S3. Use the trained floating object detection model to detect floating objects in the power transmission line channel, and output the target category, the coordinates of the rotated bounding box, and the corresponding rotation angle.

[0008] Furthermore, step S1 specifically includes the following sub-steps: S11. Obtain satellite remote sensing images of the power transmission line corridor; S12. Remove abnormal samples from the original satellite remote sensing images, convert the image format to JPEG format, and store them using RGB three channels; S13. Use the rotated bounding box annotation tool to annotate the floating targets in the image and obtain annotation information including target category, rotated bounding box coordinates and rotation angle; S14. Convert the format of the annotation information into a label format that the floating object detection model can recognize; S15. The image size is normalized using the Letterbox method; the image and its corresponding label information are enhanced using the MixUp data augmentation method. S16. Construct a rotating target bounding box dataset of floating objects in the power transmission line channel based on the enhanced image images and their corresponding label information, and divide it into training set and test set according to the proportion.

[0009] Furthermore, the floating object target includes one or more of the following: dust netting, plastic greenhouse, mulch film, and corrugated steel sheet.

[0010] Furthermore, the improvement of the YOLOv8-OBB network to construct the floating object detection model specifically includes: The improved backbone network is obtained by replacing the convolutional modules in the YOLOv8-OBB backbone network with multi-scale residual gated convolutional modules. After embedding the global-local collaborative attention module into the first upsampling module of the YOLOv8-OBB feature fusion network, an improved feature fusion network is obtained. A floating object detection model is constructed based on the improved backbone network, the improved feature fusion network, and the YOLOv8-OBB detection head network.

[0011] Furthermore, in the floating object detection model, the improved backbone network includes a multi-scale residual gated convolution module, a C2f module, an efficient multi-scale attention module, and a spatial pyramid pooling module; the improved feature fusion network includes an upsampling module, a global-local collaborative attention module, a stitching layer, a C2f module, and a convolution module; the detection head network includes three detection head output branches, which are used for the detection of floating objects at different scales; each output branch includes a target category branch for predicting the target category, a bounding box regression branch for predicting the center point coordinates and size of the rotated bounding box, and a rotation angle prediction branch for predicting the target rotation angle; The image is input into the floating object detection model. After passing through the backbone network, feature fusion network, and detection head network, the model outputs the predicted target category, the center point coordinates and size of the corresponding target's rotated bounding box, and the rotation angle of the corresponding target.

[0012] Furthermore, the multi-scale residual gated convolutional module includes a local feature extraction branch, a mesoscale feature extraction branch, and a global context feature extraction branch. The local feature extraction branch uses a 3×3 convolution to extract local texture features, the mesoscale feature extraction branch uses a 5×5 convolution to extract mesoscale structural features, and the global context feature extraction branch uses a 7×7 depthwise separable convolution to extract global context features. The output features of the three feature extraction branches are concatenated and fused using a dynamic gating mechanism. Specifically, a 1×1 convolution is first used to map the concatenated features, and then a sigmoid activation function is used to generate dynamic gating weights. These dynamic gating weights are then used to weight and fuse the output features of the three feature extraction branches to obtain fused features. Finally, the fused features and the input features of the multi-scale residual gated convolutional module are connected through a residual connection structure to obtain the output features of the multi-scale residual gated convolutional module.

[0013] Furthermore, the global-local collaborative attention module includes a global feature branch and a local feature branch. The global feature branch uses global average pooling to extract global semantic features of the input features, and the local feature branch uses convolution operations to extract local detail features of the input features. A fusion weight is obtained through a dynamic weight generation mechanism. The fusion weight is generated through convolution mapping and the Sigmoid activation function, and a weighted fusion method is used to obtain the attention enhancement features output by the global-local collaborative attention module.

[0014] Furthermore, training the floating object detection model using a training set specifically includes: The image images in the training set are input into the floating object detection model, which outputs the predicted target category, the center point coordinates and size of the corresponding target's rotated bounding box, and the corresponding target's rotation angle. The binary cross-entropy loss between the predicted target category and its corresponding ground truth target category is calculated as the classification loss. The original scene intersection-over-union (IoU) loss is calculated based on the coordinates and dimensions of the predicted rotated bounding box and its corresponding ground truth rotated bounding box coordinates. An angle error constraint term is calculated based on the predicted rotation angle and its corresponding ground truth rotation angle. An aspect ratio coefficient is calculated based on the width and height of the current ground truth rotated bounding box to obtain the target shape perception weight. The angle error constraint term and the original scene IoU loss are then weighted and combined according to the target shape perception weight to form the final dynamic angle-perceived scene IoU loss function. The total loss function is calculated based on the classification loss and the dynamic angle-perceived scene IoU loss function. With minimizing the total loss function as the optimization objective, the parameters of the floating object detection model are adjusted until the preset number of training rounds is reached, and a well-trained floating object detection model is obtained.

[0015] A second aspect of the present invention provides a device for detecting floating objects in power transmission line corridors based on satellite remote sensing imagery, comprising one or more processors and a memory, wherein the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the above-described method for detecting floating objects in power transmission line corridors based on satellite remote sensing imagery.

[0016] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, is used to implement the above-described method for detecting floating objects in power transmission line channels based on satellite remote sensing images.

[0017] Compared with the prior art, the beneficial effects of the present invention are: (1) This invention introduces a multi-scale residual gated convolution (MGRC) module into the backbone network of YOLOv8-OBB. By constructing convolutional branches with different receptive fields and combining them with a dynamic gating mechanism, it achieves adaptive fusion of local texture features, mesoscale structural features, and global contextual features. Compared with traditional convolutional structures, the MGRC module of this invention can extract floating target feature information more fully, improve the model's ability to represent targets at different scales, and enhance the ability to distinguish between targets and backgrounds in complex backgrounds, effectively improving the detection accuracy of floating targets in power transmission line channels.

[0018] (2) This invention proposes a Global-Local Collaborative Attention (GLCA) module, which integrates global semantic information and local detail information to achieve joint attention to the transmission line channel area and the floating object target area. The GLCA module of this invention can effectively enhance the feature response of the target area, suppress the interference of complex backgrounds such as vegetation, roads, buildings and shadows, and improve the model's ability to identify small targets, weak targets and occluded targets, thereby improving the detection accuracy and robustness of floating objects in remote sensing images.

[0019] (3) In view of the characteristics of floating objects in remote sensing images, such as random orientation, slender shape and large rotation angle, the present invention constructs a dynamic angle-aware SIoU loss function (DA-SIoU) based on the SIoU loss function. By introducing angle constraint terms and aspect ratio dynamic weighting mechanism, the sensitivity of the model to the angle error of the rotating bounding box is enhanced, so that the predicted box can fit the real target contour more accurately, effectively improve the positioning accuracy of rotating targets and reduce the bounding box regression error.

[0020] (4) This invention can make full use of the large-scale observation advantages of satellite remote sensing images to realize the automatic identification and accurate positioning of floating targets in power transmission line channels, improve the detection accuracy of floating targets and the positioning capability of rotating targets under complex background conditions, reduce the workload of manual inspection, improve the efficiency of power transmission line operation status monitoring, and provide reliable technical support for the safe operation of power grid transmission lines. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method for detecting floating objects in power transmission line channels based on satellite remote sensing imagery according to the present invention; Figure 2 This is a schematic diagram of the LabelImg2 software of the present invention; Figure 3 This is a schematic diagram of the network structure of the floating object detection model of the present invention; Figure 4 This is a schematic diagram of the network structure of the MGRC module of the present invention; Figure 5 This is a schematic diagram of the network structure of the GLCA module of the present invention; Figure 6 This is a schematic diagram of the DA-SIoU of the present invention; Figure 7 This is a comparison chart of detection results provided by an embodiment of the present invention; wherein, Figure 7 In the image, (a) shows the detection results of the original YOLOv8-OBB model. Figure 7 (b) in the figure represents the detection result of the floating object detection model described in this invention. Detailed Implementation

[0022] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention as detailed in the appended claims.

[0023] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The singular forms “a,” “the,” and “the” used in this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0024] It should be understood that although the terms first, second, third, etc., may be used in this invention to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first information may also be referred to as second information without departing from the scope of this invention, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0025] The present invention will now be described in detail with reference to the accompanying drawings. Unless otherwise specified, the features of the following embodiments and implementations can be combined with each other.

[0026] See Figure 1 The present invention provides a method for detecting floating objects in power transmission line corridors based on satellite remote sensing imagery, which specifically includes the following steps: S1. Acquire satellite remote sensing images of the power transmission line channel, preprocess them to construct a dataset of rotating target boxes of floating objects in the power transmission line channel, and divide them into training and test sets according to the proportion.

[0027] S11. Obtain satellite remote sensing images of power transmission line corridors.

[0028] In this embodiment, satellite remote sensing images of a power transmission line corridor in a certain area were acquired through a high-resolution satellite remote sensing platform. A total of 1,000 images were acquired, which included scene information such as vegetation, buildings, and floating objects.

[0029] S12. Remove abnormal samples from the original satellite remote sensing images, convert the image format to JPEG format, and store them using RGB three channels.

[0030] Specifically, the raw satellite remote sensing images undergo quality screening to remove abnormal samples such as damaged images, images that cannot be read normally, severely blurred images, and images with abnormal content. After manual review, the abnormal samples in the raw satellite remote sensing images are removed. Subsequently, the remaining satellite remote sensing images from different sources are uniformly converted to JPEG format and stored using RGB three-channel encoding.

[0031] S13. Use the rotated bounding box annotation tool to annotate floating targets in the image to obtain structured annotation information including target category, rotated bounding box coordinates and rotation angle.

[0032] Furthermore, the floating object targets include one or more of the following: dust nets, plastic greenhouses, mulch film, and corrugated steel sheets.

[0033] Specifically, a rotated bounding box annotation tool, such as LabelImg2 software, is used to annotate floating objects in the image, such as... Figure 2 As shown, obtaining structured annotation information such as target category, rotated bounding box coordinates, and rotation angle (i.e., target direction) can accurately describe the target's location, size, and orientation. In this embodiment, the target categories are set as fang-cw, su-ldp, di-mo, and cai-gb, corresponding to dust control netting, plastic greenhouses, mulch film, and color steel plates, respectively. After annotation is completed, the LabelImg2 software automatically generates an XML format annotation file corresponding to the image. This XML format annotation file contains structured annotation information such as target category, rotated bounding box coordinates (i.e., coordinates of the center point of the rotated bounding box), and rotation angle.

[0034] S14. Convert the format of the annotation information into a label format that the floating object detection model can recognize in subsequent steps.

[0035] Specifically, a Python program is used to convert the XML format annotation file into a TXT format label file that the floating object detection model can recognize in subsequent steps, so as to meet the training requirements of the floating object detection model.

[0036] S15. The Letterbox method is used to normalize the size of the image; the MixUp data augmentation method is used to enhance the image and its corresponding label information.

[0037] Specifically, to standardize input dimensions and enrich training samples, the Letterbox method is used to normalize the image dimensions. The input images are scaled and padded according to the original image aspect ratio, uniformly adjusting the image dimensions to 640×640 pixels, ensuring the image size is a multiple of 32 to accommodate the downsampling process of the floating object detection model. Subsequently, the MixUp data augmentation method is used to enhance the image images and their corresponding label information. Specifically, two image images and their corresponding label information are randomly selected from all image images. A fusion coefficient λ (ranging from 0 to 1) is randomly sampled from the Beta distribution. Based on this fusion coefficient λ, the pixel matrices of the two image images are linearly weighted and mixed to generate a new mixed image. Simultaneously, the labels of the two image images are also linearly weighted and fused to obtain the mixed labels, which are the label information of the mixed image. Finally, the mixed image and its corresponding labels are used as new samples to train the floating object detection model. The MixUp data augmentation method generates virtual samples through linear combination, which effectively increases the diversity of training samples in the later stage and improves the generalization ability of the floating object detection model.

[0038] S16. Construct a rotating target bounding box dataset of floating objects in the power transmission line channel based on the enhanced image images and their corresponding label information, and divide it into training set and test set according to the proportion.

[0039] Specifically, a rotating target bounding box dataset of floating objects along the power transmission line corridor is constructed based on the enhanced image images and their corresponding label information obtained in step S15. Then, the rotating target bounding box dataset is divided into a training set and a test set in an 8:2 ratio according to time sequence and scene distribution. The training set is used to train the floating object detection model and optimize its parameters; the test set is used to evaluate the performance of the floating object detection model.

[0040] S2. Improve the YOLOv8-OBB network to construct a floating object detection model, the overall structure of which is as follows: Figure 3 As shown, the floating object detection model is trained using the training set to obtain a trained floating object detection model.

[0041] In this embodiment, the YOLOv8-OBB network is improved in the following ways: ① A multi-scale residual gated convolution (MGRC) module is introduced into the YOLOv8-OBB backbone network. Specifically, the improved backbone network is obtained by replacing the convolution (Conv) modules in the backbone network with the designed MGRC modules. Different scale convolution branches are used to extract local texture features, mesoscale structural features, and large-scale contextual features of floating objects. ② A global-local collaborative attention (GLCA) module is embedded into the YOLOv8-OBB feature fusion network (Neck). This module is embedded after the first upsampling (DySample) module to obtain the improved feature fusion network, which helps to fuse global semantic information of the transmission line corridor area with local detail information of floating objects, improving the feature response intensity of the target area. ③ The dynamic angle-aware scene intersection-over-union ratio (SIoU) loss function (DA-SIoU) is used instead of the original bounding box regression loss function, which enhances the floating object detection model's ability to represent features, locate boundaries, and detect rotating targets in complex backgrounds of remote sensing images. The final floating object detection model was obtained.

[0042] A floating object detection model is constructed based on the improved backbone network, the improved feature fusion network, and the YOLOv8-OBB detection head network, such as... Figure 3 As shown. The improved backbone network includes the MGRC module, C2f module, efficient multi-scale attention (EMA) module, and spatial pyramid pooling (SPPF) module, with the C2f module being a more efficient feature extraction module. The improved feature fusion network includes the DySample module, GLCA module, concat layer, C2f module, and Conv module. The detection head network includes three output branches, each receiving different scale features from the feature fusion network for detecting floating objects of different scales (small, medium, and large). Each output branch includes a target classification branch for predicting the target category, a bounding box regression branch for predicting the center point coordinates and size of the rotated bounding box, and an angle prediction branch for predicting the target rotation angle. Image images are input into the floating object detection model, and after processing as follows... Figure 3 The floating object detection model shown outputs the predicted target category, the center point coordinates and size of the corresponding target's rotated bounding box, and the corresponding target's rotation angle after the backbone network, feature fusion network, and detection head network are constructed.

[0043] Furthermore, the network structure of the MGRC module is as follows: Figure 4As shown, the model specifically includes a local feature extraction branch, a mesoscale feature extraction branch, and a global context feature extraction branch. The local feature extraction branch uses 3×3 convolutions to extract local texture features. 3×3 convolutions have strong local perception capabilities, effectively extracting the edge contours, texture details, and local salient region information of floating objects. The mesoscale feature extraction branch uses 5×5 convolutions to extract mesoscale structural features. 5×5 convolutions effectively enhance the model's perception of the overall shape and structure information of the target. The global context feature extraction branch uses 7×7 depthwise separable convolutions (DWConv) to extract global context features. 7×7 depthwise separable convolutions can expand the receptive field to obtain overall structural information of the remote sensing image and spatial relationship information between the floating object and the background. This can be expressed by the following formula:

[0044]

[0045]

[0046] In the formula, Here, C represents the number of channels, and H and W represent the height and width of the feature map, respectively. This represents a 3×3 convolution. This represents a 5×5 convolution. This represents a 7×7 depthwise separable convolution. Represents local texture features. Indicates mesoscale structural features. This represents the global context features. Subsequently, the output features from the three feature extraction branches are concatenated to obtain the concatenated features, which can be expressed by the following formula:

[0047] In the formula, This indicates a feature concatenation operation. This represents the concatenated features. Furthermore, to achieve importance allocation for features at different scales, a dynamic gating mechanism is constructed for weighted feature fusion. Specifically, the concatenated features are first mapped using a 1×1 convolution, then dynamic gating weights are generated using the Sigmoid activation function. These dynamic gating weights are then used to weight and fuse the output features of the three feature extraction branches to obtain the fused features, which can be expressed by the following formula:

[0048]

[0049] In the formula, Represents a 1×1 convolution. This represents the Sigmoid activation function, and G represents the dynamic gating weights. This represents element-wise multiplication. This represents the fused features. The dynamic gating weights G automatically adjust the contribution of features at each scale based on the current input image content. Finally, the output features of the MGRC module are obtained through the residual connection structure.

[0050] In the formula, This indicates the characteristics output by the MGRC module.

[0051] Furthermore, the input features of the EMA module are divided into G sub-feature groups along the channel dimension. This allows different feature groups to learn their own independent semantic information and enables the subsequent attention mechanism to apply more precisely to each feature group. Each sub-feature group enters a 1×1 branch and a 3×3 branch. The 1×1 branch generates two directional attention weights through one-dimensional global average pooling in the height / width direction, shared 1×1 convolution, splitting, and using a sigmoid function. These weights are then multiplied with the atomic feature groups and processed by GroupNorm to obtain the branch output. The 3×3 branch extracts local features through 3×3 convolution. Subsequently, the two branches undergo global average pooling, Softmax, and cross-matrix multiplication to form a cross-spatial response. Spatial attention weights are obtained through Reshape and a sigmoid function. After recalibrating the sub-feature groups, they are concatenated along the channel dimension to obtain the output of the EMA module. In this embodiment, G is preferably set to 32, ensuring that the number of channels and spatial dimensions of the EMA module's input and output are consistent. The specific process of the EMA module is as follows: First, the input features of the EMA module are grouped. The input of the EMA module is the high-level feature X of the backbone network, with C, H, and W channels, respectively. First, X is divided into G sub-feature groups along the channel dimension. The size of each sub-feature group is Each sub-feature group performs the same attention calculation under mutually independent conditions.

[0052] Then, parallel multi-scale branching is performed. For any sub-feature group The input is simultaneously fed into a 1×1 branch and a 3×3 branch. In the 1×1 branch, a one-dimensional global average pooling is first performed along the width W direction to obtain a height-oriented encoding feature of size (C / G)×H×1; simultaneously, a one-dimensional global average pooling is performed along the height H direction to obtain a width-oriented encoding feature of size (C / G)×1×W. The two directional encoding features are concatenated in the spatial dimension and then input into a shared 1×1 convolution for mapping. Subsequently, they are re-split according to the original height and width directions and activated separately by Sigmoid to obtain the height-oriented attention weights. Pay attention to weights in the width direction. .Will and , Element-wise multiplication based on spatial location, where directional weights are expanded via broadcasting in the missing spatial dimension, and then normalized using GroupNorm to obtain a 1×1 branch output. In the parallel 3×3 branch, Local spatial information and local cross-channel information are extracted directly through 3×3 convolution to obtain 3×3 branch output. Its channel count and spatial dimensions are similar to Maintain consistency.

[0053] Then cross-spatial feature interaction is performed. (Separately for...) and Two-dimensional global average pooling is performed to obtain global channel descriptions for both branches, which are then normalized using Softmax. Simultaneously, the output of the other branch is flattened spatially into a feature matrix of length H×W. A first matrix multiplication is performed between the normalized global description of the 1×1 branch and the flattened feature of the 3×3 branch to obtain the first 1×(H×W) spatial response. A second matrix multiplication is then performed between the normalized global description of the 3×3 branch and the flattened feature of the 1×1 branch to obtain the second 1×(H×W) spatial response. The two spatial responses are then added element-wise to obtain the cross-spatial response S, which incorporates information from both scales. i .

[0054] Finally, attention recalibration and output are performed. The cross-spatial response S i The spatial size is re-transformed into H×W, and then activated by Sigmoid to obtain the spatial attention weight A of the i-th sub-feature group. i The spatial attention weights are broadcast along the channel dimension to C / G channels, and then combined with the atomic feature set X. i Element-wise multiplication yields the recalibrated sub-feature group Y. i After processing all G sub-feature groups, Y0, Y1, ..., Y... G-1 By splicing along the channel dimension, the final output Y of the EMA module is obtained, whose size is still C×H×W, consistent with the input X.

[0055] Furthermore, the network structure of the GLCA module is as follows: Figure 5As shown, the model includes a global feature branch and a local feature branch. The global feature branch uses global average pooling to extract global semantic features of the input features (i.e., the input features of the GLCA module). Through global average pooling, the overall spatial distribution information of the transmission line corridor area and the global relationship information between the target and the background can be obtained, enhancing the network's ability to perceive scene structures at different scales. Its output dimension is C×1×1. The local feature branch uses convolution operations to extract local detail features of the input features. The local feature branch focuses on the edge contours of floating objects, texture details, and local salient regions of the target, thereby improving the model's ability to extract fine-grained target features. Its output dimension is C×H×W, which can be expressed by the following formula:

[0056]

[0057] In the formula, here These are the input features for the GLCA module. This indicates a global average pooling operation. Represents global semantic features. This represents local detail features. Since the dimensions of global semantic features and local detail features differ, a spatial broadcast mechanism is needed to expand the C×1×1 global semantic features along the height H and width W directions to C×H×W, and then add them element-wise with the local detail features. This operation is only used for spatial dimension alignment, does not change the number of channels, and does not introduce additional learnable parameters. Subsequently, a collaborative fusion mechanism is used to jointly model global semantic information and local texture information. Specifically, to achieve adaptive fusion of global and local information, a dynamic weight generation mechanism is designed to obtain fusion weights, which are generated through convolutional mapping and the Sigmoid activation function.

[0058] In the formula, This represents the adaptive fusion weight between global semantic features and local detail features, and its value ranges from 0 to 1; This represents the Sigmoid activation function. This represents the convolutional mapping. Finally, a weighted fusion method is used to obtain the attention-enhanced features as the output features of the GLCA module.

[0059] In the formula, This represents the attention-enhanced features output by the GLCA module. When the target region occupies a large spatial area, the model automatically increases the global feature weights; when the floating target is small or has obvious edge features, the model automatically increases the local feature weights, thereby achieving adaptive allocation of feature information under different scenarios.

[0060] In this embodiment, the floating object detection model is trained using a training set. Specifically, this includes: inputting image data from the training set into the floating object detection model, outputting the predicted target category, the center point coordinates and size of the corresponding rotated bounding box, and the rotation angle of the corresponding target. The binary cross-entropy loss between the predicted target category and its corresponding real target category is calculated as the classification loss. The original SIoU loss is calculated based on the center point coordinates and size of the predicted rotated bounding box and its corresponding real rotated bounding box coordinates. An angle error constraint term is calculated based on the predicted rotation angle and its corresponding real rotation angle. An aspect ratio coefficient is calculated based on the width and height of the current real rotated bounding box, and the target shape perception weight is obtained based on the aspect ratio coefficient. Finally, the angle error constraint term and the original SIoU loss are weighted and combined according to the target shape perception weight to form the final dynamic angle-aware scene intersection-union ratio (DA-SIoU) loss function. The total loss function is calculated based on the classification loss and the dynamic angle-aware scene intersection-union ratio loss function. The parameters of the floating object detection model are adjusted with the goal of minimizing the total loss function until a preset number of training rounds are reached, resulting in a trained floating object detection model.

[0061] It should be noted that DA-SIoU is used as the loss function for bounding box regression, replacing the default bounding box regression loss function of YOLOv8-OBB. For example... Figure 6 As shown, DA-SIoU introduces an angle error constraint term and target shape-aware weights on the basis of the original SIoU loss function. The angle error constraint term describes the angle deviation between the predicted rotated bounding box and the real rotated bounding box and is converted into a differentiable loss value. The target shape-aware weights dynamically adjust the weights of the angle error constraint term according to the aspect ratio of the rotated bounding box.

[0062] like Figure 6 As shown, when performing rotated bounding box prediction and loss calculation, the following method is adopted: The parametric form describes the rotated bounding box, where This represents the coordinates of the center point of the predicted rotated bounding box, where w and h represent the width and height of the predicted rotated bounding box, respectively. This represents the predicted rotation angle, and the corresponding true rotation bounding box is... ,in This represents the coordinates of the center point of the actual rotated bounding box. and These represent the width and height of the actual rotated bounding box, respectively. Representing the true rotation angle (i.e., the corresponding label information in the training set), based on the predicted rotation angle. and the actual rotation angle Construct angle error constraint terms:

[0063] In the formula, This represents the angle error constraint term. This represents the cosine value. When the predicted rotation angle matches the actual rotation angle... The value approaches 0; as the rotation angle deviation increases, the value of the angle error constraint term gradually increases, thereby guiding the model to learn more accurate target orientation information. Considering that most floating objects such as plastic greenhouses have a large aspect ratio, the positioning error will be further amplified when the rotation angle deviates slightly. Therefore, introducing target shape-aware weights can improve the regression accuracy of the rotated bounding box.

[0064] For the actual rotated bounding box during training, let its width be... The height is The aspect ratio coefficient is defined as follows:

[0065] In the formula, r represents the aspect ratio coefficient. This indicates the operation of retrieving the maximum value. This indicates the minimum value operation. Further, a dynamic weighting factor is constructed based on the aspect ratio coefficient, which is the target shape-aware weight:

[0066] In the formula, This represents the target shape perception weight. When the target approaches a square, Approaching 1; when the target exhibits a slender structure, The angle error of slender targets is gradually increased, thereby increasing the proportion of the angle error in the loss function.

[0067] Based on the target shape perception weights, the aforementioned angle error constraint term is weighted and combined with the original SIoU loss term to form the final dynamic angle-perception scene intersection-over-union (IoU) loss function:

[0068] In the formula, This represents the cross-union ratio (CUB) loss function for dynamically angle-aware scenes. This represents the original SIoU loss term. This represents the adjustment coefficient, used to adjust the weight of the angle error constraint term in the intersection-union loss function of the dynamic angle perception scene. In this embodiment, the adjustment coefficient of the angle error constraint term is preferably set to 0.5.

[0069] The formula for calculating the total loss function is:

[0070] In the formula, Represents the total loss function. This represents classification loss.

[0071] Furthermore, during the training process of adjusting the parameters of the floating object detection model, the optimization objective is to minimize the total loss function. The backpropagation algorithm is adopted, and the AdamW optimizer is used to continuously update the parameters of the floating object detection model. The initial learning rate is set to 0.001, the weight decay coefficient is set to 0.0005, the momentum factor is set to 0.937, the batch size is set to 16, and the number of training rounds is set to 200 batches. By adaptively adjusting the learning rate, the convergence speed of the model is improved, while the parameter oscillation phenomenon during training is reduced, so that the total loss function gradually converges, thereby improving the positioning accuracy of the rotated bounding box.

[0072] S3. Use the trained floating object detection model to detect floating objects in the power transmission line channel, and output the target category, the coordinates of the rotated bounding box, and the corresponding rotation angle.

[0073] Furthermore, in this embodiment, the performance of the trained floating object detection model was validated using a test set. Mean AP, precision, recall, and F1 score were used as evaluation metrics for model performance, and the detection results on the test set were quantitatively analyzed.

[0074] The mean accuracy is used to evaluate the overall detection performance of the model. In this embodiment, mAP@0.5 (mAP50) is used for evaluation. That is, when the intersection-union ratio (IU) between the predicted bounding box and the ground truth bounding box is greater than or equal to 0.5, the predicted target is considered to have successfully matched the ground truth target. Based on this, the mean accuracy (AP) for each category is calculated, and then mAP is further calculated. The calculation formula is as follows:

[0075] In the formula, This represents the mean of the precision values. represents the average precision of the i-th floating object category, and N represents the total number of floating object categories.

[0076] Precision is used to evaluate the accuracy of test results, and its calculation formula is as follows:

[0077] In the formula, Indicates accuracy. Indicates the number of correctly detected targets. This indicates the number of false positives.

[0078] Recall rate is used to evaluate the ability to detect a target, and its calculation formula is as follows:

[0079] In the formula, Indicates recall rate, This indicates the number of targets that were missed.

[0080] The F1 score is used to comprehensively reflect the model's detection performance, and its calculation formula is as follows:

[0081] In the formula, This represents the F1 score.

[0082] To verify the effectiveness of the improvements in this invention, a comparative experiment was conducted on the same test set between the floating object detection model of this invention and the original YOLOv8-OBB model. The experimental results are shown in Table 1.

[0083] Table 1: Comparison of various performance evaluation metrics for the test set

[0084] As shown in Table 1, the experimental results indicate that, compared with the original YOLOv8-OBB model, the floating object detection model described in this invention has significantly improved precision, recall, mAP50, and F1 score, by 4.89%, 3.20%, 4.1%, and 0.04, respectively.

[0085] In addition, the comparison chart of detection results is as follows: Figure 7 As shown, in remote sensing images containing floating targets with complex backgrounds, different scales, and different rotation directions, the original YOLOv8-OBB model exhibits issues with missed detections and inaccurate localization. Figure 7 As shown in (a) of the paper. The floating object detection model described in this invention can more accurately identify various floating object targets, and the rotated bounding box has a higher degree of fit with the true contour of the target, demonstrating stronger feature extraction and rotational target localization capabilities, such as... Figure 7 As shown in (b) of the paper, the effectiveness and feasibility of the method described in this invention are thus verified.

[0086] In summary, this invention, through innovative improvements to the YOLOv8-OBB network, achieves an efficient and accurate remote sensing detection method for floating objects in power transmission line corridors. It fully demonstrates that the method described in this invention can effectively utilize satellite remote sensing imagery to achieve high-precision and robust automatic detection of floating objects in power transmission line corridors under complex backgrounds, providing reliable technical support for the safe operation and maintenance of power transmission lines.

[0087] Corresponding to the aforementioned embodiments of the method for detecting floating objects in power transmission line channels based on satellite remote sensing images, the present invention also provides embodiments of a device for detecting floating objects in power transmission line channels based on satellite remote sensing images.

[0088] The present invention provides a device for detecting floating objects in power transmission line channels based on satellite remote sensing imagery, comprising one or more processors and a memory, wherein the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the method for detecting floating objects in power transmission line channels based on satellite remote sensing imagery in the above embodiment.

[0089] The embodiments of the floating object detection device for power transmission line channels based on satellite remote sensing imagery of the present invention can be applied to any device with data processing capabilities, such as a computer. The device embodiments can be implemented through software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, in addition to the processor, memory, network interface, and non-volatile memory, the data processing device in the embodiments typically includes other hardware depending on its actual functions, which will not be elaborated further.

[0090] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0091] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0092] This invention also provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements the above-described method for detecting floating objects in power transmission line channels based on satellite remote sensing imagery.

[0093] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0094] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting floating objects in power transmission line corridors based on satellite remote sensing imagery, characterized in that, Includes the following steps: S1. Acquire satellite remote sensing images of the power transmission line channel, preprocess them to construct a dataset of rotating target boxes of floating objects in the power transmission line channel, and divide them into training and test sets according to the proportions. S2. Improve the YOLOv8-OBB network to build a floating object detection model, and use the training set to train the floating object detection model to obtain a trained floating object detection model. S3. Use the trained floating object detection model to detect floating objects in the power transmission line channel, and output the target category, the coordinates of the rotated bounding box, and the corresponding rotation angle.

2. The method for detecting floating objects in power transmission line corridors based on satellite remote sensing imagery according to claim 1, characterized in that, Step S1 specifically includes the following sub-steps: S11. Obtain satellite remote sensing images of the power transmission line corridor; S12. Remove abnormal samples from the original satellite remote sensing images, convert the image format to JPEG format, and store them using RGB three channels; S13. Use the rotated bounding box annotation tool to annotate the floating targets in the image and obtain annotation information including target category, rotated bounding box coordinates and rotation angle; S14. Convert the format of the annotation information into a label format that the floating object detection model can recognize; S15. The image size is normalized using the Letterbox method; the image and its corresponding label information are enhanced using the MixUp data augmentation method. S16. Construct a rotating target bounding box dataset of floating objects in the power transmission line channel based on the enhanced image images and their corresponding label information, and divide it into training set and test set according to the proportion.

3. The method for detecting floating objects in power transmission line corridors based on satellite remote sensing imagery according to claim 1, characterized in that, The floating object targets include one or more of the following: dust nets, plastic greenhouses, mulch film, and corrugated steel sheets.

4. The method for detecting floating objects in power transmission line corridors based on satellite remote sensing imagery according to claim 1, characterized in that, The improvement of the YOLOv8-OBB network to construct a floating object detection model specifically includes: The improved backbone network is obtained by replacing the convolutional modules in the YOLOv8-OBB backbone network with multi-scale residual gated convolutional modules. After embedding the global-local collaborative attention module into the first upsampling module of the YOLOv8-OBB feature fusion network, an improved feature fusion network is obtained. A floating object detection model is constructed based on the improved backbone network, the improved feature fusion network, and the YOLOv8-OBB detection head network.

5. The method for detecting floating objects in power transmission line corridors based on satellite remote sensing imagery according to claim 4, characterized in that, In the floating object detection model, the improved backbone network includes a multi-scale residual gated convolution module, a C2f module, an efficient multi-scale attention module, and a spatial pyramid pooling module; the improved feature fusion network includes an upsampling module, a global-local collaborative attention module, a stitching layer, a C2f module, and a convolution module; the detection head network includes three detection head output branches, which are used for the detection of floating objects at different scales; each output branch includes a target category branch for predicting the target category, a bounding box regression branch for predicting the center point coordinates and size of the rotated bounding box, and a rotation angle prediction branch for predicting the target rotation angle; The image is input into the floating object detection model. After passing through the backbone network, feature fusion network, and detection head network, the model outputs the predicted target category, the center point coordinates and size of the corresponding target's rotated bounding box, and the rotation angle of the corresponding target.

6. The method for detecting floating objects in power transmission line corridors based on satellite remote sensing imagery according to claim 4 or 5, characterized in that, The multi-scale residual gated convolutional module includes a local feature extraction branch, a mesoscale feature extraction branch, and a global context feature extraction branch. The local feature extraction branch uses a 3×3 convolution to extract local texture features, the mesoscale feature extraction branch uses a 5×5 convolution to extract mesoscale structural features, and the global context feature extraction branch uses a 7×7 depthwise separable convolution to extract global context features. The output features of the three feature extraction branches are concatenated and fused using a dynamic gating mechanism. Specifically, a 1×1 convolution is first used to map the concatenated features, then a sigmoid activation function is used to generate dynamic gating weights, which are then used to weight and fuse the output features of the three feature extraction branches to obtain fused features. Finally, the fused features and the input features of the multi-scale residual gated convolutional module are connected through a residual connection structure to obtain the output features of the multi-scale residual gated convolutional module.

7. The method for detecting floating objects in power transmission line corridors based on satellite remote sensing imagery according to claim 4 or 5, characterized in that, The global-local collaborative attention module includes a global feature branch and a local feature branch. The global feature branch uses global average pooling to extract global semantic features of the input features, and the local feature branch uses convolution operations to extract local detail features of the input features. The fusion weights are obtained through a dynamic weight generation mechanism. These fusion weights are generated by convolutional mapping and the Sigmoid activation function, and the attention enhancement features output by the global-local collaborative attention module are obtained by weighted fusion.

8. The method for detecting floating objects in power transmission line corridors based on satellite remote sensing imagery according to claim 1, characterized in that, The training of the floating object detection model using a training set specifically includes: The image images in the training set are input into the floating object detection model, which outputs the predicted target category, the center point coordinates and size of the corresponding target's rotated bounding box, and the rotation angle of the corresponding target. The binary cross-entropy loss between the predicted target category and its corresponding ground truth target category is calculated as the classification loss. The original scene intersection-over-union (IoU) loss is calculated based on the coordinates and dimensions of the predicted rotated bounding box and its corresponding ground truth rotated bounding box coordinates. An angle error constraint term is calculated based on the predicted rotation angle and its corresponding ground truth rotation angle. An aspect ratio coefficient is calculated based on the width and height of the current ground truth rotated bounding box to obtain the target shape perception weight. The angle error constraint term and the original scene IoU loss are then weighted and combined according to the target shape perception weight to form the final dynamic angle-perceived scene IoU loss function. The total loss function is calculated based on the classification loss and the dynamic angle-perceived scene IoU loss function. With minimizing the total loss function as the optimization objective, the parameters of the floating object detection model are adjusted until the preset number of training rounds is reached, and a well-trained floating object detection model is obtained.

9. A device for detecting floating objects in power transmission line corridors based on satellite remote sensing imagery, comprising one or more processors and a memory, characterized in that, The memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the method for detecting floating objects in power transmission line channels based on satellite remote sensing imagery as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, It stores a program that, when executed by a processor, is used to implement the method for detecting floating objects in power transmission line channels based on satellite remote sensing imagery as described in any one of claims 1-8.

Images