A Foreign Object Intrusion Detection Method for Power Transmission Corridors Based on ISD-Net

Abstract

This invention belongs to the interdisciplinary research field of power system automation and computer vision, specifically a foreign object intrusion detection method for power transmission corridors based on ISD-Net. The specific steps are as follows: First, a foreign object dataset for the power transmission corridor is constructed based on UAV inspection videos. Then, RGB images of the power transmission corridor are input into the parallel decoupled backbone module PDB for multi-scale feature extraction, resulting in a global semantic feature map F4 for the foreign object, a corridor saliency feature map F6, and a foreign object detail enhancement feature map F8. This invention dynamically and adaptively fuses the DAF-Neck and its MAGI submodule, enabling dynamic evaluation of scene complexity and selection of appropriate attention enhancement paths based on the actual content of the input features. This achieves more adaptive feature fusion, improving the model's detection stability and robustness in complex environments.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary research technology of power system automation and computer vision, specifically to a foreign object intrusion detection method for power transmission corridors based on ISD-Net. Background Technology

[0002] Against the backdrop of the national "14th Five-Year Plan" development strategy and the continuous advancement of smart grid construction, ensuring the safe, stable, and efficient operation of power transmission networks has become the cornerstone of supporting national economic development. However, power transmission networks are facing increasingly severe operational challenges. Frequent incidents of foreign object intrusion, such as bird nesting, kite entanglement, and construction cranes accidentally entering the network, pose a significant threat to line safety and can easily trigger chain reactions such as phase-to-phase short circuits and insulation deterioration, ultimately leading to large-scale power outages and causing incalculable economic losses and social impacts. Traditional inspection methods, primarily relying on manual foot patrols or tower climbing, suffer from low efficiency and insufficient accuracy. Therefore, utilizing drones equipped with visual sensors for automated inspections has become a mainstream trend in the industry. Thus, the power grid industry urgently needs methods that can automatically analyze drone inspection videos to detect and identify foreign object intrusions in transmission corridors, supporting the high-reliability and high-availability inspection requirements in complex scenarios along transmission corridors.

[0003] To address these challenges, deep learning-based computer vision technology has become central to image analysis in UAV power line inspections. Early research focused on directly applying general object detection models, such as the Faster R-CNN series (a two-stage detector) and the YOLO series (a single-stage detector), to foreign object intrusion detection tasks. These methods automatically extract image features through convolutional neural networks, achieving significant improvements in detection accuracy and generalization ability compared to traditional image processing methods that rely on manually designed features.

[0004] As research deepened, researchers discovered the unique characteristics of power transmission corridor scenarios. For example, target sizes vary greatly, ranging from large bird nests a few meters in size to thin lines a few centimeters in size, and the background is extremely complex, including intertwined tree branches, buildings, and the structures of the towers themselves. To address these issues, subsequent research began to specifically improve existing deep learning frameworks. Feature pyramid networks (FPNs) and their variants, such as the bidirectional feature pyramid network (BiFPN), are widely used to fuse features from different network levels to enhance the model's sensitivity to multi-scale targets. Furthermore, attention mechanisms such as the CBAM module and SE module have been introduced. By recalibrating the weights of the channel or spatial dimensions of the feature map, the model can focus on more information-rich regions and suppress interference from irrelevant background information.

[0005] While these improvements have enhanced detection performance to some extent, they mostly involve local optimizations within the existing detection framework. They do not fundamentally address the issue of small target information loss in the initial feature extraction stage of the backbone network, nor do they systematically resolve the conflict between the feature requirements of classification and localization tasks. Therefore, current technology still faces bottlenecks such as insufficient robustness in complex dynamic environments, high false negative rates for small targets, and the need to improve localization accuracy, as detailed below:

[0006] 1. Difficulty in effectively detecting foreign objects of multiple sizes simultaneously: Most backbone networks in existing technologies employ a serial downsampling structure. During the process of extracting high-level semantic features layer by layer, key recognition features such as fine edges and textures of small targets, such as distant Mars or tangled thin lines, are easily diluted or even completely lost. Subsequent feature fusion networks, such as FPN and BiFPN, cannot effectively recover the lost information, which fundamentally limits the model's ability to detect small targets.

[0007] 2. Unstable performance in foreign object intrusion detection in complex power transmission corridor environments: Power transmission corridor environments are complex and variable, with background elements such as tree branches and building outlines easily confused with foreign object targets; changes in lighting and target occlusion can significantly reduce the target's recognizability. Most existing models employ fixed attention mechanisms or feature fusion strategies, lacking the ability to dynamically adjust processing strategies based on the current image context, leading to a sharp decline in detection performance under complex or non-ideal conditions.

[0008] 3. Difficulty in achieving good results in both object identification and localization simultaneously: Object detection requires the simultaneous completion of two tasks: classification and localization. Classification identifies the object category, while localization accurately selects the object. Classification relies on global semantic features, while localization requires local spatial detail features. Existing methods typically use a compromise set of feature maps to drive both tasks simultaneously, or only separate the processing branches in the final prediction head. However, these methods do not eliminate the conflict between the two requirements at the feature level, making it difficult for the model to achieve high-precision localization and high-confidence classification simultaneously, especially when dealing with irregular or occluded objects.

[0009] Based on the above, a method for detecting foreign object intrusion in power transmission corridors based on ISD-Net is invented. Summary of the Invention

[0010] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0011] A method for detecting foreign object intrusion in power transmission corridors based on ISD-Net includes the following specific steps:

[0012] S1, Construct a foreign object dataset for power transmission corridors based on drone inspection videos;

[0013] S2, input the RGB image of the power transmission corridor into the parallel decoupled backbone module PDB (Parallel Decoupled Backbone) for multi-scale feature extraction to obtain the global semantic feature map of foreign objects F4, the corridor saliency feature map F6, and the foreign object detail enhancement feature map F8;

[0014] S3 inputs F4, F6 and F8 into the Dynamic Adaptive Fusion Neck module DAF-Neck to perform multi-scale feature fusion and adaptive enhancement, and outputs small-scale foreign object prediction map L7, medium-scale foreign object prediction map L10 and large-scale foreign object prediction map L12.

[0015] S4 inputs L7, L10 and L12 into the Task-Adaptive Synergistic Prediction Head module TASP-Head to perform task decoupling and collaborative prediction, and finally outputs a set of detection results of foreign objects in the power transmission corridor.

[0016] S5. Construct the ISD-Net model, which consists of the parallel decoupled backbone module PDB, the dynamic adaptive fusion neck module DAF-Neck, and the task adaptive collaborative prediction head module TASP-Head. Then train the ISD-Net model to obtain the optimized ISD-Net model.

[0017] S6 applies the optimized ISD-Net model to the actual UAV inspection task of the power transmission corridor and outputs the final detection results.

[0018] As a preferred embodiment of the foreign object intrusion detection method for power transmission corridors based on ISD-Net described in this invention, the specific steps of S1 are as follows:

[0019] S11 uses high-definition camera equipment mounted on a drone to conduct inspection flights over the power transmission corridor, recording inspection videos of foreign objects such as bird nests, kites, balloons, plastic bags, construction cranes, smoke / flames, etc.

[0020] S12, extract key frames from the power transmission corridor inspection video to obtain the power transmission corridor RGB image; depending on the specific scenario requirements, there are two methods: one is to sample at a fixed frequency and equal time interval, such as extracting 1 frame or 5 frames per second; the other is to use adaptive sampling based on dynamic scene changes, that is, to automatically increase the sampling rate when a significant change in the image is detected.

[0021] S13, Use the annotation tool to annotate the RGB image of the power transmission corridor, select the foreign objects in the image in the form of bounding boxes, and assign a corresponding category label to each bounding box;

[0022] S14, normalize the size of the RGB image of the power transmission corridor to obtain a size-normalized foreign object dataset of the power transmission corridor;

[0023] S15, perform data augmentation processing on the RGB images of the power transmission corridor by using random rotation, horizontal / vertical flipping, color jitter, adding random noise, and simulating complex environmental effects such as rain, fog, and occlusion, in order to expand the data scale and improve the model's generalization ability, and obtain the RGB image set of the power transmission corridor.

[0024] S16. The RGB image set of the power transmission corridor is divided into training set, validation set and test set according to the proportion, which are used for subsequent model training and evaluation.

[0025] As a preferred embodiment of the foreign object intrusion detection method for power transmission corridors based on ISD-Net described in this invention, the specific steps of S2 are as follows:

[0026] S21, input the RGB image of the power transmission corridor into the initial convolutional layer of Conv(3×3) convolution with a stride of 2 for downsampling operation, and combine BN and SiLU activation functions to obtain the initial feature map F1;

[0027] S22, F1 is split and processed in parallel heterogeneous manner to obtain the global semantic feature map of foreign objects F4, the corridor saliency feature map F6, and the foreign object detail enhancement feature map F8.

[0028] As a preferred embodiment of the foreign object intrusion detection method for power transmission corridors based on ISD-Net described in this invention, the specific steps of S3 are as follows:

[0029] S31. To address the issue of missed detection of small-scale foreign objects due to insufficient semantic information, firstly, F4 is upsampled to obtain an upsampled semantic feature map L1. Then, L1 and F6 are input together into the first MAGI submodule for processing to obtain a first-level attention-gated enhanced fusion feature map L2. Next, L2 is input into the C3k2 submodule for processing to obtain a neck feature map L3. Finally, L3 is upsampled to obtain an upsampled feature map L4, which is used for subsequent fusion with small-scale features.

[0030] S32, to utilize the details preserved by the high-resolution path and suppress interference from complex backgrounds, firstly, L4 and F8 are jointly input into the second MAGI submodule for dynamic fusion processing to obtain the second-level attention-gated enhanced fusion feature map L5; then, L5 is input into the C3k2 submodule for processing to enhance the fused small-scale features to obtain the small-scale neck feature map L6; finally, L6 is subjected to a Conv(3×3) convolution operation with a stride of 2 to obtain the small-scale foreign object prediction map L7;

[0031] S33. To feed back the precise positioning information from the lower layers to the mid-to-high-level features, firstly, L7 and the neck feature map L3 are concatenated to obtain the preliminary mesoscale fused feature map L8. Then, L8 is input into the channel compression C3k2 submodule for processing, so as to compress the number of channels from 512 to 256 while strengthening the mesoscale fused features, resulting in the mesoscale foreign object prediction map L10. Next, L10 is subjected to a Conv(3×3) convolution operation with a stride of 2 to achieve a downsampling effect, resulting in the mesoscale core fused feature map L11. Finally, L11 and F4 are concatenated to obtain the large-scale foreign object prediction map L12.

[0032] As a preferred embodiment of the foreign object intrusion detection method for power transmission corridors based on ISD-Net described in this invention, the specific steps of S4 are as follows:

[0033] S41, input L7 into the first TAF for processing to obtain small-scale fusion feature map X1; input L10 into the second TAF for processing to obtain medium-scale fusion feature map X3; input L12 into the third TAF for processing to obtain large-scale fusion feature map X5;

[0034] S42, input feature map X1 into submodule C3k2 for processing to obtain small-scale prediction pre-feature map X2; input feature map X3 into submodule C3k2 for processing to obtain medium-scale prediction pre-feature map X4; input feature map X5 into submodule C3k2 for processing to obtain large-scale prediction pre-feature map X6.

[0035] S43, input X2 into the first SPN submodule to obtain the small-scale detection result R1; input X4 into the second SPN submodule to obtain the medium-scale detection result R2; input X6 into the third SPN submodule to obtain the large-scale detection result R3;

[0036] S44, R1, R2, and R3 form a detection result set, which is the final detection result of foreign objects in the power transmission corridor.

[0037] As a preferred embodiment of the foreign object intrusion detection method for power transmission corridors based on ISD-Net described in this invention, the specific steps of S5 are as follows:

[0038] S51, Construct the ISD-Net model, which includes the parallel decoupling backbone module PDB, the dynamic adaptive fusion neck module DAF-Neck, and the task adaptive collaborative prediction head module TASP-Head;

[0039] S52. First, initialize all neural network parameters and set specific hyperparameters for the foreign object intrusion detection task in power transmission corridors, including training epochs, batch size, optimizer, and learning rate. Then, input the training and validation sets containing various types of foreign objects in power transmission corridors into the intelligent collaborative foreign object intrusion detection model for end-to-end training. During training, the network parameters are continuously updated through the backpropagation algorithm and a specially designed loss function for foreign object intrusion detection in power transmission corridors until the model's foreign object intrusion detection performance on the validation set tends to converge. At this point, training ends, and the optimized ISD-Net model is obtained.

[0040] Compared with existing technologies:

[0041] 1. Detection capability for targets of different sizes: Existing serial downsampling backbone networks lose fine edge and texture information of small foreign objects when extracting high-level semantics. This invention decouples the backbone PDB in parallel, outputting three feature maps suitable for detecting large, medium, and small-scale foreign objects respectively. Through the multi-scale context fusion submodule MCFM, detailed information of small foreign objects is fully preserved while extracting high-level semantics, solving the problem of small target information being lost in the initial stage in traditional serial backbone networks. This module is designed to fully preserve the detailed information of small foreign objects from the initial stage of the network, while also taking into account the semantic context of large-sized foreign objects, solving the problem of insufficient coverage of targets of different sizes in traditional serial structures.

[0042] 2. Adaptability in complex environments: Existing methods typically employ fixed attention mechanisms or fusion strategies, which lead to a significant drop in detection performance when faced with complex scenarios such as intertwined tree branches, building interference, lighting changes, or partial target occlusion during power transmission corridor inspections. This invention, through dynamic adaptive fusion of the DAF-Neck and its MAGI submodule, can dynamically assess the complexity of the scene and select appropriate attention enhancement paths based on the actual content of the input features, thereby achieving more adaptive feature fusion and improving the model's detection stability and robustness in complex environments.

[0043] 3. Collaboration for Classification and Localization Tasks: Traditional detection models typically use the same set of features to complete both classification and localization tasks simultaneously, or employ only simple branch separation in the detection head. This often leads to problems such as bounding box offset or mismatch between classification confidence and localization quality in practical applications. This invention utilizes the task-adaptive collaborative prediction head TASP-Head. The TAF submodule extracts task-specific features for classification and localization tasks through two structurally asymmetric paths. Based on this, the SPN submodule introduces a bidirectional calibration mechanism to promote mutual verification and optimization between classification and localization results, thereby improving both classification accuracy and localization precision. This approach is suitable for detecting foreign objects with irregular shapes or blurred boundaries. Attached Figure Description

[0044] Figure 1 This is the overall flowchart of the present invention;

[0045] Figure 2 This is a diagram of the parallel decoupling backbone module of the present invention;

[0046] Figure 3 This is a structural diagram of the multi-scale context fusion submodule of the present invention;

[0047] Figure 4 This is a diagram of the dynamic adaptive fusion neck module of the present invention;

[0048] Figure 5 This is a diagram of the attention-gated inference submodule of the present invention;

[0049] Figure 6 This is a diagram of the attention gating network of the present invention;

[0050] Figure 7 This is a diagram of the task adaptive collaborative prediction head module of the present invention;

[0051] Figure 8 This is a structural diagram of the task adaptive fusion submodule of the present invention;

[0052] Figure 9 This is a structural diagram of the collaborative prediction hub submodule of the present invention;

[0053] Figure 10 This is a diagram of the main structure of the ISD-Net of the present invention;

[0054] Figure 11 This is an RGB image of the power transmission corridor of the present invention;

[0055] Figure 12 This is an RGB image of the power transmission corridor of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0057] This invention provides a foreign object intrusion detection method for power transmission corridors based on ISD-Net. Please refer to [link / reference]. Figures 1-11 The specific steps are as follows:

[0058] S1, Construct a foreign object dataset for power transmission corridors based on drone inspection videos;

[0059] The specific steps of S1 are as follows:

[0060] S11 uses high-definition camera equipment mounted on a drone to conduct inspection flights over the power transmission corridor, recording inspection videos of foreign objects such as bird nests, kites, balloons, plastic bags, construction cranes, smoke / flames, etc.

[0061] S12, extract key frames from the power transmission corridor inspection video to obtain the power transmission corridor RGB image; depending on the specific scenario requirements, there are two methods: one is to sample at a fixed frequency and equal time interval, such as extracting 1 frame or 5 frames per second; the other is to use adaptive sampling based on dynamic scene changes, that is, to automatically increase the sampling rate when a significant change in the image is detected.

[0062] S13, Use the annotation tool to annotate the RGB image of the power transmission corridor, select the foreign objects in the image in the form of bounding boxes, and assign a corresponding category label to each bounding box;

[0063] S14, normalize the size of the RGB image of the power transmission corridor to obtain a size-normalized foreign object dataset of the power transmission corridor;

[0064] S15, perform data augmentation processing on the RGB images of the power transmission corridor by using random rotation, horizontal / vertical flipping, color jitter, adding random noise, and simulating complex environmental effects such as rain, fog, and occlusion, in order to expand the data scale and improve the model's generalization ability, and obtain the RGB image set of the power transmission corridor.

[0065] S16. The RGB image set of the power transmission corridor is divided into training set, validation set and test set according to the proportion, which are used for subsequent model training and evaluation.

[0066] The annotation tools used include, but are not limited to, LabelImg, Labelme, and CVAT. These tools annotate the location and category information of target objects in images through human-computer interaction.

[0067] S1 Implementation Example:

[0068] First, using high-definition cameras mounted on drones, an inspection flight was conducted over the power transmission corridor, recording approximately 4 hours of inspection video at a frame rate of 30fps. The video included various foreign objects intruding into the power transmission corridor, such as bird nests, kites, balloons, plastic bags, construction cranes, smoke, and flames.

[0069] Then, keyframes were extracted from the inspection video using an equal-time sampling method, at a frequency of 2 frames per second, resulting in approximately 28,800 original RGB images of the power transmission corridor. After initial screening, meaningless frames with blurry images or no foreign objects were removed, resulting in 2,000 valid images containing the aforementioned six types of foreign objects.

[0070] Next, the bounding boxes of foreign objects in the valid images are labeled using the annotation tool, and each bounding box is assigned a corresponding category label.

[0071] Next, the labeled images were normalized to a size of 640×640 pixels, resulting in a normalized dataset of foreign objects in the power transmission corridor.

[0072] Finally, data augmentation was performed on the dataset, specifically using random rotation, horizontal or vertical flipping, color jittering, adding random noise, and simulating complex environmental effects such as rain, fog, and occlusion, expanding the data size from 2000 images to 8000 images. After data augmentation, the augmented power transmission corridor RGB image set was divided into training, validation, and test sets in an 8:1:1 ratio. The training set contained 6400 images, the validation set contained 800 images, and the test set contained 800 images, used for subsequent model training and evaluation.

[0073] S2, input the RGB image of the power transmission corridor into the parallel decoupled backbone module PDB (Parallel Decoupled Backbone) for multi-scale feature extraction to obtain the global semantic feature map of foreign objects F4, the corridor saliency feature map F6, and the foreign object detail enhancement feature map F8;

[0074] The Parallel Decoupled Backbone Module (PDB) addresses the issue of extremely large target size differences in foreign object intrusion detection in power transmission corridors. By setting up three parallel functional specialization branches, it is suitable for capturing semantic context information of large foreign objects, filtering salient features of medium-sized foreign objects, and extracting high-resolution details of small foreign objects, respectively. This solves the problem of insufficient compatibility of traditional networks when dealing with the detection of large, medium and small foreign objects in power transmission corridor scenarios.

[0075] The beneficial effects of the PDB module are specifically manifested in the following ways: First, through its high-resolution feature extraction path, it fully preserves the effective identification information of small foreign object targets, improving the detection recall rate of such easily missed targets; Second, it takes into account the detection of large, medium, and small foreign objects, broadening the application scope of the foreign object intrusion detection model for power transmission corridors. Its module structure is as follows: Figure 2 As shown.

[0076] The specific steps of S2 are as follows:

[0077] S21, input the RGB image of the power transmission corridor into the initial convolutional layer of Conv(3×3) convolution with a stride of 2 for downsampling operation, and combine BN and SiLU activation functions to obtain the initial feature map F1;

[0078] S22, F1 is split and processed in parallel heterogeneous manner to obtain the global semantic feature map of foreign objects F4, the corridor saliency feature map F6, and the foreign object detail enhancement feature map F8.

[0079] In the first branch, F1 is first input into a Conv(3×3) convolution with a stride of 2 for downsampling, resulting in a downsampled semantic feature map F2. In path one, F2 is input into the MCFM submodule for processing, resulting in a multi-scale contextual feature map F3. F3 is then input into the C2PSA submodule for processing, resulting in a foreign object global semantic feature map F4. In path two, F2 is input into the lightweight attention submodule for processing, resulting in an attention-weighted feature map F5. F5 is then input into the mid-layer C3k2 submodule for processing, resulting in a corridor saliency feature map F6.

[0080] In the second branch, F1 is first processed by multi-scale convolution to obtain a multi-scale parallel feature map F7. F7 is then input into the lightweight C3k2 submodule for processing. The lightweight C3k2 submodule adopts an 8-group grouping structure and a structure combining convolution and residual connections, which enhances the continuity of feature propagation while controlling the amount of computation, thereby obtaining the foreign object detail enhancement feature map F8.

[0081] Among them: 1. Multi-scale convolution uses three different scale convolution kernels, Conv(1×1), Conv(3×3) and Conv(5×5), in parallel, and finally concatenates the results, which can simultaneously capture multi-scale information from local details to global context, effectively preserving the detailed features of small-sized foreign objects.

[0082] 2. The lightweight C3k2 submodule introduces residual connections on the basis of grouped convolution (8 groups), which greatly reduces the number of parameters while ensuring the stability of gradient flow and feature propagation. It is dedicated to high-resolution paths to enhance the ability to extract details of foreign objects.

[0083] This invention designs a multi-scale context fusion module (MCFM). Its core design concept addresses the technical challenges of detecting foreign objects in power transmission corridors, such as large differences in target scale and complex backgrounds, by constructing parallel and heterogeneous feature processing pathways. Three branches work collaboratively, enabling the model to simultaneously capture contextual information of large-sized foreign objects, salient features of medium-sized targets, and detailed structural details of small-sized foreign objects, thereby improving the overall detection performance for multi-scale foreign objects. The module employs lightweight structures such as depthwise separable convolution and grouped convolution, maintaining high accuracy while controlling computational overhead, making it suitable for embedded inspection equipment with high real-time requirements. The structure of the MCFM submodule is as follows: Figure 3 As shown.

[0084] Let the input of the MCFM submodule be the downsampled semantic feature map M0, and in the MCFM submodule, the input is F2. The operation process of the MCFM submodule is as follows [the steps of the MCFM submodule are labeled A1 and A2 below]:

[0085] A1: In branch one, M0 is first input into a grouped convolution (Group Conv(3×3) with 8 groups), and detailed features are extracted using Batch Normalization (BN) and SiLU activation functions to obtain a high-resolution detailed feature map M1. M1 is then input into the SE channel attention submodule for processing to obtain a channel-weighted detailed feature map M2. M2 is then input into a Conv(1×1) convolution with a stride of 2 for downsampling, and combined with BN processing to obtain a multi-scale detailed feature map M3.

[0086] In branch two, M0 is first fed into a global average pooling layer for global context extraction, resulting in a global context vector M4. M4 is then fed into a Conv(1×1) convolution for feature expansion, combined with Batch Normalization (BN) and the SiLU activation function, yielding an expanded context vector M5. M5 is then fed into a bilinear upsampling layer for spatial reconstruction, resulting in a reconstructed context feature map M6. M6 is then fed into a Conv(1×1) convolution combined with the Sigmoid activation function to obtain the context spatial attention M7. M7 is then multiplied element-wise with M0 to obtain a spatially weighted feature map M8. Finally, M8 is fed into a Conv(1×1) convolution with a stride of 2 for downsampling, combined with BN, to obtain the context-enhanced feature map M9.

[0087] In branch three, M0 is first input into a Conv(1×1) convolution for channel compression, and then combined with BN and SiLU activation functions to obtain the compressed feature map M10. M10 is then input into a Depthwise Conv(3×3) depthwise separable convolution with a stride of 2 for downsampling, and then combined with BN and SiLU activation functions to obtain the multi-scale pyramid feature map M11.

[0088] A2: Concatenate M3, M9, and M11 to obtain the multi-scale fused feature map M12. Input M12 into a Conv(1×1) convolution for feature refinement, and combine it with BN and SiLU activation functions to obtain the multi-scale context feature map F3.

[0089] Among them: 1. The C2PSA submodule integrates the cross-layer features of the CSP structure and the PSA attention mechanism. It enriches the information flow through cross-layer feature fusion and uses spatial attention to focus on key areas, which can efficiently extract multi-scale features and enhance the global semantic information of foreign objects.

[0090] 2. The lightweight attention submodule combines spatial attention and simplified SE channel attention, which can dynamically assign weights to the spatial location and channels of the feature map, thereby suppressing background interference, highlighting salient features in the power transmission corridor, and with low computational overhead.

[0091] 3. The middle layer C3k2 submodule is constructed using grouped convolution (8 groups). It is located in the middle layer of the network and is responsible for deepening the downsampled features. It can effectively extract the salient features of medium-sized foreign objects and control the computational cost while ensuring performance.

[0092] S2 Implementation Example:

[0093] The input is a 640×640 pixel, 3-channel RGB image of a power transmission corridor captured by a drone. First, the image undergoes initial convolutional layer processing to obtain an initial feature map F1 of 320×320 pixels and 64 channels. Then, the initial feature map F1 is fed into two parallel branches for processing:

[0094] In the first branch, F1 is input into a Conv(3×3) convolution with a stride of 2 for downsampling, resulting in a 160×160 pixel, 128-channel downsampling semantic feature map F2. In path one, F2 is input into the MCFM submodule for processing, resulting in an 80×80 pixel, 256-channel multi-scale contextual feature map F3.

[0095] In the MCFM submodule: In branch one, F2 is input to a grouped convolution (Group Conv(3×3) with 8 groups, processed using BN and SiLU activation functions to obtain a high-resolution detail feature map M1 of 160×160 pixels and 96 channels. M1 is then input to the SE channel attention submodule for processing, resulting in a channel-weighted detail feature map M2 of 160×160 pixels and 96 channels. Finally, M2 is input to a Conv(1×1) convolution with a stride of 2 for downsampling, processed using BN, to obtain a multi-scale detail feature map M3 of 80×80 pixels and 160 channels. In branch two, F2 is input to a global average pooling layer for global context extraction, resulting in a global context vector M4 of 1×1 pixels and 128 channels. M4 is then input to a Conv(1×1) convolution for feature expansion, processed using BN and SiLU activation functions, to obtain an expanded context vector M5 of 1×1 pixels and 256 channels. M5 is input to a bilinear upsampling layer for spatial reconstruction, resulting in a 160×160 pixel, 256-channel reconstructed context feature map M6. M6 is then input to a Conv(1×1) convolution combined with a Sigmoid activation function to obtain a 160×160 pixel, 128-channel context spatial attention map M7. M7 is then element-wise multiplied with the original feature map M0 to obtain a 160×160 pixel, 128-channel spatially weighted feature map M8. Finally, M8 is input to a Conv(1×1) convolution with a stride of 2 for downsampling, combined with Batch Normalization (BN), to obtain an 80×80 pixel, 64-channel context enhancement feature map M9. In branch three, F2 is input to a Conv(1×1) convolution for channel compression, combined with BN and SiLU activation functions, to obtain a 160×160 pixel, 32-channel compressed feature map M10. M10 is input into a 3×3 depthwise separable convolution with a stride of 2 for downsampling, combined with Batch Normalization (BN) and SiLU activation functions, resulting in an 80×80 pixel, 32-channel multi-scale pyramid feature map M11. Finally, M3, M9, and M11 are concatenated to obtain an 80×80 pixel, 256-channel multi-scale fused feature map M12. M12 is then input into a 1×1 Convolution for feature refinement, combined with BN and SiLU activation functions, resulting in an 80×80 pixel, 256-channel multi-scale contextual feature map F3.

[0096] F3 is input to the C2PSA submodule for processing, resulting in a 40×40 pixel, 512-channel global semantic feature map of foreign objects, F4. In path two, F2 is input to the lightweight attention submodule for processing, resulting in an 80×80 pixel, 256-channel attention-weighted feature map, F5. F5 is then input to the mid-layer C3k2 submodule for processing, resulting in an 80×80 pixel, 256-channel corridor saliency feature map, F6.

[0097] In the second branch, F1 is processed by multi-scale convolution to obtain a 320×320 pixel, 192-channel multi-scale parallel feature map F7. F7 is then input into the lightweight C3k2 submodule for processing to obtain a 160×160 pixel, 128-channel foreign object detail enhancement feature map F8.

[0098] S3 inputs F4, F6 and F8 into the Dynamic Adaptive Fusion Neck module DAF-Neck to perform multi-scale feature fusion and adaptive enhancement, and outputs small-scale foreign object prediction map L7, medium-scale foreign object prediction map L10 and large-scale foreign object prediction map L12.

[0099] The Dynamic Adaptive Fusion Neck (DAF-Neck) module is used to achieve intelligent fusion and enhancement of multi-scale transmission corridor features.

[0100] The DAF-Neck module achieves multi-scale feature fusion through a bidirectional interactive path and includes a MAGI submodule specifically designed for complex situations in power transmission corridor inspections, such as high background interference and small differences between foreground targets and the background. For example, when tree branches are intertwined or towers obstruct the view, the module dynamically adjusts the attention processing path to enhance key features. This module can adaptively select fusion strategies based on image content, improving the accuracy and adaptability of feature representation compared to the fixed feature fusion method in traditional neck networks.

[0101] The beneficial effects of the DAF-Neck module are as follows: It employs a dynamic adaptive fusion mechanism to improve the feature fusion performance of power transmission corridor images, enabling the neck network to adapt to dynamic changes in the corridor environment and providing more robust feature inputs for subsequent foreign object detection steps. Its DAF-Neck module structure is as follows: Figure 4 As shown.

[0102] The specific steps of S3 are as follows:

[0103] S31. To address the issue of missed detection of small-scale foreign objects due to insufficient semantic information, firstly, F4 is upsampled to obtain an upsampled semantic feature map L1. Then, L1 and F6 are input together into the first MAGI submodule for processing to obtain a first-level attention-gated enhanced fusion feature map L2. Next, L2 is input into the C3k2 submodule for processing to obtain a neck feature map L3. Finally, L3 is upsampled to obtain an upsampled feature map L4, which is used for subsequent fusion with small-scale features.

[0104] S32, to utilize the details preserved by the high-resolution path and suppress interference from complex backgrounds, firstly, L4 and F8 are jointly input into the second MAGI submodule for dynamic fusion processing to obtain the second-level attention-gated enhanced fusion feature map L5; then, L5 is input into the C3k2 submodule for processing to enhance the fused small-scale features to obtain the small-scale neck feature map L6; finally, L6 is subjected to a Conv(3×3) convolution operation with a stride of 2 to obtain the small-scale foreign object prediction map L7;

[0105] S33. To feed back the precise positioning information from the lower layers to the mid-to-high-level features, firstly, L7 and the neck feature map L3 are concatenated to obtain the preliminary mesoscale fused feature map L8. Then, L8 is input into the channel compression C3k2 submodule for processing, so as to compress the number of channels from 512 to 256 while strengthening the mesoscale fused features, resulting in the mesoscale foreign object prediction map L10. Next, L10 is subjected to a Conv(3×3) convolution operation with a stride of 2 to achieve a downsampling effect, resulting in the mesoscale core fused feature map L11. Finally, L11 and F4 are concatenated to obtain the large-scale foreign object prediction map L12.

[0106] The channel compression C3k2 submodule adds a channel compression operation before the standard C3k2 structure, aiming to significantly reduce the number of channels in the input features to lower computational complexity. Feature extraction is then performed using grouped convolutions and residual connections. It further optimizes computational efficiency while maintaining a lightweight design, making it ideal for deployment on resource-constrained devices and a key component for enabling real-time model execution.

[0107] The Meta-Attentive Gating Network (MAGI) submodule's core design concept lies in introducing a context-aware and dynamic decision-making mechanism for power transmission corridor inspection. This submodule contains a lightweight meta-attentive gating decision network capable of analyzing the complexity of input features in real time, such as determining whether the current feature is primarily affected by background branches or uneven lighting, and then dynamically selecting the optimal attention enhancement strategy. The significant benefits of this design include: better handling of common lighting variations and background interference in power transmission corridors; employing matching feature enhancement methods for different scenarios; and improving the detection capability for foreign objects. Furthermore, the submodule's lightweight structure helps achieve a balance between detection robustness and operational efficiency in resource-constrained airborne edge devices. The structure of the MAGI submodule is as follows: Figure 5 As shown.

[0108] Let the inputs to the MAGI submodule be feature map Z0 and feature map Z0', respectively. In the first MAGI submodule, these are L1 and F6, respectively. The execution process of the MAGI submodule is as follows [the steps of the MAGI submodule are labeled A1, A2, ... below]:

[0109] A1: First, concatenate Z0 and Z0' to obtain concatenated feature map Z1. Then, perform a Conv(1×1) convolution operation on Z1 to obtain gated feature map Z2. Finally, input Z2 into the following five branches for processing:

[0110] In the spatial priority branch, Z2 is first subjected to spatial attention processing to obtain spatial attention feature map Z3; then Z3 is subjected to channel attention processing to obtain spatial dominance enhancement feature map Z4.

[0111] In the channel-first branch, channel attention processing is first performed on Z2 to obtain the channel attention feature map Z5; then spatial attention processing is performed on Z5 to obtain the channel-dominant enhancement feature map Z6.

[0112] In the parallel weighted fusion branch, spatial attention processing is performed on Z2 to obtain spatial attention feature map Z7; at the same time, channel attention processing is performed on Z2 to obtain channel attention feature map Z8; then Z7 and Z8 are fused to obtain fused feature map Z9.

[0113] In the computation of the bypass branch, Z2 is subjected to DWConv2D (3×3) depth-separable convolution to obtain a lightweight convolutional feature map Z10.

[0114] In the weight sequence calculation branch, Z2 is input to the meta-attention gating network for processing, resulting in a four-dimensional gating path weight vector sequence [w1, w2, w3, w4].

[0115] A2: The outputs Z4, Z6, Z9, Z10 of the four parallel branches are weighted and fused with the corresponding weights w1, w2, w3, w4 generated by the meta-attention gating network to obtain the first meta-attention gating enhanced fusion feature map L2:

[0116]

[0117] The Meta-Attentive Gating Network (MAG) is a lightweight decision network. Through end-to-end learning, this network adaptively adjusts the weight distribution of multiple attention paths (space and channel) during fusion based on specific contexts common in power transmission corridors, such as lighting variations and background interference. This design maintains low computational overhead while providing adaptive weight allocation for subsequent feature weighting fusion, thus balancing detection accuracy and operational efficiency within the resource constraints of airborne edge devices, and enhancing the adaptability of the foreign object detection model to complex on-site environments. Its MAG structure is as follows: Figure 6 As shown.

[0118] The execution process of the meta-attention gating network is as follows: global average pooling is performed on Z2 to obtain a 256-dimensional global context descriptor feature vector C1; C1 is transformed by a fully connected layer to obtain a 64-dimensional compressed feature representation vector C2; C2 is activated by the ReLU function to obtain a 64-dimensional nonlinear activation feature vector C3; C3 is transformed by a fully connected layer to obtain a four-dimensional original path score logical value sequence [v1,v2,v3,v4]; the logical value is normalized by Softmax to finally obtain a four-dimensional gating path weight vector sequence [w1,w2,w3,w4].

[0119] Among them: 1. ReLU is a very commonly used activation function. Its rule is very simple: for the input value, if it is greater than zero, the output is the same; if it is less than or equal to zero, the output is zero. This one-size-fits-all nonlinear operation introduces a crucial nonlinear factor into neural networks, enabling the network to fit very complex data relationships.

[0120] S3 Implementation Example:

[0121] First, the 40×40 pixel, 512-channel F4 is upsampled by a factor of 2 to obtain an upsampled semantic feature map L1 of 80×80 pixels and 512 channels. L1 and the 80×80 pixel, 256-channel F6 are then input into the first MAGI submodule for processing to obtain a first-level attention-gated enhanced fusion feature map L2 of 80×80 pixels and 256 channels.

[0122] In the MAGI submodule, the input L1 and F6 are first concatenated along the channel dimension to obtain a concatenated feature map Z1 with 80×80 pixels and 768 channels. Then, a Conv(1×1) convolution operation is performed on Z1 to compress the number of channels to 256, resulting in a gated feature map Z2 with 80×80 pixels and 256 channels. The following five parallel branches process Z2 simultaneously: In the spatial priority branch, spatial attention processing is applied to Z2 to obtain an 80×80 pixel, 256-channel spatial attention feature map Z3; channel attention processing is applied to Z3 to obtain an 80×80 pixel, 256-channel spatial dominant enhancement feature map Z4; in the channel priority branch, channel attention processing is applied to Z2 to obtain an 80×80 pixel, 256-channel channel attention feature map Z5; spatial attention processing is applied to Z5 to obtain an 80×80 pixel, 256-channel channel dominant enhancement feature map Z6; in the parallel weighted fusion branch, spatial attention processing is applied to Z2 to obtain an 80×80 pixel, 256-channel spatial attention feature map Z7; simultaneously, channel attention processing is applied to Z2 to obtain an 80×80 pixel, 256-channel channel attention feature map. Attention feature map Z8; weighted fusion operation is performed on Z7 and Z8 to obtain fused feature map Z9 with 80×80 pixels and 256 channels; in the computation side-path branch, Z2 is processed by DWConv (3×3) depthwise separable convolution to obtain lightweight convolution feature map Z10 with 80×80 pixels and 256 channels; in the sequence weight calculation branch, Z2 is input into the meta-attention gating network for processing. First, global average pooling is performed on Z2 to obtain a 256-dimensional global context descriptor feature vector; the vector is transformed by a fully connected layer from 256 to 64 to obtain a 64-dimensional compressed feature representation vector; the vector is processed by the ReLU activation function to obtain a 64-dimensional nonlinear activation feature vector; the vector is transformed by a fully connected layer from 64 to 4 to obtain a 4-dimensional original path score logical value sequence [v1, [v1, v2, v3, v4]; finally, Softmax normalization is performed on [v1, v2, v3, v4] to obtain the four-dimensional gated path weights [w1, w2, w3, w4]. Finally, the outputs Z4, Z6, Z9, Z10 of the four branches are weighted and fused with the corresponding gated path weights to obtain the first-order attention-gated enhanced fusion feature map L2, which is 80×80 pixels and has 256 channels.

[0123] The 80×80 pixel, 256-channel L2 feature map is input into the C3k2 submodule for processing, keeping the feature map size and number of channels unchanged, to obtain the 80×80 pixel, 256-channel neck feature map L3. L3 is then upsampled by a factor of 2 to obtain the 160×160 pixel, 256-channel upsampled feature map L4.

[0124] Secondly, F8 and L4 are input together into the second MAGI submodule for processing, resulting in a 160×160 pixel, 128-channel second-order attention-gated enhanced fusion feature map L5. L5 is then input into the C3k2 submodule for processing, yielding a small-scale neck feature map L6 of 160×160×128. Finally, a Conv(3×3) convolution operation with a stride of 2 is performed on L6 to achieve downsampling, resulting in a small-scale foreign object prediction map L7 of 80×80×256.

[0125] Then: L7 and L3 are concatenated to obtain an 80×80×512 mesoscale preliminary fused feature map L8. L8 is then processed by the channel compression C3k2 submodule to obtain an 80×80×256 mesoscale foreign object prediction map L10. A Conv(3×3) convolution operation with a stride of 2 is performed on L10 to achieve downsampling, resulting in a 40×40×512 mesoscale core fused feature map L11. Finally, L11 and F4 are concatenated to obtain a 40×40×1024 large-scale foreign object prediction map L12.

[0126] Finally, the DAF-Neck module outputs three enhanced feature maps for prediction: a small-scale foreign object prediction map L7 of 80×80×256, a medium-scale foreign object prediction map L10 of 80×80×256, and a large-scale foreign object prediction map L12 of 40×40×1024.

[0127] S4 inputs L7, L10 and L12 into the Task-Adaptive Synergistic Prediction Head module TASP-Head to perform task decoupling and collaborative prediction, and finally outputs a set of detection results of foreign objects in the power transmission corridor.

[0128] In the Task Adaptive Collaborative Prediction Head (TASP-Head) module, a TAF submodule is designed to address the difference in feature requirements between classification and localization tasks in foreign object intrusion detection in power transmission corridors. This submodule performs task-specific processing on the input features, generating semantic features suitable for classification and spatial detail features suitable for localization. Additionally, an SPN submodule is designed to extract basic features through shared convolutional layers and establish an interactive calibration mechanism between the classification head and the regression head. This mechanism cross-validates the preliminary classification results with the localization results, and finally outputs consistent detection results through deep regression and deep classification units.

[0129] The TASP-Head module is designed to address the challenges of detecting foreign objects in power transmission corridors due to their diverse shapes and complex backgrounds. Through task-specific feature processing and interactive calibration of prediction results, it improves the balance between classification accuracy and positioning precision, enhancing the overall performance of foreign object intrusion detection while maintaining computational efficiency. The structure of the TASP-Head module is as follows: Figure 7 As shown.

[0130] The specific steps of S4 are as follows:

[0131] S41, input L7 into the first TAF for processing to obtain small-scale fusion feature map X1; input L10 into the second TAF for processing to obtain medium-scale fusion feature map X3; input L12 into the third TAF for processing to obtain large-scale fusion feature map X5;

[0132] S42, input feature map X1 into submodule C3k2 for processing to obtain small-scale prediction pre-feature map X2; input feature map X3 into submodule C3k2 for processing to obtain medium-scale prediction pre-feature map X4; input feature map X5 into submodule C3k2 for processing to obtain large-scale prediction pre-feature map X6.

[0133] S43, input X2 into the first SPN submodule to obtain the small-scale detection result R1; input X4 into the second SPN submodule to obtain the medium-scale detection result R2; input X6 into the third SPN submodule to obtain the large-scale detection result R3;

[0134] S44, R1, R2, and R3 form a detection result set, which is the final detection result of foreign objects in the power transmission corridor.

[0135] The Task-Adaptive Fusion (TAF) submodule generates feature representations rich in global contextual semantics for the foreign object classification task and preserves high-resolution spatial detail features for the foreign object localization task through two independent and structurally asymmetric processing paths. The benefits of this design are: resolving the contradiction between the difficulty of simultaneously capturing the strong semantic information required for foreign object classification and the fine spatial details required for foreign object localization in a unified feature map; and improving the detection accuracy for foreign objects with irregular shapes and blurred boundaries typical of power transmission corridors, such as enhancing the ability to distinguish between tangled thin wires on the corridor and distant smoke.

[0136] Generally, object detection comprises two core sub-tasks: classification, which determines the category of the target (e.g., a bird's nest or a kite), relying on strong semantic features rich in global contextual information; and localization, which accurately outlines the target's location, relying on high-resolution spatial features that preserve detail. Its module structure is as follows: Figure 8As shown.

[0137] Let the input of the TAF submodule be the small-scale foreign object prediction map K0. In the first TAF submodule, the input is L7. The operation process of the TAF submodule is as follows [the steps of the TAF submodule are labeled A1 and A2 below]:

[0138] A1: Input K0 into the Bidirectional Feature Pyramid Network (BiFPN) to obtain the bidirectional feature map K1. Then, K1 is processed by the Task Split unit, which uses two Conv(1×1) convolutions with non-shared weights to generate a separated localization feature map K2 and a separated semantic feature map K6. K2 is then input into the localization task path for processing, and K6 is input into the classification task path for processing.

[0139] (1) In the localization task path, firstly, the separated localization feature map K2 is processed by Conv2D (1×1) convolution, and combined with the SiLU activation function to generate the localization task convolutional feature map K3. Then, K3 is processed by DWConv2D (3×3) depthwise separable convolution, and the SiLU activation function is applied to obtain the localization task convolutional feature map K4. Then, K4 is processed by Conv2D (1×1) convolution to generate the localization specialized feature map K5.

[0140] (2) In the classification task path, the semantic feature map K6 is first processed by Conv2D (1×1) convolution and combined with the SiLU activation function to generate the classification task convolutional feature map K7. K7 is then processed by DWConv2D (5×5) depthwise separable convolution and the SiLU activation function is applied to obtain the classification task depthwise convolutional feature map K8. K8 is then input into the squeezing and activation network submodule SE for processing to obtain the attention-weighted classification feature map K9. Finally, K9 is processed by Conv2D (1×1) convolution to obtain the classification specialized feature map K10.

[0141] A2: Adaptive fusion of K5 and K10 is performed to obtain small-scale fused feature map X1;

[0142] Among them, the Squeeze-and-Excitation Network (SE) submodule is a lightweight and efficient channel attention mechanism. Its core idea is to teach the neural network to pay attention to which channels, i.e. which features, in the feature map are more important, and to adaptively enhance important features and suppress secondary features.

[0143] Among them, the BiFPN bidirectional feature pyramid network efficiently integrates high-level semantic features and low-level detail features through bidirectional cross-scale connections from top to bottom and bottom to top, combined with a learnable weighted fusion mechanism. High-level semantic features are used to identify target categories, while low-level detail features are used to capture target edges and textures. At the same time, it simplifies redundant structures to control computational load, ultimately improving the recognition accuracy of small targets and occluded targets, and providing more comprehensive and accurate feature support for subsequent tasks.

[0144] Among them, adaptive fusion doubles the number of feature channels while integrating multi-source information by splicing two input feature maps and applying Conv(1×1) convolution to expand the channel dimension, thereby significantly improving the feature representation capability and model capacity and providing richer feature information for subsequent layers, but it will also increase the computational overhead and the number of parameters accordingly.

[0145] The Synergistic Predictive Nexus (SPN) submodule's core design concept lies in establishing a bidirectional calibration and collaborative inference mechanism between foreign object classification and localization regression tasks. The beneficial effects of this design are: effectively solving the technical problem of high confidence in foreign object classification but low localization accuracy in traditional detection heads, such as correctly identifying kites but with offset bounding boxes; and ensuring that foreign object intrusion detection results achieve both accuracy in category semantic recognition and spatial localization accuracy through forced mutual calibration and collaborative optimization of classification and regression tasks, thereby improving the reliability and accuracy of foreign object intrusion alarm results in power transmission corridors. Its module structure is as follows: Figure 9 As shown.

[0146] Let the input of the SPN submodule be the small-scale foreign object prediction map K0. In the first SPN submodule, the input is X2. The operation process of the SPN submodule is as follows [the steps of the SPN submodule are labeled A1, A2, and A3 below]:

[0147] A1: Perform a shared convolution process of Conv2D(3×3)+BN+ReLU on B0 to obtain a shared feature map B1. Subsequently, B1 is processed by four branches: In the first and fourth branches, B1 is directly retained for subsequent interactive calibration; in the second branch, B1 is processed by Conv2D(1×1)+GAP+FC through the regression head to obtain the initial foreign object localization vector B2; at the same time, in the third branch, B1 is input into the classification head and processed by Conv2D(1×1)+GAP+FC+Softmax to obtain the initial foreign object classification vector B5.

[0148] A2: In the interactive calibration phase, B1 and B2 are multiplied element-wise to obtain the calibration regression feature map B3; at the same time, B1 and B5 are multiplied element-wise to obtain the calibration classification feature map B6.

[0149] A3: In the enhancement phase, B3 is input to the depth regression unit, which sequentially performs Conv2D(3×3)+BN+ReLU and two consecutive Conv2D(1×1)+FC processes to obtain location information B4. B6 is input to the depth classification unit, which sequentially performs SE, Conv2D(3×3), GAP, FC, and Softmax processes to obtain classification information B7. B4 and B7 together constitute the foreign object detection result R, that is, R includes foreign object location information and foreign object category information.

[0150] Among them, 1. GAP (Global Average Pooling) is a spatial pooling operation that replaces traditional fully connected layers. Its process involves averaging the spatial dimensions of each channel of the feature map, generating a channel-level global descriptor. This operation significantly reduces the number of model parameters, effectively suppresses overfitting, and enhances the model's generalization ability and interpretability by preserving channel semantic information.

[0151] 2. The Fully Connected Layer (FC) is the core component for achieving high-level feature integration. This layer maps the input feature vector to a high-dimensional hidden space through dense connections, learning complex nonlinear relationships between features. Mathematically, it is a composite operation of affine transformation and activation function, and in classification networks, it undertakes the final transformation from abstract features to category discrimination.

[0152] 3. The Softmax function is the standard normalized exponential function for multi-class classification problems. This function transforms the original logits scalar into a discrete probability distribution through exponential transformation and normalization operations, and its output satisfies the normative requirements of probability axioms. During model inference, the probability vector output by this function provides the necessary metric basis for calculating cross-entropy loss.

[0153] 4. Conv2D is a core layer in convolutional neural networks used to process two-dimensional data such as images, functioning similarly to a feature extractor. It slides a set of learnable convolutional kernels across the input image, calculating the dot product of a filter and the corresponding image region at each location, thereby detecting local features in the image, such as edges, corners, textures, or more complex patterns. This process generates new feature maps, where the value at each location represents the presence of a specific feature in the original input, achieving effective preservation and abstraction of spatial information.

[0154] S4 Implementation:

[0155] First, the L7, with a size of 80×80 pixels and 256 channels, is input into the first TAF submodule. The L7 is first processed by BiFPN to obtain a bidirectional feature map K1 with a size of 80×80 pixels and 256 channels. Then, K1 is processed by the task separation unit, using two Conv(1×1) convolutions with non-shared weights to split the feature map into two parallel and structurally asymmetric processing paths, resulting in a separation localization feature map K2 with a size of 80×80 pixels and 128 channels and a separation semantic feature map K6 with a size of 80×80 pixels and 128 channels, respectively.

[0156] In the localization task path, K2 is first processed by Conv(1×1) convolution and SiLU activation function is applied to obtain the first convolution feature map K3 with 80×80 pixels and 128 channels. Then, K3 is processed by DWConv2D(3×3) depthwise separable convolution and SiLU activation function is applied to obtain the first convolution feature map K4 with 80×80 pixels and 128 channels. Then, K4 is processed by Conv(1×1) convolution to obtain the localization specialization feature map K5 with 80×80 pixels and 64 channels.

[0157] In the classification task path, K6 is first processed by Conv(1×1) convolution and combined with the SiLU activation function to obtain the first convolutional feature map K7 with 80×80 pixels and 128 channels. K7 is then processed by DWConv2D(5×5) depthwise separable convolution and the SiLU activation function is applied to obtain the next depthwise convolutional feature map K8 with 80×80 pixels and 128 channels. K8 is then processed by SE to obtain the attention-weighted classification feature map K9 with 80×80 pixels and 128 channels. Finally, K9 is processed by Conv(1×1) convolution to obtain the final classification specialized feature map K10 with 80×80 pixels and 64 channels.

[0158] Adaptive fusion of K5 and K10 yields a small-scale fused feature map X1 with 80×80 pixels and 128 channels.

[0159] Next, X1 is input into C3k2 for processing to obtain a small-scale prediction pre-feature map X2 with 80×80 pixels and 128 channels.

[0160] Finally, X2 is input to the first SPN submodule, where it undergoes a shared convolutional process consisting of a Conv(3×3) layer, a batch normalization layer, and a ReLU activation function to obtain a shared feature map B1 with a size of 80×80 pixels and 128 channels. Subsequently, B1 is split into four processing branches: in the first and fourth branches, B1 directly retains the original feature map for subsequent interactive calibration; in the second branch, B1 undergoes a regression head process consisting of a Conv(1×1) convolution, a global average pooling layer, and a fully connected layer to obtain the initial foreign object localization vector B2; in the third branch, B1 undergoes a classification head process consisting of a Conv(1×1) convolution, a global average pooling layer, a fully connected layer, and a Softmax normalization function to obtain the initial foreign object classification vector B5. In the interactive calibration phase, the shared feature map B1 (80×80 pixels, 128 channels) in the first branch is element-wise multiplied with the initial foreign object localization vector B2 to generate a calibration regression feature map B3 (maintaining the size of 80×80 pixels and 128 channels). Simultaneously, the shared feature map B1 (80×80 pixels, 128 channels) in the fourth branch is element-wise multiplied with the initial foreign object classification vector B5 to generate a calibration classification feature map B6 (maintaining the size of 80×80 pixels and 128 channels). In the refinement phase, the calibration regression feature map B3 is processed by a deep regression unit consisting of Conv(3×3) convolution, batch normalization layer, ReLU activation function, and two consecutive Conv(1×1) convolutions followed by a fully connected operation, outputting the final location information. The calibration classification feature map B6 is processed by a deep classification unit consisting of SE processing, Conv(3×3) convolution, global average pooling layer, fully connected layer, and Softmax normalization function, outputting the final category information.

[0161] Similarly, the 80×80 pixel, 256-channel L10 is input into the second TAF submodule to obtain an 80×80×128 mesoscale fusion feature map X3. X3 is then processed by C3k2 to obtain an 80×80×128 mesoscale prediction pre-feature map X4. The 40×40 pixel, 1024-channel L12 is input into the third TAF submodule to obtain a 40×40×512 large-scale fusion feature map X5. X5 is then processed by C3k2 to obtain a 40×40×512 large-scale prediction pre-feature map X6. X4 is then processed by the second SPN submodule to obtain the mesoscale detection result. X6 is then processed by the third SPN submodule to obtain the large-scale detection result. The outputs of the three SPN submodules together constitute a detection result set containing foreign object intrusion detection information at all scales, ultimately forming the foreign object detection result for the power transmission corridor.

[0162] S5. Construct the ISD-Net model, which consists of the parallel decoupled backbone module PDB, the dynamic adaptive fusion neck module DAF-Neck, and the task adaptive collaborative prediction head module TASP-Head. Then train the ISD-Net model to obtain the optimized ISD-Net model.

[0163] The specific steps of S5 are as follows:

[0164] S51, Construct the ISD-Net model, which includes a parallel decoupled backbone module PDB, a dynamic adaptive fusion neck module DAF-Neck, and a task adaptive collaborative prediction head module TASP-Head; its main structure is as follows: Figure 10 As shown;

[0165] (1) Parallel decoupling backbone module PDB: The RGB image of the power transmission corridor is input into the parallel decoupling backbone module for processing, and multi-scale feature extraction is performed in parallel to obtain the global semantic feature map of foreign objects F4, the corridor saliency feature map F6 and the foreign object detail enhancement feature map F8;

[0166] (2) Dynamic Adaptive Fusion Neck Module DAF-Neck: Input F4, F6 and F8 into the dynamic adaptive fusion neck module to perform multi-scale feature fusion and adaptive enhancement processing to obtain small-scale foreign object prediction map L7, medium-scale foreign object prediction map L10 and large-scale foreign object prediction map L12.

[0167] (3) Task Adaptive Collaborative Prediction Head Module TASP-Head: Performs task decoupling and collaborative prediction processing on L7, L10 and L12 to obtain the set of detection results of foreign objects in the power transmission corridor;

[0168] S52. First, initialize all neural network parameters and set specific hyperparameters for the foreign object intrusion detection task in power transmission corridors, including training epochs, batch size, optimizer, and learning rate. Then, input the training and validation sets containing various types of foreign objects in power transmission corridors into the intelligent collaborative foreign object intrusion detection model for end-to-end training. During training, the network parameters are continuously updated through the backpropagation algorithm and a specially designed loss function for foreign object intrusion detection in power transmission corridors until the model's foreign object intrusion detection performance on the validation set tends to converge. At this point, training ends, and the optimized ISD-Net model is obtained.

[0169] S52 Example:

[0170] The training run consisted of 100 epochs with a batch size of 16. AdamW was chosen as the optimizer, with an initial learning rate of 1e-4. Cosine annealing was employed to decay the learning rate and improve the model's convergence stability in complex power transmission corridor environments. The loss function was designed specifically for foreign object intrusion detection in power transmission corridors, consisting of classification and regression losses. Focal Loss was used for classification to address the imbalance of foreign object categories in power transmission corridors, while CIoU Loss was used for regression to improve the accuracy of foreign object bounding box localization. During training, the model parameters were updated using backpropagation based on the total loss value for foreign object intrusion detection in power transmission corridors.

[0171] AdamW is an improved version of the Adam optimizer. It separates weight decay (L2 regularization) from gradient updates, instead of mixing weight decay into gradient calculation as in the original Adam. This decoupling allows weight decay to truly play a regularization role, effectively preventing overfitting and improving the model's generalization ability. It has become the preferred optimizer in modern deep learning, especially in Transformer models.

[0172] S6 applies the optimized ISD-Net model to the actual UAV inspection task of the power transmission corridor and outputs the final detection results.

[0173] The trained and optimized ISD-Net model was lightweighted and then deployed to an UAV-borne edge computing platform. In actual inspection tasks, the UAV used its onboard camera equipment to capture video streams of the power transmission corridor and extracted images from them according to a set sampling rate.

[0174] The captured image, after preprocessing such as size normalization, is input into the deployed ISD-Net model for foreign object detection. The model's output detection results are then filtered by confidence thresholding and post-processed with non-maximum suppression. The generated detection result image includes the location and category labeling of the foreign object target.

[0175] The detection results are transmitted back to the ground monitoring system in real time via the drone's wireless communication module, and are also stored in the airborne storage device, forming a complete processing flow from image acquisition and analysis to result reporting and storage.

[0176] S6 Implementation Example:

[0177] The trained and optimized ISD-Net model was deployed on the NVIDIA Jetson AGX Xavier edge computing platform to build an integrated foreign object intrusion detection system for power transmission corridors. During actual inspections, a drone extracted RGB images of the power transmission corridor from a 1080p@30fps video stream at a fixed sampling rate of 5 frames per second. The RGB images of the power transmission corridor are shown below. Figure 11 As shown.

[0178] Images were preprocessed and uniformly scaled to 640×640 pixels before being input into the model. Under the Jetson platform optimization environment, the model's average inference time was approximately 60 milliseconds / frame, outputting raw detection tensors at L7, L10, and L12 scales. Subsequently, non-maximum suppression post-processing with a confidence threshold of 0.25 and an intersection-over-union (IoU) threshold of 0.5 was applied to remove redundant bounding boxes, yielding refined detection results. Finally, the detection boxes and class labels were overlaid on the original image, outputting the final detection result image containing the location and class information of foreign objects in the power transmission corridor. The final detection result image containing the location and class information of foreign objects in the power transmission corridor is shown below. Figure 12 As shown.

[0179] To effectively alert maintenance personnel, the aforementioned detection result images will be transmitted in real time back to the maintenance personnel's handheld control terminal or ground monitoring center via the 4G / 5G wireless communication module on the drone, and will also be stored locally in the drone's built-in 64GB eMMC. As a basis for data backup and post-event analysis, the original video stream and detection result images will also be stored in the drone's local embedded storage device.

[0180] The above processing flow forms a data closed loop from drone inspection of power transmission corridors to real-time analysis to result reporting and storage, providing intuitive and timely foreign object intrusion alarm information for power transmission corridor operation and maintenance, and significantly improving the intelligence level of inspection operations and operation and maintenance efficiency.

[0181] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for detecting foreign object intrusion in power transmission corridors based on ISD-Net, characterized in that, The specific steps are as follows: S1, Construct a foreign object dataset for power transmission corridors based on drone inspection videos; S2, input the RGB image of the power transmission corridor into the parallel decoupled backbone module PDB, perform multi-scale feature extraction, and obtain the global semantic feature map of foreign objects F4, the corridor saliency feature map F6, and the foreign object detail enhancement feature map F8; The parallel decoupling backbone module PDB includes an initial convolutional layer and parallel first and second branches; the RGB image of the power transmission corridor is downsampled by the initial convolutional layer to obtain an initial feature map F1; In the first branch, F1 is processed by the downsampling layer to obtain the downsampled feature map F2. F2 is split into two paths: in path one, it is processed by the multi-scale context fusion submodule MCFM and the C2PSA submodule in sequence to obtain the foreign object global semantic feature map F4. In path two, it is processed by the lightweight attention submodule and the middle layer C3k2 submodule in sequence to obtain the corridor saliency feature map F6. In the second branch, F1 is processed sequentially by multi-scale convolution and lightweight C3k2 submodule to obtain foreign object detail enhancement feature map F8; S3 inputs F4, F6 and F8 into the dynamic adaptive fusion neck module DAF-Neck to perform multi-scale feature fusion and adaptive enhancement, and outputs small-scale foreign object prediction map L7, medium-scale foreign object prediction map L10 and large-scale foreign object prediction map L12. In the dynamic adaptive fusion neck module DAF-Neck, F4 is upsampled and then input together with F6 into the first-level attention-gated inference submodule MAGI for multi-branch weighted fusion. Then, it is processed by the C3k2 submodule to obtain the neck feature map L3. After upsampling, L3 and F8 are input together into the second MAGI for multi-branch weighted fusion. Then, after processing by the C3k2 submodule and convolutional downsampling, a small-scale foreign object prediction map L7 is obtained. L7 and L3 are concatenated and processed by the channel compression C3k2 submodule to obtain a medium-scale foreign object prediction map L10. After convolutional downsampling, L10 is concatenated with F4 to obtain a large-scale foreign object prediction map L12. S4 inputs L7, L10 and L12 into the task adaptive collaborative prediction head module TASP-Head to perform task decoupling and collaborative prediction, and finally outputs a set of detection results of foreign objects in the power transmission corridor; In the task-adaptive collaborative prediction head module TASP-Head, L7, L10, and L12 are independently input into three corresponding task-adaptive fusion sub-modules TAF for processing, resulting in small-scale fusion feature map X1, medium-scale fusion feature map X3, and large-scale fusion feature map X5. X1, X3, and X5 are then processed by the C3k2 sub-module to obtain small-scale prediction pre-feature map X2, medium-scale prediction pre-feature map X4, and large-scale prediction pre-feature map X6. X2, X4, and X6 are then input into three corresponding collaborative prediction hub sub-modules SPN for processing, outputting small-scale detection result R1, medium-scale detection result R2, and large-scale detection result R3, which are finally combined to form the detection result set of foreign objects in the power transmission corridor. S5. Construct the ISD-Net model, which consists of the parallel decoupled backbone module PDB, the dynamic adaptive fusion neck module DAF-Neck, and the task adaptive collaborative prediction head module TASP-Head. Then train the ISD-Net model to obtain the optimized ISD-Net model. S6 applies the optimized ISD-Net model to the actual UAV inspection task of the power transmission corridor and outputs the final detection results.

2. The method for detecting foreign object intrusion in power transmission corridors based on ISD-Net according to claim 1, characterized in that, The specific steps of S1 are as follows: S11 uses high-definition camera equipment mounted on a drone to conduct inspection flights over the power transmission corridor and record video of power transmission corridor inspections involving foreign object incidents. S12, extract key frames from the power transmission corridor inspection video to obtain the RGB image of the power transmission corridor; S13, Use the annotation tool to annotate the RGB image of the power transmission corridor, select the foreign objects in the image in the form of bounding boxes, and assign a corresponding category label to each bounding box; S14, normalize the size of the RGB image of the power transmission corridor to obtain a size-normalized foreign object dataset of the power transmission corridor; S15, perform data augmentation processing on the RGB images of the power transmission corridor by using random rotation, horizontal / vertical flipping, color jitter, adding random noise, and simulating complex environmental effects such as rain, fog, and occlusion, in order to expand the data scale and improve the model's generalization ability, and obtain the RGB image set of the power transmission corridor. S16. The RGB image set of the power transmission corridor is divided into training set, validation set and test set according to the proportion, which are used for subsequent model training and evaluation.

3. The method for detecting foreign object intrusion in a power transmission corridor based on ISD-Net according to claim 2, characterized in that, The specific steps of S2 are as follows: S21, the RGB image of the power transmission corridor is input into the initial convolutional layer of the Conv convolution with a stride of 2 for downsampling operation, and combined with the BN and SiLU activation functions to obtain the initial feature map F1; S22, F1 is split and processed in parallel heterogeneous manner to obtain the global semantic feature map of foreign objects F4, the corridor saliency feature map F6, and the foreign object detail enhancement feature map F8.

4. The method for detecting foreign object intrusion in a power transmission corridor based on ISD-Net according to claim 3, characterized in that, The specific steps of S3 are as follows: S31, firstly, F4 is upsampled to obtain the upsampled semantic feature map L1; then, L1 and F6 are input together into the first MAGI submodule for processing to obtain the first meta-attention-gated enhanced fusion feature map L2; next, L2 is input into the C3k2 submodule for processing to obtain the neck feature map L3; finally, L3 is upsampled to obtain the upsampled feature map L4, which is used for subsequent fusion with small-scale features; S32, firstly, L4 and F8 are input together into the second MAGI submodule for dynamic fusion processing to obtain the second-level attention-gated enhanced fusion feature map L5; then, L5 is input into the C3k2 submodule for processing to enhance the fused small-scale features to obtain the small-scale neck feature map L6; finally, L6 is subjected to a Conv convolution operation with a stride of 2 to obtain the small-scale foreign object prediction map L7; S33, firstly, L7 and the neck feature map L3 are concatenated to obtain a preliminary mesoscale fusion feature map L8; then, L8 is input into the channel compression C3k2 submodule for processing to compress the number of channels from 512 to 256 while enhancing the mesoscale fusion features, resulting in a mesoscale foreign object prediction map L10; next, L10 is subjected to a Conv convolution operation with a stride of 2 to achieve a downsampling effect, resulting in a mesoscale core fusion feature map L11; finally, L11 and F4 are concatenated to obtain a large-scale foreign object prediction map L12. The Meta-Attention Gated Inference Submodule (MAGI) includes parallel spatial priority branches, channel priority branches, parallel weighted fusion branches, computational bypass branches, and weight sequence computation branches. Let the inputs of MAGI be feature maps Z0 and Z0'. Z0 and Z0' are concatenated and then subjected to Conv(1×1) convolution for dimensionality reduction to obtain the gated feature map Z2. Z2 is then input into the parallel branches: In the spatial priority branch, Z2 is processed by spatial attention and channel attention in sequence to obtain the spatially dominant enhanced feature map Z4; In the channel-first branch, Z2 is processed by channel attention and spatial attention in sequence to obtain the channel-dominated enhanced feature map Z6; In the parallel weighted fusion branch, Z2 is processed by spatial attention and channel attention respectively and then weighted to obtain the fused feature map Z9; In computing the bypass branch, Z2 is processed by DWConv2D (3×3) depth separable convolution to obtain a lightweight convolutional feature map Z10; In the weight sequence calculation branch, Z2 inputs to the meta-attention gating network, which is then processed by global average pooling, dimensionality reduction of the fully connected layer, ReLU activation function, and fully connected layer mapping to a four-dimensional original path score logical value sequence [v1,v2,v3,v4]. After Softmax normalization, the output is a four-dimensional gated path weight vector sequence [w1,w2,w3,w4]. Z4, Z6, Z9, and Z10 are multiplied and weighted with the weight vectors w1, w2, w3, and w4 respectively to obtain the first-order attention-gated enhanced fusion feature map L2. Then, L2 is input to the C3k2 submodule for processing to obtain the neck feature map L3. Finally, L3 is upsampled to obtain the upsampled feature map L4, which is used for subsequent fusion with small-scale features.

5. The method for detecting foreign object intrusion in a power transmission corridor based on ISD-Net according to claim 4, characterized in that, The specific steps of S4 are as follows: S41, input L7 into the first TAF for processing to obtain small-scale fusion feature map X1; input L10 into the second TAF for processing to obtain medium-scale fusion feature map X3; input L12 into the third TAF for processing to obtain large-scale fusion feature map X5; S42, input the feature map X1 into the C3k2 submodule for processing to obtain the small-scale prediction pre-feature map X2; The feature map X3 is input into the C3k2 submodule for processing to obtain the pre-prepared feature map X4 for mesoscale prediction. The feature map X5 is input into the C3k2 submodule for processing to obtain the large-scale prediction pre-feature map X6. S43, input X2 into the first SPN submodule to obtain the small-scale detection result R1; input X4 into the second SPN submodule to obtain the medium-scale detection result R2; input X6 into the third SPN submodule to obtain the large-scale detection result R3; S44, R1, R2, and R3 are combined into a detection result set, which is the final detection result of foreign objects in the power transmission corridor; The Task Adaptive Fusion Submodule (TAF) includes parallel localization and classification task paths. Let the input of the TAF be the small-scale foreign object prediction map K0. K0 is input into the Bidirectional Feature Pyramid Network (BiFPN) to obtain a bidirectional feature map K1. K1 is then processed through two non-shared weight Conv(1×1) convolutions within the task separation unit to separate the localization feature map K2 and the semantic feature map K6. In the localization task path, K2 is processed sequentially by Conv2D(1×1) convolution, DWConv2D(3×3) depthwise separable convolution and Conv2D(1×1) convolution to obtain the localization specialized feature map K5; In the classification task path, K6 is processed sequentially by Conv2D (1×1) convolution, DWConv2D (5×5) depthwise separable convolution, the squeeze and excitation network SE module, and Conv2D (1×1) convolution to obtain the classification-specific feature map K10; K5 and K10 are spliced ​​and adaptively fused to obtain the small-scale fused feature map X1; The collaborative prediction hub submodule SPN contains four processing branches. Let the input of the collaborative prediction hub submodule SPN be the pre-prediction feature map B0. B0 undergoes shared convolution processing including Conv2D (3×3), batch normalization, and ReLU activation functions to obtain the shared feature map B1. Subsequently, B1 is split into four branches for processing: In the first and fourth branches, B1 is directly reserved for subsequent interactive calibration; In the second branch, B1 is processed by a regression head that includes Conv2D (1×1), global average pooling, and a fully connected layer to obtain the initial foreign object localization vector B2; In the third branch, B1 is processed by a classification head that includes Conv2D (1×1), global average pooling, fully connected layers and Softmax to obtain the initial foreign object classification vector B5; During the interactive calibration phase, the first branch retains B1 and B2, which are multiplied element-wise to obtain the calibration regression feature map B3. The fourth branch retains B1 and B5, which are multiplied element-wise to obtain the calibration classification feature map B6. B3 is input to the deep regression unit and is processed sequentially by Conv2D (3×3) combined with batch normalization and ReLU, and then by two consecutive Conv2D (1×1) processes combined with a fully connected layer to obtain the final location information B4. B6 is input to the deep classification unit and is processed sequentially by the SE module, Conv2D (3×3), global average pooling, a fully connected layer, and Softmax to obtain the final classification information B7. B4 and B7 together constitute the detection result.

6. The method for detecting foreign object intrusion in a power transmission corridor based on ISD-Net according to claim 5, characterized in that, The specific steps of S5 are as follows: S51, Construct the ISD-Net model, which includes the parallel decoupling backbone module PDB, the dynamic adaptive fusion neck module DAF-Neck, and the task adaptive collaborative prediction head module TASP-Head; S52. First, initialize all neural network parameters and set dedicated hyperparameters for the foreign object intrusion detection task in power transmission corridors. Then, input training and validation set data containing various types of foreign objects in power transmission corridors into the intelligent collaborative foreign object intrusion detection model for end-to-end training. During training, the network parameters are continuously updated through the backpropagation algorithm and a specially designed loss function for foreign object intrusion detection in power transmission corridors until the foreign object intrusion detection performance index of the model on the validation set tends to converge. At this point, training ends, and the optimized ISD-Net model is obtained.

Images