A power transmission line hidden danger detection method and system

Abstract

The application discloses a power transmission line hidden danger detection method and system, relates to the technical field of power transmission line monitoring, acquires a power transmission line image to be detected, carries out multi-scale feature extraction on the power transmission line image to be detected through a backbone network, obtains corresponding first, second and third scale feature maps, carries out cross-scale feature fusion on the first, second and third scale feature maps through a multi-mode attention feature network, obtains first, second and third scale fusion feature maps, and carries out collaborative decoupling processing on the first, second and third scale fusion feature maps through a hierarchical guiding detection network, and obtains corresponding detection results. The application overcomes the technical problem that the existing power transmission line hidden danger detection method adopts a general target detection algorithm, but multi-class hidden danger collaborative detection is prone to missed detection, and the detection accuracy of the power transmission line is reduced.

Description

Technical Field

[0001] This invention relates to the field of power transmission line monitoring technology, and in particular to a method and system for detecting hidden dangers in power transmission lines. Background Technology

[0002] Transmission lines are crucial carriers in power systems, responsible for long-distance power transmission. Exposed to complex natural environments, they are susceptible to lightning strikes, wind damage, bird strikes, external forces, and aging corrosion. This leads to frequent safety hazards such as insulator damage, hardware corrosion, conductor strand breakage, vibration damper slippage, and foreign object intrusion, directly threatening the safe and stable operation of the power grid. In severe cases, this can cause major accidents like line tripping and large-scale power outages. Traditional transmission line inspections rely mainly on manual tower climbing or ground-based telescope inspections. This method is not only labor-intensive, risky, and inefficient, but also limited by terrain, lighting, and viewing angles, resulting in serious issues of missed or false detections. It no longer meets the urgent needs of modern smart grids for efficient, precise, and intelligent operation and maintenance. With the rapid development of drone inspection technology and computer vision technology, intelligent hazard detection based on deep learning-based drone aerial images is gradually becoming the mainstream technology for transmission line operation and maintenance, providing a feasible path to improve inspection efficiency and diagnostic accuracy.

[0003] Currently, most existing methods for detecting hidden dangers in power transmission lines use general target detection algorithms. Although these algorithms can locate and identify power components and obvious defects, they are difficult to adapt to heterogeneous hidden danger features such as texture, geometry, and context. Collaborative detection of multiple types of hidden dangers is prone to missed detections, which reduces the detection accuracy of power transmission lines. Summary of the Invention

[0004] This invention provides a method and system for detecting hidden dangers in power transmission lines. It solves the technical problem that existing methods for detecting hidden dangers in power transmission lines mostly use general target detection algorithms. Although these algorithms can locate and identify power components and obvious defects, they are difficult to adapt to heterogeneous hidden danger features such as texture, geometry, and context. Furthermore, collaborative detection of multiple types of hidden dangers is prone to missed detections, which reduces the detection accuracy of power transmission lines.

[0005] The first aspect of this invention provides a method for detecting hidden dangers in power transmission lines, comprising:

[0006] Multiple training images are acquired, and image preprocessing is performed on each training image to generate a corresponding line feature set;

[0007] The preset initial transmission line hazard detection model is trained using the line feature set to generate a corresponding target transmission line hazard detection model. The target transmission line hazard detection model includes a backbone network, a multi-modal attention feature network, and a hierarchical guided detection network.

[0008] An image of the transmission line to be detected is acquired, and multi-scale feature extraction is performed on the image of the transmission line to be detected through the backbone network to obtain the corresponding first-scale feature map, second-scale feature map and third-scale feature map;

[0009] The multimodal attention feature network is used to perform cross-scale feature fusion on the first scale feature map, the second scale feature map and the third scale feature map to obtain the corresponding first scale fused feature map, second scale fused feature map and third scale fused feature map.

[0010] The hierarchical guided detection network is used to perform collaborative decoupling processing on the first-scale fused feature map, the second-scale fused feature map, and the third-scale fused feature map to obtain the corresponding detection results.

[0011] Optionally, the step of training the preset initial transmission line hazard detection model using the line feature set to generate the corresponding target transmission line hazard detection model includes:

[0012] The line feature set is used as input to train the preset initial transmission line hidden danger detection model to obtain the corresponding training detection data;

[0013] Calculate the loss function value of the line feature set based on the training and detection data;

[0014] When the loss function value is greater than or equal to the preset loss threshold, the network parameters of the initial transmission line hidden danger detection model are adjusted until the loss function value is less than the loss threshold.

[0015] When the loss function value is less than the loss threshold, a corresponding target transmission line hazard detection model is generated.

[0016] Optionally, the backbone network includes a first convolutional module, a multi-branch feature extraction module, a second convolutional module, a second convolutional module, and a multi-branch feature extraction module connected in sequence. The step of performing multi-scale feature extraction on the transmission line image to be detected through the backbone network to obtain corresponding first-scale feature maps, second-scale feature maps, and third-scale feature maps includes:

[0017] The first convolution module is used to extract features from the image of the transmission line to be detected to obtain the corresponding first convolution feature map.

[0018] The first convolutional feature map is decoupled and extracted using a multi-branch feature extraction module to obtain the corresponding first-scale feature map.

[0019] The second convolutional module is used to extract features from the first scale feature map to obtain the corresponding second convolutional feature map.

[0020] The second convolutional feature map is decoupled and extracted using a multi-branch feature extraction module to obtain the corresponding second-scale feature map.

[0021] The second convolution module extracts features from the second scale feature map to obtain the corresponding third convolution feature map;

[0022] The third convolutional feature map is decoupled and extracted using a multi-branch feature extraction module to obtain the corresponding third-scale feature map.

[0023] Optionally, the multi-branch feature extraction module includes two third convolutional modules, a texture analysis branch, a context aggregation branch, a geometry-aware branch, a global-aware gated weight generator, and a feature fusion layer. The step of performing multi-dimensional feature decoupling extraction on the first convolutional feature map through the multi-branch feature extraction module to obtain the corresponding first-scale feature map includes:

[0024] The third convolution module extracts features from the first convolution feature map to obtain the corresponding hidden danger perception feature map.

[0025] The global perception gating weight generator performs dynamic weight perception on the hidden danger perception feature map to obtain the corresponding first gating weight, second gating weight and third gating weight.

[0026] The texture features of the hidden danger perception feature map are extracted by the texture analysis branch to obtain the corresponding texture feature map, and the texture feature map is weighted and modulated by the first gating weight to obtain the corresponding target texture feature map.

[0027] The context aggregation branch is used to perform global context perception on the hidden danger perception feature map to obtain the corresponding context aggregation feature map, and the second gating weight is used to perform weighted modulation on the context aggregation feature map to obtain the corresponding target aggregation feature map.

[0028] Geometric features are extracted from the hazard perception feature map through the geometric perception branch to obtain the corresponding geometric feature map, and the geometric feature map is weighted and modulated using the third gating weight to obtain the corresponding target geometric feature map;

[0029] The target texture feature map, the target aggregated feature map, and the target geometric feature map are spliced ​​together by the feature fusion layer to obtain the corresponding multi-form power transmission hazard perception feature map;

[0030] The third convolution module extracts features from the multi-form power transmission hazard perception feature map to obtain the corresponding first-scale feature map.

[0031] Optionally, the global perception gating weight generator includes a global pooling module, two mapping modules, and a weight segmentation module. The step of dynamically perceiving weights on the hazard perception feature map through the global perception gating weight generator to obtain the corresponding first gating weight, second gating weight, and third gating weight includes:

[0032] The feature map of the hidden danger perception is globalized by the global pooling module to obtain the corresponding global feature vector;

[0033] The global feature vector is continuously nonlinearly mapped by two sequentially connected mapping modules to obtain the corresponding intermediate feature vector. The mapping module includes a fully connected layer and a ReLU activation function layer connected in sequence.

[0034] The intermediate feature vector is segmented by a weight segmentation module to obtain the corresponding first gate weight, second gate weight and third gate weight. The weight segmentation module includes a sigmoid activation function layer and a feature splitting layer connected in sequence.

[0035] Optionally, the texture analysis branch includes a third convolutional module, a global average pooling module, a global max pooling module, two feature fusion layers, and a first channel attention module. The step of extracting texture features from the hazard perception feature map through the texture analysis branch to obtain the corresponding texture feature map includes:

[0036] The third convolution module is used to extract features from the hazard perception feature map to obtain the corresponding first-morphology hazard perception feature map.

[0037] The first morphological hazard perception feature map is globalized by the global average pooling module to obtain the corresponding second morphological hazard perception feature map.

[0038] The first-form hazard perception feature map is sharpened by the global max pooling module to obtain the corresponding third-form hazard perception feature map.

[0039] The second form of hazard perception feature map and the third form of hazard perception feature map are spliced ​​together by the feature fusion layer to obtain the corresponding fourth form of hazard perception feature map.

[0040] The fourth form of hidden danger perception feature map is processed by the first channel attention module to obtain the corresponding channel attention weight. The channel attention weight is then multiplied by the first form of hidden danger perception feature map to obtain the corresponding fifth form of hidden danger perception feature map.

[0041] The feature fusion layer fuses the fifth-form hazard perception feature map and the first-form hazard perception feature map to obtain the corresponding texture feature map.

[0042] Optionally, the context aggregation branch includes two third convolutional modules, a multi-head self-attention module, two feature fusion layers, and a spatial attention module. The step of performing global context awareness on the hazard perception feature map through the context aggregation branch to obtain the corresponding context aggregation feature map includes:

[0043] The third convolutional module is used to extract features from the hazard perception feature map to obtain the corresponding sixth-form hazard perception feature map.

[0044] By performing contextual aggregation on the sixth-form hazard perception feature map using a multi-head self-attention module, the corresponding seventh-form hazard perception feature map is obtained.

[0045] The feature fusion layer is used to fuse the sixth form of hidden danger perception feature map and the seventh form of hidden danger perception feature map to obtain the corresponding first residual hidden danger perception feature map.

[0046] The third convolution module is used to extract features from the first residual hidden danger perception feature map to obtain the corresponding second residual hidden danger perception feature map.

[0047] The spatial attention module is used to perform spatial focusing processing on the second residual hidden danger perception feature map to obtain the corresponding spatial focusing feature map;

[0048] The first residual hazard perception feature map and the spatial focusing feature map are fused by the feature fusion layer to obtain the corresponding context aggregation feature map.

[0049] Optionally, the geometric perception branch includes a 1×1 convolutional layer, a geometric contour extraction module, and a feature fusion layer. The step of extracting geometric features from the hazard perception feature map through the geometric perception branch to obtain the corresponding geometric feature map includes:

[0050] The hazard perception feature map is extracted by a 1×1 convolutional layer to obtain the corresponding eighth-mode hazard perception feature map.

[0051] The geometric contour extraction module extracts geometric contour features from the eighth-form hazard perception feature map to obtain the corresponding ninth-form hazard perception feature map. The geometric contour extraction module includes a 3×3 deformable convolutional layer, a 1×1 convolutional layer and a ReLU activation function layer connected in sequence.

[0052] The feature fusion layer performs feature fusion on the eighth-form hazard perception feature map and the ninth-form hazard perception feature map to obtain the corresponding geometric feature map.

[0053] Optionally, the multimodal attention feature network includes a spatial pyramid fast pooling module, two sampling modules, two second convolutional modules, four feature fusion layers, and three deep multimodal attention feature refinement modules. The step of performing cross-scale feature fusion on the first-scale feature map, the second-scale feature map, and the third-scale feature map through the multimodal attention feature network to obtain the corresponding first-scale fused feature map, second-scale fused feature map, and third-scale fused feature map includes:

[0054] The third-scale feature map is extracted using the spatial pyramid fast pooling module to obtain the corresponding first intermediate-scale feature map.

[0055] The first intermediate scale feature map is upsampled by the sampling module to obtain the corresponding second intermediate scale feature map. The sampling module includes a third convolution module and a sampling operation module connected in sequence.

[0056] The second intermediate scale feature map and the second scale feature map are concatenated by the feature fusion layer to obtain the corresponding third intermediate scale feature map.

[0057] The third intermediate scale feature map is upsampled by the sampling module to obtain the corresponding fourth intermediate scale feature map, and the fourth intermediate scale feature map and the first scale feature map are spliced ​​by the feature fusion layer to obtain the corresponding fifth intermediate scale feature map.

[0058] The deep multimodal attention feature refinement module is used to refine the fifth intermediate scale feature map to obtain the corresponding first scale fusion feature map.

[0059] The second convolutional module extracts features from the first scale fusion feature map to obtain the corresponding sixth intermediate scale feature map. The feature fusion layer then concatenates the sixth intermediate scale feature map and the third intermediate scale feature map to obtain the corresponding seventh intermediate scale feature map.

[0060] The deep multimodal attention feature refinement module is used to refine the seventh intermediate scale feature map to obtain the corresponding second scale fusion feature map.

[0061] The second convolutional module extracts features from the second scale fusion feature map to obtain the corresponding eighth intermediate scale feature map. The feature fusion layer then concatenates the eighth intermediate scale feature map and the first intermediate scale feature map to obtain the corresponding ninth intermediate scale feature map.

[0062] The deep multimodal attention feature refinement module is used to refine the ninth intermediate scale feature map to obtain the corresponding third scale fusion feature map.

[0063] Optionally, the deep multimodal attention feature refinement module includes two first deep attention branches, two second deep attention branches, two 1×1 convolutional layers, two feature fusion layers, and two third convolutional modules. The step of refining the fifth intermediate-scale feature map using the deep multimodal attention feature refinement module to obtain the corresponding first-scale fused feature map includes:

[0064] The fifth intermediate scale feature map is integrated by a 1×1 convolutional layer to obtain the corresponding first depth refined feature map.

[0065] The first deep refined feature map is refined using a dual-channel texture through the first deep attention branch to obtain the corresponding second deep refined feature map.

[0066] The first deep refined feature map is subjected to two-layer geometric refinement through the second deep attention branch to obtain the corresponding third deep refined feature map;

[0067] The first deep refined feature map, the second deep refined feature map, and the third deep refined feature map are fused by the feature fusion layer to obtain the corresponding fourth deep refined feature map.

[0068] The fourth deep refined feature map is integrated by sequentially connecting the third convolutional module and the 1×1 convolutional layer to obtain the corresponding fifth deep refined feature map.

[0069] The fifth deep refined feature map is refined using a dual-channel texture through the first deep attention branch to obtain the corresponding sixth deep refined feature map;

[0070] The fifth deep refined feature map is subjected to two-layer geometric refinement through the second deep attention branch to obtain the corresponding seventh deep refined feature map;

[0071] The fifth, sixth, and seventh deep refined feature maps are fused using a feature fusion layer to obtain the corresponding eighth deep refined feature map.

[0072] The third convolution module integrates the features of the eighth deep refined feature map to obtain the corresponding first-scale fused feature map.

[0073] Optionally, the first deep attention branch includes two ghost convolution modules, a first-channel attention module, and a second-channel attention module. The step of performing dual-channel texture refinement on the first deep refined feature map through the first deep attention branch to obtain the corresponding second deep refined feature map includes:

[0074] The first deep refined feature map is transformed by the ghost convolution module to obtain the corresponding first intermediate refined feature map.

[0075] The first intermediate refined feature map is processed by the second channel attention module to obtain the corresponding second intermediate refined feature map.

[0076] The first intermediate refined feature map and the second intermediate refined feature map are multiplied by elements to obtain the corresponding third intermediate refined feature map.

[0077] The third intermediate refined feature map is subjected to depth feature transformation by the ghost convolution module to obtain the corresponding fourth intermediate refined feature map;

[0078] The fourth intermediate refined feature map is processed by the first channel attention module to obtain the corresponding fifth intermediate refined feature map.

[0079] The fourth intermediate refined feature map and the fifth intermediate refined feature map are multiplied by elements to obtain the corresponding second deep refined feature map.

[0080] Optionally, the second deep attention branch includes two deformable convolutional modules, a spatial attention module, and a first channel attention module. The step of performing two-layer geometric refinement on the first deep refined feature map through the second deep attention branch to obtain the corresponding third deep refined feature map includes:

[0081] Geometric features are extracted from the first deep refined feature map using a deformable convolution module to obtain the corresponding sixth intermediate refined feature map.

[0082] The sixth intermediate refined feature map is spatially focused by the spatial attention module to obtain the corresponding seventh intermediate refined feature map.

[0083] The sixth intermediate refined feature map and the seventh intermediate refined feature map are multiplied by element to obtain the corresponding eighth intermediate refined feature map.

[0084] Geometric features are extracted from the eighth intermediate refined feature map using a deformable convolution module to obtain the corresponding ninth intermediate refined feature map.

[0085] The ninth intermediate refined feature map is processed by the first channel attention module to obtain the corresponding tenth intermediate refined feature map.

[0086] The ninth intermediate refined feature map and the tenth intermediate refined feature map are multiplied by elements to obtain the corresponding third deep refined feature map.

[0087] Optionally, the hierarchical guided detection network includes three asymmetric interactive detection heads and a post-processing module. The step of performing collaborative decoupling processing on the first-scale fused feature map, the second-scale fused feature map, and the third-scale fused feature map through the hierarchical guided detection network to obtain the corresponding detection results includes:

[0088] The first-scale fused feature map is subjected to hierarchical collaborative prediction by an asymmetric interaction detection head to obtain the corresponding first prediction result. The asymmetric interaction detection head includes three third convolutional modules, a first prediction branch, a second prediction branch, a third prediction branch, and a fourth prediction branch.

[0089] The second-scale fused feature map is subjected to hierarchical collaborative prediction using an asymmetric interactive detection head to obtain the corresponding second prediction result.

[0090] The third-scale fused feature map is subjected to hierarchical collaborative prediction by an asymmetric interactive detection head to obtain the corresponding third prediction result.

[0091] The post-processing module filters the first prediction result, the second prediction result, and the third prediction result to obtain the corresponding detection result.

[0092] Optionally, the step of performing hierarchical collaborative prediction on the first scale fused feature map using an asymmetric interactive detection head to obtain the corresponding first prediction result includes:

[0093] The first scale fusion feature map is extracted by two sequentially connected third convolutional modules to obtain the corresponding first predicted feature map.

[0094] The first prediction branch performs hidden danger detection on the first prediction feature map to obtain the corresponding first prediction tensor. The first prediction branch includes a 1×1 convolutional layer, a third convolutional module, a feature co-refinement module, and a 1×1 convolutional layer connected in sequence.

[0095] The second prediction branch extracts features from the first prediction feature map to obtain the corresponding second prediction feature map. The second prediction branch includes a 1×1 convolutional layer and a Sigmoid activation function layer connected in sequence.

[0096] The second predicted feature map is multiplied element-wise with the first predicted feature map to obtain the corresponding third predicted feature map.

[0097] The third prediction feature map is extracted by the third convolution module to obtain the corresponding fourth prediction feature map;

[0098] The third prediction branch is used to identify potential hazards in the fourth prediction feature map to obtain the corresponding second prediction tensor. The third prediction branch includes a feature co-refinement module and a 1×1 convolutional layer connected in sequence.

[0099] The fourth prediction branch is used to assess the hazard level of the fourth prediction feature map to obtain the corresponding third prediction tensor, and the first prediction tensor, the second prediction tensor and the third prediction tensor are determined as the corresponding first prediction result.

[0100] A second aspect of the present invention provides a power transmission line hazard detection system, comprising:

[0101] The preprocessing module is used to acquire multiple training images, perform image preprocessing on each training image, and generate a corresponding line feature set;

[0102] The training module is used to train the preset initial transmission line hidden danger detection model using the line feature set to generate the corresponding target transmission line hidden danger detection model. The target transmission line hidden danger detection model includes a backbone network, a multi-modal attention feature network, and a hierarchical guided detection network.

[0103] The acquisition module is used to acquire images of the transmission line to be detected, and to perform multi-scale feature extraction on the images of the transmission line to be detected through the backbone network to obtain corresponding first-scale feature maps, second-scale feature maps and third-scale feature maps;

[0104] The cross-scale feature fusion module is used to perform cross-scale feature fusion on the first-scale feature map, the second-scale feature map and the third-scale feature map through the multi-modal attention feature network to obtain the corresponding first-scale fused feature map, second-scale fused feature map and third-scale fused feature map;

[0105] The detection module is used to perform collaborative decoupling processing on the first scale fusion feature map, the second scale fusion feature map and the third scale fusion feature map through the hierarchical guided detection network to obtain the corresponding detection results.

[0106] As can be seen from the above technical solutions, the present invention has the following advantages:

[0107] This invention acquires multiple training images, performs image preprocessing on each training image to generate corresponding line feature sets, and uses these line feature sets to train a pre-set initial transmission line hazard detection model to generate a corresponding target transmission line hazard detection model. The target transmission line hazard detection model includes a backbone network, a multi-modal attention feature network, and a hierarchical guided detection network. It acquires images of the transmission line to be detected, extracts multi-scale features from these images using the backbone network to obtain corresponding first-scale, second-scale, and third-scale feature maps, performs cross-scale feature fusion on these maps using the multi-modal attention feature network to obtain corresponding first-scale fused feature maps, second-scale fused feature maps, and third-scale fused feature maps, and finally performs collaborative decoupling processing on these maps using the hierarchical guided detection network to obtain the corresponding detection results. This invention overcomes the technical challenges of existing transmission line hazard detection methods, which often employ general target detection algorithms. While these algorithms can locate and identify power components and obvious defects, they struggle to adapt to heterogeneous hazard features such as texture, geometry, and context. Furthermore, collaborative detection of multiple hazard types is prone to missed detections, reducing the accuracy of transmission line hazard detection. Compared to traditional transmission line detection methods, this invention utilizes a backbone network to efficiently extract multi-scale and multi-type hazard features from images of the transmission line to be detected. It can simultaneously adapt to heterogeneous features such as texture, geometry, and context, significantly enhancing the feature representation capability for different hazard types. A multi-modal attention feature network then completes cross-scale feature fusion and deep refinement, effectively suppressing background interference and enhancing the features of minor hazards. This solves the problems of feature ambiguity and information loss that traditional fusion methods often cause. Finally, a hierarchical guided detection network achieves collaborative decoupling of localization, classification, and hazard assessment, eliminating feature conflicts between multiple tasks. It can simultaneously output hazard location, category, and hazard level in a single inference, making the detection results more aligned with the actual needs of power operation and maintenance, thus improving the accuracy and efficiency of transmission line hazard detection. Attached Figure Description

[0108] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0109] Figure 1 This is a flowchart of the steps of a method for detecting hidden dangers in power transmission lines provided in Embodiment 1 of the present invention;

[0110] Figure 2This is a flowchart illustrating the steps of a method for detecting hidden dangers in power transmission lines according to Embodiment 2 of the present invention.

[0111] Figure 3 This is a schematic diagram of the target transmission line hidden danger detection model provided in Embodiment 2 of the present invention;

[0112] Figure 4 This is a schematic diagram of the structure of the multi-branch feature extraction module provided in Embodiment 2 of the present invention;

[0113] Figure 5 This is a schematic diagram of the structure of the global perception gating weight generator provided in Embodiment 2 of the present invention;

[0114] Figure 6 This is a schematic diagram of the deep multimodal attention feature refining module provided in Embodiment 2 of the present invention;

[0115] Figure 7 This is a schematic diagram of the asymmetric interactive detection head provided in Embodiment 2 of the present invention;

[0116] Figure 8 This is a schematic diagram of the feature collaborative refining module provided in Embodiment 2 of the present invention;

[0117] Figure 9 This is a structural block diagram of a power transmission line hidden danger detection system provided in Embodiment 3 of the present invention. Detailed Implementation

[0118] This invention provides a method and system for detecting hidden dangers in power transmission lines. It addresses the technical problem that existing methods for detecting hidden dangers in power transmission lines often employ general target detection algorithms. While these algorithms can locate and identify power components and obvious defects, they are difficult to adapt to heterogeneous hidden danger features such as texture, geometry, and context. Furthermore, collaborative detection of multiple types of hidden dangers is prone to missed detections, which reduces the detection accuracy of power transmission lines.

[0119] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that in the optional embodiments of the present invention, the object information and other related data involved require the permission or consent of the object when the embodiments of the present invention are applied to specific products or technologies, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. That is to say, if the embodiments of the present invention involve data related to the object, it needs to be obtained with the authorization and consent of the object, the authorization and consent of the relevant departments, and in compliance with the relevant laws, regulations, and standards of the country and region. If personal information is involved in the embodiments, the acquisition of all personal information requires the consent of the individual. If sensitive information is involved, the separate consent of the information subject is required, and the embodiments also need to be implemented with the authorization and consent of the object.

[0120] Please see Figure 1 , Figure 1 This is a flowchart illustrating the steps of a method for detecting hidden dangers in power transmission lines, as provided in Embodiment 1 of the present invention.

[0121] This invention provides a method for detecting hidden dangers in power transmission lines, comprising:

[0122] Step 101: Obtain multiple training images, perform image preprocessing on each training image, and generate corresponding line feature sets.

[0123] Training images refer to power line inspection images used for training deep learning models. They include various hidden danger samples and normal line samples, covering different shooting conditions and hidden danger forms.

[0124] Image preprocessing refers to operations such as size normalization, data augmentation, and noise reduction performed on the original inspection images to unify the image format and improve data quality.

[0125] The line feature set refers to the set of feature data obtained after preprocessing and feature extraction, which includes information such as the structure of the transmission line, the texture of hidden dangers, the geometric contour, and the contextual relationship. It is used to train the hidden danger detection model.

[0126] In this embodiment of the invention, images of potential hazards in transmission lines under different seasons, lighting conditions, and weather conditions are collected through methods such as drone aerial photography, fixed camera monitoring, and manual inspection and photography. This generates multiple training images containing five categories of hazards: insulator damage, hardware corrosion, conductor damage, vibration damper slippage, and foreign object intrusion, labeled with three levels: general, severe, and critical. Each training image undergoes image preprocessing (i.e., uniformly scaling and cropping the training images to achieve consistent resolution, then performing data augmentation by adjusting brightness and contrast, adding Gaussian noise, performing affine transformation, and hybrid blurring to eliminate inconsistent image scales and expand sample diversity), generating a corresponding line feature set.

[0127] Step 102: Train the preset initial transmission line hazard detection model using the line feature set to generate the corresponding target transmission line hazard detection model. The target transmission line hazard detection model includes a backbone network, a multi-modal attention feature network, and a hierarchical guided detection network.

[0128] The initial transmission line hazard detection model refers to a basic deep learning model that has not been trained and has only completed network structure construction and parameter initialization, possessing a basic network framework for target detection and hazard classification.

[0129] The target transmission line hidden danger detection model refers to the final model obtained after training and convergence of the line feature set, which can directly perform hidden danger detection and level assessment on transmission line inspection images.

[0130] In this embodiment of the invention, the line feature set is divided into a training set, a validation set, and a test set according to a certain ratio and then input into a preset initial transmission line hazard detection model. Training detection data of hazard location, category, and level are obtained through forward propagation. Then, the loss function value composed of classification loss, regression loss, and classification loss is calculated by combining the annotation information of the line feature set. The weights and bias parameters of each network layer in the initial transmission line hazard detection model are iteratively adjusted according to the loss function value through backpropagation. The fitting ability and generalization ability of the initial transmission line hazard detection model to the line hazard features are continuously optimized. When the total loss converges to the preset loss threshold and the detection and classification accuracy of the model on the validation set reaches a stable state, the training is stopped and a target transmission line hazard detection model with end-to-end detection and classification capabilities is generated. The target transmission line hazard detection model consists of three parts: a backbone network responsible for multi-scale basic feature extraction, a multi-modal attention feature network responsible for cross-scale feature fusion and denoising refinement, and a classification-guided detection network responsible for hazard location, classification, and level assessment.

[0131] Step 103: Obtain the image of the transmission line to be detected, and perform multi-scale feature extraction on the image of the transmission line to be detected through the backbone network to obtain the corresponding first-scale feature map, second-scale feature map and third-scale feature map.

[0132] The images of the transmission lines to be inspected refer to real-time images of the transmission lines during on-site inspections that have not been analyzed using models. These images include the line itself, towers, surrounding environment, and potential hazards.

[0133] The first-scale feature map refers to the high-resolution feature map output from the shallow layer of the backbone network, which is mainly used to perceive large-scale structures and large-sized potential hazards in the line.

[0134] The second-scale feature map refers to the medium-resolution feature map output by the middle layer of the backbone network, which takes into account both details and semantics, and perceives medium-sized potential targets.

[0135] The third-scale feature map refers to the low-resolution, high-semantic feature map output from the deep layers of the backbone network, which focuses on minor hidden dangers and deep semantic information.

[0136] In this embodiment of the invention, images of the transmission line to be inspected are collected in real time by drone inspection equipment, monitoring cameras or manual shooting devices, the images of the transmission line to be inspected are preprocessed, and multi-scale feature extraction is performed on the preprocessed images of the transmission line to be inspected through a backbone network to obtain the first scale feature map, the second scale feature map and the third scale feature map corresponding to large target perception, medium target perception and small target perception.

[0137] Step 104: Perform cross-scale feature fusion on the first-scale feature map, the second-scale feature map, and the third-scale feature map using a multi-modal attention feature network to obtain the corresponding first-scale fused feature map, second-scale fused feature map, and third-scale fused feature map.

[0138] The first-scale fused feature map refers to a high-resolution feature map that has undergone cross-scale fusion and attention refinement, retaining clear details and used to detect large and clearly visible hazards.

[0139] The second-scale fused feature map refers to a medium-resolution feature map that has been fused and refined, balancing details and semantics, and is used to detect medium-sized vulnerabilities.

[0140] The third-scale fusion feature map refers to a low-resolution, high-semantic feature map that has been fused and refined, focusing on deep information and used to detect small, vague potential problems.

[0141] In this embodiment of the invention, a multi-modal attention feature network is used to perform cross-scale feature fusion on the first-scale feature map, the second-scale feature map, and the third-scale feature map to obtain a first-scale fused feature map that retains high-resolution details, a second-scale fused feature map that takes into account both semantic and local information, and a third-scale fused feature map that focuses on deep semantic information.

[0142] Step 105: Perform collaborative decoupling processing on the first-scale fused feature map, the second-scale fused feature map, and the third-scale fused feature map through a hierarchical guided detection network to obtain the corresponding detection results.

[0143] The detection result refers to the final output information of the model, which includes the coordinates of the bounding box of the hazard target, the hazard category, the hazard severity level, and the confidence score.

[0144] In this embodiment of the invention, the first-scale fused feature map, the second-scale fused feature map, and the third-scale fused feature map are input into the hierarchical guided detection network. The asymmetric interactive detection head in the hierarchical guided detection network performs parallel decoupled calculations on the first-scale fused feature map, the second-scale fused feature map, and the third-scale fused feature map for hazard location, category identification, and level assessment, respectively. Then, the prediction results of the three scales are summarized and confidence is filtered. After removing redundant boxes through non-maximum suppression, the detection results containing hazard location, hazard category, and hazard level are obtained.

[0145] In this embodiment of the invention, multiple training images are acquired, and each training image is preprocessed to generate a corresponding line feature set. The line feature set is used to train a preset initial transmission line hazard detection model to generate a corresponding target transmission line hazard detection model. The target transmission line hazard detection model includes a backbone network, a multi-modal attention feature network, and a hierarchical guided detection network. An image of the transmission line to be detected is acquired. The backbone network extracts multi-scale features from the image of the transmission line to be detected to obtain corresponding first-scale feature maps, second-scale feature maps, and third-scale feature maps. The multi-modal attention feature network performs cross-scale feature fusion on the first-scale feature maps, second-scale feature maps, and third-scale feature maps to obtain corresponding first-scale fused feature maps, second-scale fused feature maps, and third-scale fused feature maps. The hierarchical guided detection network performs collaborative decoupling processing on the first-scale fused feature maps, second-scale fused feature maps, and third-scale fused feature maps to obtain the corresponding detection result. This invention overcomes the technical challenges of existing transmission line hazard detection methods, which often employ general target detection algorithms. While these algorithms can locate and identify power components and obvious defects, they struggle to adapt to heterogeneous hazard features such as texture, geometry, and context. Furthermore, collaborative detection of multiple hazard types is prone to missed detections, reducing the accuracy of transmission line hazard detection. Compared to traditional transmission line detection methods, this invention utilizes a backbone network to efficiently extract multi-scale and multi-type hazard features from images of the transmission line to be detected. It can simultaneously adapt to heterogeneous features such as texture, geometry, and context, significantly enhancing the feature representation capability for different hazard types. A multi-modal attention feature network then completes cross-scale feature fusion and deep refinement, effectively suppressing background interference and enhancing the features of minor hazards. This solves the problems of feature ambiguity and information loss that traditional fusion methods often cause. Finally, a hierarchical guided detection network achieves collaborative decoupling of localization, classification, and hazard assessment, eliminating feature conflicts between multiple tasks. It can simultaneously output hazard location, category, and hazard level in a single inference, making the detection results more aligned with the actual needs of power operation and maintenance, thus improving the accuracy and efficiency of transmission line hazard detection.

[0146] Please see Figure 2 , Figure 2 This is a flowchart of the steps of a method for detecting hidden dangers in power transmission lines provided in Embodiment 2 of the present invention.

[0147] This invention provides a method for detecting hidden dangers in power transmission lines, comprising:

[0148] Step 201: Obtain multiple training images, perform image preprocessing on each training image, and generate corresponding line feature sets.

[0149] In this embodiment of the invention, images of potential hazards in transmission lines under different seasons, lighting and weather conditions are collected by means of drone aerial photography, fixed camera monitoring and manual inspection and shooting. Multiple training images are formed, which include five types of hazards: insulator damage, hardware corrosion, conductor damage, vibration damper slippage and foreign object intrusion, and are labeled with three levels: general, serious and critical. Image preprocessing is performed on each training image to generate a corresponding line feature set.

[0150] Step 202: Train the preset initial transmission line hidden danger detection model using the line feature set input to obtain the corresponding training detection data.

[0151] Training and detection data refers to the raw prediction data output by the initial transmission line hazard detection model after forward inference on the line feature set during the training phase, including preliminary results of location, classification, and grading.

[0152] In this embodiment of the invention, the line feature set is input into the preset initial transmission line hidden danger detection model, and the feature extraction, fusion and prediction calculation are completed sequentially through the backbone network, multi-modal attention feature network and hierarchical guided detection network according to the preset forward propagation logic, and the corresponding training detection data is output.

[0153] Step 203: Calculate the loss function value of the line feature set based on the training and detection data.

[0154] In this embodiment of the invention, the bounding box regression loss, hazard category classification loss, and hazard level assessment loss between the training detection data and the line feature set are calculated respectively. Based on preset loss weights, the bounding box regression loss, hazard category classification loss, and hazard level assessment loss are weighted and calculated to obtain the corresponding loss function values.

[0155] It is worth mentioning that the bounding box regression loss was calculated using the intersection-union ratio (IUU) loss. The hazard category classification loss was calculated using the cross-entropy loss. The hazard level assessment loss was calculated using the weighted cross-entropy loss.

[0156] Step 204: When the loss function value is greater than or equal to the preset loss threshold, adjust the network parameters of the initial transmission line hidden danger detection model until the loss function value is less than the loss threshold.

[0157] The loss threshold refers to the pre-set critical value for loss convergence, which is the basis for judging whether the model has been trained and whether the accuracy has met the standard.

[0158] Network parameters refer to the weights and biases of each convolutional layer, attention module, and fully connected layer in the model.

[0159] In this embodiment of the invention, when the loss function value is greater than or equal to the preset loss threshold, it indicates that the prediction accuracy of the current model has not yet reached the expected standard. Then, the network parameters of the initial transmission line hidden danger detection model are adjusted according to the gradient descent direction through the backpropagation algorithm until the loss function value is less than the loss threshold.

[0160] Step 205: When the loss function value is less than the loss threshold, a corresponding target transmission line hazard detection model is generated. The target transmission line hazard detection model includes a backbone network, a multi-modal attention feature network, and a hierarchical guided detection network.

[0161] In this embodiment of the invention, when the loss function value is less than the loss threshold, it indicates that the model has fully learned the texture, geometry, context features and level discrimination rules of the hidden dangers of the transmission line, and the prediction accuracy and generalization ability have reached a stable and qualified state. At this time, training and parameter updates are stopped, and the network structure and parameters that have been optimized are solidified and saved to generate a target transmission line hidden danger detection model with end-to-end hidden danger detection and classification capabilities.

[0162] Step 206: Obtain the image of the transmission line to be detected, and perform multi-scale feature extraction on the image of the transmission line to be detected through the backbone network to obtain the corresponding first-scale feature map, second-scale feature map and third-scale feature map.

[0163] Further, see Figure 3 As shown, the backbone network includes a first convolutional module, a multi-branch feature extraction module, a second convolutional module, a multi-branch feature extraction module, a second convolutional module, and a multi-branch feature extraction module connected in sequence. Step 206 includes the following sub-steps:

[0164] S11. The first convolution module is used to extract features from the image of the transmission line to be detected, and the corresponding first convolution feature map is obtained.

[0165] It should be noted that the structure of the first convolutional module is as follows: a 6×6 two-dimensional convolutional layer (i.e., Conv2d layer), a batch normalization layer (i.e., BN layer) and a SiLU activation function layer are connected in series. The stride of the convolutional layer is 4 and the padding is 1.

[0166] In this embodiment of the invention, the first convolutional module, which consists of a 6×6 convolutional layer, a batch normalization layer and a SiLU activation function layer, performs preliminary convolutional operations and basic feature extraction on the transmission line image to be detected (i.e., F1), retaining the low-level visual information such as line edges, contours and textures in the image, thereby obtaining the corresponding first convolutional feature map (i.e., F2).

[0167] S12. The first convolutional feature map is decoupled and extracted using a multi-branch feature extraction module to obtain the corresponding first-scale feature map.

[0168] Further, see Figure 4 As shown, the multi-branch feature extraction module (i.e., the MBFE-DWF module) includes two third convolutional modules, a texture analysis branch, a context aggregation branch, a geometry-aware branch, a global awareness gated weight generator, and a feature fusion layer. S12 includes the following sub-steps:

[0169] S121. The first convolution feature map is extracted by the third convolution module to obtain the corresponding hidden danger perception feature map.

[0170] It should be noted that the third convolutional module includes a 3×3 convolutional layer (i.e., a Conv2d layer), a BN batch normalization layer, and a SiLU activation function layer connected in sequence.

[0171] In this embodiment of the invention, the first convolutional feature map is channel-normalized and local feature-enhanced by a third convolutional module consisting of a 3×3 convolutional layer (i.e., Conv2d layer), a BN batch normalization layer and a SiLU activation function layer, in turn. This further focuses on the texture, contour and structural information related to the hidden dangers of the transmission line, suppresses invalid background interference, and thus obtains a hidden danger perception feature map (i.e., X1) that is adapted to subsequent multi-branch decoupling extraction.

[0172] S122. Dynamically perceive the hazard perception feature map by using a global perception gating weight generator to obtain the corresponding first gating weight, second gating weight and third gating weight.

[0173] Further, see Figure 5 As shown, the Global Aware Gated Weight Generator (GPGWG) includes a global pooling module, two mapping modules, and a weight splitting module. S122 includes the following sub-steps:

[0174] S1221. The feature map of hidden danger perception is globalized by the global pooling module to obtain the corresponding global feature vector.

[0175] It should be noted that the Global Pooling Module (GAP) refers to the network unit used to compress global information in feature maps, aggregating spatial dimensional information into channel-level global descriptions.

[0176] In this embodiment of the invention, the global pooling module performs global information aggregation and feature globalization processing on the hidden danger perception feature map, compresses the spatial dimension and extracts the overall statistical information of the channel dimension, thereby obtaining a global feature vector (i.e., u) that can characterize the global semantics of the entire feature map.

[0177] S1222. The global feature vector is continuously nonlinearly mapped by two sequentially connected mapping modules to obtain the corresponding intermediate feature vector. The mapping module includes a fully connected layer and a ReLU activation function layer connected in sequence.

[0178] In this embodiment of the invention, the global feature vector is continuously transformed and reconstructed by two sequentially connected mapping modules (FC+ReLU), thereby enhancing the expressive power and discriminative power of the global features and obtaining an intermediate feature vector (i.e., h2) with stronger semantic representation capabilities.

[0179] S1223. The intermediate feature vector is segmented by a weight segmentation module to obtain the corresponding first gate weight, second gate weight and third gate weight. The weight segmentation module includes a Sigmoid activation function layer and a feature splitting layer connected in sequence.

[0180] The first gating weight refers to the dynamic coefficients used to weight and modulate the features of the texture analysis branches.

[0181] The second gating weight refers to the dynamic coefficients used to weight and modulate the features of the context aggregation branch.

[0182] The third gating weight refers to the dynamic coefficients used to weight and modulate the geometric perception branch features.

[0183] It should be noted that the feature splitting layer refers to a network layer that uniformly divides a one-dimensional feature vector into multiple sub-vectors along the channel dimension.

[0184] In this embodiment of the invention, the intermediate feature vector is input into the weight segmentation module. First, the feature values ​​are normalized to between 0 and 1 by the Sigmoid activation function. Then, the feature is evenly split according to the channel dimension, and the single feature vector is split into three sets of independent weight parameters, thereby obtaining the first gate weight (i.e., Gate_T), the second gate weight (i.e., Gate_C), and the third gate weight (i.e., Gate_G) used for modulating the texture branch, the context branch, and the geometric branch, respectively.

[0185] S123. Extract texture features from the hidden danger perception feature map through the texture analysis branch to obtain the corresponding texture feature map, and use the first gating weight to perform weighted modulation on the texture feature map to obtain the corresponding target texture feature map.

[0186] Further, see Figure 4 As shown, the texture analysis branch includes a third convolutional module, a global average pooling module, a global max pooling module, two feature fusion layers, and a first channel attention module. S123 includes the following sub-steps:

[0187] S1231. The hazard perception feature map is extracted by the third convolution module to obtain the corresponding first-form hazard perception feature map.

[0188] In this embodiment of the invention, the third convolution module refines the local features and regularizes the channel dimensions of the hazard perception feature map, further enhancing the texture details and structural differences of the hazard area and filtering out irrelevant background information, thereby obtaining a first-form hazard perception feature map (i.e., X2) suitable for texture feature analysis.

[0189] S1232. The first-form hazard perception feature map is globalized by the global average pooling module to obtain the corresponding second-form hazard perception feature map.

[0190] It should be noted that the Global Average Pooling (GAP) module refers to a network unit used to aggregate the spatial dimensions of feature maps, and obtains global features by calculating the spatial average value of each channel of the feature map.

[0191] In this embodiment of the invention, the first morphological hazard perception feature map is subjected to global average aggregation processing by the global average pooling module, which compresses spatial redundant information, retains global statistical features related to hazard texture, and eliminates the influence of local position differences on feature expression, thereby obtaining the second morphological hazard perception feature map (i.e., X3) that focuses on global texture semantics.

[0192] S1233. The first-mode hazard perception feature map is sharpened by the global max pooling module to obtain the corresponding third-mode hazard perception feature map.

[0193] In this embodiment of the invention, the first-form hazard perception feature map is subjected to global max pooling operation (i.e., spatial dimension compression and feature sharpening of the feature map, focusing on retaining key textures, edges and other core features of the hazard area, suppressing background redundant information, and highlighting minor hazard details and feature differences) through the global max pooling module (i.e. GMP) to obtain the corresponding third-form hazard perception feature map (i.e. X4).

[0194] S1234. The second-form hazard perception feature map and the third-form hazard perception feature map are spliced ​​together by the feature fusion layer to obtain the corresponding fourth-form hazard perception feature map.

[0195] In this embodiment of the invention, the second-form hazard perception feature map and the third-form hazard perception feature map are spliced ​​together by a feature fusion layer (i.e., Conact) to obtain the corresponding fourth-form hazard perception feature map (i.e., X5).

[0196] S1235. The fourth form of hidden danger perception feature map is processed by the first channel attention module to obtain the corresponding channel attention weight. The channel attention weight is then multiplied by the first form of hidden danger perception feature map to obtain the corresponding fifth form of hidden danger perception feature map.

[0197] It should be noted that the first channel attention module is an efficient channel attention (ECA) module, which is used to perform channel adaptive weighting on the feature map.

[0198] In this embodiment of the invention, the fourth-form hazard perception feature map is processed by the first channel attention module through channel adaptive adjustment (i.e., channel feature extraction and weight allocation of the feature map) to obtain the corresponding channel attention weight (i.e., X6). The channel attention weight is then multiplied element-wise with the first-form hazard perception feature map to achieve precise feature enhancement and redundancy filtering, thereby obtaining the fifth-form hazard perception feature map (i.e., X7) that focuses on the core features of the hazard.

[0199] S1236. The feature fusion layer is used to fuse the feature map of the fifth form of hidden danger perception and the feature map of the first form of hidden danger perception to obtain the corresponding texture feature map.

[0200] In this embodiment of the invention, the fifth form of hidden danger perception feature map and the first form of hidden danger perception feature map are added together by the feature fusion layer to obtain a texture feature map (i.e., X8) rich in refined texture information.

[0201] S124. Perform global context perception on the hidden danger perception feature map through the context aggregation branch to obtain the corresponding context aggregation feature map, and use the second gating weight to perform weighted modulation on the context aggregation feature map to obtain the corresponding target aggregation feature map.

[0202] Further, see Figure 4 As shown, the context aggregation branch includes two third convolutional modules, a multi-head self-attention module, two feature fusion layers, and a spatial attention module. S124 includes the following sub-steps:

[0203] S1241. The third convolution module is used to extract features from the hazard perception feature map to obtain the corresponding sixth-form hazard perception feature map.

[0204] In this embodiment of the invention, the third convolution module extracts contextual features and regularizes channels on the hazard perception feature map, further aggregating the associated features of the hazard area and its surroundings, weakening isolated background interference, and strengthening the contextual association information between the hazard and the environment, thereby obtaining the sixth form of hazard perception feature map (i.e., X9) that is adapted to subsequent contextual feature fusion.

[0205] S1242. The sixth-form hazard perception feature map is aggregated in context using a multi-head self-attention module to obtain the corresponding seventh-form hazard perception feature map.

[0206] Multi-head self-attention modules refer to a feature processing unit based on an attention mechanism. By using multiple attention heads to compute in parallel, they capture the correlation information between different positions and channels in the feature map, realize the global aggregation and enhancement of features, and improve the effectiveness of feature representation.

[0207] In this embodiment of the invention, a multi-head self-attention module (MHSA) is used to perform global context association and information interaction on features at different locations in the sixth-form hazard perception feature map, thereby mining the intrinsic relationship between the hazard area and the surrounding environment, realizing adaptive aggregation and enhancement of features, effectively filtering redundant information, strengthening key hazard features, and thus obtaining the corresponding seventh-form hazard perception feature map (i.e., X10).

[0208] S1243. The feature fusion layer is used to fuse the feature map of the sixth form of hidden danger perception and the feature map of the seventh form of hidden danger perception to obtain the corresponding first residual hidden danger perception feature map.

[0209] In this embodiment of the invention, the sixth-form hazard perception feature map and the seventh-form hazard perception feature map are spliced ​​together by the feature fusion layer to obtain the corresponding first residual hazard perception feature map (i.e., X11).

[0210] S1244. The first residual hazard perception feature map is extracted by the third convolution module to obtain the corresponding second residual hazard perception feature map.

[0211] In this embodiment of the invention, the first residual hazard perception feature map is extracted and optimized by the third convolution module to further enhance the hazard-related features and filter redundant background information. At the same time, the feature dimensions are adapted and adjusted to improve the recognition and effectiveness of the features, and finally the corresponding second residual hazard perception feature map (i.e., X12) is obtained.

[0212] S1245. The spatial attention module is used to perform spatial focusing processing on the second residual hidden danger perception feature map to obtain the corresponding spatial focusing feature map.

[0213] The spatial attention module refers to the module used to perform spatial dimension attention weighting on feature maps. By identifying the importance of different regions in the feature map and assigning different weights, it focuses on key regions, suppresses redundant backgrounds, and improves feature targeting.

[0214] In this embodiment of the invention, the spatial attention module (SAM) is used to identify and assign weights to the feature importance of different regions in the second residual hazard perception feature map, focusing on the feature information of the region where the hazard is located and weakening the interference of irrelevant background regions to achieve spatial focusing processing of features, thereby obtaining the corresponding spatial focusing feature map (X13).

[0215] S1246. The first residual hidden danger perception feature map and the spatial focusing feature map are fused by the feature fusion layer to obtain the corresponding context aggregation feature map.

[0216] In this embodiment of the invention, the first residual hidden danger perception feature map and the spatial focusing feature map are spliced ​​together by the feature fusion layer to obtain the corresponding context aggregation feature map (i.e., X14).

[0217] S125. Geometric features are extracted from the hazard perception feature map through the geometric perception branch to obtain the corresponding geometric feature map. The geometric feature map is then weighted and modulated using the third gating weight to obtain the corresponding target geometric feature map.

[0218] Furthermore, the geometry-aware branch includes a 1×1 convolutional layer, a geometric contour extraction module, and a feature fusion layer. S125 includes the following sub-steps:

[0219] S1251. The hazard perception feature map is extracted by a 1×1 convolutional layer to obtain the corresponding eighth-mode hazard perception feature map.

[0220] In this embodiment of the invention, a 1×1 convolutional layer is used to extract features and adjust the dimensions of the hazard perception feature map, integrate, reduce and optimize the feature channels, filter redundant information and strengthen hazard-related features, thereby obtaining an eighth-form hazard perception feature map (i.e., X15) that is adapted to the subsequent detection process.

[0221] S1252. Geometric contour features are extracted from the eighth-form hazard perception feature map using the geometric contour extraction module to obtain the corresponding ninth-form hazard perception feature map. The geometric contour extraction module includes a 3×3 deformable convolutional layer, a 1×1 convolutional layer, and a ReLU activation function layer connected in sequence.

[0222] In this embodiment of the invention, the geometric contour extraction module performs targeted geometric contour feature extraction on the eighth form of hidden danger perception feature map, focusing on line-related geometric contour information, filtering irrelevant background redundancy, strengthening the geometric contour details of the line and hidden danger area, removing invalid redundant information, retaining and extracting geometric contour features related to the line and hidden danger, and thus obtaining the corresponding ninth form of hidden danger perception feature map (i.e., X18).

[0223] In another embodiment, the eighth-morphology hazard perception feature map is input into a 3×3 deformable convolutional layer to generate a multi-morphology power transmission hazard perception feature map X16. Then, the multi-morphology power transmission hazard perception feature map X16 is input into a 1×1 convolutional layer to generate a multi-morphology power transmission hazard perception feature map X17. A ReLU activation function layer is used to perform nonlinear operations on the multi-morphology power transmission hazard perception feature map X17 to obtain the corresponding ninth-morphology hazard perception feature map.

[0224] S1253. The feature fusion layer is used to fuse the feature maps of the eighth and ninth forms of hazard perception to obtain the corresponding geometric feature maps.

[0225] In this embodiment of the invention, the eighth-form hazard perception feature map and the ninth-form hazard perception feature map are stitched together by a feature fusion layer to obtain the corresponding geometric feature map (i.e., X19).

[0226] S126. The target texture feature map, target aggregate feature map and target geometric feature map are spliced ​​together by the feature fusion layer to obtain the corresponding multi-form power transmission hazard perception feature map.

[0227] In this embodiment of the invention, the target texture feature map (i.e., X20), target aggregate feature map (i.e., X21), and target geometric feature map (i.e., X22) are spliced ​​by the feature fusion layer to integrate the hidden danger texture information, global context information, and geometric contour information extracted from the three branches into one, forming a multi-form power transmission hidden danger perception feature map (i.e., X23) that has the ability to represent details, semantic association, and morphological description.

[0228] S127. The third convolution module is used to extract features from the multi-form power transmission hazard perception feature map to obtain the corresponding first-scale feature map.

[0229] In this embodiment of the invention, the third convolution module performs feature extraction and channel regularization on the multi-form power transmission hazard perception feature map, and performs noise reduction, abstraction and information integration on the fused features through convolution operation, batch normalization and nonlinear activation. Finally, the feature map X24 with unified dimension, clear semantics and suitable for subsequent cross-scale fusion is output as the first scale feature map (i.e. F3).

[0230] S13. The first scale feature map is extracted by the second convolution module to obtain the corresponding second convolution feature map.

[0231] It should be noted that the second convolutional module consists of a 3×3 two-dimensional convolutional layer, a batch normalization layer, and a SiLU activation function layer connected in sequence. This is used for downsampling, feature enhancement, and channel adjustment of the feature map.

[0232] In this embodiment of the invention, the first-scale feature map is downsampled and deep feature is extracted by the second convolution module. Feature abstraction and resolution adjustment are completed by a combination of convolution, normalization and nonlinear activation. While preserving the core information of the hidden danger, the receptive field range is increased, thereby obtaining a second convolution feature map (i.e., F4) suitable for decoupling mid-level features.

[0233] S14. The second convolutional feature map is decoupled and extracted using the multi-branch feature extraction module to obtain the corresponding second-scale feature map.

[0234] In this embodiment of the invention, the second convolutional feature map is subjected to multi-dimensional feature decoupling extraction through a multi-branch feature extraction module (that is, in accordance with the multi-dimensional decoupling logic consistent with the first scale feature extraction, parallel extraction and dynamic gating weighted fusion of texture features, context features and geometric features are carried out respectively, so as to further enhance the semantic expression of mid-level hidden dangers while maintaining the accuracy and robustness of features), and the corresponding second scale feature map (i.e. F5) is obtained.

[0235] S15. The second convolution module is used to extract features from the second scale feature map to obtain the corresponding third convolution feature map.

[0236] In this embodiment of the invention, the second convolutional module (i.e., Conv2) performs deep feature extraction and downsampling on the second-scale feature map. Through convolutional operations, batch normalization and nonlinear activation, the receptive field is further expanded and high-level semantic information is condensed. While suppressing background interference, the core features of hidden dangers are highlighted, thereby obtaining a third convolutional feature map (i.e., F6) suitable for deep multi-branch feature decoupling.

[0237] S16. The third convolutional feature map is decoupled and extracted using the multi-branch feature extraction module to obtain the corresponding third-scale feature map.

[0238] In this embodiment of the invention, the third convolutional feature map is decoupled and extracted using a multi-branch feature extraction module. The deep features are further refined to fully focus on minor hidden dangers and key semantic information, and finally a third-scale feature map (i.e., F7) with rich semantics, strong anti-interference ability, and dedicated to the perception of small targets and vague hidden dangers is obtained.

[0239] Step 207: Perform cross-scale feature fusion on the first-scale feature map, the second-scale feature map, and the third-scale feature map through a multi-modal attention feature network to obtain the corresponding first-scale fused feature map, second-scale fused feature map, and third-scale fused feature map.

[0240] Further, see Figure 3As shown, the multimodal attention feature network (i.e., Neck) includes a spatial pyramid fast pooling module, two sampling modules, two second convolutional modules, four feature fusion layers, and three deep multimodal attention feature refining modules. Step 207 includes the following sub-steps:

[0241] S21. Multi-scale feature extraction is performed on the third-scale feature map using the spatial pyramid fast pooling module to obtain the corresponding first intermediate-scale feature map.

[0242] The Spatial Pyramid Fast Pooling Module refers to a lightweight feature enhancement module that uses multi-scale pooling kernels to perform pooling operations in parallel. It is used to quickly extract local and global contextual information from feature maps and improve the multi-scale adaptability of features.

[0243] In this embodiment of the invention, multi-scale feature extraction is performed on the third-scale feature map through the Spatial Pyramid Fast Pooling Module (SPPF). By pooling operations with different receptive fields, the multi-scale contextual information in the deep feature map is fully explored. Without significantly increasing the computational load, the global representation and detail preservation capabilities of the features are enhanced, thereby obtaining a first intermediate-scale feature map (F8) with rich multi-scale semantic information.

[0244] S22. The first intermediate scale feature map is upsampled by the sampling module to obtain the corresponding second intermediate scale feature map. The sampling module includes a third convolution module and a sampling operation module connected in sequence.

[0245] In this embodiment of the invention, the first intermediate-scale feature map is input into the sampling module. First, the feature regularization and channel optimization are completed by the third convolution module to obtain the optimized first intermediate-scale feature map (i.e., F9). Then, the optimized first intermediate-scale feature map is upsampled by the sampling operation module (i.e., Upsample) to increase the resolution of the feature map to a size that matches the second-scale feature map. While enlarging the feature map, the semantic information is not lost or blurred, and finally, the resolution-adapted and feature-pure second intermediate-scale feature map (i.e., F10) is obtained.

[0246] S23. The second intermediate scale feature map and the second scale feature map are spliced ​​together by the feature fusion layer to obtain the corresponding third intermediate scale feature map.

[0247] In this embodiment of the invention, the second intermediate scale feature map and the second scale feature map are spliced ​​together by the feature fusion layer, so as to organically combine the deep enhanced semantics with the original details of the middle layer, realize the complementarity of cross-scale feature information, effectively strengthen the feature expression of the hidden dangers of the middle scale and suppress the background interference, and finally obtain the third intermediate scale feature map (i.e., F11) which has both deep semantics and middle details.

[0248] S24. The third intermediate scale feature map is upsampled by the sampling module to obtain the corresponding fourth intermediate scale feature map. The fourth intermediate scale feature map and the first scale feature map are then spliced ​​together by the feature fusion layer to obtain the corresponding fifth intermediate scale feature map.

[0249] In this embodiment of the invention, the third intermediate scale feature map is upsampled by the sampling module to obtain the corresponding fourth intermediate scale feature map F13 (that is, the feature map F12 is obtained by extracting features from the third intermediate scale feature map by the third convolution module, and the feature map F12 is upsampled by the sampling operation module to obtain the fourth intermediate scale feature map F13). The fourth intermediate scale feature map and the first scale feature map are then concatenated by the feature fusion layer to fully integrate the high-resolution detail information with the enhanced multi-scale semantic information, further strengthening the feature expression of large-scale hidden dangers and the overall line structure, and finally obtaining the fifth intermediate scale feature map (i.e., F14) which contains full-scale information, clear details and rich semantics.

[0250] S25. The deep multimodal attention feature refinement module is used to refine the feature map of the fifth intermediate scale to obtain the corresponding first scale fusion feature map.

[0251] Further, see Figure 6 As shown, the Deep Multimodal Attention Feature Refinement Module (DMMAFR) includes two first deep attention branches, two second deep attention branches, two 1×1 convolutional layers, two feature fusion layers, and two third convolutional modules. S25 includes the following sub-steps:

[0252] S251. The fifth intermediate scale feature map is integrated through a 1×1 convolutional layer to obtain the corresponding first deep refined feature map.

[0253] In this embodiment of the invention, a 1×1 convolutional layer is used to integrate and compress the channel dimension features of the fifth intermediate scale feature map. Cross-channel information interaction and redundant feature removal are completed without changing the feature map resolution, making the feature distribution more compact and the semantics more focused, thereby obtaining a first deep refined feature map (i.e., T1) that is dimensionally regular and suitable for subsequent deep attention refinement.

[0254] S252. The first deep refined feature map is refined by dual-channel texture through the first deep attention branch to obtain the corresponding second deep refined feature map.

[0255] Further, see Figure 6 As shown, the first deep attention branch includes two ghost convolution modules, a first-channel attention module, and a second-channel attention module. S252 includes the following sub-steps:

[0256] S2521. Perform depth feature transformation on the first deep refined feature map using the ghost convolution module to obtain the corresponding first intermediate refined feature map.

[0257] In this embodiment of the invention, a lightweight deep feature transformation is performed on the first deep refined feature map by using a ghost convolution module (i.e., Ghost Conv). This reduces computational redundancy while maintaining feature expressive power, and efficiently extracts details and texture change information related to potential hazards, resulting in a first intermediate refined feature map (i.e., T2) with rich detail representation.

[0258] S2522. The first intermediate refined feature map is processed by the second channel attention module to obtain the corresponding second intermediate refined feature map.

[0259] It should be noted that the second channel attention module is the Channel Attention Module (CAM).

[0260] In this embodiment of the invention, the first intermediate refined feature map is processed by the second channel attention module through channel adaptive processing (that is, by automatically learning and giving higher weights to key feature channels, weakening invalid channel information, accurately strengthening the effective feature expression related to the hidden texture, and suppressing redundancy and noise interference in the channel dimension), thereby obtaining a second intermediate refined feature map (i.e., T3) with optimized channel weights and higher feature focus.

[0261] S2523. Perform element-wise multiplication on the first intermediate refined feature map and the second intermediate refined feature map to obtain the corresponding third intermediate refined feature map.

[0262] In this embodiment of the invention, the first intermediate refined feature map and the second intermediate refined feature map are multiplied by elements, and the original feature map is weighted and modulated using attention weights to further highlight the key features of the hidden danger and suppress background and redundant information, so as to finally obtain a third intermediate refined feature map (i.e., T4) with more focused features, more prominent details and purer semantics.

[0263] S2524. The third intermediate refined feature map is subjected to deep feature transformation through the ghost convolution module to obtain the corresponding fourth intermediate refined feature map.

[0264] In this embodiment of the invention, the third intermediate refined feature map is subjected to deep feature transformation by the ghost convolution module, which further enhances the hidden details and texture representation, and continuously improves the recognizability and robustness of the features under the premise of lightweight computation, thereby obtaining the fourth intermediate refined feature map (i.e., T5) after deep secondary purification.

[0265] S2525. The fourth intermediate refined feature map is processed by the first channel attention module to obtain the corresponding fifth intermediate refined feature map.

[0266] In this embodiment of the invention, the fourth intermediate refined feature map is processed by the first channel attention module (i.e., ECA) to perform channel adaptive processing. By automatically learning the importance of channels and assigning adaptive weights, the feature channels that are highly related to potential risks are further focused on, and redundant and interfering channel information is weakened, making the feature expression more targeted and discriminative, thereby obtaining the fifth intermediate refined feature map (i.e., T6) with the channel dimension optimized again.

[0267] S2526. Perform element-wise multiplication on the fourth intermediate refined feature map and the fifth intermediate refined feature map to obtain the corresponding second deep refined feature map.

[0268] In this embodiment of the invention, the fourth intermediate refined feature map and the fifth intermediate refined feature map are multiplied by elements, and the hidden features are precisely weighted and modulated with attention weights to further enhance the effective features and suppress background interference, and finally obtain the second deep refined feature map (i.e., T7) with prominent texture details and stable feature expression.

[0269] S253. Perform double-layer geometric refinement on the first depth refined feature map through the second deep attention branch to obtain the corresponding third depth refined feature map.

[0270] Further, see Figure 6 As shown, the second deep attention branch includes two deformable convolutional modules, a spatial attention module, and a first channel attention module. S253 includes the following sub-steps:

[0271] S2531. Geometric features are extracted from the first deep refined feature map using a deformable convolution module to obtain the corresponding sixth intermediate refined feature map.

[0272] Deformable convolutional modules (or Deformable Convs) are convolutional units with adaptive offset sampling capabilities. They can flexibly match irregular target contours and are good at extracting features such as geometric deformation and positional offset.

[0273] In this embodiment of the invention, geometric features are extracted from the first deep refined feature map by a deformable convolution module. By utilizing the adaptive perception of the target contour and deformation characteristics of deformable convolution, the geometric position, contour shape and structural offset information of the hidden danger of the transmission line are accurately captured, and a sixth intermediate refined feature map (i.e., T8) with a complete geometric expression that adapts to the irregular hidden danger shape is obtained.

[0274] S2532. The sixth intermediate refined feature map is spatially focused by the spatial attention module to obtain the corresponding seventh intermediate refined feature map.

[0275] Spatial Attention Module (SAM) refers to an attention unit that automatically learns and assigns weights in the spatial dimension of the feature map to focus on key regions and suppress background interference.

[0276] In this embodiment of the invention, the sixth intermediate refined feature map is spatially focused by the spatial attention module, which automatically locates the key areas in the feature map that are related to the geometric shape of the hidden danger and strengthens the spatial position weight, suppresses the interference of irrelevant background areas, and makes the geometric features more concentrated in the spatial position of the hidden danger target, thereby obtaining the seventh intermediate refined feature map (i.e., T9) with a high degree of spatial dimension focus.

[0277] S2533. Perform element-wise multiplication on the sixth intermediate refined feature map and the seventh intermediate refined feature map to obtain the corresponding eighth intermediate refined feature map.

[0278] In this embodiment of the invention, the sixth intermediate refined feature map and the seventh intermediate refined feature map are multiplied by elements, and the spatial attention weight is used to accurately weight and strengthen the geometric features of the hidden danger, further highlighting the geometric information of the area where the hidden danger is located and suppressing background interference, and finally obtaining the eighth intermediate refined feature map (i.e., T10) with pure geometric features and accurate spatial positioning.

[0279] S2534. Geometric features are extracted from the eighth intermediate refined feature map using a deformable convolution module to obtain the corresponding ninth intermediate refined feature map.

[0280] In this embodiment of the invention, geometric features are extracted from the eighth intermediate refined feature map by a deformable convolution module, which further enhances the adaptive capture capability of hidden danger contours, deformations and positional offsets, improves the precision and recognizability of geometric features, and thus obtains a ninth intermediate refined feature map (i.e., T11) with more complete geometric expression and more robust features.

[0281] S2535. The ninth intermediate refined feature map is processed by the first channel attention module to obtain the corresponding tenth intermediate refined feature map.

[0282] In this embodiment of the invention, the ninth intermediate refined feature map is processed by the first channel attention module (i.e., ECA). By automatically learning the importance of the geometric features of each channel and assigning adaptive weights, the feature channels that are highly related to the geometric shape of the hidden danger are further focused on, and the redundant interference channels are weakened, thereby improving the pertinence and discriminability of the geometric features. Thus, the tenth intermediate refined feature map (i.e., T12) with optimized channel weights and more prominent geometric feature expression is obtained.

[0283] S2536. Perform element-wise multiplication on the ninth intermediate refined feature map and the tenth intermediate refined feature map to obtain the corresponding third deep refined feature map.

[0284] In this embodiment of the invention, the ninth intermediate refined feature map and the tenth intermediate refined feature map are multiplied by elements, and the spatial attention weight is fused with the geometric feature depth to further enhance the accurate expression of the hidden geometric features, suppress invalid interference information, and finally obtain the third deep refined feature map (i.e., T13).

[0285] S254. The first deep refined feature map, the second deep refined feature map and the third deep refined feature map are fused through the feature fusion layer to obtain the corresponding fourth deep refined feature map.

[0286] In this embodiment of the invention, the first deep refined feature map, the second deep refined feature map and the third deep refined feature map are fused by the feature fusion layer. By channel splicing and feature complementarity, the advantages of each of the three are integrated, and the texture details, geometric shape and accurate weight information are taken into account. Redundant interference is eliminated, and the organic fusion of features of different dimensions is achieved, finally obtaining the fourth deep refined feature map (i.e. T14).

[0287] S255. The fourth deep refined feature map is integrated by sequentially connecting the third convolutional module and the 1×1 convolutional layer to obtain the corresponding fifth deep refined feature map.

[0288] In this embodiment of the invention, the fourth deep refined feature map is purified by the third convolutional module, focusing on the core features of potential problems and weakening invalid interference information. Then, the output of the third convolutional module (i.e., T15) is compressed and its features are regularized by a 1×1 convolutional layer to eliminate redundant overlap between features of different branches, realize efficient integration of multi-dimensional features, and ensure the simplicity and relevance of feature expression, thus obtaining the corresponding fifth deep refined feature map (i.e., T16).

[0289] S256. The fifth deep refined feature map is refined using a dual-channel texture through the first deep attention branch to obtain the corresponding sixth deep refined feature map.

[0290] In this embodiment of the invention, the fifth deep refined feature map is subjected to dual-channel texture refinement through a first deep attention branch. This optimizes both the texture details and channel features of the fifth deep refined feature map, strengthening core texture information while suppressing invalid interference, thus achieving precise extraction and purification of texture features to obtain a sixth deep refined feature map (T22). For example, the fifth deep refined feature map T16 is input to the Ghost Conv module to generate a deep multimodal refined feature map T17; T17 is input to a channel attention module CAM to generate channel weights T18; T18 is multiplied by T17 to obtain a deep multimodal refined feature map T19; T19 is input to the Ghost Conv module for deep feature transformation to generate a deep multimodal refined feature map T20; T20 is input to an efficient channel attention ECA module to generate channel weights T21; T21 is multiplied by T20 and weighted twice to obtain a deep multimodal refined feature map T22.

[0291] S257. Perform double-layer geometric refinement on the fifth deep refined feature map through the second deep attention branch to obtain the corresponding seventh deep refined feature map.

[0292] In this embodiment of the invention, the fifth deep refined feature map is subjected to double-layer geometric refinement through a second deep attention branch. First, the geometric morphological information in the feature is initially refined, and then the geometric structure details are further optimized to remove invalid geometric interference information, thereby achieving double-layer enhancement and purification of the geometric features and obtaining the seventh deep refined feature map (i.e., T28). For example, the fifth deep refined feature map T16 is input into a deformable convolution module to generate a deep multimodal refined feature map T23. T23 is input into a spatial attention module SAM to focus on key spatial regions and generate spatial weights T24. T24 and T23 are multiplied element-wise to focus on the original geometric features and obtain a deep multimodal refined feature map T25. Then, T25 is input into a deformable convolution module again to generate a deep multimodal refined feature map T26, and a high-efficiency channel attention ECA module is used to generate channel weights T27. T27 and T26 are weighted twice to achieve channel refinement of the deep geometric features and obtain the deep refined deep multimodal refined feature map T28.

[0293] S258. The fifth, sixth, and seventh deep refined feature maps are fused through the feature fusion layer to obtain the corresponding eighth deep refined feature map.

[0294] In this embodiment of the invention, the fifth, sixth and seventh deep refined feature maps are fused by a feature fusion layer. This integrates three core features: texture details, channel weights and geometric shapes, to achieve complementary fusion of features from different branches. Redundant information is eliminated and effective features are enhanced to obtain the eighth deep refined feature map (i.e., T29).

[0295] S259. The third convolution module integrates the features of the eighth deep refined feature map to obtain the corresponding first-scale fused feature map.

[0296] In this embodiment of the invention, the eighth deep refined feature map is integrated by the third convolution module to further compress redundant information and strengthen effective features, so as to achieve the synergistic connection of features of different dimensions and obtain feature map T30 as the first scale fusion feature map (i.e. F15).

[0297] S26. The first scale fusion feature map is extracted by the second convolution module to obtain the corresponding sixth intermediate scale feature map. The sixth intermediate scale feature map and the third intermediate scale feature map are then concatenated by the feature fusion layer to obtain the corresponding seventh intermediate scale feature map.

[0298] In this embodiment of the invention, the first-scale fused feature map is extracted by the second convolution module to further focus on the core features of the hidden danger and remove redundant interference, so as to obtain a sixth intermediate-scale feature map (i.e., F16) with clear texture and high recognizability. The sixth intermediate-scale feature map and the third intermediate-scale feature map are then spliced ​​by the feature fusion layer. Through channel splicing, feature complementarity is achieved to obtain a seventh intermediate-scale feature map (i.e., F17) with multi-dimensional information.

[0299] S27. The deep multimodal attention feature refinement module is used to refine the feature map of the seventh intermediate scale to obtain the corresponding second scale fusion feature map.

[0300] In this embodiment of the invention, the seventh intermediate scale feature map is deeply refined by the deep multimodal attention feature refinement module. Combined with the multidimensional attention mechanism of this module (i.e., DMMAFR), the core features of the hidden danger are accurately extracted and invalid interference information is eliminated. The relevance and recognizability of the feature expression are further optimized, and a second scale fusion feature map (F18) with both completeness and accuracy is obtained.

[0301] S28. The second convolutional module extracts features from the second scale fusion feature map to obtain the corresponding eighth intermediate scale feature map. The feature fusion layer then concatenates the eighth intermediate scale feature map and the first intermediate scale feature map to obtain the corresponding ninth intermediate scale feature map.

[0302] In this embodiment of the invention, the second convolution module extracts features from the second-scale fused feature map, further purifies the core features related to potential hazards and removes invalid interference information, resulting in an eighth intermediate-scale feature map (i.e., F19) with more concentrated features and higher recognizability. The feature fusion layer then splices the eighth intermediate-scale feature map and the first intermediate-scale feature map to achieve complementary advantages of the two types of features, integrate the core information of texture details and geometric features, and obtain the corresponding ninth intermediate-scale feature map (i.e., F20).

[0303] S29. The deep multimodal attention feature refinement module is used to refine the feature map of the ninth intermediate scale to obtain the corresponding third scale fusion feature map.

[0304] In this embodiment of the invention, the deep multimodal attention feature refinement module is used to refine the ninth intermediate scale feature map, further screen and strengthen the core features related to the hidden danger, remove invalid interference information, improve the discriminability and stability of the features, and obtain the corresponding third scale fusion feature map (i.e., F21).

[0305] Step 208: Perform collaborative decoupling processing on the first-scale fused feature map, the second-scale fused feature map, and the third-scale fused feature map through a hierarchical guided detection network to obtain the corresponding detection results.

[0306] Further, see Figure 3 As shown, the hierarchical guided detection network includes three asymmetric interactive detection heads and a post-processing module. Step 208 includes the following sub-steps:

[0307] S31. The first-scale fused feature map is subjected to hierarchical collaborative prediction through the asymmetric interaction detection head (i.e., AIPH-SDG) to obtain the corresponding first prediction result. The asymmetric interaction detection head includes three third convolutional modules, a first prediction branch, a second prediction branch, a third prediction branch and a fourth prediction branch.

[0308] Further, see Figure 7 As shown, S31 includes the following sub-steps:

[0309] S311. The first scale fused feature map is extracted by two sequentially connected third convolutional modules to obtain the corresponding first predicted feature map.

[0310] In this embodiment of the invention, the first scale fusion feature map F15 is extracted by two third convolutional modules connected in sequence. Through the synergistic effect of the two third convolutional modules, the feature extraction is gradually deepened, the core information of the hidden danger is purified layer by layer, redundant interference is further eliminated, and the discriminativeness and compactness of the features are enhanced to obtain the first predicted feature map (i.e., H2).

[0311] S312. The first prediction feature map is subjected to hidden danger detection through the first prediction branch to obtain the corresponding first prediction tensor. The first prediction branch includes a 1×1 convolutional layer, a third convolutional module, a feature co-refinement module and a 1×1 convolutional layer connected in sequence.

[0312] The first prediction tensor refers to the output of the first prediction branch. The tensor contains the bounding box coordinates of the potential target and its corresponding confidence score.

[0313] In this embodiment of the invention, the first predicted feature map is first integrated and reduced in dimensionality using a 1×1 convolutional layer to remove redundant features and retain core information related to hazard location, resulting in feature map H3. Then, two sequentially connected third convolutional modules are used to continuously extract local features and perform nonlinear transformations on feature map H3, enhancing the boundary and contour information of the hazard target in the feature map, resulting in the corresponding feature map H5. Next, a feature co-refinement module (SARMP) further refines and enhances feature map H5, suppressing background interference and highlighting effective features of the hazard area, resulting in feature map H6. Finally, a 1×1 convolutional layer is used to predict and decode feature map H6, obtaining a first prediction tensor (A1) containing the bounding box coordinates and confidence scores of the hazard target.

[0314] It should be noted that, for reference Figure 8As shown, the Feature Collaborative Refinement Module (SARMP) includes a 1×1 convolutional layer (Conv1×1), three third convolutional modules, a 5×5 convolutional layer (Conv5×5), a 7×7 convolutional layer (Conv7×7), a first-channel attention module, a feature fusion layer, and a sigmoid activation function layer. The specific processing steps of the Feature Collaborative Refinement Module are as follows: A1. The feature map H5 is integrated using the 1×1 convolutional layer to generate a predicted refined feature map P1. A2. The predicted refined feature map P1 is refined using the third convolutional module to generate a predicted refined feature map P2. A3. The first-channel attention module captures the cross-channel interaction information of the predicted refined feature map P2 to generate channel attention weights P3. A4. The channel attention weights P3 and the predicted refined feature map P1 are multiplied element-wise to obtain the predicted refined feature map P4. A5. The predicted refined feature map P5 is generated by using the 5×5 convolutional layer to extract mid-level semantics from the predicted refined feature map P2 using a larger receptive field. A6. The predicted refined feature map P5 is deepened using the third convolutional module to generate the predicted refined feature map P6. A7. Wide-area spatial information is extracted from the predicted refined feature map P2 using a 7×7 convolutional layer with a large convolutional kernel to generate the predicted refined feature map P7. A8. The predicted refined feature map P7 is input into a Sigmoid activation function layer to generate spatial attention weights P8. A9. Element-wise multiplication is performed between the spatial attention weights P8 and the predicted refined feature map P1 to generate the predicted refined feature map P9. A10. A concat operation is performed on the predicted refined feature maps P4, P6, and P9 using a feature fusion layer to generate the predicted refined feature map P10. A11. The predicted refined feature map P10 undergoes deep integration of cross-branch information using the third convolutional module to obtain feature map P11, which serves as the corresponding feature map H6.

[0315] S313. The first predicted feature map is extracted through the second prediction branch to obtain the corresponding second predicted feature map. The second prediction branch includes a 1×1 convolutional layer and a Sigmoid activation function layer connected in sequence.

[0316] In this embodiment of the invention, the first predicted feature map is first compressed and weighted using a 1×1 convolutional layer to extract feature information representing the area of ​​concern for potential hazards, resulting in feature map H7. Then, a sigmoid activation function layer is used to normalize the feature values ​​in feature map H7 to the range of 0 to 1, generating a second predicted feature map (i.e., H8) that can highlight the key area of ​​potential hazards and suppress background interference.

[0317] S314. Perform element-wise multiplication between the second predicted feature map and the first predicted feature map to obtain the corresponding third predicted feature map.

[0318] In this embodiment of the invention, the second predicted feature map and the first predicted feature map are multiplied element by element, and the first predicted feature map is spatially weighted and modulated using the weight information of the second predicted feature map, so that the features of the hidden danger area are enhanced and the features of the background area are suppressed, thereby obtaining a third predicted feature map (i.e., H9) that highlights the key information of the hidden danger and has lower noise interference.

[0319] S315. The third prediction feature map is extracted by the third convolution module to obtain the corresponding fourth prediction feature map.

[0320] In this embodiment of the invention, the third prediction feature map is extracted by the third convolution module. Local feature enhancement and channel regularization are completed by convolution operation, batch normalization and nonlinear activation, which further highlights the core features of the hidden danger and filters redundant background information, and obtains a fourth prediction feature map (i.e. H10) suitable for subsequent classification and grading tasks.

[0321] S316. The third prediction branch is used to identify potential hazards in the fourth prediction feature map to obtain the corresponding second prediction tensor. The third prediction branch includes a feature co-refinement module and a 1×1 convolutional layer connected in sequence.

[0322] The second prediction tensor refers to the classification result tensor output by the third prediction branch, which contains the predicted probability scores for each hazard category and is used to determine the specific hazard type of the target to be detected.

[0323] In this embodiment of the invention, the semantic features related to the hazard category in the fourth prediction feature map are deeply focused and denoised by the feature collaborative refinement module to enhance the feature differences between different categories of hazards, and a feature map H11 is obtained. Then, the feature map H11 is classified and decoded by a 1×1 convolutional layer and mapped to the probability distribution of the corresponding hazard category to obtain the second prediction tensor used to characterize the hazard category.

[0324] S317. The fourth prediction feature map is evaluated for hidden danger level through the fourth prediction branch to obtain the corresponding third prediction tensor, and the first prediction tensor, the second prediction tensor and the third prediction tensor are determined as the corresponding first prediction results.

[0325] The third prediction tensor refers to the result tensor output by the fourth prediction branch, which contains the predicted probability values ​​for each level of the hidden danger.

[0326] It should be noted that the fourth prediction branch includes a feature co-refinement module and a 1×1 convolutional layer connected in sequence.

[0327] In this embodiment of the invention, the subtle features of the hidden danger in the fourth prediction feature map are extracted and judged by the fourth prediction branch to complete the assessment calculation of the severity level of the hidden danger, and the third prediction tensor containing the probabilities of three levels: general, severe and critical is output. The first prediction tensor, the second prediction tensor and the third prediction tensor are determined as the corresponding first prediction result C1.

[0328] S32. The second-scale fused feature map is subjected to hierarchical collaborative prediction using an asymmetric interactive detection head to obtain the corresponding second prediction result.

[0329] The second prediction result refers to the complete medium-scale prediction set output by the asymmetric interactive detection head for the second-scale fused feature map, which includes the hazard location tensor, classification tensor, and level assessment tensor.

[0330] In this embodiment of the invention, the second-scale fusion feature map F18 is subjected to hierarchical collaborative prediction by an asymmetric interactive detection head. Following the same hierarchical collaborative prediction logic as the first-scale fusion feature map, the parallel decoupled calculations of hazard location detection, category identification and level assessment are completed in sequence. The location, classification and level prediction tensors corresponding to the medium-scale feature are output, and finally the corresponding second prediction result C2 is obtained by integration.

[0331] S33. The third-scale fused feature map is subjected to hierarchical collaborative prediction by the asymmetric interactive detection head to obtain the corresponding third prediction result.

[0332] The third prediction result refers to the complete small-scale prediction set output by the asymmetric interactive detection head for the third-scale fused feature map, which includes the localization tensor, classification tensor, and ranking evaluation tensor.

[0333] In this embodiment of the invention, the third-scale fusion feature map F21 is subjected to hierarchical collaborative prediction by an asymmetric interactive detection head. The hierarchical collaborative prediction process is consistent with that of the first two scales. The decoupling operation of hazard location, category identification and level assessment is completed in sequence. The location tensor, classification tensor and level assessment tensor corresponding to the deep feature map of this small scale are output. After integration, the corresponding third prediction result C3 is obtained.

[0334] S34. The post-processing module filters the first prediction result, the second prediction result, and the third prediction result to obtain the corresponding detection result.

[0335] It should be noted that the post-processing module is used to uniformly organize, deduplicate, filter and output the original multi-scale prediction results of the model, thereby improving the accuracy and simplicity of the final detection results.

[0336] In this embodiment of the invention, the first prediction result, the second prediction result and the third prediction result are decoded and restored by the post-processing module, and then the non-maximum suppression algorithm is used to remove duplicate and redundant detection boxes. The hidden danger targets are screened and filtered according to the confidence threshold, and the detection results containing the precise location, category and level are integrated.

[0337] It should be noted that bounding box decoding refers to the process of restoring / converting the bounding box encoding parameters output by the network to bounding box coordinates in the image coordinate system. Non-maximum suppression algorithms are commonly used filtering algorithms to remove overlapping and redundant candidate boxes.

[0338] In this embodiment of the invention, multiple training images are acquired, and each training image is preprocessed to generate a corresponding line feature set. The line feature set is used to train a preset initial transmission line hazard detection model to generate a corresponding target transmission line hazard detection model. The target transmission line hazard detection model includes a backbone network, a multi-modal attention feature network, and a hierarchical guided detection network. An image of the transmission line to be detected is acquired. The backbone network extracts multi-scale features from the image of the transmission line to be detected to obtain corresponding first-scale feature maps, second-scale feature maps, and third-scale feature maps. The multi-modal attention feature network performs cross-scale feature fusion on the first-scale feature maps, second-scale feature maps, and third-scale feature maps to obtain corresponding first-scale fused feature maps, second-scale fused feature maps, and third-scale fused feature maps. The hierarchical guided detection network performs collaborative decoupling processing on the first-scale fused feature maps, second-scale fused feature maps, and third-scale fused feature maps to obtain the corresponding detection result. This invention overcomes the technical challenges of existing transmission line hazard detection methods, which often employ general target detection algorithms. While these algorithms can locate and identify power components and obvious defects, they struggle to adapt to heterogeneous hazard features such as texture, geometry, and context. Furthermore, collaborative detection of multiple hazard types is prone to missed detections, reducing the accuracy of transmission line hazard detection. Compared to traditional transmission line detection methods, this invention utilizes a backbone network to efficiently extract multi-scale and multi-type hazard features from images of the transmission line to be detected. It can simultaneously adapt to heterogeneous features such as texture, geometry, and context, significantly enhancing the feature representation capability for different hazard types. A multi-modal attention feature network then completes cross-scale feature fusion and deep refinement, effectively suppressing background interference and enhancing the features of minor hazards. This solves the problems of feature ambiguity and information loss that traditional fusion methods often cause. Finally, a hierarchical guided detection network achieves collaborative decoupling of localization, classification, and hazard assessment, eliminating feature conflicts between multiple tasks. It can simultaneously output hazard location, category, and hazard level in a single inference, making the detection results more aligned with the actual needs of power operation and maintenance, thus improving the accuracy and efficiency of transmission line hazard detection.

[0339] Please see Figure 9 , Figure 9 This is a structural block diagram of a power transmission line hidden danger detection system provided in Embodiment 3 of the present invention.

[0340] This invention provides a power transmission line hidden danger detection system, comprising:

[0341] Preprocessing module 301 is used to acquire multiple training images, perform image preprocessing on each training image, and generate corresponding line feature sets;

[0342] Training module 302 is used to train a preset initial transmission line hazard detection model using a line feature set to generate a corresponding target transmission line hazard detection model. The target transmission line hazard detection model includes a backbone network, a multi-modal attention feature network, and a hierarchical guided detection network.

[0343] The acquisition module 303 is used to acquire images of the transmission line to be detected, and to perform multi-scale feature extraction on the images of the transmission line to be detected through the backbone network to obtain the corresponding first-scale feature map, second-scale feature map and third-scale feature map.

[0344] The cross-scale feature fusion module 304 is used to perform cross-scale feature fusion on the first-scale feature map, the second-scale feature map and the third-scale feature map through a multi-modal attention feature network to obtain the corresponding first-scale fused feature map, the second-scale fused feature map and the third-scale fused feature map.

[0345] The detection module 305 is used to perform collaborative decoupling processing on the first-scale fused feature map, the second-scale fused feature map, and the third-scale fused feature map through a hierarchical guided detection network to obtain the corresponding detection results.

[0346] Furthermore, training module 302 includes:

[0347] The training submodule is used to train the preset initial transmission line hidden danger detection model by inputting the line feature set, and to obtain the corresponding training detection data.

[0348] The loss submodule is used to calculate the loss function value of the line feature set based on the training detection data;

[0349] The adjustment submodule is used to adjust the network parameters of the initial transmission line hidden danger detection model when the loss function value is greater than or equal to the preset loss threshold, until the loss function value is less than the loss threshold.

[0350] When the loss function value is less than the loss threshold, a corresponding target transmission line hazard detection model is generated.

[0351] Furthermore, the backbone network includes a first convolutional module, a multi-branch feature extraction module, a second convolutional module, a multi-branch feature extraction module, a second convolutional module, and a multi-branch feature extraction module connected in sequence, and the acquisition module 303 includes:

[0352] The first extraction submodule is used to extract features from the image of the transmission line to be detected through the first convolution module to obtain the corresponding first convolution feature map.

[0353] The second extraction submodule is used to perform multi-dimensional feature decoupling extraction on the first convolutional feature map through the multi-branch feature extraction module to obtain the corresponding first scale feature map.

[0354] The third extraction submodule is used to extract features from the first scale feature map through the second convolution module to obtain the corresponding second convolution feature map.

[0355] The fourth extraction submodule is used to perform multi-dimensional feature decoupling extraction on the second convolutional feature map through the multi-branch feature extraction module to obtain the corresponding second-scale feature map;

[0356] The fifth extraction submodule is used to extract features from the second-scale feature map through the second convolution module to obtain the corresponding third convolution feature map;

[0357] The sixth extraction submodule is used to perform multi-dimensional feature decoupling extraction on the third convolutional feature map through the multi-branch feature extraction module to obtain the corresponding third-scale feature map.

[0358] Furthermore, the multi-branch feature extraction module includes two third convolutional modules, a texture analysis branch, a context aggregation branch, a geometry-aware branch, a global-aware gated weight generator, and a feature fusion layer. The second extraction sub-module includes:

[0359] The first extraction unit is used to extract features from the first convolutional feature map through the third convolutional module to obtain the corresponding hidden danger perception feature map;

[0360] The weight perception unit is used to dynamically perceive the weight of the hidden danger perception feature map through the global perception gating weight generator, and obtain the corresponding first gating weight, second gating weight and third gating weight.

[0361] The texture feature extraction unit is used to extract texture features from the hidden danger perception feature map through the texture analysis branch to obtain the corresponding texture feature map, and to perform weighted modulation on the texture feature map using the first gate weight to obtain the corresponding target texture feature map;

[0362] The context-aware unit is used to perform global context awareness on the hidden danger awareness feature map through the context aggregation branch to obtain the corresponding context aggregation feature map, and to use the second gating weight to perform weighted modulation on the context aggregation feature map to obtain the corresponding target aggregation feature map.

[0363] The geometric feature extraction unit is used to extract geometric features from the hidden danger perception feature map through the geometric perception branch to obtain the corresponding geometric feature map, and then use the third gating weight to perform weighted modulation on the geometric feature map to obtain the corresponding target geometric feature map.

[0364] The first fusion unit is used to stitch together the target texture feature map, the target aggregate feature map and the target geometric feature map through the feature fusion layer to obtain the corresponding multi-form power transmission hazard perception feature map;

[0365] The second extraction unit is used to extract features from the multi-form power transmission hazard perception feature map through the third convolution module to obtain the corresponding first-scale feature map.

[0366] Furthermore, the globally aware gated weight generator includes a global pooling module, two mapping modules, and a weight splitting module. The weight-aware unit includes:

[0367] The feature globalization subunit is used to globalize the feature map of hidden danger perception through the global pooling module to obtain the corresponding global feature vector.

[0368] The first mapping subunit is used to perform continuous nonlinear mapping on the global feature vector through two mapping modules connected in sequence to obtain the corresponding intermediate feature vector. The mapping module includes a fully connected layer and a ReLU activation function layer connected in sequence.

[0369] The gated weight segmentation subunit is used to perform gated weight segmentation on the intermediate feature vector through the weight segmentation module to obtain the corresponding first gated weight, second gated weight and third gated weight. The weight segmentation module includes a Sigmoid activation function layer and a feature splitting layer connected in sequence.

[0370] Furthermore, the texture analysis branch includes a third convolutional module, a global average pooling module, a global max pooling module, two feature fusion layers, and a first channel attention module. The texture feature extraction unit includes:

[0371] The first extraction subunit is used to extract features from the hazard perception feature map through the third convolution module to obtain the corresponding first-form hazard perception feature map.

[0372] The second extraction subunit is used to perform feature globalization on the first form of hidden danger perception feature map through the global average pooling module to obtain the corresponding second form of hidden danger perception feature map.

[0373] The first pooling subunit is used to sharpen the first-form hazard perception feature map through the global max pooling module to obtain the corresponding third-form hazard perception feature map.

[0374] The first fusion subunit is used to stitch together the second-form hazard perception feature map and the third-form hazard perception feature map through the feature fusion layer to obtain the corresponding fourth-form hazard perception feature map.

[0375] The first attention unit is used to perform channel adaptive processing on the fourth form of hidden danger perception feature map through the first channel attention module to obtain the corresponding channel attention weight, and to perform element-wise multiplication operation between the channel attention weight and the first form of hidden danger perception feature map to obtain the corresponding fifth form of hidden danger perception feature map.

[0376] The second fusion subunit is used to perform feature fusion on the fifth-form hazard perception feature map and the first-form hazard perception feature map through the feature fusion layer to obtain the corresponding texture feature map.

[0377] Furthermore, the context aggregation branch includes two third convolutional modules, a multi-head self-attention module, two feature fusion layers, and a spatial attention module. The context-aware unit includes:

[0378] The third extraction subunit is used to extract features from the hazard perception feature map through the third convolution module to obtain the corresponding sixth-form hazard perception feature map.

[0379] The fourth extraction subunit is used to perform contextual aggregation on the sixth-form hazard perception feature map through the multi-head self-attention module to obtain the corresponding seventh-form hazard perception feature map.

[0380] The third fusion subunit is used to perform feature fusion on the sixth form of hidden danger perception feature map and the seventh form of hidden danger perception feature map through the feature fusion layer to obtain the corresponding first residual hidden danger perception feature map.

[0381] The fifth extraction subunit is used to extract features from the first residual hidden danger perception feature map through the third convolution module to obtain the corresponding second residual hidden danger perception feature map.

[0382] The sixth extraction subunit is used to perform spatial focusing processing on the second residual hidden danger perception feature map through the spatial attention module to obtain the corresponding spatial focusing feature map;

[0383] The fourth fusion subunit is used to perform feature fusion on the first residual hidden danger perception feature map and the spatial focusing feature map through the feature fusion layer to obtain the corresponding context aggregation feature map.

[0384] Furthermore, the geometry-aware branch includes a 1×1 convolutional layer, a geometric contour extraction module, and a feature fusion layer. The geometric feature extraction unit includes:

[0385] The seventh extraction subunit is used to extract features from the hazard perception feature map through a 1×1 convolutional layer to obtain the corresponding eighth-form hazard perception feature map.

[0386] The eighth extraction subunit is used to extract geometric contour features from the eighth form of hazard perception feature map through the geometric contour extraction module to obtain the corresponding ninth form of hazard perception feature map. The geometric contour extraction module includes a 3×3 deformable convolutional layer, a 1×1 convolutional layer and a ReLU activation function layer connected in sequence.

[0387] The fifth fusion subunit is used to perform feature fusion on the eighth-form hazard perception feature map and the ninth-form hazard perception feature map through the feature fusion layer to obtain the corresponding geometric feature map.

[0388] Furthermore, the multimodal attention feature network includes a spatial pyramid fast pooling module, two sampling modules, two second convolutional modules, four feature fusion layers, three deep multimodal attention feature refinement modules, and a cross-scale feature fusion module 304, including:

[0389] The first stitching submodule is used to stitch the second intermediate scale feature map and the second scale feature map through the feature fusion layer to obtain the corresponding third intermediate scale feature map.

[0390] The first sampling submodule is used to upsample the third intermediate scale feature map through the sampling module to obtain the corresponding fourth intermediate scale feature map, and to concatenate the fourth intermediate scale feature map and the first scale feature map through the feature fusion layer to obtain the corresponding fifth intermediate scale feature map.

[0391] The seventh extraction submodule is used to perform deep feature refinement on the fifth intermediate scale feature map through the deep multimodal attention feature refinement module to obtain the corresponding first scale fusion feature map.

[0392] The eighth extraction submodule is used to extract features from the first scale fused feature map through the second convolution module to obtain the corresponding sixth intermediate scale feature map, and to concatenate the sixth intermediate scale feature map and the third intermediate scale feature map through the feature fusion layer to obtain the corresponding seventh intermediate scale feature map.

[0393] The ninth extraction submodule is used to perform deep feature refinement on the seventh intermediate scale feature map through the deep multimodal attention feature refinement module to obtain the corresponding second scale fusion feature map.

[0394] The tenth extraction submodule is used to extract features from the second scale fusion feature map through the second convolution module to obtain the corresponding eighth intermediate scale feature map, and to concatenate the eighth intermediate scale feature map and the first intermediate scale feature map through the feature fusion layer to obtain the corresponding ninth intermediate scale feature map.

[0395] The eleventh extraction submodule is used to refine the ninth intermediate scale feature map through the deep multimodal attention feature refinement module to obtain the corresponding third scale fusion feature map.

[0396] Furthermore, the deep multimodal attention feature refinement module includes two first deep attention branches, two second deep attention branches, two 1×1 convolutional layers, two feature fusion layers, and two third convolutional modules. The seventh extraction submodule includes:

[0397] The first feature integration unit is used to integrate the features of the fifth intermediate scale feature map through a 1×1 convolutional layer to obtain the corresponding first depth refined feature map.

[0398] The first dual-channel texture refining unit is used to refine the first deep-refined feature map in a dual-channel manner through the first deep attention branch to obtain the corresponding second deep-refined feature map.

[0399] The first double-layer geometric refinement unit is used to perform double-layer geometric refinement on the first deep refinement feature map through the second deep attention branch to obtain the corresponding third deep refinement feature map.

[0400] The second fusion unit is used to perform feature fusion on the first deep refined feature map, the second deep refined feature map and the third deep refined feature map through the feature fusion layer to obtain the corresponding fourth deep refined feature map.

[0401] The second feature integration unit is used to integrate the features of the fourth deep refined feature map through the third convolutional module and the 1×1 convolutional layer connected in sequence to obtain the corresponding fifth deep refined feature map.

[0402] The second dual-channel texture refining unit is used to refine the fifth deep-refined feature map in a dual-channel manner through the first deep attention branch to obtain the corresponding sixth deep-refined feature map.

[0403] The second double-layer geometric refinement unit is used to perform double-layer geometric refinement on the fifth deep refinement feature map through the second deep attention branch to obtain the corresponding seventh deep refinement feature map.

[0404] The third fusion unit is used to perform feature fusion on the fifth, sixth and seventh deep refined feature maps through the feature fusion layer to obtain the corresponding eighth deep refined feature map.

[0405] The third feature integration unit is used to integrate the features of the eighth deep refined feature map through the third convolution module to obtain the corresponding first scale fused feature map.

[0406] Furthermore, the first deep attention branch includes two ghost convolutional modules, a first-channel attention module, and a second-channel attention module, and a first dual-channel texture refining unit, including:

[0407] The first deep feature extraction submodule is used to perform deep feature transformation on the first deep refined feature map through the ghost convolution module to obtain the corresponding first intermediate refined feature map.

[0408] The first intermediate refined feature map is processed by the second channel attention module to obtain the corresponding second intermediate refined feature map.

[0409] The first and second intermediate refined feature maps are multiplied by elements to obtain the corresponding third intermediate refined feature map.

[0410] The second deep feature extraction submodule is used to perform deep feature transformation on the third intermediate refined feature map through the ghost convolution module to obtain the corresponding fourth intermediate refined feature map.

[0411] The fourth intermediate refined feature map is processed by the first channel attention module to obtain the corresponding fifth intermediate refined feature map.

[0412] Element-wise multiplication is performed on the fourth and fifth intermediate refined feature maps to obtain the corresponding second deep refined feature map.

[0413] Furthermore, the second deep attention branch includes two deformable convolutional modules, a spatial attention module, and a first channel attention module, and a first two-layer geometric refinement unit, including:

[0414] The third deep feature extraction submodule is used to extract geometric features from the first deep refined feature map through a deformable convolution module to obtain the corresponding sixth intermediate refined feature map.

[0415] The sixth intermediate refined feature map is spatially focused by the spatial attention module to obtain the corresponding seventh intermediate refined feature map.

[0416] Perform element-wise multiplication on the sixth and seventh intermediate refined feature maps to obtain the corresponding eighth intermediate refined feature map.

[0417] The fourth deep feature extraction submodule is used to extract geometric features from the eighth intermediate refined feature map through a deformable convolution module to obtain the corresponding ninth intermediate refined feature map.

[0418] The ninth intermediate refined feature map is processed by the first channel attention module to obtain the corresponding tenth intermediate refined feature map.

[0419] Element-wise multiplication is performed on the ninth and tenth intermediate refined feature maps to obtain the corresponding third deep refined feature map.

[0420] Furthermore, the hierarchical guided detection network includes three asymmetric interactive detection heads and a post-processing module. Detection module 305 includes:

[0421] The first hierarchical collaborative prediction submodule is used to perform hierarchical collaborative prediction on the first scale fused feature map through the asymmetric interaction detection head to obtain the corresponding first prediction result. The asymmetric interaction detection head includes three third convolutional modules, a first prediction branch, a second prediction branch, a third prediction branch and a fourth prediction branch.

[0422] The second hierarchical collaborative prediction submodule is used to perform hierarchical collaborative prediction on the second-scale fused feature map through an asymmetric interactive detection head to obtain the corresponding second prediction result.

[0423] The third hierarchical collaborative prediction submodule is used to perform hierarchical collaborative prediction on the third-scale fused feature map through an asymmetric interactive detection head to obtain the corresponding third prediction result.

[0424] The filtering submodule is used to filter the first prediction result, the second prediction result, and the third prediction result through the post-processing module to obtain the corresponding detection result.

[0425] Furthermore, the first hierarchical collaborative prediction submodule includes:

[0426] The first prediction unit is used to extract features from the first scale fused feature map through two sequentially connected third convolutional modules to obtain the corresponding first prediction feature map.

[0427] The first prediction branch performs hidden danger detection on the first prediction feature map to obtain the corresponding first prediction tensor. The first prediction branch includes a 1×1 convolutional layer, a third convolutional module, a feature co-refinement module, and a 1×1 convolutional layer connected in sequence.

[0428] The second prediction unit is used to extract features from the first prediction feature map through the second prediction branch to obtain the corresponding second prediction feature map. The second prediction branch includes a 1×1 convolutional layer and a Sigmoid activation function layer connected in sequence.

[0429] The second predicted feature map is multiplied element-wise with the first predicted feature map to obtain the corresponding third predicted feature map.

[0430] The third prediction feature map is extracted by the third convolution module to obtain the corresponding fourth prediction feature map;

[0431] The third prediction branch is used to identify potential hazards in the fourth prediction feature map to obtain the corresponding second prediction tensor. The third prediction branch includes a feature co-refinement module and a 1×1 convolutional layer connected in sequence.

[0432] The third prediction unit is used to assess the hazard level of the fourth prediction feature map through the fourth prediction branch, obtain the corresponding third prediction tensor, and determine the first prediction tensor, the second prediction tensor and the third prediction tensor as the corresponding first prediction result.

[0433] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0434] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0435] 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 hidden dangers in power transmission lines, characterized in that, include: Multiple training images are acquired, and image preprocessing is performed on each training image to generate a corresponding line feature set; The preset initial transmission line hazard detection model is trained using the line feature set to generate a corresponding target transmission line hazard detection model. The target transmission line hazard detection model includes a backbone network, a multi-modal attention feature network, and a hierarchical guided detection network. An image of the transmission line to be detected is acquired, and multi-scale feature extraction is performed on the image of the transmission line to be detected through the backbone network to obtain the corresponding first-scale feature map, second-scale feature map and third-scale feature map; The multimodal attention feature network is used to perform cross-scale feature fusion on the first scale feature map, the second scale feature map and the third scale feature map to obtain the corresponding first scale fused feature map, second scale fused feature map and third scale fused feature map. The hierarchical guided detection network is used to perform collaborative decoupling processing on the first-scale fused feature map, the second-scale fused feature map, and the third-scale fused feature map to obtain the corresponding detection results.

2. The method for detecting hidden dangers in transmission lines according to claim 1, characterized in that, The step of training a preset initial transmission line hazard detection model using the line feature set to generate a corresponding target transmission line hazard detection model includes: The line feature set is used as input to train the preset initial transmission line hidden danger detection model to obtain the corresponding training detection data; Calculate the loss function value of the line feature set based on the training and detection data; When the loss function value is greater than or equal to the preset loss threshold, the network parameters of the initial transmission line hidden danger detection model are adjusted until the loss function value is less than the loss threshold. When the loss function value is less than the loss threshold, a corresponding target transmission line hazard detection model is generated.

3. The method for detecting hidden dangers in transmission lines according to claim 1, characterized in that, The backbone network includes a first convolutional module, a multi-branch feature extraction module, a second convolutional module, a second convolutional module, and a multi-branch feature extraction module connected in sequence. The step of performing multi-scale feature extraction on the transmission line image to be detected through the backbone network to obtain corresponding first-scale feature maps, second-scale feature maps, and third-scale feature maps includes: The first convolution module is used to extract features from the image of the transmission line to be detected to obtain the corresponding first convolution feature map. The first convolutional feature map is decoupled and extracted using a multi-branch feature extraction module to obtain the corresponding first-scale feature map. The second convolutional module is used to extract features from the first scale feature map to obtain the corresponding second convolutional feature map. The second convolutional feature map is decoupled and extracted using a multi-branch feature extraction module to obtain the corresponding second-scale feature map. The second convolution module extracts features from the second scale feature map to obtain the corresponding third convolution feature map; The third convolutional feature map is decoupled and extracted using a multi-branch feature extraction module to obtain the corresponding third-scale feature map.

4. The method for detecting hidden dangers in transmission lines according to claim 3, characterized in that, The multi-branch feature extraction module includes two third convolutional modules, a texture analysis branch, a context aggregation branch, a geometry-aware branch, a global awareness gated weight generator, and a feature fusion layer. The step of performing multi-dimensional feature decoupling extraction on the first convolutional feature map through the multi-branch feature extraction module to obtain the corresponding first-scale feature map includes: The third convolution module extracts features from the first convolution feature map to obtain the corresponding hidden danger perception feature map. The global perception gating weight generator performs dynamic weight perception on the hidden danger perception feature map to obtain the corresponding first gating weight, second gating weight and third gating weight. The texture features of the hidden danger perception feature map are extracted by the texture analysis branch to obtain the corresponding texture feature map, and the texture feature map is weighted and modulated by the first gating weight to obtain the corresponding target texture feature map. The context aggregation branch is used to perform global context perception on the hidden danger perception feature map to obtain the corresponding context aggregation feature map, and the second gating weight is used to perform weighted modulation on the context aggregation feature map to obtain the corresponding target aggregation feature map. Geometric features are extracted from the hazard perception feature map through the geometric perception branch to obtain the corresponding geometric feature map, and the geometric feature map is weighted and modulated using the third gating weight to obtain the corresponding target geometric feature map; The target texture feature map, the target aggregated feature map, and the target geometric feature map are spliced ​​together by the feature fusion layer to obtain the corresponding multi-form power transmission hazard perception feature map; The third convolution module extracts features from the multi-form power transmission hazard perception feature map to obtain the corresponding first-scale feature map.

5. The method for detecting hidden dangers in transmission lines according to claim 4, characterized in that, The global perception gating weight generator includes a global pooling module, two mapping modules, and a weight segmentation module. The step of dynamically perceiving weights on the hazard perception feature map using the global perception gating weight generator to obtain the corresponding first gating weight, second gating weight, and third gating weight includes: The feature map of the hidden danger perception is globalized by the global pooling module to obtain the corresponding global feature vector; The global feature vector is continuously nonlinearly mapped by two sequentially connected mapping modules to obtain the corresponding intermediate feature vector. The mapping module includes a fully connected layer and a ReLU activation function layer connected in sequence. The intermediate feature vector is segmented by a weight segmentation module to obtain the corresponding first gate weight, second gate weight and third gate weight. The weight segmentation module includes a sigmoid activation function layer and a feature splitting layer connected in sequence.

6. The method for detecting hidden dangers in transmission lines according to claim 4, characterized in that, The texture analysis branch includes a third convolutional module, a global average pooling module, a global max pooling module, two feature fusion layers, and a first channel attention module. The step of extracting texture features from the hazard perception feature map through the texture analysis branch to obtain the corresponding texture feature map includes: The third convolution module is used to extract features from the hazard perception feature map to obtain the corresponding first-morphology hazard perception feature map. The first morphological hazard perception feature map is globalized by the global average pooling module to obtain the corresponding second morphological hazard perception feature map. The first-form hazard perception feature map is sharpened by the global max pooling module to obtain the corresponding third-form hazard perception feature map. The second form of hazard perception feature map and the third form of hazard perception feature map are spliced ​​together by the feature fusion layer to obtain the corresponding fourth form of hazard perception feature map. The fourth form of hidden danger perception feature map is processed by the first channel attention module to obtain the corresponding channel attention weight. The channel attention weight is then multiplied by the first form of hidden danger perception feature map to obtain the corresponding fifth form of hidden danger perception feature map. The feature fusion layer fuses the fifth-form hazard perception feature map and the first-form hazard perception feature map to obtain the corresponding texture feature map.

7. The method for detecting hidden dangers in transmission lines according to claim 4, characterized in that, The context aggregation branch includes two third convolutional modules, a multi-head self-attention module, two feature fusion layers, and a spatial attention module. The step of performing global context awareness on the hazard perception feature map through the context aggregation branch to obtain the corresponding context aggregation feature map includes: The third convolutional module is used to extract features from the hazard perception feature map to obtain the corresponding sixth-form hazard perception feature map. By performing contextual aggregation on the sixth-form hazard perception feature map using a multi-head self-attention module, the corresponding seventh-form hazard perception feature map is obtained. The feature fusion layer is used to fuse the sixth form of hidden danger perception feature map and the seventh form of hidden danger perception feature map to obtain the corresponding first residual hidden danger perception feature map. The third convolution module is used to extract features from the first residual hidden danger perception feature map to obtain the corresponding second residual hidden danger perception feature map. The spatial attention module is used to perform spatial focusing processing on the second residual hidden danger perception feature map to obtain the corresponding spatial focusing feature map; The first residual hazard perception feature map and the spatial focusing feature map are fused by the feature fusion layer to obtain the corresponding context aggregation feature map.

8. The method for detecting hidden dangers in transmission lines according to claim 4, characterized in that, The geometric perception branch includes a 1×1 convolutional layer, a geometric contour extraction module, and a feature fusion layer. The step of extracting geometric features from the hazard perception feature map through the geometric perception branch to obtain the corresponding geometric feature map includes: The hazard perception feature map is extracted by a 1×1 convolutional layer to obtain the corresponding eighth-mode hazard perception feature map. The geometric contour extraction module extracts geometric contour features from the eighth-form hazard perception feature map to obtain the corresponding ninth-form hazard perception feature map. The geometric contour extraction module includes a 3×3 deformable convolutional layer, a 1×1 convolutional layer and a ReLU activation function layer connected in sequence. The feature fusion layer performs feature fusion on the eighth-form hazard perception feature map and the ninth-form hazard perception feature map to obtain the corresponding geometric feature map.

9. The method for detecting hidden dangers in transmission lines according to claim 1, characterized in that, The multimodal attention feature network includes a spatial pyramid fast pooling module, two sampling modules, two second convolutional modules, four feature fusion layers, and three deep multimodal attention feature refinement modules. The step of performing cross-scale feature fusion on the first-scale feature map, the second-scale feature map, and the third-scale feature map through the multimodal attention feature network to obtain the corresponding first-scale fused feature map, second-scale fused feature map, and third-scale fused feature map includes: The third-scale feature map is extracted using the spatial pyramid fast pooling module to obtain the corresponding first intermediate-scale feature map. The first intermediate scale feature map is upsampled by the sampling module to obtain the corresponding second intermediate scale feature map. The sampling module includes a third convolution module and a sampling operation module connected in sequence. The second intermediate scale feature map and the second scale feature map are concatenated by the feature fusion layer to obtain the corresponding third intermediate scale feature map. The third intermediate scale feature map is upsampled by the sampling module to obtain the corresponding fourth intermediate scale feature map, and the fourth intermediate scale feature map and the first scale feature map are spliced ​​by the feature fusion layer to obtain the corresponding fifth intermediate scale feature map. The deep multimodal attention feature refinement module is used to refine the fifth intermediate scale feature map to obtain the corresponding first scale fusion feature map. The second convolutional module extracts features from the first scale fusion feature map to obtain the corresponding sixth intermediate scale feature map. The feature fusion layer then concatenates the sixth intermediate scale feature map and the third intermediate scale feature map to obtain the corresponding seventh intermediate scale feature map. The deep multimodal attention feature refinement module is used to refine the seventh intermediate scale feature map to obtain the corresponding second scale fusion feature map. The second convolutional module extracts features from the second scale fusion feature map to obtain the corresponding eighth intermediate scale feature map. The feature fusion layer then concatenates the eighth intermediate scale feature map and the first intermediate scale feature map to obtain the corresponding ninth intermediate scale feature map. The deep multimodal attention feature refinement module is used to refine the ninth intermediate scale feature map to obtain the corresponding third scale fusion feature map.

10. The method for detecting hidden dangers in transmission lines according to claim 9, characterized in that, The deep multimodal attention feature refinement module includes two first deep attention branches, two second deep attention branches, two 1×1 convolutional layers, two feature fusion layers, and two third convolutional modules. The step of refining the fifth intermediate-scale feature map using the deep multimodal attention feature refinement module to obtain the corresponding first-scale fused feature map includes: The fifth intermediate scale feature map is integrated by a 1×1 convolutional layer to obtain the corresponding first depth refined feature map. The first deep refined feature map is refined using a dual-channel texture through the first deep attention branch to obtain the corresponding second deep refined feature map. The first deep refined feature map is subjected to two-layer geometric refinement through the second deep attention branch to obtain the corresponding third deep refined feature map; The first deep refined feature map, the second deep refined feature map, and the third deep refined feature map are fused by the feature fusion layer to obtain the corresponding fourth deep refined feature map. The fourth deep refined feature map is integrated by sequentially connecting the third convolutional module and the 1×1 convolutional layer to obtain the corresponding fifth deep refined feature map. The fifth deep refined feature map is refined using a dual-channel texture through the first deep attention branch to obtain the corresponding sixth deep refined feature map; The fifth deep refined feature map is subjected to two-layer geometric refinement through the second deep attention branch to obtain the corresponding seventh deep refined feature map; The fifth, sixth, and seventh deep refined feature maps are fused using a feature fusion layer to obtain the corresponding eighth deep refined feature map. The third convolution module integrates the features of the eighth deep refined feature map to obtain the corresponding first-scale fused feature map.

11. The method for detecting hidden dangers in transmission lines according to claim 10, characterized in that, The first deep attention branch includes two ghost convolution modules, a first-channel attention module, and a second-channel attention module. The step of performing dual-channel texture refinement on the first deep refined feature map through the first deep attention branch to obtain the corresponding second deep refined feature map includes: The first deep refined feature map is transformed by the ghost convolution module to obtain the corresponding first intermediate refined feature map. The first intermediate refined feature map is processed by the second channel attention module to obtain the corresponding second intermediate refined feature map. The first intermediate refined feature map and the second intermediate refined feature map are multiplied by elements to obtain the corresponding third intermediate refined feature map. The third intermediate refined feature map is subjected to depth feature transformation by the ghost convolution module to obtain the corresponding fourth intermediate refined feature map; The fourth intermediate refined feature map is processed by the first channel attention module to obtain the corresponding fifth intermediate refined feature map. The fourth intermediate refined feature map and the fifth intermediate refined feature map are multiplied by elements to obtain the corresponding second deep refined feature map.

12. The method for detecting hidden dangers in transmission lines according to claim 10, characterized in that, The second deep attention branch includes two deformable convolutional modules, a spatial attention module, and a first channel attention module. The step of performing two-layer geometric refinement on the first deep refined feature map through the second deep attention branch to obtain the corresponding third deep refined feature map includes: Geometric features are extracted from the first deep refined feature map using a deformable convolution module to obtain the corresponding sixth intermediate refined feature map. The sixth intermediate refined feature map is spatially focused by the spatial attention module to obtain the corresponding seventh intermediate refined feature map. The sixth intermediate refined feature map and the seventh intermediate refined feature map are multiplied by element to obtain the corresponding eighth intermediate refined feature map. Geometric features are extracted from the eighth intermediate refined feature map using a deformable convolution module to obtain the corresponding ninth intermediate refined feature map. The ninth intermediate refined feature map is processed by the first channel attention module to obtain the corresponding tenth intermediate refined feature map. The ninth intermediate refined feature map and the tenth intermediate refined feature map are multiplied by elements to obtain the corresponding third deep refined feature map.

13. The method for detecting hidden dangers in transmission lines according to claim 1, characterized in that, The hierarchical guided detection network includes three asymmetric interactive detection heads and a post-processing module. The step of performing collaborative decoupling processing on the first-scale fused feature map, the second-scale fused feature map, and the third-scale fused feature map through the hierarchical guided detection network to obtain the corresponding detection results includes: The first-scale fused feature map is subjected to hierarchical collaborative prediction by an asymmetric interaction detection head to obtain the corresponding first prediction result. The asymmetric interaction detection head includes three third convolutional modules, a first prediction branch, a second prediction branch, a third prediction branch, and a fourth prediction branch. The second-scale fused feature map is subjected to hierarchical collaborative prediction using an asymmetric interactive detection head to obtain the corresponding second prediction result. The third-scale fused feature map is subjected to hierarchical collaborative prediction by an asymmetric interactive detection head to obtain the corresponding third prediction result. The post-processing module filters the first prediction result, the second prediction result, and the third prediction result to obtain the corresponding detection result.

14. The method for detecting hidden dangers in transmission lines according to claim 13, characterized in that, The step of performing hierarchical collaborative prediction on the first-scale fused feature map using an asymmetric interactive detection head to obtain the corresponding first prediction result includes: The first scale fusion feature map is extracted by two sequentially connected third convolutional modules to obtain the corresponding first predicted feature map. The first prediction branch performs hidden danger detection on the first prediction feature map to obtain the corresponding first prediction tensor. The first prediction branch includes a 1×1 convolutional layer, a third convolutional module, a feature co-refinement module, and a 1×1 convolutional layer connected in sequence. The second prediction branch extracts features from the first prediction feature map to obtain the corresponding second prediction feature map. The second prediction branch includes a 1×1 convolutional layer and a Sigmoid activation function layer connected in sequence. The second predicted feature map is multiplied element-wise with the first predicted feature map to obtain the corresponding third predicted feature map. The third prediction feature map is extracted by the third convolution module to obtain the corresponding fourth prediction feature map; The third prediction branch is used to identify potential hazards in the fourth prediction feature map to obtain the corresponding second prediction tensor. The third prediction branch includes a feature co-refinement module and a 1×1 convolutional layer connected in sequence. The fourth prediction branch is used to assess the hazard level of the fourth prediction feature map to obtain the corresponding third prediction tensor, and the first prediction tensor, the second prediction tensor and the third prediction tensor are determined as the corresponding first prediction result.

15. A power transmission line hidden danger detection system, characterized in that, include: The preprocessing module is used to acquire multiple training images, perform image preprocessing on each training image, and generate a corresponding line feature set; The training module is used to train the preset initial transmission line hidden danger detection model using the line feature set to generate the corresponding target transmission line hidden danger detection model. The target transmission line hidden danger detection model includes a backbone network, a multi-modal attention feature network, and a hierarchical guided detection network. The acquisition module is used to acquire images of the transmission line to be detected, and to perform multi-scale feature extraction on the images of the transmission line to be detected through the backbone network to obtain corresponding first-scale feature maps, second-scale feature maps and third-scale feature maps. The cross-scale feature fusion module is used to perform cross-scale feature fusion on the first-scale feature map, the second-scale feature map and the third-scale feature map through the multi-modal attention feature network to obtain the corresponding first-scale fused feature map, second-scale fused feature map and third-scale fused feature map; The detection module is used to perform collaborative decoupling processing on the first-scale fusion feature map, the second-scale fusion feature map, and the third-scale fusion feature map through the hierarchical guided detection network to obtain the corresponding detection results.

Images