Power transmission line key target detection algorithm based on DualAxis-SEAtt attention mechanism

Abstract

The invention discloses a power transmission line key target detection algorithm based on a DualAxis-SEAtt attention mechanism, relates to an algorithm for detecting a power transmission line key target, and aims to solve the problems that power transmission line key target detection in unmanned aerial vehicle inspection is interfered by a complex background, small target recognition accuracy is low, and detection accuracy is low. And the existing algorithm precision and the reasoning efficiency are difficult to consider at the same time. According to the algorithm, Yolov8 is used as a framework, a DualAxis-SEAtt module comprising a channel attention branch, a cross attention branch and an output fusion module is embedded in a backbone network of the Yolov8, key features are screened through the channel branch, horizontal / vertical bidirectional spatial dependence is captured through the cross branch, an enhanced feature map is output through feature splicing, projection transformation and residual error connection, and the feature map is subjected to feature fusion, projection transformation and residual error connection. Accurate detection of power lines, towers, insulators and stockbridge dampers is realized.

Description

Technical Field

[0001] This invention relates to an algorithm for detecting key targets on power transmission lines, specifically an algorithm, device, storage medium, and program product for detecting key targets on power transmission lines based on the DualAxis-SEAtt attention mechanism. It is based on the fusion and optimization of channel attention mechanism and spatial attention mechanism, and incorporates a Yolov8 backbone network to improve the model's accuracy in detecting and identifying key targets, thereby meeting the accuracy requirements of UAV inspection. Background Technology

[0002] In power systems, the safe and stable operation of transmission lines is crucial. Power lines, power poles, insulators, and vibration dampers, as core components of transmission lines, directly determine the reliability and operational safety of the power grid. With the continuous expansion of my country's power grid and the increasing coverage of transmission lines, the operating environment is becoming increasingly complex. Traditional manual inspection algorithms are not only inefficient and costly, but also struggle to meet the monitoring needs of large-scale lines, let alone the requirements for real-time, high-precision monitoring across multiple time periods. Therefore, automated and intelligent monitoring technologies based on unmanned aerial vehicle (UAV) inspections have become an important development direction in the field of power system operation and maintenance.

[0003] The mainstream classification of object detection algorithms revolves around whether candidate regions are generated, and can be divided into two-stage and single-stage algorithms. Two-stage algorithms first generate candidate regions through RPN and selective search, and then perform category judgment and bounding box regression. Examples include Faster R-CNN and Mask R-CNN. They have high accuracy but are slow and difficult to adapt to real-time inspection by UAVs. Single-stage algorithms perform direct "end-to-end" detection, are fast, and can easily incorporate attention mechanisms. They are further divided into anchor-based and anchorless types. The former pre-sets anchor boxes, while the latter relies on key point detection, making them highly adaptable to irregular targets.

[0004] In the scenario of UAV inspection of power transmission lines, anchor-based single-stage target detection algorithms are more applicable due to their "end-to-end detection and fast inference speed" characteristics, among which the YOLO series models perform particularly well. Specifically, YOLOv1-YoLOv3 lay the foundation for detection by optimizing the feature extraction network and loss function, while YOLOv5-YoLOv8 continuously iterate the network structure, focusing on improving the detection accuracy of small-sized targets. In addition, some studies have attempted to incorporate attention mechanisms such as CBAM and ECA into YOLOv8 to further enhance the model's ability to capture features of power transmission components in complex backgrounds.

[0005] Existing technical solutions have two shortcomings: First, the integration of attention mechanisms is mostly a simple superposition without key optimization for the characteristics of power transmission components, resulting in unreasonable attention weight allocation. Insufficient capture of key features of power transmission components leads to a decrease in feature recognition, which significantly affects the model's detection accuracy. Second, existing solutions do not fully coordinate the relationship between attention mechanisms and model inference efficiency. Some models introduce relatively complex attention modules, which can improve detection accuracy to some extent, but significantly increase the model's computational complexity, resulting in a decrease in inference speed and failing to meet the needs of real-time inspection at the edge of UAVs. Summary of the Invention

[0006] The purpose of this invention is to provide a key target detection algorithm for power transmission lines based on the DualAxis-SEAtt attention mechanism. Addressing the problem that existing algorithms for detecting key targets on power transmission lines are easily affected by complex background interference and have insufficient accuracy in identifying small targets, this invention proposes a dual-axis attention mechanism fusion algorithm for key targets on power transmission lines. It innovatively links a dedicated texture channel, edge channel, and spatial position channel collaboratively to improve the unstable and inaccurate identification of small targets by UAVs in complex environments.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A key target detection algorithm for transmission lines based on the DualAxis-SEAtt attention mechanism is characterized by embedding a DualAxis-SEAtt attention module into the YOLOv8 backbone network, using the YOLOv8 target detection algorithm as the basic framework. The DualAxis-SEAtt module includes a channel attention branch, a cross-junction attention branch, and an output fusion module. Specifically, the algorithm includes the following steps:

[0009] S1: Simultaneously input the feature map of the input transmission line image into the channel attention branch and the cross-intersection attention branch of the DualAxis-SEAtt module to achieve parallel enhancement of channel features and spatial features;

[0010] S2: The channel attention branch performs global feature compression, weight learning, and feature transformation on the feature map, filters out the key channel features of the transmission line target, and compresses the number of channels to C / 4, where C is the number of channels in the original feature map;

[0011] S3: The cross-cutting attention branch generates a query vector Q based on the output features of the channel attention branch. It combines the original input features to generate a key vector K and a value vector V. In the transmission line detection environment, Q represents the query feature, which is used to locate key targets such as conductors and insulators; K represents the key feature, which stores feature information at different locations in the image to respond to the query; V represents the value feature, which provides detailed features at each location for reconstructing the final representation. The spatial dependencies of key targets in the transmission line are captured through horizontal and vertical bidirectional attention calculations.

[0012] S4: The output features of the channel attention branch and the cross attention branch are concatenated along the channel dimension, the original number of channels is restored by projection transformation, and residual connections are introduced to fuse with the initial input features to output an enhanced feature map;

[0013] S5: Based on the enhanced feature map, complete the detection of key targets of power lines, power poles, insulators, and vibration dampers for power transmission lines.

[0014] The processing flow for the S2 channel attention branch specifically includes the following steps:

[0015] S21: Feature compression stage: Global average pooling is used to process the input [B,C,H,W] dimension feature map, compressing the feature map dimension to [B,C,1,1]; where B is the batch size, C is the number of channels, H is the feature map height, and W is the feature map width.

[0016] S22: Weight learning stage: First, the number of channels is reduced from C to C / 16 through 1×1 convolution. After processing by the ReLU activation function, the number of channels is restored to C through 1×1 convolution. Then, the channel weights in the range [0,1] are generated using the Sigmoid activation function.

[0017] S23: Feature transformation stage: A combination of "1×1 convolution + batch normalization + GELU activation" is used to compress the number of channels in the feature map to C / 4, and output a feature map of [B,C / 4,H,W] dimensions.

[0018] The processing flow for the cross-attention branch in step S3 specifically includes the following steps:

[0019] S31: Feature Mapping Stage: The query vector Q is generated by a 1×1 convolution of the [B,C / 4,H,W] feature map output from the channel attention branch, with the channel dimension adjusted to C / 8. This vector is used to represent the semantic query information of key targets such as wires and insulators in the input image. The key vector K and value vector V are generated by a 1×1 grouped convolution with 4 groups from the original input [B,C,H,W] feature map, with the channel dimension reduced to C / 8. This vector is used to provide feature location mapping and detailed feature expression of key targets, respectively. Finally, Q, K, and V vectors with dimensions of [B,C / 8,H,W] are obtained.

[0020] S32: Dual-axis attention calculation stage: When calculating horizontal attention, the Q and K vectors are transposed to [B,W,H,C / 8], a similarity matrix is ​​constructed and normalized by softmax to obtain the horizontal attention weights, which are multiplied with the transposed V vector matrix to obtain the horizontal weighted features, and then converted back to the [B,C / 8,H,W] dimension;

[0021] S33: During vertical attention calculation, the Q, K, and V vectors are transposed to [B,H,W,C / 8], and the vertical weighted features are obtained following the same process as above; Feature fusion stage: The weighted features of [B,C / 8,H,W] output in the horizontal and vertical directions are added and fused point by point to output a fused feature map of [B,C / 8,H,W] dimensions.

[0022] Step S4, which concatenates the output features of the channel attention branch and the cross-attention branch along the channel dimension, specifically includes the following steps:

[0023] S41: Feature concatenation: The [B,C / 4,H,W] feature output by the channel attention branch and the [B,C / 8,H,W] feature output by the cross attention branch are concatenated along the channel dimension to obtain the fused feature [B,3C / 8,H,W].

[0024] S42: Projection Transformation: Through 1×1 convolution and batch normalization operations, the dimensions of the fused feature channels are restored to [B,C,H,W].

[0025] S43: Residual connection: Weighted summation of the projected features and the initial input feature map.

[0026] The parameters for algorithm training were set as follows: initial learning rate of 0.001, batch size of 16, number of iterations of 500, and original image size of 640×640 pixels.

[0027] The dataset used for algorithm training includes images of four types of power transmission components: power lines, power poles, insulators, and vibration dampers. Valid training samples are generated through label annotation, horizontal / vertical flipping, and color transformation data augmentation.

[0028] It also includes S6 to evaluate the object detection performance of the DualAxis-SEAtt attention mechanism, using accuracy. Recall rate Mean accuracy To ensure the comprehensiveness of the algorithm evaluation, the values ​​of the three evaluation indicators are all between 0 and 1. The higher the value, the better the detection effect. The calculation formulas are as follows:

[0029] ;

[0030] ;

[0031] ;

[0032] ;

[0033] In the formula, For the accuracy of the model; The number of samples that are key targets of actual power transmission lines and that the model correctly detects the target; The model incorrectly identified the number of samples that were actually key targets of non-transmission lines as targets. The recall rate of the model; This represents the number of samples that are actually key targets of transmission lines but were not detected by the model. The average precision for a single category, where Specific categories representing key targets of transmission lines; The value of recall rate; To correspond to the recall rate The precision value under the given range; The total number of categories of key targets for transmission lines, which are divided into four categories: power lines, power poles, insulators, and vibration dampers.

[0034] The working mechanism of this key target detection algorithm for transmission lines based on the DualAxis-SEAtt attention mechanism revolves around the core logic of "parallel feature enhancement - dual-axis feature fusion - accurate detection and evaluation": Using the Yolov8 framework as the basic framework, the feature map of the transmission line image is first synchronously input into the channel attention branch and cross-junction attention branch of the DualAxis-SEAtt module to achieve parallel enhancement of channel features and spatial features. The channel attention branch dynamically filters key channel features of transmission line targets, including power lines, towers, insulators, and vibration dampers, through global feature compression, weight learning, and feature transformation, and reduces the dimension to C / 4, suppressing redundant channel information and providing high-quality input for spatial feature capture. The cross-attention branch generates a query vector Q based on the channel-filtered features, and combines it with the original features to generate a key vector K and a value vector V. Through horizontal and vertical bidirectional attention calculations, it comprehensively captures the spatial dependencies of the target, breaking through the limitation of traditional spatial attention's single-direction capture. The output fusion module concatenates the channel and spatial features along the dimension, restores the original channel dimension through projection transformation, and introduces residual connections to retain the initial feature information, outputting an enhanced feature map. The enhanced feature map is fed into the Yolov8 Neck and Head networks to complete target category judgment and bounding box regression, achieving accurate detection. Finally, the detection performance is quantitatively evaluated through metrics such as precision, recall, and mean precision, verifying the effectiveness of the algorithm under complex working conditions.

[0035] Compared with the prior art, the beneficial effects of the present invention are:

[0036] 1. The DualAxis-SEAtt attention module designed in this invention breaks through the limitations of single-dimensional optimization in traditional attention mechanisms. The channel attention branch dynamically allocates channel weights based on the key target features of the transmission line, accurately screening key channel features of power lines, towers, insulators, and vibration dampers. The cross-intersection attention branch captures the spatial dependency of the target through horizontal / vertical bidirectional calculations, solving the problems of unreasonable attention weight allocation and low feature recognition in existing technologies. Experiments verify that the algorithm's mAP50 and Precision are superior to traditional attention mechanisms such as SEAttention, ECA, and CBAM, and the recognition accuracy of small targets in complex backgrounds is greatly improved.

[0037] 2. The channel attention branch adopts a lightweight structure of "dimensionality reduction-activation-dimensionality increase". The cross-intersecting attention branch reduces the amount of computation through grouped convolution and dimensionality compression. The residual connection of the output fusion module retains feature information without adding a lot of extra computation. This solves the problem of high computational complexity and reduced inference speed caused by the introduction of complex attention modules in existing technologies, and is suitable for the real-time inspection needs of UAV edge terminals.

[0038] 3. The algorithm is designed with a dual-axis attention mechanism tailored to the characteristics of key targets on power transmission lines. The training dataset covers actual inspection conditions such as complex lighting and background occlusion and has been augmented with targeted data. The parameter settings are also tailored to the UAV inspection scenario. Compared with general target detection algorithms, it has higher adaptability to the detection of core components of power transmission lines and meets the needs of large-scale automated and high-precision inspection of power transmission lines. Attached Figure Description

[0039] Figure 1 This is a diagram illustrating the overall framework of the DualAxis-SEAtt attention mechanism, a key target detection algorithm for transmission lines based on the DualAxis-SEAtt attention mechanism of this invention.

[0040] Figure 2 The flowchart of the channel attention branch of the key target detection algorithm for transmission lines based on the DualAxis-SEAtt attention mechanism of the present invention is shown below.

[0041] Figure 3 The flowchart of the cross-attention branch of the key target detection algorithm for transmission lines based on the DualAxis-SEAtt attention mechanism of the present invention is shown below.

[0042] Figure 4 The image shows the target detection results of the DualAxis-SEAtt attention mechanism, which is a key target detection algorithm for transmission lines based on the DualAxis-SEAtt attention mechanism of this invention.

[0043] Figure 5 This is an attention thermal imaging image of the Yolov8 model of a key target detection algorithm for transmission lines based on the DualAxis-SEAtt attention mechanism, as described in this invention. Detailed Implementation

[0044] The technical solutions of the present invention will now be described in detail with reference to the accompanying drawings.

[0045] like Figure 1-5 As shown, this invention uses the YOLOv8 target detection algorithm as a framework and proposes a target detection algorithm based on the DualAxis-SEAtt attention mechanism. By embedding the DualAxis-SEAtt module into the YOLOv8 backbone network, an end-to-end feature enhancement system of "feature selection - dual-axis enhancement - fusion preservation" is constructed. This mechanism selects key feature channels through channel attention branches, captures bidirectional spatial dependencies through cross-interference attention branches, and integrates features and preserves original information through a fusion output module. Ultimately, it improves the detection accuracy of key targets in power transmission lines such as power lines, insulators, power poles, and vibration dampers, adapting to the detection needs of complex backgrounds and multi-scale targets.

[0046] Specifically, the core module of the DualAxis-SEAtt attention mechanism references the channel attention mechanism and the spatial attention mechanism, and the specific process is as follows:

[0047] First, the feature map is input, and then passed to the channel attention branch and cross-interference attention branch of the DualAxis-SEAtt module to achieve parallel enhancement of channel and spatial features.

[0048] Secondly, the channel attention branch filters out key channel features of transmission line targets through global feature compression, weight learning, and feature calibration. To reduce the computational burden on subsequent models, the number of channels is reduced to C / 4. The cross-attention branch first generates a query vector Q based on the output of the channel attention branch, and combines it with the original features to generate a key vector K and a value vector V. Through horizontal and vertical bidirectional attention calculations, it captures the spatial dependencies of the horizontal and vertical distribution of key transmission line targets. Finally, the fusion output module concatenates the features from the two branches, restores the original number of channels through projection transformation, and introduces residual connections to fuse with the initial input features to avoid feature loss. The final output features are enhanced feature maps, ensuring compatibility with the YOLOv8 Backbone and Neck network structures. Based on the enhanced feature maps, the detection of key transmission line targets such as power lines, power poles, insulators, and vibration dampers is completed, facilitating subsequent model computation. The overall framework of the DualAxis-SEAtt attention mechanism is as follows: Figure 1 As shown.

[0049] The channel attention branch is the core module of the DualAxis-SEAtt attention mechanism for mining channel features. Its main function is to dynamically allocate channel weights, mine the correlations between channels in the feature map, strengthen key channel features for transmission line target detection, suppress redundant channel information, and simultaneously adapt the feature dimension. This provides high-quality input for the spatial feature capture of the subsequent cross-attention branch, while also adapting to the Yolov8 model's requirements for filtering scattered features such as edges and textures in transmission line inspection. The channel attention branch process is as follows: Figure 2 As shown, the channel attention branch is mainly divided into three stages: "feature compression - weight learning - feature transformation".

[0050] The feature compression stage aims to preserve global features and avoid limitations imposed by local information. This stage employs global average pooling to process the input [B,C,H,W] dimensional feature map, aggregating the spatial information of each channel into a single value, compressing the feature map dimension to [B,C,1,1]. This achieves the transformation from spatial dimensional information to channel dimensional information, allowing the model to more comprehensively grasp the global feature responses of each channel and providing a foundation for subsequent weight learning.

[0051] The weight learning stage employs a lightweight "dimensionality reduction-activation-dimensionality increase" structure to accurately learn the importance weights of each channel. First, a 1×1 convolution reduces the number of channels from C to C / 16, reducing computation while promoting information exchange between different channels. Next, the ReLU activation function is used to process the dimensionality-reduced features, suppressing invalid responses from redundant channels and enhancing feature discriminative power. Finally, a 1×1 convolution restores the number of channels to C, and the Sigmoid activation function generates channel weights ranging from [0,1]. Weights closer to 1 indicate a higher contribution of the corresponding channel to transmission line target detection.

[0052] The feature transformation stage aims to adapt to the feature processing requirements of the subsequent cross-attention branch, while further optimizing feature quality. This stage employs a combination of "1×1 convolution + batch normalization (BN) + GELU activation": the 1×1 convolution compresses the number of channels in the feature-calibrated [B,C,H,W] dimension feature map to C / 4, reducing computational cost while reorganizing channel information; the BN operation stabilizes the feature value distribution and accelerates model training convergence; the GELU activation function, with its smoother gradient changes, effectively preserves the detailed structural features of key targets in the transmission line, ultimately outputting a [B,C / 4,H,W] dimension feature map, providing a suitable and high-quality input for the cross-attention branch.

[0053] The cross-attention branch is the core module of the DualAxis-SEAtt attention mechanism for mining spatial features. It aims to overcome the limitations of traditional spatial attention mechanisms, which can only capture local or single-directional spatial dependencies. By independently calculating attention weights in both horizontal and vertical directions, it comprehensively captures the spatial correlation information of key targets on transmission lines in the feature map, providing accurate spatial feature support for subsequent feature fusion and target detection. The cross-attention branch process is as follows: Figure 3 As shown.

[0054] The cross-attention branch is mainly divided into three stages: "feature mapping - dual-axis attention calculation - feature fusion".

[0055] In the feature mapping stage, Q (query), K (key), and V (value) vectors are generated based on the output features of the channel attention branch and the original input features. The query vector Q is generated by a 1×1 convolution of the [B,C / 4,H,W] feature map output by the channel attention branch, adjusting the channel dimension to C / 8. This reduces computational cost while preserving key channel calibration features. The key vector K and value vector V are generated by a 1×1 grouped convolution with 4 groups from the original input [B,C,H,W] feature map, similarly reducing the channel dimension to C / 8. Grouped convolution reduces the number of parameters and strengthens local correlations within channels, ultimately yielding Q, K, and V vectors with dimensions of [B,C / 8,H,W].

[0056] Dual-axis attention computation is the core innovation of this branch, which is processed in parallel in the horizontal and vertical directions: In the horizontal attention computation, the Q and K vectors are first transposed to [B,W,H,C / 8]. A similarity matrix is ​​constructed through matrix multiplication and normalized by softmax to obtain the horizontal attention weights, capturing the positional associations of different columns in the same row. Then, it is multiplied with the transposed V vector matrix to obtain the horizontal weighted features, and finally converted back to the [B,C / 8,H,W] dimension. The vertical attention computation process is the same, only the Q, K, and V vectors are transposed to [B,H,W,C / 8] to capture the positional associations of different rows in the same column. The dual-axis design can simultaneously obtain the linear and local spatial distribution features of the target, avoiding information omission.

[0057] In the feature fusion stage, the weighted features of [B,C / 8,H,W] output in the horizontal and vertical directions are superimposed and fused by point-by-point addition. Without the need for additional parameters, the effective information of the dual-axis spatial attention is fully preserved, and the final output is a fused feature map in the [B,C / 8,H,W] dimensions, which together with the output features of the channel attention branch enters the subsequent output fusion stage.

[0058] Output fusion is a key step in the DualAxis-SEAtt attention mechanism to achieve collaborative enhancement of "channel-space" features. Its core function is to integrate the output features of the channel attention branch and the cross-intersection attention branch, while preserving the original feature information and strengthening the representation ability of key features, so as to provide complete and efficient feature input for the detection and segmentation tasks of the subsequent YOLOv8-seg model.

[0059] The output fusion calculation is based on the output features of two major branches and achieves deep integration of features in three steps: First, feature splicing is performed, which splices the [B,C / 4,H,W] features output by the channel attention branch and the [B,C / 8,H,W] features output by the cross attention branch along the channel dimension to obtain the fused feature with dimensions [B,3C / 8,H,W]. This step can completely retain the key features after channel screening and the dual-axis spatial correlation features, avoiding the loss of information from single-branch features, and is especially suitable for the joint requirements of "channel texture features + spatial location features" in power transmission components.

[0060] Next, a projection transformation is performed. Through 1×1 convolution and batch normalization, the concatenated [B,3C / 8,H,W] feature channel dimensions are restored to the original input [B,C,H,W]. The 1×1 convolution is responsible for the linear mapping of the channel dimensions, ensuring that the feature dimensions are compatible with the subsequent model structure. The BN operation stabilizes the feature value distribution and avoids the feature distribution shift after fusion from affecting the model training convergence.

[0061] Finally, residual connections are introduced to perform a weighted summation of the projected features and the initial input feature map. The residual design can effectively preserve the basic information of the original features, prevent the loss of feature representation integrity due to attention weighting, and strengthen the contribution of key information in the fused features through weight adjustment.

[0062] To adapt to the task of UAV inspection and identification of power transmission components, this paper constructs a dedicated dataset containing four core components: power lines, power poles, insulators, and vibration dampers. The original images are taken from the Internet, covering working conditions such as complex lighting and background occlusion. After labeling and data augmentation such as horizontal / vertical flipping and color transformation, 10,998 valid images were obtained. The number of samples in each category is shown in Table 1.

[0063] Table 1 shows the number of images categorized by target category.

[0064] Power transmission component categories Sample size (sheets) power lines 3489 Power poles 2831 insulator 3590 Vibration damper 1088

[0065] Experimental environment and parameter settings

[0066] To ensure the reliability and effectiveness of the comparative experiments, the algorithm was performed on the same device, and the parameters used for the same network were kept consistent. The hardware, software environment, and training parameter settings are shown in Table 2.

[0067] Table 2 Experimental Environment and Parameter Settings

[0068] name parameter Language environment Python 3.0.8 GPU NVIDIA RTX 4090 Pre-trained weights yolov8n.pt Initial Learning Path 0.001 Batch size 16 Number of iterations 500 Original image size 640×640

[0069] To evaluate the object detection performance of the DualAxis-SEAtt attention mechanism, this paper uses three commonly used object detection metrics: Precision, Recall, and mAP (mean AP) to ensure comprehensive algorithm evaluation. The values ​​of all three metrics range from 0 to 1, with higher values ​​indicating better detection performance. The calculation formulas are as follows:

[0070]

[0071]

[0072]

[0073]

[0074] In the formula, For the accuracy of the model; The actual key targets of the transmission line are power lines, power poles, insulators, and vibration dampers, and the model correctly detects the number of samples of the target. The model incorrectly identified the number of samples that were actually key targets of non-transmission lines as targets. The recall rate of the model; This represents the number of samples that are actually key targets of transmission lines but were not detected by the model. The average precision for a single category, where Specific categories representing key targets of transmission lines; The value of recall rate; To correspond to the recall rate The precision value under the given range; This represents the total number of categories of key targets for power transmission lines, corresponding to four categories: power lines, power poles, insulators, and vibration dampers.

[0075] To verify the target detection performance of the proposed DualAxis-SEAtt attention mechanism, this experiment embeds the DualAxis-SEAtt attention mechanism module with other attention mechanism modules into the Yolov8 model for comparison. The attention mechanisms compared are SEAttention, ECA, CBAM, and GAM. The experimental results are shown in Table 3.

[0076] Table 3. Results of comparative experiments on different attention mechanisms

[0077] Model mAP50 Precision Recall Yolov8+DualAxis-SEAtt 0.89004 0.92931 0.81508 Yolov8+SEAttention 0.88674 0.91156 0.82837 Yolov8+ECA 0.88094 0.92679 0.81473 Yolov8+CBAM 0.87667 0.91808 0.80687 Yolov8+GAM 0.88513 0.92746 0.81025

[0078] The comparative experimental results are shown in Table 3 above. The left side of the table lists the types of attention mechanisms, and the right side lists the target detection results under different attention mechanisms. It is clear that different attention mechanisms perform differently when performing the task of detecting critical targets on power transmission lines.

[0079] (1) From the perspective of mAP50 analysis: Yolov8+DualAxis-SEAtt ranked first with 0.89004, which improved by 0.33, 0.91, 1.34 and 0.49 percentage points respectively compared with SEAttention (0.88674), ECA (0.88094), CBAM (0.87667) and GAM (0.88513). Its dual-axis collaborative design of "channel filtering + cross-shaped spatial locking" effectively balances classification and positioning accuracy and overcomes the limitations of traditional single-dimensional optimization.

[0080] (2) From the perspective of precision: The DualAxis-SEAtt attention mechanism showed excellent anti-false detection ability with 0.92931, which is 1.77 and 1.12 percentage points higher than SEAttention (0.91156) and CBAM (0.91808), respectively. It reduces the risk of background false activation by constructing target spatial dependence and filtering specific features through dual axes.

[0081] (3) From the perspective of recall: The DualAxis-SEAtt attention mechanism has a recall of 0.81508, which is slightly lower than SEAttention (0.82837), but higher than ECA (0.81473), CBAM (0.80687) and GAM (0.81025). It achieves a balance of "low false detection and high recall" by strengthening the features of small / occluded targets through spatial branching and amplifying weak signals through channel branching.

[0082] Analysis of the overall experimental results shows that Yolov8+DualAxis-SEAtt exhibits the best overall performance: both mAP50 and Precision are higher than those of comparable models such as SEAttention and ECA. Although Recall is slightly lower than SEAttention, it remains stable above 0.8, meeting the requirements for transmission line inspection. This advantage stems from the core design of the DualAxis-SEAtt attention mechanism, which breaks through the limitations of traditional single-dimensional approaches. It dynamically allocates weights to channel branches to filter key channel information, and uses bidirectional cross-branching to capture spatial features, ultimately adapting to complex inspection conditions such as strong light and occlusion.

[0083] The experimental results of the proposed algorithm and the attention heatmap of the Yolov8 model are shown in the figure below. Figure 4 , Figure 5 As shown, the results demonstrate that the DualAxis-SEAtt attention mechanism significantly improves the performance of key target detection tasks in transmission lines under complex environments.

Claims

1. A key target detection algorithm for transmission lines based on the DualAxis-SEAtt attention mechanism, characterized in that, Based on the YOLOv8 object detection algorithm, a DualAxis-SEAtt attention module is embedded in the YOLOv8 backbone network. The DualAxis-SEAtt module includes channel attention branches, cross-junction attention branches, and an output fusion module, specifically including the following steps: S1: Simultaneously input the feature map of the input transmission line image into the channel attention branch and the cross-intersection attention branch of the DualAxis-SEAtt module to achieve parallel enhancement of channel features and spatial features; S2: The channel attention branch performs global feature compression, weight learning, and feature transformation on the feature map, filters out the key channel features of the transmission line target, and compresses the number of channels to C / 4, where C is the number of channels in the original feature map; S3: The cross-cutting attention branch generates a query vector Q based on the output features of the channel attention branch. It combines the original input features to generate a key vector K and a value vector V. In the transmission line detection environment, Q represents the query feature, which is used to locate key targets such as conductors and insulators; K represents the key feature, which stores feature information at different locations in the image to respond to the query; V represents the value feature, which provides detailed features at each location for reconstructing the final representation. The spatial dependencies of key targets in the transmission line are captured through horizontal and vertical bidirectional attention calculations. S4: The output features of the channel attention branch and the cross attention branch are concatenated along the channel dimension, the original number of channels is restored by projection transformation, and residual connections are introduced to fuse with the initial input features to output an enhanced feature map; S5: Based on the enhanced feature map, complete the detection of key targets of power lines, power poles, insulators, and vibration dampers for power transmission lines.

2. The transmission line critical target detection algorithm based on the DualAxis-SEAtt attention mechanism according to claim 1, characterized in that, The processing flow for the S2 channel attention branch specifically includes the following steps: S21: Feature compression stage: Global average pooling is used to process the input [B,C,H,W] dimension feature map, compressing the feature map dimension to [B,C,1,1]; where B is the batch size, C is the number of channels, H is the feature map height, and W is the feature map width. S22: Weight learning stage: First, the number of channels is reduced from C to C / 16 through 1×1 convolution. After processing by the ReLU activation function, the number of channels is restored to C through 1×1 convolution. Then, the channel weights in the range [0,1] are generated using the Sigmoid activation function. S23: Feature transformation stage: A combination of "1×1 convolution + batch normalization + GELU activation" is used to compress the number of channels in the feature map to C / 4, and output a feature map of [B,C / 4,H,W] dimensions.

3. The transmission line critical target detection algorithm based on the DualAxis-SEAtt attention mechanism according to claim 1, characterized in that, The processing flow for the cross-attention branch in step S3 specifically includes the following steps: S31: Feature Mapping Stage: The query vector Q is generated by a 1×1 convolution of the [B,C / 4,H,W] feature map output from the channel attention branch, with the channel dimension adjusted to C / 8. This vector is used to represent the semantic query information of key targets such as wires and insulators in the input image. The key vector K and value vector V are generated by a 1×1 grouped convolution with 4 groups from the original input [B,C,H,W] feature map, with the channel dimension reduced to C / 8. This vector is used to provide feature location mapping and detailed feature expression of key targets, respectively. Finally, Q, K, and V vectors with dimensions of [B,C / 8,H,W] are obtained. S32: Dual-axis attention calculation stage: When calculating horizontal attention, the Q and K vectors are transposed to [B,W,H,C / 8], a similarity matrix is ​​constructed and normalized by softmax to obtain the horizontal attention weights, which are multiplied with the transposed V vector matrix to obtain the horizontal weighted features, and then converted back to the [B,C / 8,H,W] dimension; S33: During vertical attention calculation, the Q, K, and V vectors are transposed to [B,H,W,C / 8], and the vertical weighted features are obtained following the same process as above; Feature fusion stage: The weighted features of [B,C / 8,H,W] output in the horizontal and vertical directions are added and fused point by point to output a fused feature map of [B,C / 8,H,W] dimensions.

4. The transmission line critical target detection algorithm based on the DualAxis-SEAtt attention mechanism according to claim 1, characterized in that, Step S4, which concatenates the output features of the channel attention branch and the cross-attention branch along the channel dimension, specifically includes the following steps: S41: Feature concatenation: The [B,C / 4,H,W] feature output by the channel attention branch and the [B,C / 8,H,W] feature output by the cross attention branch are concatenated along the channel dimension to obtain the fused feature [B,3C / 8,H,W]. S42: Projection Transformation: Through 1×1 convolution and batch normalization operations, the dimensions of the fused feature channels are restored to [B,C,H,W]. S43: Residual connection: Weighted summation of the projected features and the initial input feature map.

5. The transmission line critical target detection algorithm based on the DualAxis-SEAtt attention mechanism according to any one of claims 1 to 4, characterized in that, The parameters for algorithm training were set as follows: initial learning rate of 0.001, batch size of 16, number of iterations of 500, and original image size of 640×640 pixels.

6. The transmission line critical target detection algorithm based on the DualAxis-SEAtt attention mechanism according to any one of claims 1 to 4, characterized in that, The dataset used for algorithm training includes images of four types of power transmission components: power lines, power poles, insulators, and vibration dampers. Valid training samples are generated through label annotation, horizontal / vertical flipping, and color transformation data augmentation.

7. The transmission line critical target detection algorithm based on the DualAxis-SEAtt attention mechanism according to claim 1, characterized in that, It also includes S6 to evaluate the object detection performance of the DualAxis-SEAtt attention mechanism, using accuracy. Recall rate Mean accuracy To ensure the comprehensiveness of the algorithm evaluation, the values ​​of the three evaluation indicators are all between 0 and 1. The higher the value, the better the detection effect. The calculation formulas are as follows: ； ； ； ； In the formula, For the accuracy of the model; The number of samples that are key targets of actual power transmission lines and that the model correctly detects the target; The model incorrectly identified the number of samples that were actually key targets of non-transmission lines as targets. The recall rate of the model; This represents the number of samples that are actually key targets of transmission lines but were not detected by the model. The average precision for a single category, where Specific categories representing key targets of transmission lines; The value of recall rate; To correspond to the recall rate The precision value under the given range; The total number of categories of key targets for transmission lines, which are divided into four categories: power lines, power poles, insulators, and vibration dampers.

Images