Power transmission line foreign matter detection method based on CE-YOLOv11

Abstract

The invention discloses a CE-YOLOv11-based power transmission line foreign matter detection method, which comprises the following steps: acquiring an image data set covering multiple types of typical power transmission line foreign matters, fusing depth separable convolution and a universal reverse bottleneck structure UIB in a backbone network, constructing a feature extraction network C3k2UIB on the basis of a feature extraction module C3k2, and extracting the foreign matters from the feature extraction network C3k2. In the feature fusion stage, an attention-guided feature fusion module EC2PSA is constructed, aiming at the characteristics of a transmission line foreign matter detection task, the transmission line foreign matter detection positioning precision and the small target detection performance are optimized by using a loss function powerfull-iou, and multiple rounds of training are performed under fixed size input. Through the improvement of a network structure and a feature extraction mechanism, the processing speed is effectively improved while the detection precision is ensured, the feature extraction and recognition capability of a small target is enhanced, and the robustness and generalization capability of a model in a complex environment are improved, so that the accuracy and real-time performance of transmission line foreign matter detection are improved, the detection time is shortened, and the detection efficiency is improved. And a more reliable technical support is provided for safe and stable operation of a power system.

Description

TECHNICAL FIELD

[0001] The present application relates to a power line foreign object detection method based on CE-YOLOv11, in particular to a power line foreign object detection method based on CE-YOLOv11 with robustness and practicality in complex environment and small target detection tasks, belonging to the technical field of power line foreign object detection. BACKGROUND

[0002] As a key link of power transmission, the safety and stability of power lines are crucial for the normal operation of the entire power system. However, power lines often face various external objects, such as bird nesting, litter falling, broken kites, and out-of-control balloons, which can cause line short circuits and even large-scale power outages. If a short circuit occurs in a remote forest area, it may also cause a fire. In order to avoid such accidents, regular inspection of power lines is particularly important.

[0003] Most existing power line foreign object detection methods are based on traditional convolutional neural network (CNN) architecture. Such methods have high dependence on image quality in practical applications, and when the input image has problems such as illumination variation, occlusion, blur, or complex background, the robustness and anti-interference ability of the model are poor. In complex scenarios, traditional CNNs are prone to insufficient local feature extraction and insufficient multi-scale information fusion, making it difficult to accurately identify foreign objects with different shapes and significant size differences. Due to the complex environment around the power line, trees, birds, kites, and other foreign objects often have partial occlusion or similar background to the power line, making it difficult for the model to effectively extract features from the occluded or low-contrast regions, thereby affecting the detection accuracy. In addition, convolutional networks are usually limited to local regions during feature extraction, lacking comprehensive utilization of global context information in images, and easily ignoring the global structural relationships in power scenarios, resulting in inaccurate foreign object detection results.

[0004] Although some existing methods attempt to introduce attention mechanisms or improve network structures, there are still deficiencies. For example, the attention mechanism design of some methods is not reasonable enough to adaptively emphasize key feature regions, leading to unsatisfactory detection results in cases of power line interlacing or complex background; other methods fail to fully consider the characteristics of power line scenarios in network structure improvement, making the model's performance still limited when facing diversification and complex environments. In view of the problems of insufficient balance between speed and accuracy, weak small target detection ability, and inaccurate recognition under occlusion and deformation in existing methods, the present application urgently needs to provide a power line foreign object detection method to overcome the above problems. SUMMARY

[0005] The technical problem to be solved by the present application is to provide a power line foreign matter detection method based on CE-YOLOv11, which has robustness and practicality in complex environments and small target detection tasks.

[0006] To solve the above technical problems, the technical solution of the present application is: a power line foreign matter detection method based on CE-YOLOv11, which is characterized in that: the power line foreign matter detection method based on CE-YOLOv11 comprises: acquiring an image data set covering multiple typical power line foreign matters, and dividing the image data set into a training set, a validation set and a test set in a ratio of 7:2:1 after data enhancement;

[0007] In the backbone network, a depth separable convolution and a general reverse bottleneck structure UIB are fused, and on the basis of a feature extraction module C3k2, a feature extraction network C3k2_UIB is constructed.

[0008] In the feature fusion stage, an attention-guided feature fusion module EC2PSA is constructed to fuse the semantic information in the detail branch with the context features.

[0009] The loss function powerful-iou is used to optimize the power line foreign matter detection positioning accuracy and small target detection performance.

[0010] Under the condition of fixed size input, a plurality of rounds of model training are performed, and adaptive optimization strategy is used to dynamically update the model parameters to obtain a detection model.

[0011] The detection model is used to detect the power line foreign matter.

[0012] Preferably, the image data set covers four types of foreign matters, including bird nests, kites, balloons and suspended garbage; and the data enhancement generates 6400 training sample pictures from the image data set.

[0013] Preferably, the depth separable convolution is a superposition combination of a depth convolution Depthwise Convolution and a point-by-point convolution Pointwise Convolution decomposed from a standard convolution operation.

[0014] The depth convolution Depthwise Convolution applies a local convolution kernel to each channel of the input data, which reduces the number of parameters while ensuring spatial feature extraction capability; and the point-by-point convolution Pointwise Convolution is calculated on all channels through a 1x1 convolution kernel, which realizes the fusion of cross-channel information and the adjustment of feature dimension.

[0015] Preferably, the universal reverse bottleneck structure UIB first increases the number of channels of the input feature map through the expansion layer to enhance the feature expression ability and expand the feature space; then uses the depth separable convolution to realize efficient spatial feature extraction in a high-dimensional space; then reduces the number of channels in the compression layer to reduce redundant information and control the calculation overhead; finally, the output feature map maintains the same number of channels as the input feature map, taking into account the needs of lightweight and efficient feature representation.

[0016] The feature extraction module C3k2 is a reference model YOLOv11, and the input features of the feature extraction module C3k2 include two parts, one part is directly transmitted through ordinary convolution operation, and the other part is subjected to deep feature extraction through several C3K or Bottleneck structures. Finally, the two parts of input features are spliced and fused through 1x1 convolution.

[0017] Preferably, the attention-guided feature fusion module EC2PSA uses the synergy of global context feature extraction PSABlock and local detail compensation EMA to realize efficient integration of multi-scale and multi-semantic features in the foreign object detection task of power transmission lines; the generated multi-scale features are divided into complementary detail branches and context branches after convolution processing;

[0018] The context branch uses global context feature extraction PSABlock to extract information in different feature subspaces in parallel through a multi-scale adaptive perception mechanism, improving the network's understanding ability of global context;

[0019] The detail branch uses local detail compensation EMA to retain fine-grained local feature information, making up for the deficiency of global context feature extraction PSABlock in small target detection, and enhancing the continuity and consistency of features in the spatial and channel dimensions.

[0020] Preferably, the local detail compensation EMA includes branch fusion and attention-guided dynamic weighting.

[0021] The branch fusion is to splice the two branches after processing by the global context feature extraction PSABlock and the local detail compensation EMA through the Concat function operation, and to realize the fusion of detail information and context information by using convolution to compress the channel dimension.

[0022] The attention-guided dynamic weighting inputs the features after branch fusion into an attention weight distribution mechanism to obtain dynamic weight coefficients through Sigmoid function activation.

[0023] Preferably, the loss function powerful-iou is:

[0024]

[0025] wherein L PIoU denotes the improved loss PloU,L IoU denotes the traditional loss IoU, P denotes the comprehensive position offset, denotes the boundary difference regular term;

[0026] wherein, , wherein L IoU denotes the traditional loss IoU, denotes the loss function IoU, denotes a prediction box, denotes a target box.

[0027] Preferably, the formula of the comprehensive position offset is:

[0028]

[0029] wherein P denotes the comprehensive position offset, denotes a target box width, denotes a target box height; denotes a horizontal distance between a left edge of a prediction box and a left edge of a target box; denotes a horizontal distance between a right edge of a prediction box and a right edge of a target box; denotes a vertical distance between an upper edge of a prediction box and an upper edge of a target box; denotes a vertical distance between a lower edge of a prediction box and a lower edge of a target box.

[0030] Preferably, the fixed size is: the input image size is 640*640, each model training period is 100 rounds, the Batch-size is 16, the SGD optimizer is used, the initial learning rate is 0.01, the weight decay coefficient is 0.0005, and the momentum is 0.937.

[0031] Preferably, the evaluation index of the detection model includes the precision P, the recall rate R, the average precision mean, the model parameter quantity, the number of floating point operations per second, and the model weight size.

[0032] Accuracy is the percentage of correctly predicted samples in the total prediction samples in the prediction result, and the formula is:

[0033]

[0034] wherein, denotes the accuracy, denotes the number of correctly predicted samples in the prediction result, denotes the number of incorrectly predicted samples in the prediction result;

[0035] Recall rate is the percentage of correctly predicted samples in the total samples, and the formula is:

[0036]

[0037] In the formula, represents the recall rate, represents the number of correctly predicted samples in the prediction results, represents the number of incorrectly predicted samples in the prediction results, represents the number of samples that cannot be predicted;

[0038] The average precision mean is an index that combines the accuracy P and the recall rate R to reflect the overall performance, and the formula is:

[0039]

[0040] In the formula, represents the average precision of the i-th class; i represents an integer of 1, 2,..., and i represents the average precision of the i-th class; represents the number of classes; represents the average precision mean of all classes;

[0041] The experimental environment for several rounds of training is as follows: Windows operating system, NVIDIA GeForce RTX 4060Ti graphics card, 16GB video memory, 12th Gen Intel Core i5-12490F processor, PyTorch 2.4.1 deep learning framework, and CUDA 12.4 GPU acceleration library. TM

[0042] The advantages of the present application are as follows: through the improvement of the network structure and the feature extraction mechanism, the processing speed is effectively improved while ensuring the detection accuracy, the feature extraction and recognition ability of small targets are enhanced, and the robustness and generalization ability of the model in complex environments are improved, thereby improving the accuracy and real-time performance of the foreign matter detection of the power transmission line, shortening the detection time, and providing more reliable technical support for the safe and stable operation of the power system. BRIEF DESCRIPTION OF DRAWINGS

[0043] The present application will be further described in detail below in combination with the drawings and specific embodiments.

[0044] Figure 1 is a flowchart of a power transmission line foreign matter detection method based on CE-YOLOv11.

[0045] Figure 2 ​is a schematic diagram of a feature extraction network C3k2_UIB in a power transmission line foreign matter detection method based on CE-YOLOv11.

[0046] Figure 3 is a schematic diagram of an attention-guided feature fusion module EC2PSA in a power transmission line foreign matter detection method based on CE-YOLOv11.

[0047] Figure 4 is a schematic diagram of a loss function powerful-iou in a power transmission line foreign matter detection method based on CE-YOLOv11.

[0048] Figure 5 is a schematic diagram of a feature extraction module C3k2 in a power transmission line foreign matter detection method based on CE-YOLOv11. DETAILED DESCRIPTION

[0049] The power transmission line foreign matter detection method based on CE-YOLOv11 includes: acquiring an image data set covering multiple typical power transmission line foreign matters, performing data enhancement through random rotation, brightness disturbance, blur processing and scale zooming to improve the robustness of the model, and dividing the training set, the validation set and the test set in proportion; in the backbone network, the depth separable convolution and the general reverse bottleneck structure UIB are fused, the feature extraction network C3k2_UIB is constructed based on the feature extraction module C3k2 to reduce the computational complexity and enhance the multi-scale expression ability; in the feature fusion stage, the attention-guided feature fusion module EC2PSA is constructed, the semantic information in the detail branch and the context features are adaptively fused to improve the feature capture ability of the model for small foreign matters on the surface of the power transmission line; the loss function powerful-iou is used to optimize the power transmission line foreign matter detection positioning accuracy and small target detection performance; a certain number of rounds of model training are performed under the condition of fixed size input, and the adaptive optimization strategy is used to dynamically update the model parameters to obtain a detection model; and the detection model is used to detect the foreign matters of the power transmission line.

[0050] Through the improvement of the network structure and the feature extraction mechanism, the processing speed is effectively improved while ensuring the detection accuracy, the feature extraction and recognition ability for small targets is enhanced, and the robustness and generalization ability of the model in complex environments are improved, thereby improving the accuracy and real-time performance of the power transmission line foreign matter detection, shortening the detection time, and providing more reliable technical support for the safe and stable operation of the power system.

[0051] The image dataset covers four categories of foreign objects, including bird nests, kites, balloons and hanging garbage; data augmentation generates 6400 training sample pictures from the image dataset, and divides the training sample pictures into training set, validation set and test set in the ratio of 7:2:1. The dataset contains 1600 pictures per category, including bird nests, kites, balloons and hanging garbage. Considering the complex and changeable lighting, climate and imaging conditions in the actual operation environment of the power transmission line, it is difficult to ensure ideal detection environment, and at the same time, in order to increase the number of samples and increase the feature information that the model can learn, the data augmentation method is used to expand the original data. Specifically, more diversified training samples are generated by randomly rotating, brightness disturbance, blurring and random scaling operations on the image.

[0052] The depth separable convolution is to decompose the standard convolution operation into the superposition combination of depthwise convolution and pointwise convolution. The depthwise convolution applies a local convolution kernel to each channel of the input data, which reduces the number of parameters while ensuring spatial feature extraction capability. The pointwise convolution is calculated through a 1x1 convolution kernel on all channels to realize cross-channel information fusion and feature dimension adjustment.

[0053] In the design of the backbone network, the depth separable convolution and the general reverse bottleneck UIB structure are fused, and a lightweight and efficient feature extraction network C3k2_UIB (as shown in Figure 5 ) is constructed based on the feature extraction module C3k2 (as shown in Figure 2 ). The network significantly reduces the number of parameters and computational complexity through depth separable convolution, improves the operation efficiency while maintaining the convolution expression ability; at the same time, the UIB structure can realize flexible balance between feature compression and channel expansion, and enhance cross-channel information interaction; combined with the multi-scale information aggregation advantage of the C3k2 module, the overall model effectively improves the multi-scale expression ability and detection precision of the features while reducing the computational complexity.

[0054] Depthwise separable convolution is an optimized convolution method that decomposes the standard convolution operation into depthwise convolution and pointwise convolution. Depthwise convolution applies a local convolution kernel to each channel of the input data separately, thereby significantly reducing the number of parameters while maintaining the ability to extract spatial features. Pointwise convolution is a calculation performed on all channels through a 1x1 convolution kernel, thereby achieving cross-channel information fusion and feature dimension adjustment. This process not only effectively integrates multi-channel information from the depthwise convolution output, but also enhances the overall feature expression ability while reducing redundancy and compensating for the lack of channel interaction in depthwise convolution. This combined operation significantly reduces computational complexity and model parameter size while maintaining feature extraction accuracy.

[0055] As shown in Figure 5 , a general reverse bottleneck UIB structure is introduced to further optimize the feature extraction process. The general reverse bottleneck UIB is: first, the channel number of the input feature map is increased through the expansion layer to enhance the feature expression ability and expand the feature space; then, depthwise separable convolution is used to efficiently extract spatial features in high-dimensional space; then, the channel number is reduced in the compression layer to reduce redundant information and control computational overhead; finally, the output feature map maintains the same channel number as the input feature map, thereby balancing the needs of lightweight and efficient feature representation. Through this "expansion-extraction-compression-output" design, the network is not only lightweight, but also retains sufficient feature expression ability, thereby effectively improving the model's ability to recognize multi-scale targets in power line foreign object detection.

[0056] The feature extraction module C3k2 is the baseline model YOLOv11. The input features of the feature extraction module C3k2 include two parts, one of which is directly transmitted through ordinary convolution operation, and the other is extracted through several C3K (when the C3K parameter is set to True) or Bottleneck structure. Finally, the two parts of the input features are spliced and fused through 1x1 convolution. The structural characteristics of the feature extraction module C3k2 are: the multi-scale convolution kernel C3K in the feature extraction module C3k2, K is the adjustable kernel size, such as 3x3, 5x5, etc. This design can expand the receptive field, enabling the model to capture more extensive contextual information. Although large convolution kernels improve large target detection capabilities, using large convolution kernels not only may lead to overfitting or computational redundancy, but also has no obvious gain in extracting local details of small targets (<32x32 pixels), and even causes feature blurring due to a too large receptive field. The feature extraction network C3k2_UIB is to replace the C3k module of the feature extraction module C3k2 with a UIB module (as shown in Figure 2 ).

[0057] The high-efficiency feature expression capability of the depth separable convolution and the general reverse bottleneck UIB is combined, and is embedded into the feature extraction module C3k2 to form a feature extraction network C3k2_UIB backbone network structure, so that a better speed and accuracy balance is realized in the foreign matter detection task of the power transmission line.

[0058] The attention-guided feature fusion module EC2PSA utilizes the synergy of the global context feature extraction PSA Block and the local detail compensation EMA to realize efficient integration of multi-scale and multi-semantic features in the foreign matter detection task of the power transmission line. The generated multi-scale features are divided into complementary detail branches and context branches after convolution processing. The context branch utilizes the global context feature extraction PSA Block to extract information in different feature subspaces in parallel through a multi-scale adaptive perception mechanism, thereby improving the network's understanding ability of the global context. The detail branch utilizes the local detail compensation EMA to retain fine-grained local feature information, make up for the deficiency of the global context feature extraction PSA Block in small target detection, and enhance the continuity and consistency of the features in the spatial and channel dimensions.

[0059] The local detail compensation EMA includes branch fusion and attention-guided dynamic weighting. The branch fusion is to splice the two branches after processing by the global context feature extraction PSA Block and the local detail compensation EMA through the Concat function operation, and to compress the channel dimension by convolution to realize the fusion of the detail information and the context information. The attention-guided dynamic weighting inputs the features after the branch fusion into an attention weight distribution mechanism to obtain dynamic weight coefficients through the Sigmoid function activation.

[0060] Specifically, the feature fusion module EC2PSA (as shown in Figure 3 ) realizes efficient integration of multi-scale and multi-semantic features in the foreign matter detection task of the power transmission line through the synergy mechanism of the "global context feature extraction PSA Block + local detail compensation EMA". This module maintains strong sensitivity to small targets in complex backgrounds, significantly improving the model's ability to capture fine-grained foreign matters and detection accuracy.

[0061] In the feature fusion stage, the attention-guided feature fusion module EC2PSA (Efficient Context-aware Channel Partitioned Self-Attention) is used to realize the fusion of the detail branch and the context branch, which includes the following processes: 1. Feature branch splitting (Split): The input multi-scale features are first processed by convolution and then divided into two complementary branches: 2. Detail branch: retaining local texture, edge information and spatial fine-grained features, focusing on capturing small and medium-sized targets and local structures in foreign objects on the power line. 3. Context branch: emphasizing global scene and long-range dependence, retaining overall spatial distribution and semantic information, used to perceive the relative relationship between foreign objects and power lines and background.

[0062] Global semantic enhancement PSABlock: In the context branch, the global semantic enhancement PSABlock is introduced, which extracts information in different feature subspaces in parallel through a multi-scale adaptive perception mechanism, improving the network's understanding of global context, especially in complex backgrounds.

[0063] Local detail compensation EMA: In the detail branch, EMA (Efficient Multi-scale Attention) is introduced to enhance the continuity and consistency of features in both spatial and channel dimensions. Local detail compensation EMA retains fine-grained local feature information, making up for the shortcomings of global semantic enhancement PSABlock in small target detection. Local detail compensation EMA can maintain the clarity of edges and textures when capturing small-scale foreign objects such as power line surface attachments.

[0064] Branch fusion (Concat + Conv): After processing by global semantic enhancement PSABlock and local detail compensation EMA, the two branches are concatenated through Concat operation and the channel dimension is compressed using convolution, thereby realizing the fusion of detail information and context information.

[0065] Attention-guided dynamic weighting (Sigmoid + Re-weight): The fused features are input into the attention weight distribution mechanism, and the dynamic weight coefficients are obtained by activating the Sigmoid function. Different scales and branches of features are adaptively weighted; key areas and key information (such as foreign objects on the surface of the power line) are strengthened, while redundant and irrelevant features (such as background noise) are suppressed.

[0066] Output enhanced features: The features output by the final feature fusion module EC2PSA have both global context information and local detail features, outputting the results of small foreign object features on the surface of the power line, maintaining accurate understanding of the overall power transmission environment, and thus providing better input features for the subsequent detection head.

[0067] To improve the positioning accuracy and small target detection performance in the task of foreign matter detection of power transmission lines, a Powerful-IoU (PIoU) loss function (as shown in Figure 4 ) is introduced based on the traditional IoU loss. The loss function powerful-iou is used to optimize the positioning accuracy and small target detection performance of foreign matter detection of power transmission lines. The loss function measures the relative difference between the two boundary positions while calculating the overlap between the predicted box and the target box, thereby achieving more accurate regression constraints.

[0068] Based on the traditional IoU loss , a boundary offset term is introduced to construct the PloU loss. The loss function powerful-iou is:

[0069]

[0070] In the formula, L PIoU represents the improved loss PloU, L IoU represents the traditional loss IoU, and P represents the comprehensive position offset. represents the boundary difference regularization term, which is used to punish the misalignment of the predicted box edge and the target box, and is more tolerant to small offset and more severe to large offset.

[0071] wherein, is used to constrain the overlap between the predicted box and the target box. In the formula, L IoU represents the traditional loss IoU, represents the loss function IoU, represents the predicted box (anchor bbox), and the target box (target bbox).

[0072] The boundary difference between the predicted box and the target box in the horizontal and vertical directions is calculated, and the comprehensive position offset P is obtained after normalization. The formula of the comprehensive position offset P is:

[0073]

[0074] In the formula, P represents the comprehensive position offset, represents the target box width, represents the target box height; represents the horizontal distance between the left edge of the predicted box and the left edge of the target box; represents the horizontal distance between the right edge of the predicted box and the right edge of the target box; represents the vertical distance between the upper edge of the predicted box and the upper edge of the target box; represents the vertical distance between the lower edge of the prediction box and the lower edge of the target box.

[0075] This indicator describes the relative distance of the four boundaries of the prediction box and the target box, which can comprehensively depict the positioning error.

[0076] The improved loss PIoU significantly accelerates the convergence speed and reduces the "drifting box" phenomenon by providing a more explicit convergence direction for the prediction box through the boundary difference term. At the same time, it not only relies on the overlap area, but also directly constrains the alignment degree of the four edges with the target box, thereby greatly improving the boundary fitting degree. In addition, the improved loss PIoU has excellent adaptability to small target detection, especially for small volume but position difference sensitive foreign objects (such as kites, balloons), it can effectively improve the positioning accuracy.

[0077] Under the condition of fixed size input, the model is trained for several rounds, and the adaptive optimization strategy is used to dynamically update the model parameters to obtain a high-precision and strong-robustness final detection model. After the training of the detection model is completed, the precision and lightweight performance of the final detection model are evaluated on the test set. Several rounds of training are carried out under fixed size input, and the fixed size is: the input image size is 640x640, each model training period is 100 rounds, Batch-size is 16, SGD optimizer is used, initial learning rate is 0.01, weight decay coefficient is 0.0005, and momentum is 0.937. The experimental environment for several rounds of training is: Windows operating system, NVIDIA GeForce RTX 4060Ti graphics card with 16GB video memory, processor is 12th Gen Intel Core i5-12490F, deep learning framework is PyTorch 2.4.1, and GPU acceleration library is CUDA 12.4. TM

[0078] The detection model is used to detect foreign objects on the transmission line. The detection model of the application uses precision P (Precision), recall R (Recall), mean average precision mAP (mean Average Precision), model parameter quantity (Params), floating point operation number per second GFLOPs (GFLOPs), and model weight size as the evaluation indicators of the model. The smaller the model parameter quantity (Params) and the floating point operation number per second GFLOPs (GFLOPs), the smaller the computing power required by the model, the lower the performance requirement of the hardware, and the easier the deployment in edge devices. The larger the precision value represents the more accurate the model detection, and the larger the recall value represents the lower the model miss rate.

[0079] Accuracy is the percentage of correctly predicted samples in the total predicted samples in the prediction result, and the formula is:​

[0080]

[0081] wherein, P represents the accuracy rate, TP represents the number of correctly predicted samples in the prediction result, FN represents the number of incorrectly predicted samples in the prediction result.

[0082] Recall is the percentage of correctly predicted samples in the prediction result in the total samples, and the formula is:

[0083]

[0084] wherein, R represents the recall rate, TP represents the number of correctly predicted samples in the prediction result, FN represents the number of incorrectly predicted samples in the prediction result, FN represents the number of samples that cannot be predicted.

[0085] The average precision mean is an index reflecting the global performance by combining the accuracy rate P and the recall rate R, and the formula is:

[0086]

[0087] wherein, P represents the average precision of the i-th class; i represents an integer of 1, 2, …; C represents the number of classes; P represents the average precision of all classes.

[0088] Ablation experiments were conducted in conjunction with the present invention, as shown in Table 1, which includes five sets of algorithms (baseline model YOLOv11, model YOLOv11+C3k2_UIB, model YOLOv11+EMA, model YOLOv11+C3k2_UIB+EMA, and the present invention model YOLOv11+C3k2_UIB+EMA+PIoU). In Table 1, P represents precision, the proportion of truly positive samples among those predicted as positive by the model, measuring accuracy; R represents recall, the proportion of all truly positive samples correctly predicted by the model, measuring recall rate; mAP50 represents the mean precision at an IoU threshold of 0.5, which is the most important core indicator in the COCO evaluation criteria; mAP50-90 represents the mean precision from an IoU threshold of 0.5 to 0.95 (step size 0.05), which is a more rigorous and comprehensive evaluation indicator; Parameters represents the total number of parameters in the model, directly measuring the size and complexity of the model; GFLOPs represents the billion floating-point operations required for one forward inference, measuring the computational complexity and speed of the model.

[0089] Table 1. Ablation Experiment Comparison Table

[0090] Algorithm P R mAP50 mAP50-90 Parameters GFLOPs YOLOv11 0.896 0.834 0.913 0.666 2,582,932 6.3 YOLOv11+C3k2_UIB 0.905 0.853 0.914 0.654 2,646,460 7.9 YOLOv11+EMA 0.904 0.828 0.907 0.661 2,593,300 6.4 YOLOv11+ C3k2_UIB +EMA 0.923 0.823 0.921 0.661 2,554,428 7.6 YOLOv11+ C3k2_UIB +EMA+PIoU 0.925 0.859 0.923 0.672 2,554,428 7.6

[0091] In ablation experiments, the baseline model YOLOv11 achieved high detection accuracy while maintaining low computational cost. Introducing the C3k2_UIB module improved both precision and recall, indicating that this module effectively enhances feature extraction capabilities, but also increases the number of parameters and computational complexity. Further addition of the EMA attention mechanism, while not significantly changing the number of parameters, made it more sensitive to small target detection, resulting in improved mAP@50-90. Combining C3k2_UIB with EMA significantly improved the model's precision, demonstrating that the fusion of global context and local detail features can enhance the characterization of foreign objects on transmission lines.

[0092] Finally, by introducing the PioU loss function, the model achieved optimal performance in key metrics such as precision, recall, and mAP, while maintaining a good parameter size and computational efficiency. Overall, the synergistic optimization of C3k2_UIB, EMA, and PioU in this invention not only improves the balance between speed and accuracy but also significantly enhances the model's robustness and practicality in complex environments and small object detection tasks.

[0093] The specific embodiments described herein are merely illustrative of the spirit of the application. Various modifications or changes in the specific embodiments described herein can occur to those skilled in the art to which the application pertains without departing from the spirit of the application, and it is understood that such modifications or changes are to be considered as within the scope of the application as defined by the appended claims.

Claims

1. A method for detecting foreign objects in power transmission lines based on CE-YOLOv11, characterized in that... include: Obtain an image dataset covering multiple types of typical foreign objects in power transmission lines. After data augmentation, divide the image dataset into training, validation, and test sets in a 7:2:1 ratio. In the backbone network, depthwise separable convolution and general inverse bottleneck structure UIB are integrated, and a feature extraction network C3k2_UIB is constructed based on the feature extraction module C3k2. In the feature fusion stage, an attention-guided feature fusion module EC2PSA is constructed to fuse semantic information from detail branches with contextual features; The powerful-iou loss function is used to optimize the positioning accuracy of foreign object detection and small target detection performance in power transmission lines. The model is trained for several rounds under fixed input conditions, and the model parameters are dynamically updated using an adaptive optimization strategy to obtain the detection model. The aforementioned detection model is used to detect foreign objects in power transmission lines.

2. The method for detecting foreign objects in power transmission lines based on CE-YOLOv11 as described in claim 1, characterized in that: The image dataset covers four categories of foreign objects, including bird nests, kites, balloons, and hanging trash. Data augmentation generates 6,400 training sample images from the image dataset.

3. The method for detecting foreign objects in power transmission lines based on CE-YOLOv11 as described in claim 1, characterized in that: The depthwise separable convolution is a combination of decomposing the standard convolution operation into a superposition of depthwise convolution and pointwise convolution. The depthwise convolution applies a local convolution kernel to each channel of the input data, reducing the number of parameters while ensuring spatial feature extraction capabilities; the pointwise convolution performs calculations on all channels using a 1×1 convolution kernel.

4. The method for detecting foreign objects in power transmission lines based on CE-YOLOv11 as described in claim 1, characterized in that: The general inverse bottleneck structure UIB first increases the number of channels in the input feature map through an expansion layer; then, it uses the depthwise separable convolution to achieve efficient spatial feature extraction in high-dimensional space; next, it reduces the number of channels in a compression layer; finally, the output feature map maintains a number of channels that matches the input feature map. The feature extraction module C3k2 is based on the YOLOv11 model. The input features of the feature extraction module C3k2 consist of two parts. One part is directly passed through ordinary convolution operations, and the other part is extracted through deep features by several C3K or Bottleneck structures. Finally, the two parts of input features are concatenated and fused through 1x1 convolution.

5. The method for detecting foreign objects in power transmission lines based on CE-YOLOv11 as described in claim 1, characterized in that: The attention-guided feature fusion module EC2PSA utilizes the synergy of global context feature extraction PSABlock and local detail compensation EMA; the generated multi-scale features are divided into complementary detail branches and context branches after convolution processing; The context branch extracts PSABlock using global context features and extracts information in parallel in different feature subspaces through a multi-scale adaptive perception mechanism. The detail branch utilizes Local Detail Compensation (EMA) to preserve fine-grained local feature information, thus compensating for the shortcomings of Global Context Feature Extraction (PSABlock) in small object detection.

6. The method for detecting foreign objects in power transmission lines based on CE-YOLOv11 as described in claim 5, characterized in that: The Local Detail Compensation (EMA) includes dynamic weighting of branch fusion and attention-guided calculations; The branch fusion is achieved by concatenating two branches after global context feature extraction PSABlock and local detail compensation EMA, and then using the Concat function operation to merge the detail information and context information by using convolution to compress the channel dimension. The attention-guided dynamic weighting inputs the branch-fused features into the attention weight allocation mechanism, and obtains dynamic weight coefficients through activation by the Sigmoid function.

7. The method for detecting foreign objects in power transmission lines based on CE-YOLOv11 as described in claim 1, characterized in that: The loss function powerful-iou is: In the formula, L PIoU Indicates the loss after improvement, PloU, L IoU This represents the traditional loss IoU, and P represents the overall position offset. Indicates the boundary difference regularization term; in, In the formula, L IoU Indicates traditional loss IoU, Represents the loss function IoU. This indicates setting a prediction box. Represents the target bounding box.

8. The method for detecting foreign objects in power transmission lines based on CE-YOLOv11 as described in claim 7, characterized in that: The formula for the comprehensive position offset is: In the formula, P represents the overall position offset. Indicates the width of the target box. Indicates the height of the target bounding box; This represents the horizontal distance between the left edge of the predicted bounding box and the left edge of the target bounding box; This represents the horizontal distance between the right edge of the predicted bounding box and the right edge of the target bounding box; This represents the vertical distance between the top edge of the predicted bounding box and the top edge of the target bounding box; This represents the vertical distance between the bottom edge of the predicted bounding box and the bottom edge of the target bounding box.

9. The method for detecting foreign objects in power transmission lines based on CE-YOLOv11 as described in claim 1, characterized in that: The fixed dimensions are as follows: the input image size is 640×640, each model training cycle is 100 rounds, the batch size is 16, the SGD optimizer is used, the initial learning rate is 0.01, the weight decay coefficient is 0.0005, and the momentum is 0.

937.

10. The method for detecting foreign objects in power transmission lines based on CE-YOLOv11 as described in claim 1, characterized in that: The evaluation metrics for the detection model include precision P, recall R, mean precision, number of model parameters, number of floating-point operations per second (billions) and model weights. accuracy It is the percentage of correctly predicted samples out of the total predicted samples in the prediction results, and the formula is: In the formula, Indicates accuracy rate. This indicates the number of correctly predicted samples in the prediction results. This indicates the number of incorrectly predicted samples in the prediction results; Recall rate It is the percentage of correctly predicted samples out of the total sample in the prediction results, and the formula is: In the formula, Indicates recall rate, This indicates the number of correctly predicted samples in the prediction results. This indicates the number of incorrectly predicted samples in the prediction results. This indicates samples that could not be predicted. Mean precision is a metric that combines accuracy (P) and recall (R) to reflect overall performance. The formula is: In the formula, Indicates the first The average precision of a single class; i represents an integer from 1, 2, ...; Indicates the number of categories; This represents the mean precision across all categories. The experimental environment for several rounds of training was as follows: Windows operating system, NVIDIA GeForce RTX4060Ti graphics card with 16GB of video memory, and 12th Gen Intel® Core processor. TM The processor is an i5-12490F, the deep learning framework is PyTorch 2.4.1, and the GPU acceleration library is CUDA 12.4.

Images