A power transmission line insulator defect detection method based on LDC-YOLOV11

Abstract

The application discloses the technical field of deep learning target detection and insulator detection, and particularly relates to a power transmission line insulator defect detection method based on LDC-YOLOV11, which acquires power transmission line insulator defect detection image data sets and divides them into a training set, a verification set and a test set; a lightweight dynamic feature extraction module LDFE is designed in the main network part, which extracts the global context features of insulator defects under complex inspection backgrounds while reducing model parameters; a channel compensation up-sampling module CUC is designed in the neck network part, which improves the up-sampling reconstruction quality of low-resolution feature maps in the feature pyramid with a small parameter increment; a coordinate-aware lightweight interactive detection head LAI-Detect lightweight detection head is designed, which enhances the positioning and discrimination ability of the lightweight model for weak texture and tiny defect targets; based on this, the application enhances the ability of the model to extract insulator multi-scale defect features in the inspection image, so that the model can realize lightweight deployment on the basis of ensuring detection accuracy.

Description

Technical Field

[0001] This invention relates to the field of deep learning target detection and insulator detection technology, specifically a method for detecting defects in transmission line insulators based on LDC-YOLOV11. Background Technology

[0002] Transmission lines bear the crucial responsibility of efficiently and stably transmitting electrical energy generated by power plants to various power-consuming areas. Insulators, as key components in transmission lines providing electrical insulation and mechanical support, directly affect the safety and stability of the entire power system. Exposed to complex outdoor environments for extended periods, insulators are highly susceptible to lightning strikes, pollution buildup, mechanical stress, and extreme weather conditions, leading to defects such as skirt damage and flashover burns.

[0003] With the continuous expansion of transmission line scale and the increasingly complex operating environment, insulator defect detection faces severe challenges. On the one hand, insulator strings exhibit multi-scale, elongated distribution characteristics in aerial images, while flashover points or minor damage targets occupy only a small pixel area, representing a typical challenge for small target detection. On the other hand, the background of transmission line corridors is complex, and defect features are highly similar to background noise, easily leading to missed or false detections. Traditional insulator inspection mainly relies on manual tower climbing or telescope inspection, but transmission lines typically span vast geographical areas, and manual inspection has inherent drawbacks such as low efficiency, high risk, and strong subjectivity.

[0004] In recent years, the widespread adoption of drone technology has provided a new technical means for power transmission line inspection. Equipped with high-resolution visible light cameras, drones can acquire aerial images of insulators at close range and from multiple angles, significantly improving the flexibility and coverage of inspections. However, relying on manual interpretation of the massive amounts of image data collected by drones is not only time-consuming and labor-intensive but also fails to meet real-time requirements. Furthermore, the insulator defects in aerial images are often small, with complex backgrounds and dramatic scale variations, making accurate and reliable automatic detection difficult with traditional image processing methods. With the rapid development of deep learning technology, target detection methods based on convolutional neural networks (CNNs), recurrent neural networks (RNNs), and Transformer architectures have provided new solutions for automatic insulator defect detection. These methods can automatically learn deep feature representations of defects from a large number of aerial images, effectively overcoming the limitations of traditional methods that rely on manually designed features and have poor generalization capabilities.

[0005] Methods based on convolutional neural networks (CNNs), such as the classic Faster R-CNN, SSD, and the early YOLO series, utilize hierarchical convolutional operations to construct feature pyramids, enabling automatic feature representation of insulator location and macroscopic defects such as large-area damage and missing parts in aerial images. This effectively overcomes the limitation of poor generalization ability of traditional manually designed features.

[0006] Methods based on recurrent neural networks (RNNs) are often used to help determine the coherence and correlation of defective regions in inspection videos by introducing contextual semantics or temporal correlation information.

[0007] The Transformer-based approach utilizes the Self-Attention mechanism to establish global dependencies between long-distance pixels in an image, demonstrating stronger modeling capabilities when capturing global contextual information and large defect features.

[0008] However, existing deep learning-based methods still face many challenges when applied to insulator defect detection: First, most general target detection models are not good at extracting features of small defects, and these small defects are easily drowned out by background noise from various environments during feature extraction; Second, the scale difference between the insulator body and the defect is large, and feature fusion networks cannot effectively take into account multi-scale features, and edge detail features are easily lost; Third, complex natural backgrounds and weather interference can easily cause feature confusion and image quality degradation, making it difficult to identify and locate weak textures and small defects, and detection accuracy still has significant shortcomings.

[0009] Therefore, there is an urgent need to design a target detection model specifically for detecting defects in insulators of transmission lines, to solve the problems of severe background noise interference in complex inspection environments, easy loss of edge detail features in the multi-scale feature upsampling stage, and difficulty in accurately identifying and locating weak texture and small defect targets, while taking into account the requirements of detection accuracy and lightweight deployment of the model.

[0010] While single-stage detection methods have achieved some success in insulator defect detection research, they also face challenges such as extremely small defect targets and severe background noise interference. To address these issues, researchers have proposed several improvements. For example, some scholars have proposed single-stage algorithms optimized based on attention mechanisms, introducing collaborative attention or deformable convolutions into the YOLO network to reduce prediction bias caused by small and dense targets, thus improving the algorithm's performance in complex backgrounds. On the other hand, other scholars have enhanced the model's ability to extract local defect features of insulators by designing specific multi-scale feature fusion networks. However, these algorithms still cannot solve the problems of severe background noise interference in complex inspection environments, the easy loss of edge detail features during multi-scale feature upsampling, and the difficulty in accurately identifying and locating weak textures and small defect targets, while also balancing the model's detection accuracy with lightweight deployment requirements.

[0011] Based on the above, a method for detecting defects in transmission line insulators based on LDC-YOLOV11 is invented. Summary of the Invention

[0012] In view of the problems existing in the above and / or the existing method for detecting defects in transmission line insulators based on LDC-YOLOV11, the present invention is proposed.

[0013] Therefore, the purpose of this invention is to provide a defect detection method for transmission line insulators based on LDC-YOLOV11. This method aims to reduce the number of model parameters and computational load, while addressing to some extent the problems of severe background noise interference in complex inspection environments, easy loss of edge detail features during multi-scale feature upsampling, and difficulty in accurately identifying and locating weak textures and small defects. It also takes into account both the detection accuracy of the model and the need for lightweight deployment.

[0014] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0015] A method for detecting defects in transmission line insulators based on LDC-YOLOV11 includes the following steps:

[0016] S1: Obtain the image dataset of defect detection of insulators of transmission lines. After image enhancement preprocessing such as motion jitter blurring, rain, and snowfall, a high-quality dataset is obtained. Then, it is divided into training set, validation set and test set according to a 7:2:1 ratio.

[0017] S2: In the backbone network, a lightweight dynamic feature extraction module LDFE is designed to expand the global receptive field to suppress complex background noise in the inspection image;

[0018] S3: In the neck network, a channel-compensated upsampling module (CUC) is designed to improve the restoration quality of transmission line insulator defect detection images during the upsampling process;

[0019] S4: Design a coordinate-aware lightweight interactive detection head LAI-Detect to enhance the lightweight model's ability to locate and distinguish targets with weak textures and minor defects;

[0020] S5: Input the training set images from the dataset into the LDC-YOLOV11 model for iterative training until the model loss function converges, and save the training weights that are optimal for multi-scale feature extraction and localization.

[0021] S6: Evaluate the effectiveness of the model in insulator defect detection, analyze the differences between the LDC-YOLOV11 model and mainstream target detection models in multiple indicators, and verify the superiority of the algorithm in terms of accuracy, lightweight and generalization.

[0022] As a preferred embodiment of the LDC-YOLOV11-based method for detecting defects in transmission line insulators according to the present invention, the specific steps of S1 are as follows:

[0023] S11: Obtain the China Power Line Insulator Dataset CPLID, which contains 1600 images. In addition, 400 images were collected through actual shooting and public network collection, which together constitute a high-quality dataset containing 2000 images.

[0024] S12, apply motion blur, rain, and snowfall image enhancement preprocessing to the image;

[0025] S13. The preprocessed images are divided into a training set of 1400 images, a validation set of 400 images, and a test set of 200 images for subsequent experiments.

[0026] As a preferred embodiment of the transmission line insulator defect detection method based on LDC-YOLOV11 described in this invention, in step S2, the lightweight dynamic feature extraction module LDFE consists of an input CBS unit, a Split unit, three cascaded GatedLSConvBlock units, a Concat unit, and an output CBS unit.

[0027] As a preferred embodiment of the LDC-YOLOV11-based method for detecting defects in transmission line insulators according to the present invention, the operation steps of the lightweight dynamic feature extraction module LDFE are as follows:

[0028] Within a single GatedLSConvBlock, for the input feature tensor Normalization is performed, where For batch size, For the number of channels, For feature map height, The width of the feature map;

[0029] Expand the number of channels using a point convolution;

[0030] The transformed features are divided into three parts along the channel dimension according to a specific ratio: gated features g, identity features i, and features to be convolved d.

[0031] As a preferred embodiment of the LDC-YOLOV11-based method for detecting defects in transmission line insulators according to the present invention, the specific steps of S3 are as follows:

[0032] S31, for a given input feature map Where C is the number of channels, H and W are the height and width, respectively. The channel compensation upsampling module CUC upsamples it by a factor of 2 to obtain... ;

[0033] S32, using depthwise separable convolution pairs for the magnified image. Perform feature extraction;

[0034] S33, the channel compensation upsampling module CUC uses the channel rearrangement operation AC to regroup and transpose the feature maps in the channel dimension, so that the features of different groups can be fully mixed;

[0035] S34 performs a ShiftChannelMix operation on the features after channel rearrangement. By combining channel segmentation with multi-directional cyclic shifting, different channels can perceive information from different spatial locations in the feature map.

[0036] As a preferred embodiment of the LDC-YOLOV11-based method for detecting defects in transmission line insulators according to the present invention, the specific steps of S4 are as follows:

[0037] S41, The input to LAI-Detect is a list of feature maps from different layers of the backbone network. These correspond to three detection scales;

[0038] S42, for the first Input feature maps at each scale ,in, The initial feature transformation is performed through independent 3×3 convolutional layers to adapt to the detection requirements of different scales;

[0039] S43, after transformation, features at all scales are extracted and channel dimensions are adjusted using depthwise separable convolutions and 1×1 convolutions to obtain the output. This enables networks to learn general feature representations across scales and tasks;

[0040] S44, Two independent prediction branches are fed in parallel: one is the bounding box coordinate distribution prediction branch, and the other is the classification confidence calibration branch that integrates regression distribution information.

[0041] S45: Concatenate the prediction results from the two branches to generate the final output features of the detection head. .

[0042] As a preferred embodiment of the LDC-YOLOV11-based method for detecting defects in transmission line insulators according to the present invention, the specific steps of S5 are as follows:

[0043] S51, input the training set images into the LDC-YOLOV11 model equipped with LDFE, CUC, and LAI-Detect modules for iterative training;

[0044] S52 uses a validation set to monitor model loss in real time during training. After the loss function converges, the training effect is quantitatively evaluated using precision, recall, and average precision metrics. The optimal training weights for multi-scale feature extraction and localization are saved.

[0045] As a preferred embodiment of the LDC-YOLOV11-based method for detecting defects in transmission line insulators according to the present invention, in step S6, one or more models selected from FasterNet, Mobilenetv4, EfficientViT3, EfficientFormerv2, and LSNet are used as feature extraction networks for the target detection model, and their performance is compared under the same test environment and dataset.

[0046] Compared with existing technologies:

[0047] 1. Through the gating dynamic feature regulation and big-small kernel collaborative mechanism of the LDFE module, it can realize noise adaptive suppression and global context awareness in complex inspection background, effectively enhance the ability to extract small defect features, and avoid small target defect features being submerged by surrounding redundant background pixels.

[0048] 2. By using the CUC channel compensation upsampling module's channel rearrangement and multi-directional cyclic shift operations, it can achieve high-fidelity restoration of edge details and effective fusion of multi-scale features during the upsampling process, significantly improving the quality of feature map restoration and suppressing the loss of edge detail information in the feature fusion stage;

[0049] 3. Through the coordinate perception interaction mechanism and regression distribution information calibration of the LAI-Detect detection head, it can achieve synergistic optimization of classification confidence and localization accuracy, enhance the ability of lightweight models to distinguish and locate targets with weak texture and small defects, and reduce the risk of missed detection and false detection;

[0050] 4. Through the lightweight collaborative design of LDFE, CUC and LAI-Detect, the number of model parameters and computational complexity are effectively reduced. While maintaining high detection accuracy, it meets the requirements of lightweight deployment at the edge, achieving the best balance between detection accuracy and inference efficiency. Attached Figure Description

[0051] Figure 1 This is a flowchart of the method steps of the present invention;

[0052] Figure 2 This is a schematic diagram of the LDC-YOLOV11 structure of the present invention;

[0053] Figure 3 This is a schematic diagram of the backbone network module (LDFE) structure of the present invention;

[0054] Figure 4 This is a schematic diagram of the GatedLSConvBlock structure of the present invention;

[0055] Figure 5 This is a schematic diagram of the Channel Compensation Upsampling Module (CUC) of the present invention;

[0056] Figure 6 This is a schematic diagram of the coordinate-aware lightweight interactive detection head (LAI-Detect) structure of the present invention;

[0057] Figure 7 This is a comparison chart of the prediction results of the present invention;

[0058] Figure 8 This is a visualization of the receptive field changes in this invention;

[0059] Figure 9 The results are the thermal diagram analysis results of this invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0061] This invention provides a method for detecting defects in transmission line insulators based on LDC-YOLOV11. Please refer to [link / reference]. Figures 1-9 This includes the following steps:

[0062] S1: Obtain the image dataset for defect detection of insulators of transmission lines. After preprocessing with motion jitter blurring, rain, and snow image enhancement, a high-quality dataset is obtained. Then, it is divided into training set, validation set, and test set in a 7:2:1 ratio.

[0063] Specifically, the dataset consists of 1600 images from the China Power Line Insulator Dataset (CPLID) and 400 images collected from actual photographs and public networks, forming a high-quality dataset of 2000 images. The dataset is then preprocessed using the Imgaug image enhancement library and divided into a training set of 1400 images, a validation set of 400 images, and a test set of 200 images for subsequent experiments.

[0064] S2: In the backbone network, a lightweight dynamic feature extraction module LDFE is designed to expand the global receptive field to suppress complex background noise in the inspection image; the lightweight dynamic feature extraction module LDFE consists of an input CBS unit, a Split unit, three cascaded GatedLSConvBlock units, a Concat unit, and an output CBS unit.

[0065] The Lightweight Dynamic Feature Extraction (LDFE) module's core is the GatedLSConvBlock unit, an improvement upon LSConv. LDFE reduces computational overhead while utilizing a gating mechanism for adaptive dynamic feature control across channels, effectively suppressing background noise. Simultaneously, the introduced big.mesh kernel collaboration mechanism enhances the model's receptive field, improving its understanding of global contextual information. LDFE's adaptive capability allows the model to more accurately focus on key target pixels, preventing small target defect features from being diluted by surrounding redundant background pixels during standard sampling, thus ensuring the ability to extract features from small target defects.

[0066] Specifically, within a single GatedLSConvBlock, for the input feature tensor First, layer normalization is performed, where For batch size, For the number of channels, For feature map height, The feature map width is then expanded by a point convolution to increase the number of channels. To improve computational efficiency and preserve some of the original feature flow, the transformed features are divided into three parts along the channel dimension according to a specific ratio: gated features g, identity features i, and features to be convolved d. For the segmented feature branches, GatedLSConvBlock feeds them into the LSConv module for deep spatial feature extraction. LSConv's processing is divided into two stages: large kernel perceptron and small kernel aggregation. LSConv simulates operators in the human visual system, dynamically generating convolutional kernels through a large kernel perceptron (LKP) and using a small kernel aggregator (SKA) for adaptive feature fusion. This allows for the dynamic generation of large-size convolutional kernels based on input features, achieving improved dynamic feature extraction and context-adaptive aggregation capabilities of the model with relatively low computational cost.

[0067] S3: In the neck network, a channel-compensated upsampling module (CUC) is designed to improve the restoration quality of transmission line insulator defect detection images during the upsampling process;

[0068] The core design idea of ​​the Channel-Compensation Upsampling Convolution (CUC) module is to achieve multi-level feature interaction under zero parameter increase conditions. It combines channel shifting and channel mixing to further enhance the flow of information between channels, achieving adaptive fusion of multi-directional contextual information. Ultimately, it significantly enhances the ability to express the features of small target defects while maintaining parameter efficiency. Specifically, for a given input feature map... CUC upsampled it by 2 times and obtained Using depthwise separable convolution pairs to enlarge the image. For feature extraction, depthwise separable convolution decomposes standard convolution into two independent steps: depthwise convolution and pointwise convolution, significantly reducing the number of parameters and computational complexity. Subsequently, the module employs the channel rearrangement operation AC (AdjustChannel) to regroup and transpose the feature maps along the channel dimension, allowing for thorough mixing of features from different groups. CUC then performs a ShiftChannelMix operation on the rearranged features. Through the combination of channel segmentation and multi-directional cyclic shift operations, different channels can perceive information from different spatial locations in the feature maps, effectively expanding the receptive field of the features, enhancing the model's ability to capture local details, and achieving zero-parameter feature enhancement.

[0069] S4: Design a coordinate-aware lightweight interactive detection head LAI-Detect to enhance the lightweight model's ability to locate and distinguish targets with weak textures and minor defects;

[0070] The coordinate-aware, lightweight interactive detection head LAI-Detect takes a list of feature maps from different layers of the backbone network as input. This corresponds to three detection scales. Specifically, for the first... Input feature maps at each scale ,in, First, initial feature transformation is performed using independent 3×3 convolutional layers to adapt to detection requirements at different scales. After transformation, features at all scales are then extracted and channel dimension adjusted using depthwise separable convolutions and 1×1 convolutions to obtain the output. This enables the network to learn general feature representations across scales and tasks. To achieve information exchange between the classification and regression branches, the module executes two branches in parallel: a bounding box coordinate distribution prediction branch and a classification confidence prediction branch that compares the regression distribution information with the actual classification confidence prediction. Finally, the bounding box coordinate distribution prediction and the classification confidence prediction that compares the regression distribution information with the actual classification confidence prediction are concatenated to obtain the output of the detection head. LAI-Detect correlates classification confidence with coordinate information. Predictions with accurate localization and concentrated distribution receive a positive bonus, while predictions with ambiguous localization may receive a negative adjustment or be suppressed.

[0071] S5: Input the training set images from the dataset into the LDC-YOLOV11 model for iterative training until the model loss function converges, and save the training weights that are optimal for multi-scale feature extraction and localization.

[0072] The training set images were input into an improved LDC-YOLOV11 model network equipped with LDFE, CUC, and LAI-Detect modules for iterative training. An early termination strategy was employed to suppress overfitting, and multi-threaded data loading was configured to accelerate image preprocessing. The AdamW optimizer, combined with a weight decay mechanism, was used to alleviate overfitting, and a step-wise learning rate and momentum parameter were set to update the network gradient. During training, the model loss was monitored in real time using a validation set. After the loss function converged, the training effect was quantitatively evaluated using precision, recall, and mean precision metrics. The optimal training weights for multi-scale feature extraction and localization were saved. Specific training parameter settings were as follows: input image size (Imgsize) was set to 640×640, training epochs to 200, and patience to 50 epochs to prevent overfitting and conserve computational resources. The number of threads for data loading (Works) was set to 8 to accelerate preprocessing. The optimizer used was AdamW, with a weight decay parameter (WeightDecay) set to 0.0001 to prevent overfitting, an initial learning rate (Lr0) set to 0.0001, a final learning rate factor (Lrf) set to 1.0, and a momentum parameter set to 0.9. Model performance evaluation metrics mainly included precision, recall, and mean precision. To verify the effectiveness of LDFE, LAI-Detect, and CUC, ablation experiments were conducted, and mAP50 and mAP50:95 were used as performance evaluation metrics to measure the effect of each improvement.

[0073] S6: Evaluate the effectiveness of the model in insulator defect detection, analyze the differences between the LDC-YOLOV11 model and mainstream target detection models in multiple indicators, and verify the superiority of the algorithm in terms of accuracy, lightweight and generalization.

[0074] Multiple models, including FasterNet, MobileNetv4, EfficientViT3, EfficientFormerv2, and LSNet, were selected as feature extraction networks for the object detection model, and their performance was comprehensively compared under the same testing environment and dataset. To further evaluate the generalization ability of the LDC-YOLOV11 model, the generalization verification experiment was conducted using the publicly available insulator defect image dataset IDID (Insulator DefectImage Dataset). This strongly demonstrates the reliability of the proposed method in handling insulator defect detection tasks under complex environmental interference. Furthermore, receptive field visualization analysis and heatmap visualization analysis verified the detection effect and reliability of the improved algorithm in complex industrial environments.

[0075] In this embodiment, after the above steps, the structure diagram of the LDC-YOLOV11 model constructed by the present invention is as follows. Figure 2 As shown.

[0076] Step S1 is based on a self-built image dataset for detecting insulator defects in transmission lines, using the COCO format. A portion of the data was collected from the China Power Line Insulator Dataset (CPLID), while the remainder was obtained through photography and public network collection. Given that existing research methods primarily focus on insulator defect detection under favorable weather conditions, with insufficient attention paid to scenarios under adverse weather conditions such as snowfall, and the difficulty in collecting real-world adverse weather data, the original aerial images underwent three weather transformations—shaking blur, rain, and snowfall—for algorithm training to improve the model's inspection accuracy under various climatic conditions. Before training, the dataset was proportionally divided into training, validation, and test sets.

[0077] After image enhancement preprocessing such as jitter blurring, rain, and snowfall, a high-quality dataset was obtained. Then, it was divided into training, validation, and test sets in a 7:2:1 ratio. Specifically, the CPLID dataset containing 1600 images of Chinese power line insulators was obtained, and 400 images were collected through actual shooting and public network collection, together forming the high-quality dataset containing 2000 images. Finally, it was divided into a training set of 1400 images, a validation set of 400 images, and a test set of 200 images for subsequent experiments.

[0078] Step S1 Example: Obtain the image dataset for defect detection of transmission line insulators, and use the image enhancement library Imgaug to perform three weather transformations on the original aerial images: shake blur, rain, and snowfall. Divide the dataset into training, validation, and test sets proportionally. The core information of the dataset and the acquisition methods are summarized in Table 1.

[0079] Table 1 Summary of Insulator Defect Dataset Information

[0080]

[0081] In step S2, the lightweight dynamic feature extraction module LDFE is as follows: Figure 3 As shown, specifically:

[0082] Its core is the GatedLSConvBlock unit, which is an improvement on LSConv.

[0083] A schematic diagram of the GatedLSConvBlock structure is shown below. Figure 4As shown, this module improves the structure by introducing LSConv and combining it with GatedCNNBlock. LSConv simulates the operators of the human visual system, dynamically generating convolutional kernels through a large kernel perceptron (LKP) and using a small kernel aggregator (SKA) for adaptive feature fusion. It can dynamically generate large-size convolutional kernels based on input features, thereby improving the model's dynamic feature extraction and context-adaptive aggregation capabilities with relatively low computational cost.

[0084] The improved GatedLSConvBlock is integrated to form a brand-new backbone network module LDFE. Specifically, the input feature X first passes through a 1x1 convolutional layer. (Including convolutional layers, batch normalization layers, and SiLU activation function) Perform channel projection and uniformly divide the feature list into two parts along the channel dimension. .in, The identity mapping path is preserved to ensure the smooth flow of gradients; This is then used as the main path input, and progressive feature transformation is performed through n concatenated GatedLSConvBlocks (n is 3 by default) to generate a series of feature maps {F}. Finally, the features of all paths (including the initially retained features) are... Main path starting point (Including all LDFE outputs) are concatenated along the channel dimension to form a comprehensive feature representation that integrates multi-level, multi-receptive-field information. Finally, this comprehensive feature is passed through a 1x1 convolutional layer. Perform fusion and channel number adjustment to output the final features. LDFE reduces computational overhead while employing a gating mechanism for adaptive dynamic feature regulation between channels, effectively suppressing background noise. Simultaneously, it utilizes a big.mesh kernel-big.small kernel collaboration mechanism to expand the model's receptive field, enhancing the network's ability to capture global contextual information. LDFE's adaptive capability allows the model to more accurately focus on key target pixels, preventing small target defect features from being overwhelmed by surrounding redundant background pixels during standard sampling, thus ensuring the ability to extract features from small target defects.

[0085] Example of step S2:

[0086] Specifically, within a single GatedLSConvBlock, for the input feature tensor First, layer normalization is performed, and then the number of channels is expanded by a point convolution. In order to improve computational efficiency and retain some original feature flow, the transformed features are divided into three parts in the channel dimension according to a specific ratio: gated feature g, identity feature i, and feature d to be convolved. The complete operation of this part is expressed as Equation (1).

[0087]

[0088] in, For a 1×1 point convolution operation, The layer normalization operation is represented by Split, which represents the channel splitting function.

[0089] For the segmented feature branch d, GatedLSConvBlock feeds it into the LSConv module for deep space feature extraction. The LSConv processing is divided into two stages: large kernel perception and small kernel aggregation.

[0090] Big nucleus sensing stage:

[0091] In order to obtain the global shape and long-distance spatial dependence of the insulator target at low cost, the big kernel perception stage uses 7×7 big kernel depth convolution to model the spatial relationship, and combines two point convolutions (1×1 convolution) and activation functions to dynamically generate a set of dynamic aggregate kernel weights w related to the input content, as shown in formula (2).

[0092]

[0093] in, and These represent the weight matrices corresponding to the two channel-adjusted point convolutions, This represents a 7×7 large kernel depthwise convolution. This represents the ReLU activation function.

[0094] Small nucleus aggregation stage:

[0095] Adaptive dynamic weights generated using LKP (in For each spatial location of feature d, the local neighborhood of the g-th channel feature. Convolution operations are performed internally to extract high-dimensional dynamic semantic feature information. The process is as shown in formula (3).

[0096]

[0097] in, This represents a dynamic convolution operation, indicating that in the case of grouping, the generated dynamic weights w are used to... Convolution operations are performed. This process is carried out in parallel across all groups and all spatial locations. Features enhanced by LSConv. Channel splicing and fusion are performed with the identity feature i. Simultaneously, a gated signal is generated using the gated branch feature g through an activation function, and the spliced ​​feature map is multiplied and weighted. GatedLSConvBlock can adaptively amplify important features in the insulator defect region based on the input image, suppressing redundant and complex background interference. Finally, the fused features are projected through point convolution and combined with the input feature X to form a residual connection output, as shown in formula (4).

[0098]

[0099] in, This represents a 1×1 point convolution. This represents the GeLU activation function.

[0100] A series of feature maps {F} are generated by progressive feature transformation through n concatenated GatedLSConvBlocks (n is 3 by default). Finally, the features of all paths (including the initially preserved features) are transformed. Main path starting point (Including all LDFE outputs) are concatenated along the channel dimension to form a comprehensive feature representation that integrates multi-level, multi-receptive-field information. Finally, this comprehensive feature is passed through a 1x1 convolutional layer. Perform fusion and channel number adjustment to output the final features. .

[0101] The channel compensation upsampling module (CUC) in step S3, such as Figure 5As shown, the core design idea of ​​CUC is to achieve multi-level feature interaction under zero parameter increase conditions. It combines channel shifting and channel mixing to further enhance the information flow between channels, realize the adaptive fusion of multi-directional contextual information, and ultimately significantly enhance the ability to express the feature characteristics of small targets while maintaining parameter efficiency. Specifically, the feature recovery process based on the channel compensation upsampling module begins with the spatial dimension expansion of the input low-resolution feature map. The geometric resolution is expanded by using a spatial interpolation algorithm through an upsampling layer to lay the foundation for high-resolution feature expression. Subsequently, the feature map is fed into a depthwise separable convolutional (DWConv) layer to perform efficient local spatial feature extraction while maintaining extremely low computational overhead. In order to achieve accurate mapping and alignment of the channel dimension, the feature map is further projected onto the target channel space through an Adjust Channel (AC) layer using a 1×1 convolution, and then uniformly split in the channel dimension to decouple several independent channel subsets. These channel subsets are fed into a Shift Channel Mix unit, where positive and negative cyclic shifts of specific steps are applied along the height and width directions, including positive height shift, negative height shift, positive width shift, and negative width shift. This dynamically compensates for the geometric alignment bias introduced by upsampling in different spatial directions. Finally, these channel features containing multi-directional spatial position compensation information are concatenated along the channel dimension and then integrated and linearly mapped globally through a standard terminal convolutional (Conv) layer. This eliminates the inter-channel phase misalignment caused by cyclic shifts while outputting a high-quality feature map with aligned features and fully compensated channel information.

[0102] Example of step S3:

[0103] For a given input feature map (Where C is the number of channels, and H and W are the height and width), CUC is upsampled by 2 times to obtain Using depthwise separable convolution pairs to enlarge the image. For feature extraction, depthwise separable convolution decomposes standard convolution into two independent steps: depthwise convolution and pointwise convolution, significantly reducing the number of parameters and computational complexity. Subsequently, the module employs the channel rearrangement operation AC (Adjust Channel) to regroup and transpose the feature maps along the channel dimension, allowing features from different groups to be fully mixed. Let the number of groups be G, and the channel rearrangement process be as shown in formula (5).

[0104]

[0105] in, The output of a depthwise separable convolution. (⋅) represents the tensor reshaping operation. (⋅) is a transpose operation that swaps the positions of the grouping dimension and the channel dimension, converting the original grouped block-based channel organization structure into a cross-group staggered arrangement.

[0106] CUC performs a ShiftChannelMix operation on the channel-rearranged features. By combining channel segmentation with multi-directional cyclic shifting, different channels can perceive information from different spatial locations in the feature map, effectively expanding the receptive field of the features, enhancing the model's ability to capture local details, and achieving zero-parameter feature enhancement. Specifically, it receives input features that are uniformly divided into four parts along the channel dimension, i.e. And apply cyclic displacement operations in different directions to each sub-feature to simulate the diffusion and interaction of features in space. Apply a positive height offset, which means shifting the entire input feature map down by one pixel; Apply a negative height offset (move upwards) to Apply a positive width offset, that is, shift the entire input feature map one pixel to the right. Apply a negative width offset (shift to the left). Finally, the four offset sub-feature maps are re-stitched along the channel dimension to complete feature fusion. The translation operation is uniformly implemented using... The (⋅) function represents the height displacement process as shown in formula (6) and the width displacement process as shown in formula (7).

[0107]

[0108]

[0109] in, dw represents the height dimension (value is 2), s represents the width dimension (value is 3), and s represents the displacement step size (default is set to 1). The index [c,i,j] represents the element in the c-th channel, i-th row, and j-th column.

[0110] The coordinate-aware lightweight interaction detection head (LAI-Detect) in step S4 is as follows: Figure 6 As shown, specifically, a coordinate-aware lightweight interactive detection head, LAI-Detect (Location-Aware Interactive Detection Head), was designed. This head integrates coordinate confidence calculation, breaking down the information barrier between classification and regression tasks. It explicitly utilizes the bounding box distribution information predicted by the regression task to adaptively calibrate the classification confidence, thereby improving the model's accuracy in locating and classifying insulator defects. Specifically, the forward propagation process begins with processing the input feature map... The basic feature extraction process involves passing the feature map sequentially through a convolutional (Conv) layer, a depthwise separable convolutional (DWConv) layer, and another convolutional (Conv) layer. This process reconstructs local spatial and channel features while controlling computational complexity. The features are then split into two core prediction branches: an upper branch and a lower branch. In the upper branch, which predicts the bounding box coordinate distribution, the split features are mapped through convolutional layers and scaling (Scale_i) layers, directly outputting a regression probability distribution feature representing the bounding box position. In the classification confidence prediction branch calibrated by regression distribution information below, the processing flow presents a multi-path fusion feature calibration mechanism: on the one hand, the backbone features directly generate the initial classification confidence score through the convolutional layer. On the other hand, the characteristics of the regression probability distribution The dynamic calibration introduced into this branch is first transformed into a probability distribution through a Softmax layer. The system then inputs the data into the Probability Ranking (PR) unit to extract the top k feature vectors with the highest probabilities. The vector is fed in parallel into a weighted mean (WM) unit to calculate the statistical weighted features. Subsequently, and Concat the data along the channel dimension to construct spatial statistics. After adjusting the dimensions through a reshape operation, the data is input into a multilayer perceptron (MLP) for nonlinear mapping, thereby generating regression distribution quality assessment weights. Finally, quality assessment weights. Compared with the initial classification confidence Element-wise addition (+) is performed to accurately calibrate the classification confidence using regression distribution information, and the final classification score is output. The final classification score. Compared with the regression distribution characteristics above At the end, a concatenated concatenation (Concat) is performed to finally output the decoupled and constructed detection feature map. .

[0111] Example of step S4:

[0112] The input to LAI-Detect is a list of feature maps from different layers of the backbone network. This corresponds to three detection scales. Specifically, for the feature map input at the i-th scale... First, initial feature transformation is performed using independent 3×3 convolutional layers to adapt to detection requirements at different scales. After transformation, features at all scales are then extracted and channel dimension adjusted using depthwise separable convolutions and 1×1 convolutions to obtain the output. This enables networks to learn general feature representations across scales and tasks.

[0113] To enable information exchange between the classification and regression branches, the module executes two branches in parallel: the bounding box coordinate distribution prediction branch and the classification confidence prediction branch, which uses the regression distribution information for accurate prediction.

[0114] Bounding box coordinate distribution prediction: A 1×1 convolutional layer maps the dimensions of Xic to 4R, where R is the number of intervals in the Distribution Focal Loss (DFL) regression (R=16 in this design). This feature map represents the probability distribution of the four coordinates of the bounding box (center x, y, width w, height h) across R discrete intervals. Subsequently, a learnable scaling factor is used... i Fine-tuning the coordinate distribution yields the predicted coordinate distribution. The decoded coordinates can be obtained through DFL calculations.

[0115] Relatively accurate classification confidence prediction based on regression distribution information: using a 1×1 convolutional layer... Mapped to preliminary classification scores ,in This represents the number of categories. Traditional detection heads are used directly here. As a classification confidence score, the coordinate information of a bounding box (i.e., the central tendency and reasonableness of the predicted coordinate distribution) is highly correlated with the confidence score of its class. Therefore, a probability ranking (PR) and a weighted mean (WM) method for calculating the coordinate information were designed to regress the distribution information of each coordinate. The weighted result Q is obtained by calculating the weighted mean. i Then the weighted result Q i Preliminary classification score Perform weighted calibration.

[0116] Specifically, such as Figure 6 As shown in the regression distribution information, the classification confidence prediction branch is more accurate. Applying the Softmax operation, we obtain the normalized distribution prediction. And generate different weights for each ranking vector: let the ranking be j∈[0,K−1] (j=0 is the highest probability value). The weight corresponding to each ranking The calculation process is as shown in formula (8).

[0117]

[0118] in, This is an adjustable hyperparameter that controls the rate at which weights decay as ranking decreases. A larger value results in a higher relative weight for higher ranking probabilities. This weighting mechanism gives greater influence to high-probability values, making the weighted mean focus more on the high-certainty portion of the distribution, thus more accurately reflecting the confidence level of the coordinates.

[0119] Based on weight and Calculate the weighted mean The calculation process is as shown in formula (9).

[0120] Will and the obtained weighted mean By splicing the data, statistical features can be formed. This feature not only includes the main probability distribution of bounding box prediction, but also obtains the contribution of high-confidence regions to coordinate selection through weighted mean. Next, this statistical feature... Dimensional reshaping is performed and applied to an MLP layer (the weights and biases of the last layer of the MLP are initialized to zero to ensure that the original classification scores are not affected in the early stages of training), outputting a single-channel coordinate-weighted result. The calibrated classification confidence score (Sifinal) is calculated using the initial classification score. and The sum is obtained as shown in formula (10).

[0121]

[0122] Finally, the output of the detection head is obtained by concatenating the bounding box coordinate distribution prediction with the classification confidence prediction based on the regression distribution information. This operation correlates classification confidence with coordinate information: predictions with accurate positioning and concentrated distribution receive a positive bonus, while predictions with ambiguous positioning may receive a negative adjustment or be suppressed. Using a weighted mean method, by emphasizing the role of high-ranking probability values, it more sensitively distinguishes between high-certainty and low-certainty coordinate predictions. When the coordinate probability distribution predicted by the bounding box regression module is relatively concentrated, the calculated... A positive value improves the final classification score; conversely, a negative value indicates a discrete coordinate distribution. A negative value reduces the classification score of the anchor box.

[0123] The specific training parameter settings in step S5 are as follows: the input image size (Imgsize) is set to 640×640, the number of training epochs is set to 200, and the patience value is set to 50 epochs to prevent overfitting and save computational resources. The number of threads for data loading (Works) is set to 8 to accelerate preprocessing. The optimizer used is AdamW, the weight decay parameter (WeightDecay) is set to 0.0001 to prevent overfitting, the initial learning rate (Lr0) is set to 0.0001, the final learning rate factor (Lrf) is 1.0, and the momentum parameter is set to 0.9. The model performance evaluation metrics mainly include precision, recall, and mean precision.

[0124] in:

[0125] Object detection models typically employ a series of standardized evaluation metrics to quantitatively measure their detection accuracy and speed. Among these, metrics for evaluating detection accuracy include precision (P), recall (R), average precision (AP), and mean average precision (mAP), while frames per second (FPS) serves as a measure of speed.

[0126] Accuracy is defined as the proportion of true positive bounding boxes in the detection results, that is, the ratio of the number of samples correctly detected as targets to the total number of samples predicted as targets. Recall, on the other hand, represents the proportion of correctly identified targets among all targets, that is, the ratio of the number of samples correctly detected as targets to the total number of samples that were actually targets.

[0127] The relationship between precision and recall is usually represented by a PR curve. By observing the area enclosed by the PR curve and the coordinate axes, we can compare the average precision (AP) of different models on a specific class. Based on this, mAP is defined as the mean AP of all classes in the dataset, reflecting the overall performance of the model and its generalization ability. It is an important metric for evaluating multi-class object detection tasks.

[0128] Step S6 involves a comprehensive evaluation of the effectiveness of the LDC-YOLOv11 model in the insulator defect detection task, specifically as follows:

[0129] S6.1: To verify the effectiveness of LDFE, LAI-Detect, and CUC, ablation experiments were conducted based on a dataset of transmission line insulator defect detection images. mAP50 and mAP50:95 were used as performance evaluation metrics to measure the effect of each improvement. The experimental results are shown in Table 2, where “√” indicates that the corresponding improvement measures were incorporated into the model training. The experimental data in the table show that by introducing LDFE, CUC, and LAI-Detect into the baseline model, the performance metrics of LDC-YOLOV11 were optimized to varying degrees.

[0130] Table 2 Ablation Experiment Results of LDC-YOLOV11 Model

[0131]

[0132] First, LDFE demonstrates the most outstanding performance in balancing performance improvement and lightweight design, increasing mAP50 from the baseline of 89.0% to 91.1%, while reducing the number of model parameters from 20.0M to 16.2M, resulting in a computational reduction of approximately 23.4%. The combination of LDFE and CUC achieves mAP@0.5:0.95 at 65.0% and improves accuracy to 94.3%, forming an effective synergistic improvement. Integrating all three modules achieves optimal performance across several key metrics: mAP50 reaches 91.8%, and mAP50:95 is improved to 65.8%. LAI-Detect enhances the discriminative capability of the detection head, increasing mAP50 to 90.2%. CUC focuses on improving detection accuracy, achieving a detection precision of 93.8%, effectively improving the model's accuracy in locating small defects. In collaborative experiments with multiple modules, the various improvement strategies exhibit good compatibility. The combination of LDFE and LAIDetect further reduces the number of parameters to 15.9M while achieving an mAP@0.5:0.95 of 64.4%, a 0.6 percentage point improvement compared to using LAIDetect alone. Experiments demonstrate that the method integrating LDFE, LAIDetect, and CUC modules exhibits excellent performance in both accuracy and lightweight design. It significantly improves the accuracy of insulator defect detection while achieving lightweight design through structural optimization, reducing the number of parameters to 15.9M compared to the baseline model, and reducing floating-point computation to 77% of the baseline model. This fully demonstrates that the synergistic effect between the modules effectively enhances the model's feature representation ability and generalization performance under complex conditions, achieving a better balance between accuracy, speed, and resource consumption.

[0133] S6.2: Comparison of different lightweight backbone networks. In the comparison with other advanced lightweight backbone networks, FasterNet, Mobilenetv4, EfficientViT3, EfficientFormerv2, LSNet, and other models were selected as feature extraction networks for object detection models, and their performance was comprehensively compared under the same test environment and dataset. As shown in Table 3.

[0134] Table 3 Performance Comparison of Different Lightweight Backbone Networks

[0135]

[0136] Specifically, in terms of detection accuracy, the LDFE module achieved the highest mAP50:95 score of 64.4%, significantly outperforming other comparative networks. While FasterNet and Mobilenetv4 had lower parameter and computational costs, their mAP50:95 scores of 58.9% and 58.7% respectively indicated limited feature extraction capabilities when dealing with complex and subtle targets such as insulator defects. The lightweight Transformer-based models EfficientViT3 and EfficientFormerv2 demonstrated outstanding computational efficiency, but their relatively low recall rates of 78.0% and 77.8% respectively may have resulted in insufficient detection of minute defects, thus limiting their overall accuracy. LSNet achieved high mAP50 and recall rates, but its parameter count was relatively high. LDFE, while achieving similarly high precision and recall, achieved the highest precision of 93.4% among all models, demonstrating the superior feature discrimination ability of the improved method. In terms of model efficiency, LDFE's parameter count was at a moderate level compared to the comparative models. Although its computational cost is higher than that of purely lightweight CNNs or Transformer variants, the significant improvement in accuracy it brings is far greater. Experimental results show that, under the premise of controllable parameter and computational costs, building the feature extraction backbone network with the designed LDFE module achieves the best balance between lightweight design and performance. The LDFE backbone network outperforms other lightweight baseline networks in overall detection performance, better meeting the needs of insulator defect detection in transmission lines.

[0137] S6.3: Comparative experiments were conducted with other excellent object detection algorithms using the same testing environment and dataset. Specific experimental results are shown in Table 4.

[0138] Table 4 compares the experimental results of different models.

[0139]

[0140] In terms of overall performance, LDC-YOLOV11 strikes a balance between detection accuracy and lightweight model. On the mAP50:95 metric, which comprehensively reflects model detection accuracy, LDC-YOLOV11 achieved the highest value of 65.8%, outperforming all comparable models. Simultaneously, LDC-YOLOV11's recall rate reached 86.5%, the highest among all comparable models, indicating its strongest coverage of defects in the dataset and the lowest false negative rate.

[0141] In terms of model efficiency, LDC-YOLOV11 demonstrates the best deployment advantages. Its parameter count is only 15.9M, and its computational cost is 52.9 GFLOPs. Compared to the similarly lightweight YOLOv10, LDC-YOLOV11 achieves a 3.6 percentage point higher mAP50:95 with comparable parameter count. Compared to the more accurate YOLOv8, LDC-YOLOV11 achieves a 4.7% improvement in mAP50:95 while reducing computational cost by approximately 32.8%. This experiment demonstrates that while reducing model complexity, a more efficient feature extraction mechanism enhances the ability to capture minute defects in complex backgrounds. Furthermore, LDC-YOLOV11, with its lower computational cost and parameter count, further surpasses the high-threshold localization accuracy metric of mAP50:95, indicating that LDC-YOLOV11 possesses higher application value while maintaining good overall detection performance.

[0142] S6.4: Compare the image detection results of different models. Figure 7 The comparison of prediction results between the benchmark model YOLOv11, RT-DETR, and LDC-YOLOV11 is presented. Through comparative visualization of the prediction results, it is clearly observed that LDC-YOLOV11 outperforms RT-DETR and the benchmark model YOLOv11 in handling insulator defect detection tasks. This is reflected not only in improvements in accuracy and mAP values, but also in a significant improvement in actual prediction results.

[0143] In the prediction results of the baseline model YOLOv11, due to the influence of background noise from tower structure and severe weather conditions, the detection accuracy was relatively low, and significant missed detections occurred. RT-DETR improved the ability to extract defect features to some extent, with fewer missed detections. The bounding box range and category confidence scores were improved compared to YOLOv11, but there is still room for optimization in the accurate localization of small defects. LDC-YOLOV11, while maintaining a lightweight parameter scale, showed better detection performance and stability in the face of background noise caused by various complex environments. Specifically, in the case where the insulator shell and flashover color were highly similar as shown in the first column of the figure, LDC-YOLOV11 accurately identified all the tiny flashover points on the insulator string and did not misdetect shadows as defects. In the snowy environment shown in the second column of the figure, LDC-YOLOV11 successfully and accurately captured the tiny damage on the edges of each insulator with the highest confidence and avoided the missed detections of the baseline model and RT-DETR. In environments with rainfall, motion blur, and background interference from tower materials (as shown in columns 3 and 4 of the figure), the overall detection performance of LDC-YOLOV11 is better than the other two comparison results, with complete detection frame coverage.

[0144] Comprehensive analysis shows that LDC-YOLOV11, while reducing the number of model parameters, steadily enhances the anti-interference capability against complex hidden dangers in transmission lines, fully meeting the dual requirements of high precision and real-time performance in actual power inspection.

[0145] S6.5: Generalization Validation Experiments on Public Datasets. To further evaluate the generalization ability of the LDC-YOLOV11 model, a generalization validation experiment was conducted using the publicly available insulator defect image dataset IDID (Insulator Defect Image Dataset). IDID contains 1600 aerial images of insulators, covering 1152 instances of damage and 2448 instances of flashover. These images are derived from actual power line inspection operations and cover aerial images of insulators taken under various geographical locations, aircraft types, and lighting conditions. Compared to the transmission line insulator defect detection image dataset, IDID has significant domain differences in shooting equipment, geographical background, and annotation style, effectively evaluating the model's robustness and adaptability in unseen domains. Zero-shot evaluation was performed on the pre-trained LDC-YOLOV11 model; that is, no IDID data was introduced during the training phase, and experiments were conducted directly on its test subset to calculate various detection metrics. The experimental results are shown in Table 5.

[0146] Table 5 Comparison of generalization performance of each model on the IDID dataset

[0147]

[0148] S6.6: Visualization analysis of receptive fields and visualization analysis of heatmaps.

[0149] Figure 8 This presentation shows a visual comparison of the effective receptive field between the baseline model and LDC-YOLOV11 at the same feature layers. The visualization results show that the improved model (e.g., ...) has significantly different receptive fields. Figure 8 (a) The effective receptive field distribution is relatively limited, with low response intensity in the central region. This means that the model mainly focuses on a small local space when extracting features. This lack of a receptive field often leads to a lack of robust feature representation when facing complex backgrounds such as insulator strings or scale fluctuations caused by motion blur, thus affecting the accuracy of localization. In contrast, the improved LDC-YOLOV11 (such as...) Figure 8 (b) The receptive field performs better, the area of ​​the high response region is significantly expanded, and the central response point performs more evenly.

[0150] The results of the thermogram analysis are as follows: Figure 9 It can be seen that the baseline model, when dealing with insulator defects, lacks sufficient focus on the key target areas. For example... Figure 9 The first row of heatmap comparison results shows that the baseline model, when processing insulator string detection, partially disperses its attention across the entire insulator surface, resulting in a diffuse activation range and significant background interference. The improved model, however, accurately focuses on the specific defect locations on the insulator discs. In detecting interference from complex tower material backgrounds, the improved model... Figure 9 The second row of heatmaps shows that the baseline model exhibits more ineffective activations at the tower's metal components, indicating its difficulty in decoupling target features from cluttered background textures. In contrast, the improved model adaptively filters background noise, with activation regions strictly constrained to the insulator's physical contours, demonstrating stronger background suppression. In multi-target detection under snowfall conditions, [the model is described in the original text]. Figure 9 The comparison of the third row of heatmaps shows that, due to the interference of snowflake noise and the superimposed background of white ceramic tile color, the model's attention center is inconsistent, and the response of some detection boxes is low. The improved model has a stronger ability to capture local details, and the high-contribution heatmap accurately anchors the defect location indicated by each detection box, with a more concentrated heatmap distribution. To address the strong reflective interference in outdoor environments, [the following is a separate, unrelated sentence:] ... Figure 9 As can be seen from the comparison of the heatmaps in the fourth row, when the baseline model processes samples affected by reflective points, its attention is severely drawn to high-brightness noise, easily mistaking reflective points for flashover features. The model cannot effectively suppress this high-brightness noise. In contrast, the improved model effectively avoids background interference from snowfall and surface reflective points, and the heatmaps exhibit highly structured and focused features, resulting in clear identification results.

[0151] Overall, the improved algorithm LDC-YOLOV11, through the LDFE, CUC, and LAI-Detect modules, significantly reduces the number of parameters while exhibiting more effective target focusing and feature discrimination capabilities, verifying the detection performance and reliability of the improved algorithm in complex industrial environments.

[0152] In this context, bold text indicates the optimal value.

[0153] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for detecting defects in transmission line insulators based on LDC-YOLOV11, characterized in that, Includes the following steps: S1: Obtain the image dataset for defect detection of insulators of transmission lines. After preprocessing with motion jitter blurring, rain, and snow image enhancement, a high-quality dataset is obtained. Then, it is divided into training set, validation set, and test set in a 7:2:1 ratio. S2: In the backbone network, a lightweight dynamic feature extraction module LDFE is designed to expand the global receptive field to suppress complex background noise in the inspection image; S3: In the neck network, a channel-compensated upsampling module (CUC) is designed to improve the restoration quality of transmission line insulator defect detection images during the upsampling process; S4: Design a coordinate-aware lightweight interactive detection head LAI-Detect to enhance the lightweight model's ability to locate and distinguish targets with weak textures and minor defects; S5: Input the training set images from the dataset into the LDC-YOLOV11 model for iterative training until the model loss function converges, and save the training weights that are optimal for multi-scale feature extraction and localization. S6: Evaluate the effectiveness of the model in insulator defect detection, analyze the differences between the LDC-YOLOV11 model and mainstream target detection models in multiple indicators, and verify the superiority of the algorithm in terms of accuracy, lightweight and generalization.

2. The method for detecting defects in transmission line insulators based on LDC-YOLOV11 according to claim 1, characterized in that, The specific steps of S1 are as follows: S11: Obtain the China Power Line Insulator Dataset CPLID, which contains 1600 images. In addition, 400 images were collected through actual shooting and public network collection, which together constitute a high-quality dataset containing 2000 images. S12, apply motion blur, rain, and snowfall image enhancement preprocessing to the image; S13. The preprocessed images are divided into a training set of 1400 images, a validation set of 400 images, and a test set of 200 images for subsequent experiments.

3. The method for detecting defects in transmission line insulators based on LDC-YOLOV11 according to claim 1, characterized in that, In S2, the lightweight dynamic feature extraction module LDFE consists of an input CBS unit, a Split unit, three cascaded GatedLSConvBlock units, a Concat unit, and an output CBS unit.

4. The method for detecting defects in transmission line insulators based on LDC-YOLOV11 according to claim 3, characterized in that, The operation steps of the lightweight dynamic feature extraction module LDFE are as follows: Within a single GatedLSConvBlock, for the input feature tensor Normalization is performed, where For batch size, For the number of channels, For feature map height, The width of the feature map; Expand the number of channels using a point convolution; The transformed features are divided into three parts along the channel dimension according to a specific ratio: gated features g, identity features i, and features to be convolved d.

5. The method for detecting defects in transmission line insulators based on LDC-YOLOV11 according to claim 1, characterized in that, The specific steps of S3 are as follows: S31, for a given input feature map Where C is the number of channels, H and W are the height and width, respectively. The channel compensation upsampling module CUC upsamples it by a factor of 2 to obtain... ; S32, using depthwise separable convolution pairs for the magnified image. Perform feature extraction; S33, the channel compensation upsampling module CUC uses the channel rearrangement operation AC to regroup and transpose the feature maps in the channel dimension, so that the features of different groups can be fully mixed; S34 performs a ShiftChannelMix operation on the features after channel rearrangement. By combining channel segmentation with multi-directional cyclic shifting, different channels can perceive information from different spatial locations in the feature map.

6. The method for detecting defects in transmission line insulators based on LDC-YOLOV11 according to claim 1, characterized in that, The specific steps of S4 are as follows: S41, The input to LAI-Detect is a list of feature maps from different layers of the backbone network. These correspond to three detection scales; S42, for the first Input feature maps at each scale ,in, The initial feature transformation is performed through independent 3×3 convolutional layers to adapt to the detection requirements of different scales; S43, after transformation, features at all scales are extracted and channel dimensions are adjusted using depthwise separable convolutions and 1×1 convolutions to obtain the output. This enables networks to learn general feature representations across scales and tasks; S44, Two independent prediction branches are fed in parallel: one is the bounding box coordinate distribution prediction branch, and the other is the classification confidence calibration branch that integrates regression distribution information. S45: Concatenate the prediction results from the two branches to generate the final output features of the detection head. .

7. The method for detecting defects in transmission line insulators based on LDC-YOLOV11 according to claim 1, characterized in that, The specific steps of S5 are as follows: S51, input the training set images into the LDC-YOLOV11 model equipped with LDFE, CUC, and LAI-Detect modules for iterative training; S52 uses a validation set to monitor model loss in real time during training. After the loss function converges, the training effect is quantitatively evaluated using precision, recall, and average precision metrics. The optimal training weights for multi-scale feature extraction and localization are saved.

8. The method for detecting defects in transmission line insulators based on LDC-YOLOV11 according to claim 1, characterized in that, In step S6, one or more models from FasterNet, Mobilenetv4, EfficientViT3, EfficientFormerv2, and LSNet are selected as the feature extraction network for the object detection model, and their performance is compared under the same test environment and dataset.

Images