Power transmission line insulator equipment detection method and system based on YOLO

Abstract

The invention discloses a YOLO-based power transmission line insulator equipment detection method and system, and solves the problems of low insulator small target detection performance and poor detection precision under a complex background when a target detection model is adopted to realize insulator equipment detection, and the method comprises the following steps: constructing an enhanced YOLOv11 target detection model; performing cross-scale fusion on the P2 feature and the original P3 feature, and enhancing the fused feature; feature alignment is carried out on the features in each feature scale, mixed features are generated according to the features of the unified scale, and feature enhancement is carried out on the mixed features by using a deformable convolutional layer; performing multi-scale feature enhancement on each detection branch, wherein part of convolution branches adopt a lightweight convolution strategy; and the insulator equipment is detected according to the enhanced YOLOv11 target detection model. By enhancing the detection head of the target detection model, the small target identification precision of the target detection method under a complex background is improved.

Description

Technical Field

[0001] This invention relates to the field of power transmission line inspection technology, and in particular to a method and system for testing power transmission line insulators based on YOLO. Background Technology

[0002] During UAV inspections of power transmission lines, insulator defect identification typically relies on high-quality insulator target detection methods to extract equipment areas and then distinguish whether the equipment is faulty. For example, patent CN120807415A describes a method for detecting power transmission line insulator defects based on an improved YOLOv8s architecture. This method involves acquiring original images of insulator defects; constructing an improved YOLOv8s model for insulator defect detection: replacing the Conv modules in layers P1, P3, P5, and P7 of the backbone network with APEDown modules; replacing the Conv modules in layers P16 and P19 of the neck network with SCDSConv modules; replacing the Detect module in the head network with the GSDetect module; replacing the BCE loss function with an improved ATFL loss function; and using the trained model to detect insulator defects and determine the defect type.

[0003] However, in practice, because power transmission lines typically span long distances, the images acquired by drones often contain complex backgrounds such as cities, mountains, and rivers, which affects the high-precision detection of insulators. Furthermore, due to the limited safe operating distance of drones for power transmission lines, insulators occupy a relatively small portion of the inspection images, which also significantly impacts the target detection performance of insulators. Summary of the Invention

[0004] The purpose of this invention is to solve the problems of low performance and poor accuracy in detecting small targets on insulators in complex backgrounds when using target detection models for insulator equipment detection. This invention provides a YOLO-based method and system for detecting insulator equipment on transmission lines, which improves the accuracy of small target identification in complex backgrounds by enhancing the detection head of the target detection model.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A YOLO-based testing method for transmission line insulators includes the following steps: S1: Construct an enhanced YOLOv11 object detection model: S1.1: Add a new high-resolution P2 feature, fuse the P2 feature with the original P3 feature across scales, and enhance the fused feature; S1.2: Align the features within each feature scale, generate hybrid features based on features of the same scale, and enhance the hybrid features using deformable convolutional layers; S1.3: Perform multi-scale feature enhancement on each detection branch, with some convolutional branches employing lightweight convolution strategies; S2: Insulator equipment is inspected based on the enhanced YOLOv11 target detection model.

[0006] This invention provides a YOLO-based method for detecting insulators on power transmission lines. It constructs a high-resolution feature branch based on a 1 / 4 downsampling rate, fully utilizing the rich spatial details and texture information preserved in the shallow features of the network. The high-resolution P2 features are fused and enhanced with semantically informative P3 features, significantly improving the model's ability to perceive small targets on insulators. The gradient propagation path is optimized to ensure the stability of the training process. Deformable convolution is used to enhance the perception of detailed features such as insulator texture and the overall structure of the insulator target. This improves the accuracy of small target recognition in complex backgrounds.

[0007] Preferably, S1.2 includes: setting a sub-module for each feature scale, with each sub-module independently processing feature inputs at four different scales; each sub-module receiving two inputs: a single-scale feature Px and a set of features at all scales; using a bilinear interpolation algorithm to uniformly scale all input features in the set of scale features to the scale of the feature Px corresponding to the current sub-module; concatenating the features at the uniform scale by channel, and using a convolutional layer to perform channel dimensionality reduction and fusion to form a hybrid feature.

[0008] Preferably, step 1.2 includes: using deformable convolutional layers to perform spatial-level feature enhancement on the hybrid features, and then performing dimensionality transformation through 1×1 convolutional layers; adding the input single-scale features and the enhanced features element-wise to achieve residual connections, thereby obtaining the final enhanced features.

[0009] Preferably, S1.3 includes: extracting multi-scale features in parallel for each detection branch using four different scale convolutional kernels of 1×1, 3×3, 5×5 and 7×7, wherein the 3×3, 5×5 and 7×7 convolutional kernels adopt the lightweight convolutional form of GhostNet; concatenating the parallel multi-scale features in the channel dimension, and using 1×1 convolution to reduce the dimensionality of the concatenated high-dimensional features; and using channel attention to perform weighted optimization on the dimensionality-reduced and integrated features to obtain the output enhanced multi-scale features.

[0010] Preferably, the lightweight convolution measurement includes: performing channel dimensionality reduction using a 1×1 lightweight ordinary convolution; extracting and restoring features from the dimensionality-reduced features using a lightweight depthwise separable convolution, wherein the kernel size of the depthwise separable convolution is consistent with that of the original ordinary convolution; and concatenating the features produced in the channel dimensionality reduction stage and the feature restoration stage according to the channels to obtain the output feature map.

[0011] As a preferred embodiment, the method also includes: collecting images of transmission line insulators at different voltage levels under different scenarios, constructing an image dataset, and using the image dataset to train an enhanced insulator target detection model; acquiring real-time transmission line inspection images, scaling the inspection images and inputting them into the trained enhanced insulator target detection model, and outputting the location and classification information of the insulators.

[0012] A YOLO-based testing system for transmission line insulators includes: The data acquisition module collects images of transmission line insulators at different voltage levels under different scenarios, including the location bounding boxes and category labels of the insulators. The YOLO target detector, which includes three detection heads of different sizes, performs target detection on images of power transmission line insulator equipment. The detection head enhancement module is installed between the feature pyramid and the detection head of the YOLO target detector to enhance the detection head of the YOLO target detector.

[0013] Preferably, the detection head enhancement module includes: The high-resolution extension module adds P2-level features, fuses the P2-level features with the P3-level features of the YOLO object detector, and enhances the fused features. The cross-scale feature interaction module performs multi-scale feature alignment, concatenates features of the same scale by channel, uses convolutional layers to perform channel dimensionality reduction and fusion to form hybrid features, and performs residual connections based on the hybrid features. The multi-scale feature enhancement module adopts a parallel multi-scale convolutional kernel structure to adaptively aggregate multi-level features.

[0014] Preferably, the high-resolution extension module includes two C3k2 modules, each comprising three consecutive convolutional blocks. The first convolutional block is a 1×1 convolution that performs a dimensionality transformation and outputs C1 and C2. The second convolutional block is a bottleneck convolutional block that receives C1 and outputs C3. The third convolutional block is also a bottleneck convolutional block that receives C2 and outputs C4. The bottleneck convolutional block comprises two convolutional layers with residual connections between the input and output.

[0015] Preferably, the cross-scale feature interaction module includes four parallel sub-modules, each of which independently processes feature inputs at four different scales; the multi-scale feature enhancement module includes four independent feature enhancement units, each of which adopts a unified architecture and integrates multi-scale convolution operations and channel attention mechanisms to achieve multi-scale feature extraction and aggregation.

[0016] Therefore, this invention has the following beneficial effects: it introduces shallow features to enhance the extraction of local detail information, employs deformable convolution to achieve efficient interaction between cross-scale features, enriches contextual feature representation, and uses a parallel multi-scale convolutional kernel structure to adaptively aggregate multi-level features. A multi-level feature optimization system is constructed, achieving effective complementarity between semantic information and detail features. While maintaining high detection speed, it significantly improves detection accuracy, providing a reliable technical solution for the intelligent and automated inspection of power transmission lines. Attached Figure Description

[0017] Figure 1 This is a flowchart of the steps in the YOLO-based transmission line insulator equipment testing method of the present invention.

[0018] Figure 2 This is a schematic diagram of the architecture of the YOLO-based transmission line insulator equipment detection system in this invention.

[0019] Figure 3 This is a schematic diagram of the architecture of the YOLOv11 target detector in this invention.

[0020] Figure 4 This is a schematic diagram showing the insertion position of the detection head enhancement module in the YOLO target detector in this invention.

[0021] Figure 5 This is a schematic diagram of the high-resolution extension module in this invention.

[0022] Figure 6 This is a schematic diagram of the cross-scale feature interaction module in this invention.

[0023] Figure 7 This is a schematic diagram of the multi-scale feature enhancement module in this invention. Detailed Implementation

[0024] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0025] Example 1: This embodiment provides a YOLO-based method for testing transmission line insulators, such as... Figure 1As shown, the operation process is as follows: Step 1, construct an enhanced YOLOv11 target detection model; Step 2, perform detection on the insulator equipment based on the enhanced YOLOv11 target detection model.

[0026] The YOLO-based transmission line insulator equipment detection method provided in this embodiment is mainly used to solve the problem of insulator equipment target detection in complex backgrounds and enhance the accuracy of target detection method in small target identification in complex backgrounds.

[0027] The following examples and specific application scenarios further illustrate the technical solution and effects of the present invention. The following examples are explanations of the present invention, but the present invention is not limited to the following examples.

[0028] Step 1: Construct an enhanced YOLOv11 object detection model.

[0029] First, the basic YOLOv11 object detection model adopts a feature pyramid structure and outputs three feature layers P2, P3 and P4 at different scales. Each scale corresponds to a detection head, and the detection head outputs a predicted bounding box from the feature layer.

[0030] Specifically, the basic YOLOv11 object detection model has a detection head system containing three detection heads of different scales, corresponding to downsampling resolutions of 1 / 8, 1 / 16, and 1 / 32, respectively, and labeled as P3, P4, and P5 level features.

[0031] Based on the basic YOLOv11 target detection model, an enhanced target detection model for transmission line insulators is constructed to enhance the detection head.

[0032] Specifically: (1.1) High-resolution extension.

[0033] A high-resolution extension branch is introduced. Based on the original detection head system, a new high-resolution feature branch based on a 1 / 4 downsampling rate is added, labeled as P2 level features. P2 features retain richer shallow features such as local details, and can capture the fine texture patterns of insulator equipment, significantly improving the detection performance of small-sized insulator targets.

[0034] To perform high-resolution expansion, directly introducing high-resolution P2 features may cause gradient conflict problems. Therefore, this embodiment fuses P2 and P3 features across scales and uses two C3k2 modules to enhance the fused features, thereby stabilizing training and improving feature quality while preserving details.

[0035] (1.2) Cross-scale feature interaction.

[0036] A multi-scale feature alignment mechanism is established to unify features within each feature scale. These unified features are then concatenated by channel and 1×1 convolutional layers are used to perform channel dimensionality reduction and fusion to form hybrid features.

[0037] A cross-scale feature interaction mechanism based on residual connections is constructed. A 3×3 deformable convolutional layer is used to perform spatial-level feature enhancement on the mixed features to adaptively capture multi-scale contextual information. Finally, a 1×1 convolutional layer is used for channel mixing, and combined with the residual connection mechanism, the final enhanced features are output, realizing full interaction of cross-scale features.

[0038] (1.3) Multi-scale feature enhancement.

[0039] Multi-scale convolution and channel attention mechanisms are applied to each detection branch to achieve targeted enhancement of the feature representation capabilities of each branch.

[0040] Multi-scale convolution processing includes extracting multi-scale features in parallel using convolution kernels of four different scales: 1×1, 3×3, 5×5, and 7×7. For the 3×3, 5×5, and 7×7 convolution operations, a lightweight convolutional form from GhostNet is used, thereby significantly reducing the number of parameters and computational overhead while maintaining the multi-scale receptive field.

[0041] The multi-scale features extracted in parallel are concatenated along the channel dimension, and a 1×1 convolution is used to reduce the dimensionality of the concatenated high-dimensional features. Finally, the dimensionality-reduced features are fed into the channel attention module for weighted optimization, highlighting key information and suppressing redundant features, ultimately outputting the enhanced multi-scale features.

[0042] The multi-scale features output by each detection branch are input into the corresponding detection head, thereby achieving detection head enhancement.

[0043] Step 2: Detect the insulator equipment using the enhanced YOLOv11 target detection model.

[0044] Images of power transmission lines are obtained through drone cameras; these images are scaled up and input into an enhanced insulator equipment target detection model; the enhanced insulator equipment target detection model outputs the location and classification information of the insulator equipment.

[0045] This embodiment also provides a YOLO-based transmission line insulator equipment testing system, such as... Figure 2 As shown, it includes: a data acquisition module, a YOLO target detector, and a detection head enhancement module. The data acquisition module is connected to the YOLO target detector, and the detection head enhancement module is installed between the feature pyramid of the YOLO target detector and the detection head to enhance the detection head of the YOLO target detector.

[0046] Specifically, the data acquisition module constructs an image dataset for insulator equipment to train and validate the target detection model, while simultaneously acquiring real-time images of insulator equipment as input to the enhanced target detection model. The image dataset contains images of transmission line insulator equipment at different voltage levels under different scenarios, and includes the location bounding boxes and category labels of the insulator equipment.

[0047] The detection head enhancement module includes a high-resolution extension module, a cross-scale feature interaction module, and a multi-scale feature enhancement module. The high-resolution extension module is connected to the feature pyramid of the YOLO target detector, the cross-scale feature interaction module is connected to the high-resolution extension module, and the multi-scale feature enhancement module is connected to the cross-scale feature interaction module and the detection head of the YOLO target detector, respectively.

[0048] The high-resolution extension module introduces shallow features to enhance the ability to extract local details; the cross-scale feature interaction module uses deformable convolution to achieve efficient interaction between cross-scale features and enrich the contextual feature representation; the multi-scale feature enhancement module adopts a parallel multi-scale convolution kernel structure to adaptively aggregate multi-level features.

[0049] This detection system aims to address the problem of insufficient feature representation capabilities in the detection of small targets on insulator equipment. It can be directly inserted as a plug-and-play module between the feature pyramid and detection head of most YOLO series target detectors, exhibiting excellent compatibility and significantly improving the performance of small target detection on insulators. This effectively enhances the performance of small target detection on insulators, better meeting the accuracy requirements of practical defect identification.

[0050] YOLO object detector, such as Figure 3 As shown, the insertion position and method of the detection head enhancement module are as follows: Figure 4 As shown. Specifically, as... Figure 3 As shown, the first convolutional layer is designated as layer 0, and P2, P3, P4, and P5, as inputs, correspond to the features of layers 2, 16, 19, and 22, respectively. The output is connected to the original detection head and expanded into four detection heads.

[0051] The high-resolution extension module adds P2-level features, fuses the P2-level features with the P3-level features of the YOLO object detector, and enhances the fused features.

[0052] like Figure 5 As shown, the high-resolution extension module includes two C3k2 modules, which enhance the fused features. Specifically, each of the two C3k2 modules uses a convolutional layer to adjust the dimensions to reduce computational cost.

[0053] The C3k2 module consists of three consecutive convolutional blocks. The first convolutional block performs a dimensionality transformation using a 1×1 convolution, and its output is split into two parts, C1 and C2. The second and third convolutional blocks are bottleneck convolutional blocks, receiving C2 and C3 respectively to generate C3 and C4. The bottleneck convolutional blocks consist of two convolutional layers with residual connections between the input and output. Finally, the features from C1 to C4 are concatenated by channel and then transformed using a 1×1 convolution. This setup allows the module to perform complex feature extraction patterns while maintaining computational complexity and parameter size.

[0054] The cross-scale feature interaction module aligns features across multiple scales, concatenates features of the same scale by channel, and uses convolutional layers for channel dimensionality reduction and fusion to form hybrid features. Residual connections are then performed based on these hybrid features. The module receives a set of multi-scale input features with different spatial resolutions, maintaining the detailed information of shallow features and the semantic information of high-level features to achieve cross-scale feature interaction. It introduces deformable convolutions to dynamically adjust the sampling point positions, more accurately capturing rich contextual features of targets at different scales, and enhancing the feature diversity and discriminative ability of features at different scales.

[0055] like Figure 6 As shown, the cross-scale feature interaction module comprises four parallel sub-modules, each independently processing feature inputs at four different scales: P2, P3, P4, and P5. Each sub-module receives two distinct inputs: the first input is a single-scale feature Px, and the second input is a feature set Pall encompassing all scales. All features in Pall are scaled using bilinear interpolation to uniformly adjust to the scale of the corresponding feature Px in the current sub-module (e.g., the P3 sub-module will be uniformly scaled by 1 / 16 of the scale). The two outputs are concatenated along the channel dimension, and a 1×1 convolutional layer is used to perform channel compression and feature fusion, forming a hybrid feature.

[0056] The multi-scale feature enhancement module employs a parallel multi-scale convolutional kernel structure to adaptively aggregate multi-level features. Specifically, the multi-scale feature enhancement module uses a parallel branch structure, utilizing convolutional kernels of different scales to aggregate multi-scale information, thereby capturing finer local details and richer contextual semantics. Through efficient aggregation using adaptive weights, it enriches the feature representation hierarchy, achieves robust feature enhancement, and improves generalization ability and detection accuracy in complex environments.

[0057] like Figure 7 As shown, the multi-scale feature enhancement module includes four independent feature enhancement units, avoiding the feature smoothing problem caused by cross-scale feature interactions and enhancing the diversity representation ability of the features carried by each branch. Each feature enhancement unit adopts a unified design architecture, integrating multi-scale convolution operations and channel attention mechanisms to achieve multi-scale feature extraction and aggregation.

[0058] Within each feature enhancement unit, four different scale convolutional kernels—1×1, 3×3, 5×5, and 7×7—are used. Small-scale kernels are responsible for enhancing local details, while large-scale kernels are responsible for enhancing semantic information. For the 3×3, 5×5, and 7×7 convolutional branches, a lightweight convolutional strategy similar to GhostNet is introduced.

[0059] The YOLO-based method and system for detecting transmission line insulators provided in this embodiment have the following advantages: (1) By introducing high-resolution extension, a high-resolution feature branch based on a 1 / 4 downsampling rate was constructed to make full use of the rich spatial details and texture information retained in the shallow features of the network. Feature alignment and cross-layer fusion strategies were adopted to alleviate the gradient conflict problem that may be caused by directly introducing high-resolution features. The high-resolution P2 features were fused and enhanced with the semantically information-rich P3 features, which significantly improved the model's ability to perceive small targets on insulators and optimized the gradient propagation path to ensure the stability of the training process.

[0060] (2) By introducing a cross-scale feature interaction mechanism, an efficient interaction mechanism between local detail features and global semantic features was constructed. Deformable convolution was used to enhance the perception of detail features such as insulator texture and overall structure. Cross-level information fusion effectively enhanced the model's ability to represent targets at different scales and significantly improved its ability to detect small targets on insulators.

[0061] (3) By introducing a multi-scale feature enhancement mechanism, the feature diversity is expanded by adopting a parallel multi-branch structure. Diverse features are extracted by convolution kernels with different receptive fields, and key features are adaptively enhanced by combining an attention mechanism. This effectively enriches the expressive power of the feature space and improves the detection performance of insulator targets under complex background interference.

[0062] (4) Based on the efficient YOLOv11 framework, improvements were made. Through a carefully designed detection head enhancement module, the inadequate fusion of global semantic features and local detail features in traditional target detection networks was overcome. A multi-level feature optimization system was constructed, achieving effective complementarity between semantic information and detail features. While maintaining a high detection speed, the detection accuracy was significantly improved, providing a reliable technical solution for the intelligent and automated inspection of transmission lines. As a plug-and-play module, the detection head enhancement module can be embedded in most insulator target detection models based on the YOLO series target detectors, and has a wide range of application scenarios.

[0063] Example 2: This embodiment, based on Embodiment 1, provides a YOLO-based detection method for transmission line insulators, and provides a detailed description of the high-resolution extension, cross-scale feature interaction, and multi-scale feature enhancement in Embodiment 1.

[0064] Specifically, a YOLO-based testing method for transmission line insulators includes the following steps: Step 1: Data collection.

[0065] Images of insulators for transmission lines at different voltage levels under various scenarios are collected, and the location bounding boxes and category labels of the insulators are included.

[0066] The acquired images are used to train the subsequent enhanced YOLOv11 object detection model end-to-end. The training process includes data augmentation, parameter tuning, and other steps.

[0067] The insulator target detection network was tested using the test set in the transmission line insulator equipment target detection dataset constructed in step one, and mAP@50 and mAP@50:90 were used as insulator target detection performance evaluation indicators.

[0068] The trained enhanced YOLOv11 target detection model is deployed into the intelligent computing unit of the inspection drone to perform online inspection tasks of insulator equipment.

[0069] Step 2: Construct a basic YOLOv11 object detection model.

[0070] Construct a basic YOLOv11 object detection model, including a backbone network, a feature pyramid, and a detection head.

[0071] Step 3: High-resolution expansion.

[0072] Based on the detection head system of P3, P4, and P5 layers included in the basic YOLOv11 object detection model, a new high-resolution feature branch P2 based on a 1 / 4 downsampling rate is added.

[0073] After processing by the deep model, features P3, P4, and P5 suffer significant loss of local detail, with smaller insulators experiencing greater visual feature loss than larger ones. In contrast, features P2, derived directly from shallow network layers, retain richer local detail (such as texture and color), allowing the YOLOv11 object detection model to capture the texture patterns and structural features of insulator devices. Therefore, P2 effectively enhances the basic YOLOv11 object detection model's ability to detect small targets like insulators.

[0074] However, directly introducing high-resolution P2 feature branches may cause conflicts during gradient calculation. This is mainly because P2 features are located in the shallow layers of the network, close to the input of the basic YOLOv11 object detection model. To balance computational efficiency, this stage typically uses lightweight convolutional structures with large spans and few channels. Deeper features such as P3, P4, and P5, however, originate from the feature pyramid and have already undergone depth processing; their feature distribution differs significantly from that of the shallow P2 features. Directly performing cross-scale feature interactions can disrupt the stability of the training process and affect model convergence. Furthermore, directly optimizing P2 features can place an excessive learning burden on the lightweight convolutional layers on the input side.

[0075] Therefore, this embodiment adopts a fusion strategy when scaling up to high resolution: first, P2 features rich in detail information are fused with P3 features rich in semantics across levels, and feature alignment is used to ensure spatial consistency of features at different scales.

[0076] Subsequently, the fused features are enhanced. This high-resolution expansion step (i.e., the high-resolution expansion module) allows the model to effectively optimize the gradient propagation path while maintaining the integrity of the original details. This ensures training stability and gives the enhanced features both rich detailed information and stronger semantic representation capabilities.

[0077] Step 4: Cross-scale feature interaction.

[0078] Four parallel, independent submodules are used to independently process feature inputs at four different scales: P2, P3, P4, and P5. Each submodule receives two different inputs: the first input is the feature Px corresponding to a single scale, and the second input is a feature set Pall containing all scales. All features in Pall are scaled using a bilinear interpolation algorithm to uniformly adjust to the scale of the feature Px corresponding to the current submodule (for example, the P3 submodule will uniformly scale by 1 / 16 of the scale).

[0079] The two outputs are concatenated according to the channel dimension, and a 1×1 convolutional layer is used to perform channel compression and feature fusion to form a hybrid feature.

[0080] The hybrid features will be enhanced through a 3×3 deformable convolutional layer. Subsequently, a 1×1 convolutional layer will be used to transform the dimensions, optimize the feature representation capability, and form enhanced features.

[0081] Finally, the single-scale feature Px of the original input is added element-wise with the enhanced feature to achieve residual connection.

[0082] Deformable convolution, by introducing learnable spatial transformation parameters, can adaptively adjust the shape of the receptive field according to input features, significantly improving the model's ability to model small objects. While conventional convolution can only achieve regular sampling using a pre-defined square, deformable convolution adds an offset prediction branch, generating a pair of two-dimensional offset coordinates for each point in the conventional sampling grid, allowing the conventional sampling points to shift their positions based on local context information.

[0083] This embodiment uses a 3×3 deformable convolutional layer with 9 sampling points. The position offset of each of the 9 sampling points is calculated, and features are extracted based on the offset points instead of the regular sampling points.

[0084] Optionally, a grouped convolutional model can be added, which divides the features into multiple groups, with each group performing the same operation. All features are then aggregated together, and feature fusion and dimensionality transformation are achieved through 1×1 convolutions.

[0085] Deformable convolution's adaptive receptive field can more accurately cover insulator targets of different scales and shapes, significantly enhancing the ability to model targets and effectively improving the flexibility of feature representation. It can also establish cross-scale feature correspondences, promoting effective information interaction between features at different levels. This enables the model to better handle complex situations such as scale variations, pose changes, and partial occlusion of insulators in images, thereby extracting more discriminative feature representations and providing a more reliable feature foundation for subsequent detection decisions.

[0086] The deformable convolution takes the form of: for each group g (a total of G groups), the kernel sampling point k (the total number of sampling points is K) is calculated. The calculation method is: the sum of the fixed offset between the center position and the kth sampling point and the offset learned at the kth sampling point of the g group at position p, multiplied by the slice of the feature map in the g group, multiplied by the modulation scalar of the kth sampling point in the g group, and multiplied by the linear mapping weight that is independent of the position corresponding to the group.

[0087] The above method is used to calculate each sampling point in each group, and the sum of the final calculation results is the deformable convolution.

[0088] The offset learned at position p, the k-th sampling point in the g-th group, is typically a two-dimensional vector and is usually not an integer. Therefore, bilinear interpolation is required to obtain the final feature value. The modulation scalar at the k-th sampling point in the g-th group is usually normalized using the Softmax function along the sampling point dimension.

[0089] Step 5: Multi-scale feature enhancement.

[0090] For each detection branch, feature enhancement is performed using four different scales of convolution kernels: 1×1, 3×3, 5×5, and 7×7. Small-scale convolution kernels are responsible for enhancing local details, while large-scale convolution kernels are responsible for enhancing semantic information.

[0091] Since the size of the convolution kernel directly affects computational efficiency, a lightweight convolution strategy similar to GhostNet is introduced for convolutional branches such as 3×3, 5×5, and 7×7.

[0092] GhostNet's lightweight convolution strategy divides traditional convolution computation into two stages. The first stage is intrinsic feature generation, using a 1×1 lightweight ordinary convolution to achieve channel dimensionality reduction, with a reduction ratio set to 8 in this example. The second stage is feature dimensionality restoration. Since the input features suffer some information loss after dimensionality reduction in the first stage, this stage receives the reduced features and uses a lightweight depthwise separable convolution to achieve feature extraction and dimensionality restoration. The kernel size of the depthwise separable convolution is the same as the original ordinary convolution. The output dimension of the second stage is the total output dimension minus the output dimension of the first stage. The features generated in the first and second stages are concatenated by channel to form a complete output feature map. This feature map is then input into the detection head of the basic YOLOv11 object detection model to enhance the detection head.

[0093] In this embodiment, the convolution modified by GhostNet can maintain the insulator target detection performance basically unchanged while significantly reducing the number of parameters.

[0094] After obtaining multi-scale features, they are concatenated along the channel dimension to form a hybrid feature. A 1×1 convolution is used to fuse the hybrid feature into channels, and the fused feature is fed into the Squeeze-and-Excitation channel attention module (hereinafter referred to as the attention module). This module adaptively learns the channel weights through a compression-excitation mechanism to enhance key features and suppress redundant features, thereby improving the expressive power of the features.

[0095] The core idea of ​​the Squeeze-and-Excitation attention module is to enhance the model's ability to filter information by adaptively adjusting the channel feature responses through modeling the interdependencies between channels. This module can be divided into two processes: compression and activation. In the compression stage, global pooling is used to compress the two-dimensional spatial vectors of each channel of the feature map into a single scalar. This operation aggregates global information for each channel, capturing the global distribution of features for each channel. In the activation stage, the feature channel vectors are handled by a gating mechanism consisting of two fully connected layers. The first fully connected layer performs channel dimensionality reduction and uses the ReLU function for non-linear transformation; the second fully connected layer restores the dimension; subsequently, the Sigmoid function maps the feature channel vectors to the [0,1] interval, forming channel importance weights. Finally, the channel importance weights are multiplied by the input features channel by channel to achieve feature enhancement.

[0096] SE attention effectively amplifies the contribution of important feature channels while suppressing redundant channels, thereby improving the model's feature representation capability. In this embodiment, residual connections are introduced on top of SE attention.

[0097] The technical solution and technical effects of the present invention will be further illustrated below through specific examples and application scenarios.

[0098] (1) The experimental environment configuration is shown in Table 1.

[0099] Table 1 Experimental Environment Configuration (2) YOLOv11n was selected as the basic model and the target detection dataset for transmission line insulators was selected. This dataset contains UAV inspection images from complex backgrounds such as mountains, rivers, and cities.

[0100] (3) The experimental setup is shown in Table 2.

[0101] Table 2 Experimental Setup (4) Evaluation indicators.

[0102] This embodiment selects four commonly used evaluation metrics in the field of target detection: precision (P), recall (R), mAP@50, and mAP@50:95.

[0103] Accuracy is the quotient of the number of correctly predicted positive samples divided by the sum of the number of correctly predicted positive samples and the number of incorrectly predicted positive samples. Accuracy reflects the reliability of the model's predictions; a higher accuracy indicates fewer false positives.

[0104] Recall is the quotient of the number of correctly predicted positive samples divided by the sum of the number of correctly predicted positive samples and the number of positive samples that the model failed to predict. Recall reflects the model's ability to detect all cases; the higher the recall, the fewer false negatives.

[0105] mAP@50 represents the average accuracy across all categories when the intersection-union ratio between the predicted bounding box and the ground truth bounding box is 0.5. mAP@50:95 represents the average of the average accuracy across all categories when the intersection-union ratio threshold between the predicted bounding box and the ground truth bounding box ranges from 0.5 to 0.95 (in increments of 0.05).

[0106] (5) Experimental results.

[0107] Table 3 Experimental Results As shown in Table 3, the experimental results demonstrate that the YOLO-based transmission line insulator detection method proposed in this embodiment effectively improves the model's performance in detecting small insulator targets in complex scenarios. In the transmission line insulator target detection dataset used in this embodiment, small target samples (defined as those whose pixel area accounts for less than 2% of the total image pixel area) account for approximately two-thirds of the total. mAP@50:95, as a more stringent metric for measuring the overlap between the predicted bounding box and the ground truth bounding box, directly reflects the improvement in model localization accuracy. In the high-proportion small insulator target dataset, this metric improves by 2.98%, fully demonstrating that the improved model has more accurate predictions for small insulator targets. This experimental result confirms the effectiveness and superiority of the method provided in this embodiment in improving the performance of small insulator target detection. It exhibits high accuracy and robustness, effectively solving the problem of transmission line insulator target detection and providing a reliable technical solution for addressing the localization problem in small target detection under complex backgrounds.

[0108] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Other variations and modifications are possible without departing from the technical solutions described in the claims.

Claims

1. A method for testing transmission line insulators based on YOLO, characterized in that, include: S1: Construct an enhanced YOLOv11 object detection model: S1.1: Add a new high-resolution P2 feature, fuse the P2 feature with the original P3 feature across scales, and enhance the fused feature; S1.2: Align the features within each feature scale, generate hybrid features based on features of the same scale, and enhance the hybrid features using deformable convolutional layers; S1.3: Perform multi-scale feature enhancement on each detection branch, with some convolutional branches employing lightweight convolution strategies; S2: Insulator equipment is inspected based on the enhanced YOLOv11 target detection model.

2. The method for testing transmission line insulators based on YOLO according to claim 1, characterized in that, S1.2 includes: using four parallel sub-modules to independently process feature inputs at four different scales; each sub-module receives two inputs: a single-scale feature Px and a set of features at all scales, and uses a bilinear interpolation algorithm to uniformly scale all input features in the set of scale features to the scale of the feature Px corresponding to the current sub-module; the features at the uniform scale are concatenated by channel, and channel dimensionality reduction and fusion are performed using convolutional layers to form hybrid features.

3. The method for testing transmission line insulators based on YOLO according to claim 2, characterized in that, The 1.2 section includes: using deformable convolutional layers to perform spatial-level feature enhancement on the hybrid features, and then performing dimensional transformation through 1×1 convolutional layers; adding the input single-scale features and the enhanced features element-wise to achieve residual connections, and obtaining the final enhanced features.

4. The method for testing transmission line insulators based on YOLO according to claim 1, characterized in that, S1.3 includes: using four different scale convolutional kernels of 1×1, 3×3, 5×5 and 7×7 to extract multi-scale features in parallel for each detection branch, wherein the 3×3, 5×5 and 7×7 convolutional kernels adopt the lightweight convolutional form of GhostNet; concatenating the parallel multi-scale features in the channel dimension, and using 1×1 convolution to reduce the dimension of the concatenated high-dimensional features; and using channel attention to perform weighted optimization on the reduced-dimensional features to obtain the output enhanced multi-scale features.

5. A method for testing transmission line insulators based on YOLO according to claim 1 or 4, characterized in that, The lightweight convolution measurement includes: using a 1×1 lightweight ordinary convolution for channel dimensionality reduction; using a lightweight depthwise separable convolution to extract and restore features from the dimensionality-reduced features, with the kernel size of the depthwise separable convolution being the same as that of the original ordinary convolution; and concatenating the features produced in the channel dimensionality reduction stage and the feature restoration stage according to the channels to obtain the output feature map.

6. A method for testing transmission line insulators based on YOLO according to claim 1, 2, 3, or 4, characterized in that, Also includes: Images of transmission line insulators at different voltage levels under different scenarios were collected to construct an image dataset, which was then used to train an enhanced insulator target detection model. Acquire real-time transmission line inspection images, scale the inspection images and input them into the trained enhanced insulator equipment target detection model, and output the location and classification information of the insulator equipment.

7. A YOLO-based transmission line insulator equipment testing system, employing the YOLO-based transmission line insulator equipment testing method described in any one of claims 1-6, characterized in that, include: The data acquisition module collects images of transmission line insulators at different voltage levels under different scenarios, including the location bounding boxes and category labels of the insulators. The YOLO target detector, which includes three detection heads of different sizes, performs target detection on images of power transmission line insulator equipment. The detection head enhancement module is installed between the feature pyramid and the detection head of the YOLO target detector to enhance the detection head of the YOLO target detector.

8. The YOLO-based transmission line insulator equipment testing system according to claim 7, characterized in that, The detection head enhancement module includes: The high-resolution extension module adds P2-level features, fuses the P2-level features with the P3-level features of the YOLO object detector, and enhances the fused features. The cross-scale feature interaction module performs multi-scale feature alignment, concatenates features of the same scale by channel, uses convolutional layers to perform channel dimensionality reduction and fusion to form hybrid features, and performs residual connections based on the hybrid features. The multi-scale feature enhancement module adopts a parallel multi-scale convolutional kernel structure to adaptively aggregate multi-level features.

9. The YOLO-based transmission line insulator equipment testing system according to claim 8, characterized in that, The high-resolution extension module includes two C3k2 modules. Each C3k2 module contains three consecutive convolutional blocks. The first convolutional block is a 1×1 convolution that performs a dimensionality transformation and outputs C1 and C2. The second convolutional block is a bottleneck convolutional block that receives C1 and outputs C3. The third convolutional block is a bottleneck convolutional block that receives C2 and outputs C4. The bottleneck convolutional block contains two convolutional layers, and there is a residual connection between the input and output.

10. A YOLO-based transmission line insulator equipment testing system according to claim 8 or 9, characterized in that, The cross-scale feature interaction module includes four parallel sub-modules, each of which independently processes feature inputs at four different scales; the multi-scale feature enhancement module includes four independent feature enhancement units, each of which adopts a unified architecture and integrates multi-scale convolution operations and channel attention mechanisms to achieve multi-scale feature extraction and aggregation.

Images