Cloud–Edge Collaborative Strategy for Insulator Recognition and Defect Detection Model Using Drone-Captured Images

Abstract

In modern power systems, drones are increasingly being utilized to monitor the condition of critical power equipment. However, limited computing capacity is a key factor limiting the application of drones. To optimize the computational load on drones, this paper proposes a cloud–edge collaborative intelligence strategy to be applied to insulator identification and defect detection scenarios. Firstly, a low-computation method deployed at the edge is proposed for determing whether insulator strings are present in the captured images. Secondly, an efficient insulator recognition and defect detection method, I-YOLO (Insulator-YOLO), is proposed for cloud deployment. In the neck network, we integrate an I-ECA (Insulator-Enhanced Channel Attention) mechanism based on insulator characteristics to more comprehensively fuse features. In addition, we incorporated the insulator feature cross fusion network (I-FCFN) to enhance the detection of small-sized insulator defects. Experimental results demonstrate that the cloud–edge collaborative intelligence strategy performs exceptionally well in insulator-related tasks. The edge algorithm achieved an accuracy of 97.9% with only 0.7 G FLOPs, meeting the inspection requirements of drones. Meanwhile, the cloud model achieved a $mA{P}_{50}$ of 96.2%, accurately detecting insulators and their defects.

1. Introduction

With the advancement of drone technology, drones have become an essential means for assessing the performance of power equipment [1]. The utilization of drones significantly expands the range for detecting power equipment and enhances the flexibility of monitoring, thereby effectively maintaining the stable operation of power systems. However, limited computing capacity remains a key factor restricting the application of drones. To facilitate deployment on drone platforms, recent studies [2,3,4] have implemented lightweight deep learning models on drones for object detection tasks, achieving promising results in insulator identification and defect detection scenarios. Nonetheless, the computational complexity of these models is still relatively high. Therefore, it is crucial to decompose the task load of drones and design low-computation intelligent algorithms to enhance their applications.

This paper aims to optimize the computational load on drones, with a particular focus on insulator identification and defect detection scenarios. As critical components of power transmission lines, insulators play a key role in maintaining transmission stability in the event of a fault. Therefore, certain existing deep learning methods aim to enhance the precision of identifying insulators and their defects. These approaches include constructing different backbone networks [5], incorporating additional attention mechanisms [6], and designing plugin modules with specific functionalities [3]. These methods often feature high parameter and computational requirements, which are suitable only for server deployment and fail to leverage the flexibility of drone monitoring. Additionally, to enable the deployment of intelligent algorithms on drones, some studies focus on lightweighting general deep learning object detection models to reduce parameter counts and computational complexity. These efforts include designing lightweight plugin modules [7], utilizing backbones with fewer parameters [8], and replacing standard convolutional modules with more lightweight alternatives [4]. Despite utilizing general deep learning frameworks, these methods still possess significant computational complexity and impact the model’s ability to detect effectively when optimizing parameters and reducing complexity. To leverage the flexibility of drone monitoring and enhance the precision in identifying insulators and their defects, some work has adopted cloud–edge collaborative strategies to maximize the advantages of both drone and server [9,10]. However, with regard to the detection accuracy on the server, these methods still impose substantial computational tasks on the drone.

To reduce the computational burden on drones, this study introduces a cloud–edge collaborative intelligent approach for insulator identification and defect detection. The main contributions of this work are as follows:

• A low-computation insulator coarse recognition method is proposed for edge deployment, which successfully simplifies the task on the drone to a binary classification problem of determining the presence of insulators in captured images, yielding good classification performance.
• An effective cloud-deployed method for insulator recognition and defect detection, named I-YOLO, is proposed. It designs an insulator feature cross-fusion network, which can efficiently identify small insulator defects.
• An improved attention mechanism based on insulator features, called I-ECA, is developed to comprehensively integrate features but without increasing the model’s parameter count.

The structure of this article is as follows: Section 2 reviews the related work; Section 3 introduces the proposed cloud–edge collaborative intelligence strategy for insulator identification and defect detection, with a detailed explanation of each functional module; Section 4 presents experimental comparisons to evaluate the effectiveness of the proposed approach; Section 5 analyzes the strengths and limitations of the method, evaluates the generalization performance of the I-ECA model, and explores potential future research directions; finally, Section 6 summarizes the main findings of the paper.

2. Related Work

Cloud-based insulator recognition and defect detection models. To enhance the accuracy of recognizing specific targets, models deployed on servers are often designed with a large number of parameters and high computational complexity. Zheng et al. [11] proposed a target detection model based on an improved fusion single shot multibox detector to address poor detection performance in infrared insulator images under complex backgrounds. This model improves feature extraction through a feature enhancement module and multi-scale feature maps while employing clustering algorithms to tackle the challenge posed by the large aspect ratios of insulators. Gao et al. [12] proposed a detection network that integrates the BN-CBAM with a feature fusion module to detect defects in small insulator strings. To improve the precision of identifying insulators and detecting their defects, Zhao et al. [13] used an improved Faster R-CNN based on the Feature Pyramid Network (FPN) to locate insulators in complex backgrounds, employing a threshold segmentation algorithm, line detection, and vertical projection for fault detection. To tackle the issue of small insulator fault regions, Zhang et al. [14] incorporated a DC-FPN into the YOLOv3 network. Zhang et al. [15] enhanced the YOLOv5 algorithm by integrating the squeeze-and-excitation (SE) module to solve insulator defect detection. Ou et al. [16] improved the Faster R-CNN model by refining the feature extraction network and adding extra detection box aspect ratios for detecting various electrical equipment. Fu et al. [17] introduced an I2D-Net that incorporates multiple novel modules to enhance feature extraction from shallow layers, including a TFFN, an RFA+ module, and a CPM to suit various scales and backgrounds. He et al. [5] proposed an improved YOLOv8 algorithm, MFI-YOLO, for detecting multiple insulator fault types, utilizing a C2F network built with GhostNet and MSA-GhostBlock for feature extraction in complex backgrounds, and a multi-scale feature fusion structure called ResPANet to improve detection accuracy in multi-target scenarios. To detect severely occluded insulators in power lines, Luo et al. [6] proposed an occlusion insulator detection system based on YOLOX, using an improved SPP module and the AFF-BiFPN for effective extraction of occluded and small defect information, and a coarse extraction method based on adaptive anchor frames to enhance detection performance. To address significant changes in target scale in insulators, Wang et al. [3] proposed the multiscale channel information (MCI)-global-local attention (GLA), a plugin designed for YOLO series models, consisting of the MCI extraction module and the GLA based on context information module (GLA-CI). The MCI module extracts and utilizes multi-scale feature map information comprehensively, while the GLA-CI module captures global context information and local spatial details, enhancing learning capabilities. A comparative analysis of cloud-based insulator recognition and defect detection models is shown in

Table 1. Cloud-based comparison of insulator recognition and defect detection models.

Type	Methods	Improvement
Cloud	Zheng’s method [11]	Feature enhancement module, multi-scale feature maps
	Gao’s method [12]	BN-CBAM and feature fusion module
	Improved Faster R-CNN [13]	Improved FPN
	Improved YOLOv3 [14]	DC-FPN
	Improved YOLOv5 [15]	SE
	Improved Faster R-CNN [16]	Improved feature extraction network, additional anchor aspect ratio
	I2D-Net [17]	TFFN, RFA+, CPM
	MFI-YOLO [5]	MSA-GhostBlock, ResPANet
	Improved YOLOX [6]	AFF-BiFPN, adaptive anchor frame coarse extraction
	Plug-in for YOLO series [3]	MCI, GLA-CI

, highlighting their improvements. These methods generally involve incorporating the latest attention mechanisms into the backbone, more powerful feature extraction networks, or adaptive anchor frame algorithms in the model head to address the challenge of detecting insulators with large aspect ratios, without specifically considering their characteristics, often leading to missed or false detections.

Edge-based insulator recognition and defect detection models. In such tasks, models deployed on edge devices are typically required to have fewer parameters and lower computational complexity. Yang et al. [18] enhanced the FPN structure of YOLOv3 to a Bidirection-Fusion FPN to improve the detection of small objects, incorporating EIoU and smooth-EIoU loss functions to speed up convergence during the training phase, thus enhancing both processing speed and recognition accuracy. To tackle the difficulty of identifying small-scale insulator defects in images, Hao et al. [19] proposed an improved YOLOv4-based insulator defect detection model named ID-YOLO. This model uses a CSP-ResNeSt as the backbone, introducing a multiscale Bi-SimAM-FPN for more effective feature fusion. Li et al. [7] introduced a PEDNet based on YOLOv4-tiny to enhance detection accuracy and speed in infrared images of electrical equipment. This model proposes a novel GIAM to focus the network’s attention on salient regions, integrating an FEFN in the backbone for comprehensive feature integration. To address low detection efficiency caused by large model parameters, Feng et al. [8] developed an improved YOLOv4 model—YOLOv4++. It utilizes MobileNetv1 as the backbone and substitutes standard convolution with DSConv. Lu et al. [4] proposed an IDD-YOLO designed for deployment on UAV hardware. This model adopts GhostNet for the backbone, incorporates a lightweight attention mechanism (LCSA) for more comprehensive feature capture, and uses PANet for feature transformation. As shown in

Table 2. Edge-based comparative analysis of insulator recognition and defect detection models.

Type	Methods	Improvement
Edge	Improved YOLOv3 [18]	Bidirection-Fusion FPN, EIoU and smooth-EIoU loss functions
	ID-YOLO [19]	CSP-ResNeSt, Bi-SimAM-FPN
	PEDNet [7]	GIAM, FEFN
	YOLOv4++ [8]	MobileNetv1, DSConv
	IDD-YOLO [4]	GhostNet, LCSA, PANet, GSConv and C3Ghost

these algorithms typically employ lightweight networks in the backbone to reduce parameter count and use the latest attention mechanisms in the neck to improve detection accuracy. Although general lightweight algorithms exhibit strong generalization capabilities, insulators are often located in complex terrains such as wilderness and mountainous areas. Moreover, extreme weather conditions can further increase the likelihood of missed detections and false detections of insulators and their defects [4].

Cloud–edge collaborative insulator recognition and defect detection models. Currently, various approaches have started exploring cloud–edge collaborative strategies to tackle tasks related to insulator identification and fault detection. Song et al. [10] introduced an intelligent cloud–edge collaborative method designed for insulator identification and fault detection within the power industrial IoT. They developed an algorithm for estimating potential insulator orientations on the drones, complemented by a defect detection method on the servers. However, images captured by UAVs tend to be large, resulting in numerous iterations needed for the orientation estimation algorithm. Furthermore, if it were possible to directly determine whether an image contains potential insulator targets, it could further reduce the drone’s energy consumption. To address the issues of significant resource consumption and high latency inherent in traditional centralized cloud computing, Wei et al. [9] designed a lightweight SSD object recognition network on the edge. On the cloud, they employed three distinct detection algorithms and utilized a multi-model fusion algorithm to obtain the coordinates and confidence level of self-explosion regions in insulators. A comparative analysis of cloud–edge collaborative insulator recognition and defect detection models is presented in

Table 3. Cloud–edge comparative analysis of collaborative insulator recognition and defect detection models.

Type	Methods	Edge Improvement	Cloud Improvement
Cloud–edge	Song’s method [10]	Insulator orientation estimation algorithm	Insulator defect detection algorithm
	Wei’s method [9]	Lightweight SSD object recognition network	Multi-model fusion algorithm
	Ours	Insulator coarse recognition algorithm	I-YOLO

, highlighting their improvements. However, directly applying object detection models intended for generic objects fails to adequately account for the unique large aspect ratio of insulators, which poses challenges in improving recognition accuracy.

3. Methods

This section first introduces the design of the cloud–edge collaborative intelligence strategy and its application in insulator string identification and fault detection scenarios. Next, it provides a detailed explanation of the insulator coarse recognition algorithm deployed on the drone. Finally, it specifically describes the methods for identifying insulators and detecting defects deployed on the server insulator YOLO (I-YOLO), along with the proposed insulator efficient channel attention mechanism (I-ECA) and insulator feature cross fusion network (I-FCFN).

3.1. Intelligence Strategy for Cloud–Edge Collaboration

As shown in Figure 1, this paper proposes a cloud–edge collaborative intelligence strategy to address the issue of significant computational load imposed by deep learning algorithms on drones, applying it to scenarios of insulator.

In the insulator coarse recognition algorithm on the drone, we first extract the H channel from the images captured by the drone and perform block processing and threshold-based segmentation. Next, each patch is filtered based on insulator features, and the potential location of the target is detected. Then, these patches are stitched together, and the presence of insulators in the image is determined along with a rough estimation of their locations based on the annotations of the entire image.

Once the drone transmits images containing potential insulator targets back to the server, the server employs the I-YOLO model designed in this paper to accurately locate and detect the insulators and their defects. On the one hand, to comprehensively integrate insulator features, we incorporate an insulator-specific enhanced attention mechanism—insulator efficient channel attention mechanism (I-ECA). On the other hand, we add an insulator feature cross fusion network (I-FCFN) to improve the detection of small-sized insulator defects.

3.2. Insulator Coarse Recognition on the UAV

In practical transmission line scenarios, the individual pieces of the same insulator share similar features and exhibit significant differences from the surrounding background information. This makes insulators in drone-captured images appropriate for segmentation through threshold-based approaches, such as the Otsu algorithm [20]. According to the study in [10], for insulator detection tasks, the H channel of an image is more suitable for threshold-based segmentation compared to other channels. As shown in Figure 2, the insulator exhibits relatively clustered features.

Although the aforementioned method can obtain the potential location information of insulators, it is highly sensitive to background noise. Therefore, we propose a novel regional filtering module to enhance the Otsu threshold segmentation algorithm. Firstly, as observed in Figure 2, compared to background noise, the pixel distribution in certain regions of the insulator is denser and exhibits an overall rectangular shape. This characteristic remains consistent even when the insulator is divided into multiple segments. Secondly, the insulator has a large aspect ratio, and this rectangular region approximates an ellipse. Based on the above analysis, we designed an insulator coarse recognition algorithm following the concept of segmenting, filtering, and restoring small regions. The steps are as follows:

1.

Extract the H channel from the images captured by the drone and divide it into 16 equal sub-images, which are then stored in set $A_0$.

2.

We employ a three-threshold Otsu segmentation algorithm to adaptively segment the sub-images in $A_0$, with the segmented regions denoted as $R^{\left(i\right)}=\{r_1^{\left(i\right)},r_2^{\left(i\right)},r_3^{\left(i\right)},r_4^{\left(i\right)}\}$. Furthermore, our experiments reveal that insulators are often located within the $r_4^{\left(i\right)}$ region.

3.

After determining the largest connected component R obtained from the segmentation of each sub-image, the region is fitted using the least squares ellipse fitting algorithm. The relevant parameters are calculated: the major axis length L, the minor axis length S, and the orientation $\theta$. The fitting result is returned as $R^{\prime }$. The general quadratic equation of an ellipse is given by the following:

$$a{x}^2+bxy+c{y}^2+dx+ey+f=0$$

(1)

Points within the region R are collected to construct the data matrix D. The parameter vector $P={[a,b,c,d,e,f]}^T$ is solved using the least squares method:

$$DP=0$$

(2)

The orientation angle $\theta$ of the ellipse is calculated as follows:

$$\theta ={\displaystyle \frac{1}{2}}arctan\left({\displaystyle \frac{b}{a-c}}\right)$$

(3)

The center point of the ellipse $(x_c,y_c)$ is determined as follows:

$$x_c={\displaystyle \frac{bd-2ae}{b^2-4ac}},y_c={\displaystyle \frac{be-2cd}{b^2-4ac}}$$

(4)

L and S of the ellipse are given by the following:

$${\lambda }_1=a{cos}^2\theta +bcos\theta sin\theta +c{sin}^2\theta$$

(5)

$${\lambda }_2=a{sin}^2\theta -bcos\theta sin\theta +c{cos}^2\theta$$

(6)

$$f^{\prime }=f+d{x}_c+e{y}_c+a{x}_c^2+b{x}_c{y}_c+c{y}_c^2$$

(7)

$$L=\sqrt{-{\displaystyle \frac{f^{\prime }}{{\lambda }_1}}},S=\sqrt{-{\displaystyle \frac{f^{\prime }}{{\lambda }_2}}}$$

(8)

4.

As noted in [10], the length of an insulator is significantly greater than its width, with $aspectratio>4$. Therefore, if the $L/S$ ratio of an elliptical region is small, for instance, $L/S\approx 1$, it indicates that the region does not contain an insulator. Additionally, since this study adopts a region-based segmentation method, we require $L/S>2$ to consider that a region may potentially contain an insulator. Moreover, insulators in the captured images are typically large or medium-sized objects. Hence, we remove small regions with an area of less than 100 to eliminate noise.

5.

Considering that images captured by drones may contain multiple insulators oriented in different directions, we store the direction $L/S>2$ of the region identified by the $\theta$ from the previous step into the set $A_1$. In practical scenarios, the number of insulators is usually small, and their orientations are typically similar. Therefore, we assume that if the difference between adjacent elements in the set $A_1$ exceeds a certain threshold $\pi /12$, it indicates the presence of insulators in different orientations.

As shown in Algorithm 1, the coarse recognition algorithm for insulators successfully combines the regional filtering module with the Otsu threshold segmentation method. By employing a strategy of small region segmentation, we effectively reduce the number of iterations required by the regional filtering module. This approach decreases both memory and computational consumption on the drone, meeting the detection needs in engineering applications.

Algorithm 1: Coarse Recognition of Insulators

3.3. I-YOLO Object Detection Algorithm on the Cloud

The network structure of I-YOLO is illustrated in Figure 3. To extract complete feature information from input images and facilitate the cross-fusion of features in the network’s neck section, we primarily use the Darknet series of networks as the backbone of I-YOLO. As depicted in the structural diagrams of the relevant components, the backbone predominantly comprises a residual structure, incorporating convolutional layers, batch normalization layers, and leaky ReLU as its basic components.

To better handle the feature information of insulator targets and their defect locations, the neck section of I-YOLO incorporates the proposed I-ECA module and conducts a cross-fusion of features. The current attention mechanisms are not designed for targets like insulators, which have large aspect ratios, affecting model accuracy. Therefore, we improved the Efficient Channel Attention Mechanism based on the characteristics of insulators. Additionally, I-YOLO cleverly integrates multi-scale semantic information by cross-fusing features from both shallow and deep networks, enabling the model to effectively handle small-sized insulator defects. These enhancements enable I-YOLO to excel in the task.

We primarily use basic components and $1\times 1$ convolutional layers in the head. This section is responsible for performing the tasks of object classification and regression, and it retains the detection boxes with the highest confidence as the output results of the model.

3.4. Insulator Efficient Channel Attention Mechanism

Although current attention mechanism modules are able to consider both the channel and spatial information of input feature maps, they overlook the large aspect ratio characteristic of targets like insulators. To improve the handling of insulators by the attention mechanism without increasing model parameters, we propose the I-ECA module, which utilizes Global Average Pooling (GAP) and Global Max Pooling (GMP) to account for both global information and the most critical details in insulator images. The basic principle of the I-ECA module is illustrated in Figure 4.

First, due to the uniform features and large aspect ratio of insulators in captured images, the features of insulators become more prominent after feature extraction in the backbone network. Therefore, we apply GAP and GMP to the input feature map $F_i$ to extract not only global information but also the most important details. Second, we use a one-dimensional convolution (Conv1D) for processing to facilitate interaction involving channel information. Then, the processed channel features are element-wise added together, and the fusion result is mapped into attention weights using the sigmoid activation function $\sigma$. Finally, the input feature map $F_i$ is element-wise multiplied with the attention weights along the channel dimension to produce the processed output feature map $F_o$. The formulaic expression of the I-ECA module is as follows:

$$F_o=\sigma (Conv1D\left(GAP\left(F_i\right)\right)\oplus Conv1D\left(GMP\left(F_i\right)\right))\otimes F_i$$

(9)

3.5. Insulator Feature Cross Fusion Network

To better detect small-sized insulator defects in complex backgrounds, we improved the FPN structure in the neck section and proposed an insulator feature cross fusion network (I-FCFN). The structure of I-FCFN is illustrated in Figure 5. First, the original FPN structure is retained; second, for the two head branches shown in Figure 6, we upsample the feature maps corresponding to the next set of residual components in the backbone and perform feature fusion with the input feature maps from the original FPN structure. By integrating multi-scale semantic information from both shallow and deep networks, the I-FCFN structure can more effectively identify insulator defects in images.

4. Experiments

This section first introduces the Chinese Power Line Insulator Dataset (CPLID) [21]; secondly, it outlines the experimental setup and evaluation metrics; next, it describes the comparative experiments between existing attention mechanisms and the proposed I-ECA in insulator recognition and defect detection tasks. Following this, it presents the comparative experiments between commonly used object detection models and the proposed I-YOLO in insulator tasks. Finally, we perform ablation studies to evaluate and analyze the generalization capability of the proposed method.

4.1. Dataset

We use the publicly available Chinese Power Line Insulator Dataset (CPLID) [21] as the benchmark dataset, with its details presented in

Table 4. Detailed description of the dataset.

		Normal Insulators	Defective Insulators
Image	Number	600	248
Insulator Scale	Small	0	0
	Medium	88	0
	Large	985	248
Number of Insulators in Images	Num = 1	247	248
	Num = 2	254	0
	Num = 3	81	0
	Num = 4	15	0
	Num = 5	3	0

. First, the dataset contains images of both normal and defective insulators. Specifically, normal insulators consist of 600 images of defect-free insulators captured in real scenarios, while defective insulators include 248 synthesized images of insulators with defects. Second, based on the COCO evaluation metrics for categorizing large, medium, and small objects ($Large\ge 96\times 96>Medium\ge 32\times 32>Small$), we analyzed the target distribution in CPLID. In both types of images, most insulators belong to the large-object category, which further highlights the characteristic of insulators having a high length-to-width ratio. Finally, in the normal insulator images, there are typically multiple insulators present, some of which are partially occluded. In contrast, the defective insulator images, being synthesized, usually contain only one insulator.

Some insulator images from the CPLID are shown in Figure 6. Due to the limited number of images in the dataset, we performed data augmentation on both the training and test sets after dividing them. The augmented images are shown in Figure 7. First, we divided the images of normal insulators and defective insulators into training and test sets in an 8:2 ratio. Secondly, we applied data augmentation operations to each image in both the training and test sets, such as rotation, motion blur, Gaussian blur, color jitter, scaling, and horizontal flipping. In addition, we simulated various real-world scenarios, such as rainy, foggy, and snowy conditions, to enhance the robustness of the model. The final training set consists of 9492 images, while the test set contains 2380 images.

4.2. Experiment Setup

Our experiments were carried out on a platform featuring an Nvidia RTX2080Ti GPU. The software setup consisted of Ubuntu 18.04 and Python 3.7.10, utilizing PyTorch 1.11.0 as the deep learning framework. To ensure fair comparison across experiments, we used consistent hyperparameter settings and data augmentation throughout. We utilized the Adam optimizer for training, initializing the learning rate at 0.001, using a batch size of 8, and setting the total number of epochs to 100.

4.3. Evaluation Metrics

To evaluate the model’s performance in detecting insulators and their defects, we used the mean Average Precision (mAP) as the evaluation metric. The mAP is calculated based on the average of the Average Precision (AP) values for each category, where AP is determined by recall and precision. Their formulas are as follows:

$$Recall={\displaystyle \frac{T_P}{T_P+F_N}}$$

(10)

$$Precision={\displaystyle \frac{T_P}{T_P+F_P}}$$

(11)

In this context, $T_P$, $F_P$, and $F_N$ represent the number of true positives, false positives, and false negatives in the detection results, respectively. The Average Precision (AP) is derived from the area under the precision–recall curve. Additionally, the COCO evaluation standard provides a more detailed breakdown of AP values based on the Intersection over Union (IoU) between predicted and ground truth boxes: $A{P}_{50}$ and $A{P}_{75}$ are the average precisions calculated at IoU thresholds of 0.50 and 0.75, respectively, and $A{P}_{50:95}$ is the average of the average precisions computed at multiple IoU thresholds ranging from 0.50 to 0.95, with a step size of 0.05.

4.4. Comparative Experiments with Mainstream Edge-Based Lightweight Algorithms

As shown in

Table 5. Performance comparison of our edge-based method and mainstream edge-based lightweight algorithms on insulator recognition and defect detection tasks.

Method	$\mathit{mA}{\mathit{P}}_{50}$	Params. (M)	FLOPs (G)
YOLOv8n	94.1%	3.2	8.7
YOLOv11n	96.4%	2.6	6.5
Song	97.3% (Acc.)	0	1.4
Ours	97.9% (Acc.)	0	0.7

Bold numbers represent the best value in each column of metrics.

, we compared our edge-based insulator coarse recognition algorithm with mainstream edge-based lightweight algorithms on the augmented CPLID dataset. It is worth noting that both our method and Song’s method [10] were evaluated using classification accuracy (Acc.). Our proposed edge-based method outperforms YOLOv8n [22], YOLOv11n [23], and Song’s method. More importantly, while achieving comparable accuracy, the floating point operations (FLOPs) of our method are only 0.7 G, which is significantly lower than those of other methods. Additionally, since we adopted a traditional algorithm, our method also achieves the best performance in terms of parameter size. This makes our method fully suitable for deployment on edge devices. Furthermore, the recognition results of our edge-based insulator coarse recognition algorithm are shown in Figure 8.

4.5. Comparative Experiments with Popular Attention Mechanisms

The comparative results of the proposed I-ECA with popular attention mechanisms for insulator recognition and defect detection tasks are shown in

Table 6. Comparison of I-ECA with mainstream attention mechanisms for insulator recognition and defect detection tasks.

Attention	$\mathit{mA}{\mathit{P}}_{50:95}$	$\mathit{mA}{\mathit{P}}_{50}$	$\mathit{mA}{\mathit{P}}_{75}$
ECA	67.1%	95.4%	81.0%
I-ECA	67.8%	95.1%	82.7%

Bold numbers represent the best value in each column of metrics.

. The experimental results indicate that in the I-YOLO model, which retains the I-FCFN module, our I-ECA module, designed specifically for the unique high aspect ratio of insulators, outperforms other mainstream attention mechanisms and is better suited to this task. We utilized the same hyperparameters and configurations across all experiments. Compared to other modules, the proposed I-ECA module exhibits superior performance in terms of $mA{P}_{50:95}$, $mA{P}_{50}$, and $mA{P}_{75}$ metrics.

Figure 8. Effect of the Coarse Recognition Algorithm for Insulators: (a) RGB. (b) Hue. (c) Combined Image of Otsu Threshold Segmentation Results. (d) Coarse Recognition Result of Insulators.

Figure 8. Effect of the Coarse Recognition Algorithm for Insulators: (a) RGB. (b) Hue. (c) Combined Image of Otsu Threshold Segmentation Results. (d) Coarse Recognition Result of Insulators.

4.6. Comparison Experiment with Mainstream Object Detection Models

As shown in

Table 7. Comparison of I-YOLO with mainstream detection models for insulator recognition and defect detection tasks.

Method	$\mathit{mA}{\mathit{P}}_{50:95}$	$\mathit{mA}{\mathit{P}}_{50}$	$\mathit{mA}{\mathit{P}}_{75}$
Miao	57.9%	89.1%	68.6%
Song	62.5%	95.6%	74.0%
Tao	59.3%	91.2%	70.2%
YOLOv8n	61.7%	94.1%	72.1%
YOLOv11n	67.3%	96.4%	81.1%
I-YOLO	69.8%	96.2%	84.5%

Bold numbers represent the best value in each column of metrics.

, we compared the performance of the proposed I-YOLO with mainstream detection models on the augmented CPLID dataset. The experimental results indicate that our proposed I-YOLO model performs better in detection effectiveness than other detection models. We evaluated the state-of-the-art methods, including Miao’s method [24], Song’s method [10], Tao’s method [21], YOLOv8n [22], and YOLOv11n [23]. Miao’s method employs the SSD model to accomplish the insulator detection task; both Song’s and Tao’s methods use a VGG16-based Faster RCNN. Additionally, in Song’s method, the number and size of anchors were redesigned, introducing new aspect ratios of 1:8, 1:4, 4:1, and 8:1, with a new area scale of 64, while Tao’s method remains unchanged. YOLOv8n is the most lightweight model in the YOLOv8 series, while YOLOv11n represents the latest YOLO model to date. Compared to other models, the proposed I-YOLO demonstrates superior performance in insulator recognition and defect detection tasks. Our $mA{P}_{50}$ is comparable to that of YOLOv11n, but we achieved improvements of 2.5% and 3.4% in $mA{P}_{50:95}$ and $mA{P}_{75}$, respectively. Figure 9 illustrates the detection effectiveness of the I-YOLO model, where it can be clearly seen that the proposed model effectively detects defects within insulators. Moreover, our method achieves excellent detection performance even for partially occluded insulators.

4.7. Ablation Study

To demonstrate the efficacy of the proposed I-ECA module and I-FCFN, ablation experiments were performed, as presented in

Table 8. Ablation experiment results.

I-ECA	I-FCFN	$\mathit{mA}{\mathit{P}}_{50:95}$	$\mathit{mA}{\mathit{P}}_{50}$	$\mathit{mA}{\mathit{P}}_{75}$
-	-	64.3%	94.9%	73.5%
✓	-	67.8%	95.1%	82.7%
-	✓	67.5%	95.5%	81.9%
✓	✓	69.8%	96.2%	84.5%

Bold numbers represent the best value in each column of metrics.

. The experimental results indicate that utilizing either the I-ECA module or the I-FCFN alone enhances the detection accuracy of the I-YOLO model, further verifying the efficacy of the modules introduced in this study. Specifically, in terms of the $mA{P}_{50:95}$, $mA{P}_{50}$, and $mA{P}_{75}$ metrics, the I-ECA module has a more significant impact on detection accuracy compared to the I-FCFN. Moreover, when combined with the I-FCFN, the I-YOLO model demonstrated improvements in the performance metric.

5. Discussion

The cloud–edge collaborative intelligence strategy proposed in this paper demonstrates significant advantages in drone-based insulator recognition and defect detection tasks. By employing the edge insulator coarse recognition algorithm and the cloud-based I-YOLO model, the study achieved higher accuracy on benchmark datasets. To verify the generalizability of the proposed I-ECA module, it was integrated into various models. As shown in

Table 9. Evaluation of method generalization capability.

Method	$\mathit{mA}{\mathit{P}}_{50:95}$	$\mathit{mA}{\mathit{P}}_{50}$	$\mathit{mA}{\mathit{P}}_{75}$
Faster R-CNN	58.3%	90.2%	69.4%
Faster R-CNN+	59.1%	92.6%	71.8%
SSD	56.7%	88.1%	68.3%
SSD+	57.3%	91.0%	69.1%

, the Faster R-CNN+ and SSD+ models with the I-ECA module outperform the original Faster R-CNN and SSD models. Furthermore, these models utilize ResNet50 as their backbone.

Although most images in the benchmark dataset were captured under clear weather conditions, we utilized data augmentation techniques to cover real-world scenarios such as rainy, foggy, and snowy weather, thereby enhancing the generalization capability of our method. Furthermore, we compared the proposed edge-based method with mainstream edge-based lightweight algorithms. The results demonstrate that our method not only achieves the lowest computational complexity but also delivers higher accuracy.

We plan to conduct the following research in the future: Firstly, build an insulator dataset under complex scenarios to enhance model robustness. Secondly, continue optimizing intelligent algorithms at both the edge and cloud levels to enrich algorithm functionalities and improve detection accuracy. Then, complete the deployment of drones to optimize system efficiency and energy consumption. Lastly, we plan to extensively test the cloud–edge collaborative intelligence strategy for use in other domains, such as defect detection in industrial production environments and vehicle and pedestrian detection on roads.

6. Conclusions

In order to reduce computational energy consumption during the process of insulator recognition and defect detection using drones, we propose a cloud–edge collaborative intelligence strategy. By deploying a coarse insulator recognition algorithm on the drone, we can determine whether the captured images contain insulator targets. Meanwhile, the I-YOLO model deployed on the server, equipped with the I-ECA module and I-FCFN, enables superior detection of insulators and their small-scale defects, thereby enhancing the model’s robustness in challenging environments. The experimental results demonstrate that the cloud–edge collaborative strategy performs exceptionally well on the insulator string dataset. The edge-based algorithm achieves an accuracy of 97.9% with only 0.7 G FLOPs, meeting the inspection requirements of drones. Meanwhile, the cloud-based model achieves an $mA{P}_{50}$ of 96.2%, enabling accurate detection of insulators and their defects.