YOLOLS: A Lightweight and High-Precision Power Insulator Defect Detection Network for Real-Time Edge Deployment

Abstract

Real-time insulator defect detection is critical for ensuring the reliability and safety of power transmission systems. However, deploying deep learning models on edge devices presents significant challenges due to limited computational resources and strict latency constraints. To address these issues, we propose YOLOLS, a lightweight and efficient detection model derived from YOLOv8n and optimized for real-time edge deployment. Specifically, YOLOLS integrates GhostConv to generate feature maps through stepwise convolution, reducing computational redundancy while preserving representational capacity. Moreover, the C2f module is restructured into a ResNet–RepConv architecture, in which convolution and Batch Normalization layers are fused during inference to reduce model complexity and enhance inference speed. To further optimize performance, a lightweight shared-convolution detection head significantly reduces parameter count and computational cost without compromising detection accuracy. Additionally, an auxiliary bounding box mechanism is incorporated into the CIoU loss function, improving both convergence speed and localization precision. Experimental validation on the CPLID dataset demonstrates that YOLOLS achieves a 42.4% reduction in parameters and a 48.1% decrease in FLOPs compared to YOLOv8n while maintaining a high mAP of 91%. Furthermore, when deployed on Jetson Orin NX, YOLOLS achieves 44.6 FPS, ensuring real-time processing capability. Compared to other lightweight YOLO variants, YOLOLS achieves a better balance between accuracy, computational efficiency, and inference speed, making it an optimal solution for real-time insulator defect detection in resource-constrained edge computing environments.

1. Introduction

Insulators are vital components in power transmission systems, providing both electrical insulation and mechanical support under challenging environmental conditions [1]. However, common defects such as string-dropping and physical damage pose serious risks to grid stability and safety [2]. Therefore, timely defect detection and maintenance are thus of critical importance [3]. Traditional manual inspections characterized by low efficiency, high cost, and a substantial risk of oversight have long been the standard practice [4]. The advent of high-performance computing and deep learning has led to more efficient and accurate methods for insulator defect detection [5]. In particular, combining unmanned aerial vehicle (UAV) imagery with deep learning algorithms has emerged as the predominant method, enabling aerial imagery to be processed for the automatic identification of insulator defects [6]. Compared to classical, machine learning-based target detection techniques, deep learning typically offers higher recognition accuracy and speed [7], resulting in notable successes in power transmission defect detection tasks [8]. Moreover, as UAVs and edge computing devices become integral to transmission line inspections, real-time defect detection has gained increasing importance by facilitating immediate anomaly identification and prompt maintenance to mitigate potential safety risks and enhance grid reliability.

Deep learning-based object detection algorithms can be broadly categorized into two-stage and one-stage methods. Two-stage approaches, such as RCNN [9], SPPNet [10], Faster R-CNN [11], Mask R-CNN [12], and R-FCN [13], first generate candidate regions and then refine detections. Although highly accurate, these methods often suffer from substantial computational overhead and large parameter counts, making them less suitable for edge-computing applications. In contrast, one-stage algorithms, including YOLO [14], SSD [15], CenterNet [16], and DSSD [17], directly predict both categories and locations from input images, achieving faster detection speeds and improved real-time performance [18], which makes one-stage methods particularly attractive for insulator defect detection in the field. Recent adaptations of these methods have been tailored specifically to the challenges of insulator defect detection, refining the balance between accuracy and speed for practical applications.

Building on one-stage algorithms, recent research has targeted the lightweight design of detection models to facilitate deployment on resource-constrained devices. Modifications in one-stage YOLO-based frameworks include using alternative activation functions, such as Mish, combined with dynamic convolution operations to enhance feature representation and accelerate model convergence [19]. Qiu et al. [20] employed the MobileNet network to replace YOLOv4’s feature extraction backbone and used Laplace sharpening to improve insulator detection accuracy. Lan et al. [21] integrated the Ghost module into the YOLOv5 backbone and introduced the CBAM attention mechanism into the neck, enhancing accuracy and detection speed while reducing model size. Building on this trend, Li et al. [22] developed LiteYOLO-ID by designing a novel lightweight convolutional module (EGC, or ECA-GhostNet-C2f) and constructing EGC-CSPGhostNet to replace the standard YOLOv5S backbone. They also proposed a lightweight neck network (EGC-PANet), which reduced parameters by 47.13%, increased mean average precision (mAP@0.5) by 1%, and achieved a detection speed of 20.2 FPS on an NVIDIA Jetson TX2 NX. Zhang et al. [23] advanced a lightweight YOLOv7-based algorithm by introducing the DSC-SE module, which combines depthwise separable convolutions with the SE channel attention mechanism, into the backbone, replacing standard convolutions with grid-sensitive convolutions in the feature fusion step and employing EIoU-loss. Although this approach reduced floating-point operations (GFLOPs) by 54.5%, model size by 37.8%, and improved accuracy by 4.9%, its detection speed remained at 13 FPS on an NVIDIA Jetson Nano, which is insufficient for real-time applications. More recently, Liu et al. [24] proposed YOLO-PowerLite, a lightweight YOLO model optimized for transmission line defect detection. Their network introduces the C2f_AK module and a bidirectional feature pyramid, merging multiple convolutions in the decoupling head. This approach achieved the same accuracy as YOLOv8n while reducing parameters by 42.3%, FLOPs by 30.9%, and model size by 40.4%, though without an improvement in detection speed.

While previous studies have predominantly concentrated on YOLO-based frameworks, recent work on power insulator defect detection highlights promising non-YOLO alternatives. For example, specialized methods such as lightweight Faster R-CNN variants [25] and FusionNet [26] have demonstrated competitive performance, particularly in scenarios with complex backgrounds and variable target scales. In addition, transformer-based architectures and strategies incorporating super-resolution modules have been proposed to overcome the challenges of detecting small insulator defects, providing new avenues for integrating diverse detection paradigms [27]. Moreover, approaches addressing issues of sample imbalance and incomplete annotations using positive-unlabeled learning and focal loss have been developed to further refine detection precision in real-world scenarios [28]. These findings underscore the importance of comparing and combining both YOLO and non-YOLO approaches for a comprehensive insulator monitoring solution.

Recent investigations have addressed these challenges by incorporating advanced loss functions, such as SIoU [29], as well as innovative feature representation modules, including the MCI-GLA plug-in [30] and 3D attention-focused architectures [31]. Additionally, methods like ML-YOLOv5 [18] and SnakeNet [32] further illustrate the potential of integrating both YOLO-based and alternative strategies to enhance detection under varied conditions. Foundational methods such as GhostConv and RepConv have also demonstrated specific advantages in power infrastructure monitoring. GhostConv efficiently expands feature maps with fewer parameters [33], while RepConv fuses convolution and Batch Normalization during inference to reduce computational overhead without sacrificing accuracy [34]. Such analyses are essential for understanding the technical benefits and limitations of these modules in the context of insulator defect detection, a topic that has received limited attention in earlier studies. Furthermore, several works have optimized algorithms for deployment on UAV platforms and edge devices, achieving lower power consumption and reduced inference latency through model segmentation, efficient layer mapping, and lightweight backbones [35]. These studies collectively demonstrate a multifaceted strategy that enhances insulator detection by synergistically improving computational efficiency, feature extraction, and loss function design.

Despite these advancements, insulator defect detection still demands even higher computational efficiency and resource optimization in edge environments. Real-time performance on devices with limited hardware remains a significant challenge, as many existing models struggle to balance accuracy and inference speed under these constraints. To address this need for lightweight solutions, we propose YOLOLS, a lightweight, high-speed model derived from YOLOv8n that is specifically optimized for real-time insulator defect detection in resource-constrained settings. Our model leverages recent innovations in model re-parameterization, loss function design, and feature representation to balance detection accuracy with minimal computational overhead. The main contributions of this paper are as follows:

• Our approach replaces standard convolutions with GhostConv to achieve efficient feature map expansion, thereby reducing parameter count and computational cost. Additionally, we introduce a novel restructuring of the C2f module into a ResNet–RepConv configuration, which enables the fusion of convolution and Batch Normalization during inference, further streamlining the network for lightweight feature extraction.
• We developed a shared convolutional detection head, a weight-sharing mechanism across multiple detection scales that minimizes redundant computations and fosters a unified representation, to enhance multi-scale feature alignment and improve detection robustness.
• By integrating an auxiliary bounding box, e.g., Inner-IoU, into the existing CIoU loss framework, our model becomes more sensitive to minor discrepancies between predicted and ground truth boxes, expediting convergence and improving localization accuracy, especially for small or partially occluded insulators.
• We extensively validate YOLOLS on the CPLID and IDID datasets, encompassing a wide range of real-world insulator scenarios. Comparative experiments confirm that our model outperforms mainstream algorithms in both accuracy and inference speed under edge-computing constraints. These results confirm that the architectural modifications and loss function enhancements in YOLOLS provide a robust, rapid, and cost-effective solution for real-time insulator defect detection in challenging, resource-limited environments.

The remainder of this paper is organized as follows. Section 2 reviews related work and details the proposed approach, including the network architecture modifications, module construction, and improvements to the loss function. Section 3 presents comprehensive experimental evaluations and comparisons with state-of-the-art methods. Finally, Section 4 concludes the paper and outlines potential directions for future research.

2. Proposed Methodology and Model Construction

2.1. Proposed Model Architecture

The YOLOv8 series exhibits significant advancements in both accuracy and inference speed compared to earlier YOLO models, rendering it a popular choice for various detection scenarios [36]. Among its variants, as shown in Figure 1, the YOLOv8n model achieves optimal computational efficiency and parameter economy while retaining competitive accuracy, rendering it particularly suitable for edge-device deployment. Building on this foundation, we adopt YOLOv8n as our baseline framework and introduce targeted architectural refinements to optimize performance under resource-limited conditions. As illustrated in Figure 2, our enhanced design focuses on restructuring the backbone, neck, and head modules to accommodate resource-constrained conditions. Specifically, we replace standard convolutional layers with GhostConv, a lightweight operator that generates additional feature maps by applying linear transformations to a subset of standard convolution outputs. This two-stage approach eliminates redundant computations while preserving representational capacity. Next, we modify the C2f module by integrating ResNet-style residual connections and RepConv re-parameterization techniques. This ResNet–RepConv architecture benefits from residual skip connections during training and merges convolution and Batch Normalization operations into a single branch at inference. Consequently, the network achieves higher throughput without sacrificing performance. In the detection head, we introduce a shared-convolution paradigm in which multiple detection scales reuse a single set of convolutional filters. This strategy not only curtails the overall parameter count but also aligns multi-scale features, thereby streamlining the detection pipeline. Finally, to improve bounding-box regression, we replace the standard CIoU loss with Inner-CIoU, an enhanced loss function that incorporates auxiliary bounding boxes to fine-tune geometric alignment during optimization. By emphasizing high-IoU regions, Inner-CIoU accelerates convergence and refines localization accuracy.

Collectively referred to as YOLOLS, these architectural enhancements balance speed and accuracy under resource-limited conditions. Moreover, the design accommodates potential future optimizations, such as integer quantization or hardware-specific acceleration libraries (e.g., TensorRT), to further reduce latency on embedded platforms.

2.2. Inner-CIoU

YOLOv8 employs CIoU [37] to more comprehensively capture the geometric relationships between predicted boxes and target boxes. As depicted in Figure 3, the CIoU loss depends on factors such as box overlap, center-point distance, and aspect ratio. When the anchor box varies in position relative to the target box, different similarity measures arise. The CIoU formulation is presented in Figure 3 and Equations (1)–(3):

$$L_{\mathrm{CIoU}}=1- \mathrm{IoU}+{\displaystyle \frac{{\rho }^2\left(\mathbf{b},{\mathbf{b}}^{\mathrm{gt}}\right)}{c^2}}+\alpha \upsilon$$

(1)

where IoU denotes the intersection-over-union between the predicted bounding box b and the ground truth bounding box b^gt. ${\rho }^2\left(\mathbf{b},{\mathbf{b}}^{\mathrm{gt}}\right)$ is the squared Euclidean distance between the centers of the predicted box and the ground truth box. c represents the diagonal length of the smallest enclosing box that tightly encloses both b and b^gt. $\upsilon$ is a term measuring the discrepancy in aspect ratios between the predicted box and the ground truth box. It is computed as:

$$\upsilon ={\displaystyle \frac{4}{\pi }}{\left(\arctan {\displaystyle \frac{w^{\mathrm{gt}}}{h^{\mathrm{gt}}}}-\arctan {\displaystyle \frac{w}{h}}\right)}^2$$

(2)

where (w, h) and (w^gt, h^gt) are the width and height of the predicted box and the ground truth box, respectively. $\alpha$ is a weighting factor that dynamically balances the IoU term and the aspect ratio term:

$$\alpha ={\displaystyle \frac{\upsilon }{(1-\mathrm{IoU})+\upsilon }}$$

(3)

Putting all these components together, CIoU improves on the traditional IoU-based losses (and its variants, such as DIoU) by taking both the distance between centers and the aspect ratio into consideration, leading to more stable and accurate bounding box regression in deep learning-based object detection systems.

Although CIoU offers a more holistic measure of box overlap, its convergence can be sluggish. To overcome this limitation, we introduce the Inner-IoU concept [38], which leverages auxiliary bounding boxes to refine the loss calculation (as shown in Equations (4)–(14)). This innovation enhances the model’s localization precision and accelerates the regression of predicted boxes. The calculation formulas of Inner-IoU are as follows:

$$b_r^{gt}=x_c^{gt}+{\displaystyle \frac{w^{gt}\times ratio}{2}}$$

(4)

$$b_t^{gt}=y_c^{gt}-{\displaystyle \frac{h^{gt}\times ratio}{2}}$$

(5)

$$b_b^{gt}=y_c^{gt}+{\displaystyle \frac{h^{gt}\times ratio}{2}}$$

(6)

$$b_l=x_c-{\displaystyle \frac{w\times ratio}{2}}$$

(7)

$$b_r=x_c^{gt}+{\displaystyle \frac{w\times ratio}{2}}$$

(8)

$$b_t=y_c-{\displaystyle \frac{h\times ratio}{2}}$$

(9)

$$b_b=y_c^{gt}+{\displaystyle \frac{h\times ratio}{2}}$$

(10)

$$inter=(min(b_r^{gt},b_r)-max(b_l^{gt},b_l))(min(b_b^{gt},b_b)-max(b_t^{gt},b_t))$$

(11)

$$union=(w^{gt}\times h^{gt})\times {(ration)}^2+(w\times h)\times {(ration)}^2-inter$$

(12)

$${\mathrm{IoU}}_{\mathrm{inner}}={\displaystyle \frac{inter}{union}}$$

(13)

$${\mathrm{L}}_{\mathrm{inner}-\mathrm{CIoU}}{=\mathrm{L}}_{\mathrm{CIoU}}{+\mathrm{IoU} - \mathrm{IoU}}_{\mathrm{inner}}$$

(14)

In this study, the parameter ratio is set to 0.7, ensuring that the auxiliary bounding box is smaller than the actual target box. Although this results in a narrower effective regression range, the gradient’s absolute value is larger than that obtained using standard IoU, thus promoting faster convergence for high-IoU samples. Consequently, the model converges faster and achieves more precise localization, particularly for small or partially occluded objects. This approach is especially advantageous for insulator defect detection, where unobstructed, high-quality views of targets may be limited.

2.3. Lightweight Shared Convolutional Detection Head

In the original YOLOv8 architecture, the detection head significantly contributes to the model’s overall parameter count and computational load. This component typically relies on multiple 3 × 3 and 1 × 1 convolution layers to predict bounding boxes and classification probabilities, adding approximately 898k parameters—over 30% of the network’s total. Such complexity may hinder real-time performance, particularly on devices with limited GPU or CPU capacity, especially when deployed on resource-constrained edge devices.

To address this challenge, we propose a Shared Convolution strategy that applies a single set of convolutional filters across multiple detection scales., as illustrated in Figure 4. This weight-sharing mechanism trims redundant computations and yields a more compact model without degrading feature extraction quality. Furthermore, we replace Batch Normalization (BN) with Group Normalization (GN). BN often performs sub-optimally when batch sizes are small, a common condition in edge-computing environments [39]. In contrast, GN normalizes feature responses along group-wise channels rather than across the mini-batch, making it less sensitive to fluctuations arising from small or variable batch sizes. This advantage helps maintain robust performance in both localization and classification tasks while preserving overall model accuracy. By combining shared convolutions with GN, our redesigned detection head markedly reduces parameter usage and inference time, making it better suited for latency-critical applications. This optimization proves indispensable in scenarios where on-device processing must be executed in real-time, such as during field inspections of power lines.

2.4. ResNet–RepConv

The C2f module in YOLOv8 draws on residual network designs by incorporating N Bottleneck blocks. Although these Bottleneck blocks, with their multiple convolution operations, can capture more complex features, they also increase model complexity and reduce detection speed. As depicted in Figure 5, we address this limitation by removing the Bottleneck structure. To compensate for potential performance loss, we incorporate RepConv [40] into the gradient-flow branch. RepConv strengthens both feature extraction and gradient flow, thereby maintaining detection performance while lowering model complexity and improving inference speed.

RepConv applies the core principle of re-parameterization. During training, it uses a multi-branch convolutional structure. During inference, its branch parameters are re-parameterized onto a single main branch, and convolution layers are fused with normalization layers, thus reducing computational overhead. The following equations illustrate its operation:

$$\mathrm{Conv}\left(x\right)=W\left(x\right)+b$$

(15)

The calculation formula for the BN is as follows:

$${\mu }_B={\displaystyle \frac{1}{m}}{\displaystyle \sum _{i=1}^m{x}_i}$$

(16)

$${\sigma }_B^2={\displaystyle \frac{1}{m}}{\displaystyle \sum _{i=1}^m{(x_i-{\mu }_B)}^2}$$

(17)

$${\widehat{x}}_i={\displaystyle \frac{x_i-{\mu }_B}{\sqrt{{\sigma }_B^2+\epsilon }}}$$

(18)

$$\mathrm{BN}(x)=\gamma \cdot {\widehat{x}}_i+\beta$$

(19)

Here, γ and β are learnable parameters during the training process, while ε is a small positive constant. Substituting the convolution results into the BN formula and simplifying yields:

$$\mathrm{BN}(\mathrm{Conv}(x))={\displaystyle \frac{\gamma \cdot W(x)}{\sqrt{{\sigma }_B^2+\epsilon }}}+{\displaystyle \frac{\gamma (b-{\mu }_B)}{\sqrt{{\sigma }_B^2+\epsilon }}}+\beta$$

(20)

where the term ${\displaystyle \frac{\gamma \cdot W(x)}{\sqrt{{\sigma }_B^2+\epsilon }}}$ can be viewed as $w_1\left(x\right)$ and ${\displaystyle \frac{\gamma \left(b-{\mu }_B\right)}{\sqrt{{{\sigma }_B}^2+\epsilon }}}+\beta$ as $b_1$. Thus, the combined operations of convolution and Batch Normalization can be expressed more concisely as a single convolution:

$$\mathrm{BN}\left(\mathrm{Conv}\left(x\right)\right)=w_1 x+b_1$$

(21)

Through this re-parameterization strategy, RepConv maintains a relatively complex multi-branch structure during training for enhanced feature learning and condenses these operations into a single branch during inference. Consequently, the network benefits from richer representational capabilities while minimizing the computational burden in real-world deployment.

2.5. GhostConv

Unlike standard convolution operations, GhostConv [41] divides the process into two distinct stages, as depicted in Figure 6. First, a smaller feature map is generated using a standard convolution, which lowers the computational cost. Then, depthwise separable group convolution is applied to this intermediate feature map, producing additional feature channels. Finally, the two resulting feature maps are concatenated to form the final output. By segmenting the convolution operation into stepwise procedures, GhostConv reduces parameter count and computational overhead while maintaining the same number of output feature maps as a conventional convolution layer [42]. This design not only preserves performance but also saves computational resources, thereby improving model detection speed—an advantage particularly valuable for real-time or resource-constrained applications.

3. Experimental Results and Discussion

3.1. Experimental Details

To evaluate the effectiveness of the improved model, we use the Chinese Power Line Insulator Dataset (CPLID) [43], an open-source dataset (available at: https://github.com/InsulatorData/InsulatorDataSet (accessed on 23 March 2025)) designed explicitly for detecting power line insulator defects. This dataset includes 600 normal insulator images and 248 synthetically generated defective insulator images, yielding a total of 848 images. It features two labels: “insulator string” and “insulator defect”. As depicted in Figure 7, we employ various data augmentation techniques—such as random cropping, flipping, and color perturbations—to enhance model generalization. This process triples the number of original images. The expanded dataset is then divided into a 7:2:1 ratio, resulting in a training set of 1780 images, a test set of 509 images, and a validation set of 255 images.

Our experiments utilize two distinct hardware platforms—one for high-performance model training and another for resource-efficient edge deployment—to accommodate the full lifecycle of real-time power insulator defect detection. On the training side, we employ a server with an RTX 4090D GPU (24 GB VRAM) and an Intel(R) Xeon(R) Platinum 8474C CPU, manufactured by Intel Corporation, Penang, Malaysia, running Ubuntu 20.04. This setup leverages Python 3.8, PyTorch 1.11.0, and CUDA 11.3, enabling large-scale computations, rapid prototyping, and accelerated deep-learning workflows. The specific training hyperparameters used in our experiments are detailed in

Table 1. Training parameters.

Parameter	Value
Epochs	200
Batch	32
Image Size	640 × 640
Optimizer	SGD
Initial learning rate	0.01
Workers	8

. For on-site or near real-time deployment, we turn to a Jetson Orin NX (16 GB) module, also running Ubuntu 20.04 but optimized with TensorRT 8.5.2.2 and CUDA 11.4. This edge-oriented hardware offers a careful balance between inference speed and power consumption, making it especially suitable for field inspections or mobile robotic platforms (e.g., UAVs). By integrating TensorRT optimizations—such as layer fusion, kernel auto-tuning, and precision calibration (FP16 or INT8)—the model can achieve sub-25-ms inference latency for a single image, facilitating near real-time insulator defect detection even under stringent resource constraints. Figure 8 provides a high-level overview of our technical pipeline for real-time power insulation inspections. The workflow begins with data acquisition via UAVs, which capture images of insulators from various angles. These images are then preprocessed (e.g., resized, normalized, or augmented) before being fed into the trained YOLOLS model on the Jetson Orin NX device. Through on-device inference, defects are identified in real time, and bounding boxes or other relevant annotations are overlaid on the original images. The resulting predictions can either be stored locally for post-inspection analysis or transmitted wirelessly to a central control station for immediate operator review.

By adopting this integrated approach, spanning high-performance offline training and resource-optimized edge inference, we achieve a robust, scalable solution for real-time power insulator defect detection. This methodology not only minimizes downtime and labor expenses but also significantly mitigates the risk of catastrophic failures in power systems, showcasing the practical value and economic benefits of edge computing in industrial inspection settings.

3.2. Evaluation Metrics

To assess the detection performance of our algorithm, we use the mean Average Precision (mAP). Equations (22)–(25) define the calculation of mAP:

$$\mathrm{P}={\displaystyle \frac{TP}{TP+FP}}$$

(22)

$$\mathrm{R}={\displaystyle \frac{TP}{TP+FN}}$$

(23)

$$\mathrm{AP}={\displaystyle {\int }_0^{1}P}(R)dR\times 100\%)$$

(24)

$$\mathrm{mAP}={\displaystyle \frac{{\displaystyle \sum _{i=0}^{m}A}P}{m}}\times 100\%$$

(25)

where FP denotes the number of false-positive predictions, FN indicates the number of false-negative predictions, and TP represents the number of true positives. AP measures the detection accuracy for a single category, while mAP is the mean value of AP across multiple categories.

We also incorporate GFLOPs, Params, and FPS to evaluate model complexity and computational efficiency. GFLOPs (Giga Floating-Point Operations) and Params (number of parameters) quantify the model’s computational cost, whereas FPS indicates the model’s detection speed. Their respective equations appear in (26)–(28):

$$\mathrm{GFLOPs}=W\times H\times K^2\times C_{in}\times C_{out}$$

(26)

$$\mathrm{Params}=C_{in}\times C_{out}\times K^2$$

(27)

$$\mathrm{FPS}={\displaystyle \frac{1}{T}}$$

(28)

In these formulas, W and H represent the width and height of the feature map, C_in and C_out correspond to the input and output channels, K denotes the convolution kernel size, and T is the processing time per image.

3.3. Ablation Experiments

To systematically assess how each proposed modification affects accuracy, model size, and inference speed in insulator defect detection, we performed eight ablation experiments.

Table 2. Ablation experiment results.

Model	Inner-CIoU	LSCDH	GhostConv	ResNet–RepConv	P (%)	R (%)	mAP (%)	Params (M)	GFLOPs	FPS (Jetson Orin NX)
Yolov8n (Baseline)	–	–	–	–	92.7	86.6	91.1	3.01	8.1	36.2
A	√		–	–	96.0	85.7	92.3	3.01	8.1	36.3
B	–	√	–	–	94.4	87.7	91.9	2.30	6.5	38.3
C	–	–	√	–	92.7	85.6	91.2	2.73	7.5	34.5
D	–	–	–	√	95.3	84.4	91.1	2.20	6.0	41.7
E	√	√	–	–	93.4	88.2	92.5	2.36	6.5	38.6
F	√	√	√	–	93.9	82.7	90.9	2.09	6.1	36.2
YOLOLS	√	√	√	√	92.2	86.9	91.0	1.27	3.9	44.6

summarizes the incremental contributions of our proposed modules—Inner-CIoU, Lightweight Shared Convolutional Detection Head (LSCDH), GhostConv, and ResNet–RepConv—when integrated with the baseline YOLOv8n model.

In Table 2, a “√” indicates that the corresponding module has been incorporated into the baseline model. Model A replaces CIoU with Inner-CIoU, increasing mAP by 1.2% without impacting computational cost or inference speed. This result demonstrates that auxiliary bounding boxes enhance localization precision by improving overlapping emphasis. Incorporating LSCDH significantly reduces parameters (down to 2.30 M from 3.01 M) and lowers FLOPs from 8.1 to 6.5, while maintaining competitive accuracy (91.9%). This highlights LSCDH’s capability to streamline the detection pipeline. Substituting standard convolutions with GhostConv reduces parameters to 2.73 M and FLOPs to 7.5 but slightly lowers mAP (91.2%) and FPS (34.5), suggesting a trade-off between stepwise convolution savings and potential inference overhead. Replacing the C2F module with ResNet–RepConv achieves the largest reduction in parameters (2.20 M) and FLOPs (6.0), increasing FPS to 41.7 without affecting accuracy (91.1%). Model E integrates both Inner-CIoU and LSCDH, achieving the highest mAP (92.5%) with a modest improvement in FPS (38.6). In Model F, adding GhostConv further cuts parameters (2.09 M) and FLOPs (6.1), but the FPS and mAP both declines compared to Model E, highlighting the balance between lightweight design and performance.

Building on Model E, YOLOLS also replaces C2F with ResNet–RepConv. This change slashes FLOPs to 3.9—less than half that of YOLOv8n—and trims parameters to 1.27 M, only 40% of the baseline. The FPS leaps to 44.6 on Jetson Orin NX, an improvement of 8.4 FPS over YOLOv8n, while maintaining a mAP of 91%. The 0.1% drop in accuracy is negligible given the significant gains in speed and resource efficiency. These ablations confirm that each component (Inner-CIoU, LSCDH, GhostConv, and ResNet–RepConv) individually contributes to either accuracy or efficiency. Their combined impact makes YOLOLS a lightweight yet powerful detector for insulator defects, ensuring high-speed inference and robust accuracy in resource-constrained environments.

3.4. Comparative Experiments

To benchmark YOLOLS against state-of-the-art lightweight detection models, we compare it with YOLOv4tiny, Yolov5s, YOLOv7tiny, Yolov8s, Yolov8n, Yolov9t [44], and YOLOv10n [45]. In addition, we compared the non-YOLO models with our model. The results are summarized in

Table 3. Comparative experimental results.

Model	mAP (%)	Params (M)	GFLOPs	FPS (Jetson Orin NX)
Faster R-CNN	77.9	28.5	948.4	1.5
Efficientdet	80.8	6.6	12.6	4.7
SSD	65.4	14.3	67.0	11.6
CenterNet	84.9	32.7	109.7	10.0
YOLOv4tiny	84.4	6.1	16.5	36.3
YOLOv5s	90.8	2.5	7.1	36.3
YOLOv7tiny	89.1	6.0	13.2	20.0
YOLOv8s	94.2	11.1	28.4	19.1
YOLOv8n	91.1	3.0	8.1	36.2
YOLOv9t	91.7	2.0	7.6	20.9
YOLOv10n	90.6	2.3	6.5	34.7
YOLOv8-ACCW [46]	90.9	2.8	7.5	5.9
YOLO-POWER [24]	91.1	1.6	5.2	9.5
YOLOLS(OURS)	91.0	1.3	3.9	44.6

The best results are highlighted in bold.

.

As shown in Table 3, while YOLOLS’s mAP (91.0%) is slightly lower than the highest value (94.2% for Yolov8s), it outperforms most other models in absolute terms, remaining competitive despite having the fewest parameters (1.27 M). With an FPS of 44.6, YOLOLS outpaces all other contenders, including the comparably small Yolov9t (20.9 FPS) and Yolov10n (34.7 FPS). This speed advantage is particularly crucial for real-time edge applications. YOLOv8s, although highly accurate, consumes 11.1 M parameters and 28.4 GFLOPs, resulting in only 19.1 FPS on Jetson Orin NX—unsuitable for strict real-time scenarios. In contrast, YOLOLS processes more than twice as many frames (44.6 FPS) with fewer parameters and significantly lower FLOPs. Additionally, the number of parameters and the computational effort of the non-YOLO models are on the large side, resulting in very slow detection.

Figure 9 visually compares each model’s accuracy, speed, and parameter size. YOLOLS strikes an optimal balance, ensuring cost-effective detection with minimal hardware demands and meeting real-time constraints in edge environments.

3.5. Robustness Experiment of Model

In practical field settings, visual inputs may suffer from challenging environmental conditions like direct glare or fog occlusion. To examine model robustness, we introduced high-exposure and synthetic fog perturbations.

Table 4. Detection performance under challenging conditions.

Model	Enhancement Operation	P (%)	R (%)	mAP (%)
YOLOv8n	High-exposure	91.6	79.8	90.4
	Synthetic Fog	88.8	79.7	85.3
YOLOLF(OUR)	High-exposure	89.1	81.4	90.4
	Synthetic Fog	89.4	77.3	84.3

outlines the performance drops for both YOLOLS and YOLOv8n under these scenarios. As shown in Table 4, under high-exposure conditions, both YOLOLS and YOLOv8n achieve an mAP of 0.904, demonstrating strong adaptability to high-intensity lighting environments with minimal performance degradation. However, under synthetic fog conditions, the performance of both models declines significantly due to visual obstructions. YOLOLS experiences a 6.7% drop in mAP, whereas YOLOv8n shows a 5.8% decline. Despite this reduction, YOLOLS maintains a relatively high mAP of 0.843, confirming that its detection accuracy remains robust even in adverse weather conditions.

Figure 10 presents the detection results of both models across different scenarios. From left to right, the images depict the detection results on original images, high-exposure environments, and synthetic fog conditions, respectively. In Figure 10, red bounding boxes highlight missed detections. It is evident that in synthetic fog environments, smaller insulator defect targets are more likely to be obscured, leading to false negatives. However, in most cases, both models effectively identify insulators and defect regions, and their detection performance remains comparable. Despite its lightweight design, YOLOLS maintains robustness comparable to YOLOv8n, effectively handling complex environmental conditions. It demonstrates strong adaptability to high-exposure environments and maintains high detection accuracy under foggy conditions. These results confirm that YOLOLS is suitable for real-world insulator defect detection applications, ensuring efficient and reliable performance in challenging scenarios.

In addition, to validate the robustness and generalizability of our model, we conducted comparative experiments on another public power insulator dataset, IDID [47]. We compared our model with various detection architectures, including those specifically designed for insulator defect detection. The IDID dataset comprises 1600 high-resolution images of transmission line insulators, annotated with three categories: flashover damage insulator, broken insulator, and insulator string. As shown in

Table 5. Experimental results of different detection models on the public IDID dataset.

Models	P (%)	R (%)	mAP@50 (%)
Improved YOLOv7 [19]	-	-	88.7
YOLOv6-L [19]	-	-	83.4
Cascade R-CNN [18]	-	-	91.4
Dynamic R-CNN [18]	-	-	93.8
YOLOV7+MCI-GLA [18]	-	-	96.1
YOLOV3 (Darknet) [28]	-	-	94.8
SSD	89.3	73.8	84.1
Faster R-CNN	58.6	88.8	81.4
YOLOv8n	92.7	88.5	94.6
YOLOLS (OURS)	94.3	88.3	94.1

, the results for the first six models were directly sourced from the literature, while the remaining results were obtained from our experiments. Our findings indicate that our model is highly competitive, achieving high performance while maintaining a lightweight architecture.

Lastly, we generate feature-layer heatmaps using HiResCAM [48] for the 10th, 12th, 14th, 16th, and 18th layers of YOLOv8n and YOLOLS to visualize each network’s attention distribution (Figure 11). The brightness variations in these heatmaps indicate the degree of attention the model assigns to specific regions. As illustrated in Figure 11, both YOLOLS and YOLOv8n successfully detect insulators with high accuracy. Although YOLOLS exhibits a slightly lower confidence score, both models focus on the critical features of insulators, resulting in comparable overall detection performance. For insulator defect detection, the performance of both networks remains nearly identical. However, YOLOLS demonstrates a more precise focus on defect characteristics, indicating that despite its lightweight architecture, its defect detection capability is not compromised.

4. Conclusions

To address the real-time detection requirements for power insulator defects on edge devices, we introduced YOLOLS, a lightweight object detection model derived from YOLOv8n and optimized for real-time insulator defect detection on edge devices. YOLOLS is designed to address the limitations of resource-constrained environments by integrating several key innovations. Our model employs GhostConv for efficient feature extraction and a restructured C2f module via a ResNet–RepConv configuration to streamline inference and reduce computational costs. In addition, a shared convolutional detection head minimizes redundant computations across scales, while an auxiliary Inner-IoU mechanism refines the CIoU loss function to enhance localization accuracy, particularly for small or occluded defects. Extensive experiments on the CPLID datasets demonstrate that YOLOLS achieves a 91% mAP and processes at 44.6 FPS on the Jetson Orin NX, significantly reducing both parameter count, and FLOPs compared to YOLOv8n. In addition, we verified the generalization ability as well as the stability of the model through robustness experiments. These results confirm that YOLOLS is a robust, efficient, and practical solution for the real-time insulator detection requirements for edge devices, making it an efficient and practical solution for resource-constrained power industrial applications.

Despite these promising outcomes, further research should concentrate on fully integrating lightweight, efficient, and reliable models with edge computing frameworks to enable rapid defect detection and immediate response in power grids. Future work will focus on optimizing the detection model for a combination of cloud, edge, and end detection strategies, thereby reducing latency and enhancing scalability while maintaining robust performance under diverse operating conditions. In addition, there remains a critical need to develop integrated alert systems that leverage the fast-processing capabilities of edge platforms, triggering immediate notifications and rapid remedial actions upon defect detection. Expanding the application of the model to monitor other critical transmission components, such as power towers, fittings, transformers, and conductors, will further demonstrate its versatility in power infrastructure monitoring. Addressing these aspects will help overcome current limitations and significantly enhance the practical impact of real-time, edge-based defect detection in power industrial applications.