SML-YOLO: a lightweight insulator defect detection algorithm for transmission lines under harsh weather conditions Authors: Guangyang Ren, Haosen Wang Published: 2026 ABSTRACT Efficient defect detection of insulators in transmission lines is critical to ensuring the reliability and safety of power systems. However, adverse outdoor conditions such as rain, fog, and low illumination can reduce the contrast between targets and background, resulting in blurred defect details. Meanwhile, insulators often coexist with towers, conductors, and other elements within complex backgrounds, which can lead to inaccurate target segmentation and subsequently cause missed or false detections. This study introduces a lightweight insulator defect detection algorithm, SML-YOLO, designed for harsh weather scenarios. Built upon the YOLO11 framework, the algorithm achieves performance optimization through three core innovations. Specifically, StarNet-nano is proposed as the backbone network, thereby improving the perception of features without compromising the model’s lightweight architecture to improve the detection of small-scale defects. In addition, a multi-branch modulated feature pyramid network (MMFPN) is developed to optimize feature fusion while adhering to lightweight objectives, and an LSCD detection head is employed to balance model compactness with high detection accuracy. Experimental results demonstrate that SML-YOLO reduces model size by 21.1% and parameter count by 20.9%, while achieving a 2.4% improvement in detection accuracy and a 23.8% increase in detection speed. Moreover, it exhibits strong adaptability under adverse weather conditions. These findings confirm that the proposed algorithm combines lightweight design, high detection performance, and robust generalization capability. It not only provides an efficient solution for automated inspection of transmission lines, reducing manual inspection workload and minimizing maintenance downtime, but is also well-suited for deployment in resource-constrained UAV applications. Introduction The power grid, as the core infrastructure for national energy transmission and distribution, plays a vital role in sustaining social productivity and ensuring uninterrupted public services. Its safe and stable operation constitutes a critical foundation for the sustainable development of the economy and society1. Insulators serve a crucial role in high-voltage power transmission systems and high-speed railway catenary systems, not only provide electrical insulation and isolation to prevent leakage currents but also deliver mechanical support to withstand complex external forces such as wind load and ice accretion. They serving as an essential safety barrier ensuring reliable electric power transmission2. Accurate detection of insulator defects is an effective measure to prevent serious accidents such as line trips and large-scale blackouts, while also providing indispensable technical support for the rapid restoration of energy supply and transportation following extreme weather events such as heavy rain and ice accretion3. Therefore, timely identification and precise localization of insulator defects are crucial for ensuring the normal operation of transmission lines. Conventional manual inspection suffers from low efficiency, harsh environmental conditions, difficult accessibility, and significant safety risks, rendering it inadequate for the demands of smart grid development. With the rapid development of smart grids and the widespread application of computing technologies and intelligent devices, vision-based inspection of transmission lines has gradually become the dominant method4. Among these methods, unmanned aerial vehicles (UAVs) have been widely adopted owing to their advantages of low cost, short inspection cycles, and high mobility5. By processing UAV-acquired inspection images, potential hazards can be effectively identified, thereby reducing accident risks and enhancing grid reliability6. Furthermore, image processing and analysis not only enable the detection and localization of electrical components but also provide technical support for UAV autonomous inspection, navigation control, automatic camera focusing, and defect recognition. However, UAV-based inspection inevitably generates massive amounts of image data, which poses increasing challenges for subsequent processing and analysis. To further advance the automation and intelligence of transmission line inspection, numerous deep learning–based approaches have been proposed for electrical equipment recognition and insulator defect detection. For instance, Wang et al.7 adopted ResNeSt as the backbone and optimized the enhanced Region Proposal Network (RPN) to improve the feature capture efficiency for small defects. Yang et al.8 modified YOLOv3 by introducing the EIoU loss function, which enhanced the overlap between predicted bounding boxes and ground-truth boxes while accelerating model convergence. Wang et al.9 proposed ML-YOLOv5, which combined depthwise separable convolution with an enhanced feature fusion module, achieving a significant improvement in detection accuracy. Similarly, Qiang et al.10 extended YOLOv8s to detect not only insulator defects but also external hazards (e.g., bird nests), thereby enhancing the comprehensiveness of its application. Focusing on small-object detection, Tao et al.11 proposed SnakeNet under the YOLOv8 framework, which notably improved localization accuracy in complex scenarios. Deng et al.12 proposed RML-YOLO, which incorporated hierarchical scale fusion features and cross-layer feature fusion strategies, thus achieving higher detection precision in multi-scale scenarios. Likewise, Lu et al.13 introduced a context conversion module into the YOLOv8 backbone to better capture contextual information, and added partial convolutions to reduce redundant computation. Scapinello et al.14 integrated YOLOv8 with an adaptive context refinement (ACR) mechanism: YOLOv8 enabled real-time detection, while the ACR mechanism leveraged multi-scale contextual information around detected objects to refine classification and localization results. Meanwhile, Transformer-based methods have also demonstrated strong potential in insulator defect detection. For example, Liu et al.15 developed a high-resolution cross-modal translation module (HICT) and a multi-domain multi-scale enhancement module (MMA) to improve the model’s generalization ability in complex environments. Lu et al.16 proposed an enhanced deformable attention mechanism (EDAM), which incorporates gated activation functions to improve the capture of key feature points. Lu et al.17 designed a multi-scale feature fusion Transformer (MSFFT), significantly boosting the recognition accuracy of small defects. Finally, Zhang et al.18 introduced EFENet, an edge feature–enhanced Transformer network, which achieves high detection precision through frequency-varied multi-scale feature fusion. Most of the aforementioned studies primarily focus on improving model accuracy, while paying insufficient attention to the design requirements and processing constraints of mobile platforms. For instance, Lu et al.19 combined GhostNet with a lightweight channel–spatial attention (LCSA) module, which reduces model redundancy while preserving feature extraction capability. Li et al.20 proposed LiteYOLO-ID by improving YOLOv5s, constructing a lightweight defect detection model that effectively lowers computational costs. Shen et al.21 integrated a channel–spatial attention module into the YOLOv5 backbone to enhance its representation learning ability. An improved YOLOv8 model based on bidirectional weighted feature fusion was presented in22, which maintains high detection efficiency under complex environments. In23, LDDFSF-YOLO11 was introduced to target small-scale features, achieving enhanced small-object detection performance while retaining a lightweight design. YOLOv5s-TC and LiteYOLO-ID simplify network structures but sacrifice feature representation capability, while LDDFSF-YOLO11 enhances small-object detection at the cost of increased architectural complexity. Transformer-based lightweight solutions have also emerged, such as AdIn-DETR24, which employs an adapter-enhanced architecture to detect shed defects of composite and glass insulators, and it even achieves competitive results on relatively small datasets. Moreover, Li et al.25 proposed A2MADA-YOLO, an attention-aligned multi-scale adversarial domain adaptation model tailored for foggy scenarios, which improves model robustness while controlling its size. Overall, most existing research tends to prioritize either improving accuracy or pursuing lightweight design, but often overlooks the balance between these two aspects. As a result, models optimized solely for accuracy usually become overly complex, whereas lightweight models frequently compromise detection precision. In this study, we propose a lightweight object detection model dedicated to insulator defect recognition. The model achieves 94.6% mAP with only 2.04 M parameters, 6.3GFLOPs, and a model size of 4.1 MB, rendering it well-suited for deployment on mobile and resource-constrained devices. This work advances insulator defect detection research by balancing accuracy and efficiency. This study makes the following key contributions: 1) A lightweight and high-performance insulator defect detection framework, SML-YOLO, is proposed to address the challenges of small-object detection and complex backgrounds under adverse weather conditions. 2) A novel StarNet-nano backbone and a multi-branch modulated feature pyramid network (MMFPN) are introduced to enhance fine-grained feature representation and adaptive multi-scale feature fusion while maintaining model compactness. 3) A lightweight shared convolutional detection (LSCD) head is developed to reduce parameter redundancy and computational cost, enabling efficient deployment on resource-constrained platforms while preserving detection accuracy. Materials and methods YOLO11 basic model YOLO11 represents an efficient single-stage object detection framework, released in five configurations: n, s, m, l, and x. These versions share the same overall architecture, while their depth and width progressively increase. The architectural diagram is presented in Fig. 1. Fig. 1YOLO11 network structure. YOLO11 network structure. The backbone of YOLO11 primarily relies on the Conv module and the C3k2 module to extract multi-scale features, thereby achieving a lightweight design while maintaining rich gradient flow. The Conv module progressively downsamples feature maps, reducing their spatial resolution while preserving key information across multiple scales (large, medium, and small). The C3k2 module, derived from the C2f module in YOLOv8, adopts a more compact dual-convolution structure combined with optimized cross-stage residual connections. This design not only accelerates inference speed but also enhances the model’s applicability in real-world deployment. The compressed feature maps are then processed by the SPPF module, which employs multi-scale pooling operations to comprehensively capture contextual information, enabling more effective feature representation. The extracted features are then forwarded to the C2PSA module, which integrates two convolutional layers with a multi-head self-attention mechanism to further strengthen feature modeling capability. The refined features are passed to the neck network for multi-level feature fusion and upsampling. Finally, the detection head employs a decoupled design, where classification and regression branches are separated to improve detection flexibility. However, while this design enhances overall performance, it inevitably introduces additional parameters and computational cost. Improved YOLO11 method SML-YOLO In real-world scenarios, especially on mobile devices with limited resources, model compactness is crucial. A lightweight design not only accelerates inference but also reduces hardware dependency and power consumption, while alleviating storage overhead, thereby facilitating deployment and updates. Motivated by this, we propose a lightweight yet accurate object detection model tailored for mobile platforms, which we term SML-YOLO. The detailed architectural diagram is illustrated in Fig. 2. Fig. 2SML-YOLO network structure. SML-YOLO network structure. StarNet-nano In YOLO11, the CSPDarknet backbone leverages a feature-splitting and merging mechanism to effectively extract low-level features such as edges, textures, and gradients, while reducing redundancy. However, its ability to preserve fine-grained information is limited in small-object detection, which constrains detection accuracy. To address this limitation, we propose StarNet-nano as a lightweight backbone alternative to YOLO11. The design is derived from the core Star operation introduced in StarNet26, which maps input features into a high-dimensional nonlinear space, thereby enhancing feature representation without significantly increasing parameter count. Essentially, the Star operation functions as a cross-channel feature pairwise product kernel, closely related to polynomial kernel functions. Through inter-layer stacking, the Star operation implicitly expands feature dimensions in an exponential manner, enabling efficient transformation from low-dimensional computation to high-dimensional representation. Its general form can be expressed as:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{1}^{T}x*{w}_{2}^{T}x.$$\end{document} This property makes the Star operation inherently well-suited for small-object detection tasks in resource-constrained environments. StarNet-nano is constructed as a lightweight variant of StarNet-s1 to s4, retaining the same four-stage hierarchical structure as the YOLO11 backbone but with fewer convolutional layers and reduced module complexity. The overall architecture consists of four stages: one Star block in the first stage, two in the second, four in the third, and one in the fourth. Compared with StarNet-s1, StarNet-nano eliminates seven Star blocks, thereby reducing parameter count while improving inference accuracy. The Star block, serving as the core unit of StarNet-nano, is composed of depthwise convolution, pointwise convolution, element-wise multiplication, nonlinear activation, residual connection, and a final 1 × 1 convolution. It adopts a dual-branch design: one branch extracts spatial details, while the other captures semantic information. The two branches are dynamically fused via element-wise multiplication, highlighting target-relevant features “like a magnifying glass.” Meanwhile, residual connections preserve the original features, mitigating the loss of fine details in deeper layers. Specifically, the input feature is first processed by depthwise convolution and batch normalization, then split into two branches. Each branch undergoes channel expansion and nonlinear transformation through a 1 × 1 convolution and ReLU6 activation27. Their outputs are fused via element-wise multiplication, followed by refinement through pointwise and depthwise convolutions. The residual connection is then added to obtain the enhanced representation, and finally, a 1 × 1 convolution generates the output.The transformation of the Star block can be expressed as:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F}_{fused}=c\left(BN\left(DWConv\left({F}_{input}\right)\right)\right)\odot\:ReLU6\left(c\left(BN\left(DWConv\left({F}_{input}\right)\right)\right)\right)$$\end{document}2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F}_{output}=c\left(DWConv\left(BN\left(c\left({F}_{fused}\right)\right)\right){+F}_{input}\right)$$\end{document} Here, BN denotes BatchNorm2d, and c represents a convolutional layer used for channel dimension alignment. Building upon this design, the Star block further incorporates a Convolutional Gated Linear Unit (CGLU) mechanism28. This mechanism introduces dynamic channel-wise feature control at the end of the Star block: the output feature is split into two parts along the channel dimension, where one part is passed through a Sigmoid activation to generate gating weights, which are then applied to the other part via element-wise multiplication to achieve channel selection. The process can be formulated as:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Z}_{CGLU}={X}_{1}\odot\:{\upsigma\:}\left({X}_{2}\right)$$\end{document} where ⊙ denotes element-wise multiplication and σ(⋅) is the Sigmoid activation function. CGLU enhances feature representation by emphasizing salient features while suppressing redundant information with minimal computational overhead. Moreover, its dynamic regulation allows the network to adaptively adjust to input features, thereby improving its ability to handle complex cross-transformations. The combination of residual connections and CGLU further alleviates the vanishing gradient problem and ensures stable convergence during training. Overall, StarNet-nano, building upon the efficient mapping capability of the Star operation, integrates Star blocks with the CGLU dynamic gating mechanism to achieve a favorable balance between feature representation and computational cost. This design not only enhances the accuracy of YOLO11 in small-object detection but also preserves lightweight characteristics and efficient inference, demonstrating strong potential for practical applications. Figure 3. Fig. 3SC Block structure. SC Block structure. Multi-branch modulated feature pyramid network In the original YOLO11 architecture, the neck network employs Path Aggregation Network (PANet) as its core, which achieves multi-level feature fusion through upsampling and concatenation, while the C3k2 module is used to enhance cross-scale interaction. However, this structure exhibits two major limitations in complex scenarios. First, PANet processes all feature maps equally, lacking targeted enhancement for critical features; shallow-layer noise can thus be incorrectly introduced, leading to a higher false detection rate. Second, feature reuse efficiency remains limited, as cross-layer transmission lacks adaptive regulation. Consequently, shallow details (e.g., fine crack textures) are often overwhelmed by deep semantic features, restricting the accuracy of small-object detection. To overcome these bottlenecks, this study proposes a Multi-branch Modulated Feature Pyramid Network (MMFPN), which integrates the Multi-branch Auxiliary Feature Pyramid Network (MAFPN)29 with the Modulated Fusion Module (MFM)30 to replace the original PANet and construct an enhanced neck structure. Specifically, MAFPN aggregates shallow detail features and deep semantic features through both bottom-up and top-down pathways, thereby providing a strong foundation for refined representation of small and medium-scale objects. MFM, in turn, substitutes the traditional concatenation operation with an attention-based adaptive weighting mechanism, avoiding the redundancy and background interference associated with equal channel treatment. The synergy of MAFPN and MFM forms an intelligent multi-branch modulated neck architecture that not only preserves the multi-scale fusion advantages of PANet but also significantly improves feature utilization and fusion effectiveness. In implementation, MAFPN consists of two core modules: Shallow Auxiliary Fusion (SAF) and Advanced Auxiliary Fusion (AAF). The SAF module focuses on shallow-layer features, employing skip connections to deliver fine-grained details to deeper layers, thereby mitigating the weakening or loss of small-object features during transmission. The AAF module, designed for medium-object detection, integrates multi-level feature sources to generate semantically enriched deep features, strengthening both object boundary delineation and class discrimination. Notably, both modules employ the MFM to replace the conventional concatenation mechanism. MFM first extracts global contextual information to generate channel weights, then dynamically highlights critical channels while suppressing redundant ones. This adaptive fusion strategy enhances feature selectivity and computational efficiency, leading to more effective and discriminative feature representations. Fig. 4Module Structure in MMFPN. Module Structure in MMFPN. Shallow auxiliary fusion module The goal of the SAF module is to mitigate the issue of shallow detail features being overwhelmed by deep semantic information, as shown in Fig. 4a. Its input consists of three types of features: the downsampled shallow feature \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{n-1}$$\end{document}, the main-path feature Pₙ, and the upsampled deep feature \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}{\prime\:}_{n+1}$$\end{document}. These are adaptively fused in the MFM to generate the enhanced feature P′ₙ. This bidirectional cross-layer integration not only ensures that fine-grained textures of small objects are effectively propagated to deeper layers, but also prevents shallow information from being entirely dominated by high-level semantics. Given the backbone output feature map Pₙ∈RC×H×W (where C denotes channel dimension and H, W represent spatial dimensions), the output feature map processed by SAF is P′ₙ∈RC×H×W. The computation process is defined as:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{P}}^{{\prime\:}}_{n1}=\left\{\begin{array}{c}MFM\left(c\left(Down\left({P}_{n-1}\right)\right),{P}_{n}\right),\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:n=5\\\:MFM\left(c\left(Down\left({P}_{n-1}\right)\right),{P}_{n},Up\left({P}{\prime\:}_{n+1}\right)\right),\:\:n={3,4}\end{array}\right.$$\end{document} where Down and Up denote downsampling and upsampling operations, respectively. Advanced auxiliary fusion module The AAF module is designed for medium-object detection, with an emphasis on multi-source feature integration, as shown in Fig. 4b. Its input includes four types of features: the high-resolution shallow feature \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}{\prime\:}_{n+1}$$\end{document}, the neighboring semantic feature\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P{\prime\:}}_{n-1}$$\end{document}, the main-path feature \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}{\prime\:}_{n1}$$\end{document}, and the deeper feature \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P{\prime\:}{\prime\:}}_{n-1}$$\end{document}. After appropriate downsampling, upsampling, and convolutional processing, these features are adaptively fused through the MFM to produce a more semantically complete representation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P{\prime\:}{\prime\:}}_{n}$$\end{document}. This design not only preserves boundary and texture information but also strengthens semantic discriminability. Let the high-level feature maps obtained from the bottom-up pathway be Pₙ∈RC×H×W, (n = 3,4,5). The AAF output deep feature map is denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P{\prime\:}{\prime\:}}_{n}$$\end{document}∈RC×H×W. Its computation is given by:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P{\prime\:}{\prime\:}}_{n}=\left\{\begin{array}{c}MFM\left({{P}}^{{\prime\:}}_{n1},\:c\left(Up\left({P}{\prime\:}_{n+1}\right)\right)\right),\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:n=3\\\:MFM\left(c\left(Down\left({{P}^{{\prime\:}}}_{n-1}\right)\right),c\left(Down\left({{P}^{{\prime\:}{\prime\:}}}_{n-1}\right)\right),{P}^{{\prime\:}}_{n1},\:c\left(Up\left({P}{\prime\:}_{n+1}\right)\right)\right),\:\:n=4\\\:MFM\left(c\left(Down\left({{P}^{{\prime\:}}}_{n-1}\right)\right),c\left(Down\left({{P}^{{\prime\:}{\prime\:}}}_{n-1}\right)\right),{P}^{{\prime\:}}_{n1}\right),\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:n=5\end{array}\right.$$\end{document} Modulated fusion module The MFM is designed to replace the conventional concatenation mechanism with a dynamic weighting strategy, as shown in Fig. 4c. Its inputs are F1 (the original neck feature) and F2 (the auxiliary feature). Specifically, the fusion process operates as follows: The combined features F1 + F2 are first processed via global average pooling (GAP) to extract contextual information. The resulting vector is passed through a multi-layer perceptron (MLP) and a Softmax function to generate a channel weight matrix A. The weight matrix is applied through element-wise multiplication (⊙) with the input features, adaptively enhancing critical channels while suppressing redundant ones. Finally, a convolutional layer aggregates the weighted features to produce the fused representation. Formally, the computation can be expressed as:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{F}}_{{r}{c}}^{1}=A\odot\:{F}_{1}+{A}\odot\:{F}_{2}$$\end{document} Compared with the naive concatenation operation, MFM significantly reduces redundant information, improves computational efficiency, and maintains sensitivity to small objects even in cluttered backgrounds. In summary, the improved MMFPN neck architecture achieves optimization on three levels: SAF ensures stable transmission of shallow details and enhances small-object detection; AAF strengthens cross-scale integration and improves the representation of medium-sized objects; and MFM dynamically assigns channel weights to reduce redundancy and increase fusion efficiency. Through their synergy, the proposed neck not only improves detection accuracy for small and medium-scale insulator defects but also maintains high inference efficiency, thereby providing high-quality input features for the detection head. Lightweight shared convolutional detection head The detection head is a critical component of object detection models, responsible for both classification and localization. In YOLO11, a decoupled head architecture is utilized, enabling independent processing of classification and regression tasks. While this improves detection flexibility and stability, prior research has shown that the detection head is also the primary source of computational overhead31. Excessive parameters not only increase memory consumption but also slow down inference, which limits deployment in lightweight-oriented scenarios such as UAV-based inspections. To address this issue, we introduce a Lightweight Shared Convolutional Detection Head (LSCD), which balances accuracy and efficiency while reducing parameter size and computational cost. The LSCD is composed of three key steps. First, the input multi-scale feature maps P3, P4, and P5 (corresponding to small, medium, and large defects, respectively) are compressed using 1 × 1 convolutions, providing a foundation for lightweight processing. Second, shared convolution is applied, where a single convolutional kernel is used across P3, P4, and P5, thereby significantly reducing the number of parameters and FLOPs. However, this sharing may result in a loss of accuracy. To compensate, we integrate Group Normalization (GN): the convolved features are divided into groups along the channel dimension and normalized within each group32. This not only enhances classification and localization performance but also avoids the training instability of Batch Normalization (BN) under small batch sizes. The normalization process can be formulated as follows:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:L=\frac{{x}_{i}-{u}_{i}}{\sqrt{{\sigma\:}_{i}^{2}+{\upvarepsilon\:}}}\times\:\gamma\:+\beta\:$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{i}$$\end{document} is the input feature, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{u}_{i}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{i}$$\end{document} denote the mean and variance within each group, γ and β are learnable parameters, and ε is a small constant for numerical stability. Finally, a scale layer is applied to rescale the features, mitigating inconsistencies in object scaling across detection heads. This ensures improved accuracy for small, medium, and large objects while further enhancing computational efficiency under reduced parameter and FLOP budgets. Figure 5. Fig. 5LSCD head architecture. LSCD head architecture. Experiments and result analysis Dataset In this study, we employed the Insulator Defect Image Dataset (IDID) constructed by Lewis et al.33, which contains 1,596 high-resolution images of insulator strings across three categories: intact insulators, shell damage, and flashover damage. The dataset is publicly available on IEEE DataPort. To enhance the model’s learning capability and improve its representation of insulator features, multiple data augmentation strategies were applied. First, adjustments in brightness, color, and contrast were performed, expanding the dataset to 2,800 images. Second, to simulate complex illumination conditions encountered in real inspection scenarios—such as direct sunlight, metallic reflection, backlighting, and low-light environments—an additional 2,800 images were generated. Furthermore, to increase scene diversity, 2,100 images representing adverse weather conditions, including fog, rain, and snow, were synthesized. As a result, the final dataset consists of 7,300 images. Within the training set, the class distribution is 34.5% intact insulators, 19.5% shell-damaged insulators, and 50% flashover-damaged insulators. The dataset is divided into training, validation, and test subsets, as detailed in Tables 1 and 2, with a ratio of 7:2:1 (Table 1). It should be noted that Table 1 reports the number of images, whereas Table 2 presents the number of annotated instances. Since a single image may contain multiple insulator objects, the total number of instances exceeds the number of images. Figure 6. Fig. 6Example images from the Insulator Defect Image Dataset (IDID) provided by EPRI and licensed under CC BY 4.0. Example images from the Insulator Defect Image Dataset (IDID) provided by EPRI and licensed under CC BY 4.0. Table 1Division of our datasets.DatasetTrainValidationTestOur dataset51101460730 Table 1Division of our datasets.DatasetTrainValidationTestOur dataset51101460730 Division of our datasets. Table 2Instance-level category distribution of the dataset.DatasetTrainValidationTestintact843524101206shell-damaged36471042521flashover64121832916 Table 2Instance-level category distribution of the dataset.DatasetTrainValidationTestintact843524101206shell-damaged36471042521flashover64121832916 Instance-level category distribution of the dataset. Training settings In deep learning research, hardware and software configurations substantially influence experimental performance, often resulting in variation between models. To ensure reproducibility, this study provides detailed information on the computing environment, summarized in Table 3. Table 3Computer parameters.ConfigurationParameterOperating systemWindows11Accelerated environmentCuda 11.3LanguagePython 3.10FrameworkPytorch 1.11.0CPUIntel Xeon Silver 4210R CPU running @ 2.40 GHzGPURTX3090 (24 GB) Table 3Computer parameters.ConfigurationParameterOperating systemWindows11Accelerated environmentCuda 11.3LanguagePython 3.10FrameworkPytorch 1.11.0CPUIntel Xeon Silver 4210R CPU running @ 2.40 GHzGPURTX3090 (24 GB) Computer parameters. Furthermore, model hyperparameters can significantly impact experimental outcomes34. In the enhanced YOLO11 model, training was conducted with a batch size of 32 for 200 epochs. Prior to being fed into the network, all input images were resized to 640 × 640 pixels. The SGD optimizer was adopted, and the close_mosaic parameter was set to 10. The initial learning rate was set to 0.01, with decay scheduled via cosine annealing. The optimizer utilized SGD with a momentum of 0.937 and weight decay of 5e − 4. A warm-up phase of 3 epochs was applied. Data augmentation methods, including Mosaic and MixUp, were both enabled during training. Evaluation metrics In this study, we adopted widely used evaluation metrics for object detection, including precision, recall, mean Average Precision (mAP), number of parameters, and model size. The corresponding formulas are defined as follows:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:P=\frac{{T}P}{TP+FP}$$\end{document}8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:R=\frac{{T}P}{TP+FN}$$\end{document}9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:mAP=\frac{1}{N}\sum\:_{i=1}^{N}\left({\int\:}_{0}^{1}{P}_{i}\left(R\right)dR\right)$$\end{document}10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:FPS=\frac{1}{{T}_{preprocess}+{T}_{inference}+{T}_{postproces}}$$\end{document} True Positive (TP): The number of correctly detected objects. A predicted bounding box is considered a TP if its Intersection over Union (IoU) with the ground-truth bounding box exceeds a predefined threshold. False Positive (FP): The number of incorrectly detected objects. A predicted bounding box is considered an FP if it has no matching ground-truth object or if its IoU with the ground-truth bounding box falls below the threshold. Ablation experiments To evaluate the impact of the SC Block, MMFPN, and LSCD detection head on the detection performance of SML-YOLO, an ablation study was conducted. The detailed results are presented in Tables 4 and 5, where S, M, and L denote the SC Block, MMFPN, and LSCD detection head, respectively. Table 4Comparison of ablation experiment results.SMLPrecisionRecallmAP@50mAP@50:95ParamsGFLOPsweightFPS93.081.892.2 ± 0.06280.9 ± 0.0942.586.55.2189√92.784.991.9 ± 0.07481.6 ± 0.1052.46.45.0217√√93.787.493.7 ± 0.05783.8 ± 0.0882.46.85.0221√√√93.983.594.6 ± 0.04984.6 ± 0.0732.046.34.1234 Table 4Comparison of ablation experiment results.SMLPrecisionRecallmAP@50mAP@50:95ParamsGFLOPsweightFPS93.081.892.2 ± 0.06280.9 ± 0.0942.586.55.2189√92.784.991.9 ± 0.07481.6 ± 0.1052.46.45.0217√√93.787.493.7 ± 0.05783.8 ± 0.0882.46.85.0221√√√93.983.594.6 ± 0.04984.6 ± 0.0732.046.34.1234 Comparison of ablation experiment results. Table 5Comparison of ablation experiment results.ModelmAP@50Mean Difft-statp-valueSignificanceBaseline92.2 ± 0.062----S91.9 ± 0.074-0.3-8.52< 0.001YesS + M93.7 ± 0.057+ 1.842.16< 0.001YesS + M +L94.6 ± 0.049+ 0.921.37< 0.001Yes Table 5Comparison of ablation experiment results.ModelmAP@50Mean Difft-statp-valueSignificanceBaseline92.2 ± 0.062----S91.9 ± 0.074-0.3-8.52< 0.001YesS + M93.7 ± 0.057+ 1.842.16< 0.001YesS + M +L94.6 ± 0.049+ 0.921.37< 0.001Yes Comparison of ablation experiment results. The ablation study results demonstrate that the SC Block, MMFPN, and LSCD detection head each contribute to the performance improvement of SML-YOLO. The baseline model, without any of these modules, achieves an mAP@50 of 92.2, with a precision of 93.0 and a recall of 81.8. To further validate the effectiveness of each module, statistical significance analysis is conducted. Specifically, paired t-tests are performed between successive configurations based on ten independent runs with different random seeds. The results, summarized in Table 5, indicate that all module contributions are statistically significant (p < 0.05), confirming that the observed improvements are consistent rather than arising from random variation. The relatively large t-statistics can be attributed to the low variance across repeated runs, reflecting the high stability of the model under controlled experimental conditions. From the ablation results, it is observed that introducing the SC Block alone leads to a slight performance degradation, whereas combining it with MMFPN results in a significant improvement. The final configuration further achieves statistically significant gains. These findings indicate that the effectiveness of the proposed method arises from the complementary interaction among different modules rather than from any single component. In terms of module-specific contributions, the SC Block expands the implicit feature space through stacked Star operations, enabling richer and more expressive feature representations while maintaining computational efficiency. Although a slight decrease in accuracy is observed, the number of parameters is reduced by 6.9%, and the model size decreases by 3.8%. By incorporating MMFPN, the model’s feature fusion capability is significantly enhanced, allowing more effective extraction of multi-scale features and improved representation of objects at different scales. This results in an increase of 1.5% in mAP@50 and 2.9% in mAP@50:95, along with improvements of 0.7% in precision and 5.6% in recall. Meanwhile, the number of parameters and model size are further reduced by 6.9% and 3.8%, respectively. Finally, the integration of the LSCD detection head further improves both efficiency and accuracy by employing shared convolution, group normalization, and a scaling mechanism to harmonize multi-scale features. With all three modules incorporated, the model achieves a final mAP@50 of 94.6, while reducing the number of parameters by 20.9% and the model size by 21.1%, demonstrating a favorable trade-off between detection performance and computational cost. The impact of training duration is further analyzed. As shown in Figure. 7(b), the validation mAP reaches its peak around 180–200 epochs, and the performance difference between 200 and 300 epochs is negligible, indicating no meaningful improvement. This suggests that the model has already converged within 200 epochs, and further training primarily leads to marginal fluctuations rather than meaningful performance gains. To evaluate the practical efficiency of the proposed model, inference latency and memory usage are measured under identical hardware conditions. All experiments are conducted on an NVIDIA RTX 3090 GPU with a batch size of 1. As shown in Table 6, SML-YOLO achieves lower inference latency and reduced memory consumption compared to the baseline model, while maintaining superior detection performance. These results demonstrate that the proposed method is not only lightweight in terms of model complexity (Params and GFLOPs), but also efficient in real-world deployment scenarios. Overall, SML-YOLO achieves an effective balance between detection accuracy and computational efficiency, making it well-suited for deployment in resource-constrained environments such as UAV-based inspection systems. Fig. 7(a) Training and validation loss of SML-YOLO (200 vs. 300 epochs); (b) Validation mAP@50 of SML-YOLO (200 vs. 300 epochs). (a) Training and validation loss of SML-YOLO (200 vs. 300 epochs); (b) Validation mAP@50 of SML-YOLO (200 vs. 300 epochs). Table 6Efficiency comparison under identical hardware conditions.ModelLatency (ms)Memory (MB)YOLO1124191SML-YOLO16185 Table 6Efficiency comparison under identical hardware conditions.ModelLatency (ms)Memory (MB)YOLO1124191SML-YOLO16185 Efficiency comparison under identical hardware conditions. Statistical significance testing and visualization To ensure that the observed performance improvements are not attributable to random variation, we conducted ten independent experimental runs comparing YOLO11n and the proposed SML-YOLO on the same dataset. Paired t-tests were performed on both mAP@50 and mAP@50:95 metrics, and the corresponding box plots are presented in Fig. 7. For mAP@50, the paired t-test indicates a statistically significant improvement achieved by SML-YOLO. Specifically, the model attains a mean performance gain of 2.393 over YOLO11n, with a t-statistic of 21.37 and a p-value < 0.001. The 95% confidence interval [2.324, 2.462] further confirms the consistency and reliability of the improvement across repeated trials. Similarly, for mAP@50:95, the proposed method demonstrates a statistically significant enhancement, yielding a mean gain of 3.650, a t-statistic of 34.45, and a p-value < 0.001. The corresponding 95% confidence interval [3.601, 3.699] indicates stable and reproducible performance gains. All statistical results are summarized in Table 7. Each value is computed based on ten independent runs with different random seeds under a paired t-test framework. The relatively large t-statistics can be attributed to the low variance observed across repeated experiments, reflecting the high stability and consistency of the proposed method rather than any statistical artifact. Figure 8. Table 7Statistical significance analysis (10 runs, paired t-test).MetricMean Difft-statp-value95% CImAP@502.39321.37< 0.001[2.324, 2.462]mAP@50:953.65034.45< 0.001[3.601, 3.699] Table 7Statistical significance analysis (10 runs, paired t-test).MetricMean Difft-statp-value95% CImAP@502.39321.37< 0.001[2.324, 2.462]mAP@50:953.65034.45< 0.001[3.601, 3.699] Statistical significance analysis (10 runs, paired t-test). Fig. 8Boxplot comparison of YOLO11n and SML-YOLO on our dataset in terms of mAP@50 and mAP@50:95. (a) Comparison of mAP@50 accuracy curves; (b) Comparison of mAP@50:95 accuracy curves. Boxplot comparison of YOLO11n and SML-YOLO on our dataset in terms of mAP@50 and mAP@50:95. (a) Comparison of mAP@50 accuracy curves; (b) Comparison of mAP@50:95 accuracy curves. Comparative experiments To further validate the performance advantages of the proposed insulator defect detection model, we trained several popular neural networks35–42, and their results are summarized in Table 8. Considering that differences in model size across detection algorithms may significantly influence detection performance, we deliberately selected lightweight models with a comparable number of parameters to our proposed method for comparison. This strategy ensures a fair and effective evaluation, allowing a more objective assessment of SML-YOLO’s performance advantages under comparable computational resource conditions. Table 8Comparison of detection results of different models.Model P R mAP@50ParamsGFLOPsWeightFPSYOLOv592.680.391.42.195.94.5177YOLOv793.483.290.36.0112.38.5105YOLOv892.681.692.53.018.15.9167YOLOv987.579.987.71.776.73.9121YOLOv1091.781.291.62.768.46.1134YOLO1193.081.892.22.586.55.2189YOLO1292.480.292.82.526.25.2126YOLO1391.886.393.12.446.15.4157Ours93.983.594.62.046.34.1234 Table 8Comparison of detection results of different models.Model P R mAP@50ParamsGFLOPsWeightFPSYOLOv592.680.391.42.195.94.5177YOLOv793.483.290.36.0112.38.5105YOLOv892.681.692.53.018.15.9167YOLOv987.579.987.71.776.73.9121YOLOv1091.781.291.62.768.46.1134YOLO1193.081.892.22.586.55.2189YOLO1292.480.292.82.526.25.2126YOLO1391.886.393.12.446.15.4157Ours93.983.594.62.046.34.1234 Comparison of detection results of different models. The experimental results demonstrate that among the compared models, the proposed SML-YOLO achieves the best performance in terms of detection precision and FPS, while maintaining a competitive balance in parameter count (Params), computational cost (GFLOPs), and model size (Weights). A detailed comparison reveals that although YOLOv9 has the smallest parameter scale and model size, its detection precision, recall, and mAP@50 are significantly inferior to those of SML-YOLO. Similarly, YOLOv5, despite exhibiting the lowest computational overhead and strong lightweight characteristics, still fails to achieve high detection accuracy. From a quantitative perspective, SML-YOLO achieves an mAP@50 of 94.6%, representing a 2.4% improvement over the baseline model YOLO11n. In addition, precision (P) and recall (R) are improved by 0.9% and 1.7%, respectively. At the same time, while enhancing detection performance, the model also achieves a favorable balance with computational efficiency, as its model size and parameter count are reduced by 21.1% and 20.9%, respectively, compared with the baseline. These findings demonstrate that SML-YOLO is highly adaptable for deployment on UAV-based edge devices with limited computational resources. In summary, SML-YOLO demonstrates superior overall performance and strong application potential in insulator defect detection under complex environments. Presentation of detection results In engineering practice, insulators are prone to degradation caused by environmental factors and material aging. Defects such as fractures and flashovers are typically small in size, posing significant challenges for accurate detection. To evaluate the defect detection performance of the proposed SML-YOLO algorithm, experiments were conducted under various environmental conditions, with a focus on comparing detection results before and after the algorithmic improvements. To validate the effectiveness of each module in the proposed model and to visually demonstrate how each proposed module improves defect localization43, Fig. 9 presents detection outputs from four representative test images. We compare five models: baseline YOLO11, YOLO11 + StarNet-nano, YOLO11 + MMFPN, YOLO11 + LSCD, and the proposed SML-YOLO model. The baseline YOLO11 achieves competitive performance but struggles with small-scale defects and complex backgrounds, especially under adverse weather conditions. By introducing the StarNet-nano backbone, SML-YOLO enhances fine-grained feature representation while maintaining a lightweight architecture, enabling more accurate detection of small and subtle defects. The multi-branch modulated feature pyramid network (MMFPN) further improves multi-scale feature fusion through adaptive weighting, effectively suppressing background noise and strengthening cross-scale information flow. In addition, the lightweight shared convolutional detection (LSCD) head reduces parameter redundancy while preserving classification and localization accuracy. Overall, the proposed SML-YOLO model demonstrates superior performance with more precise bounding boxes, higher confidence scores, and improved robustness across diverse and challenging scenarios, validating the effectiveness of integrating lightweight design with advanced feature modulation mechanisms. Fig. 9Test results of defect detection using the proposed model. Test results of defect detection using the proposed model. As illustrated in Fig. 10, a comparison between the original YOLO11n model and the improved SML-YOLO clearly demonstrates the superior detection performance of SML-YOLO under normal clear-weather conditions. Fig. 10Detection results under normal weather. Detection results under normal weather. As shown in Fig. 11, the SML-YOLO algorithm was assessed for its capability to detect insulator defects under four representative environmental scenarios, namely low illumination, fog, rain, and snow, using a set of corresponding test images. Fig. 11Performance comparison under different weather scenarios. Performance comparison under different weather scenarios. As illustrated in Figs. 10e–f and 11f–h, when small defect targets are embedded within complex backgrounds, the YOLO11n algorithm demonstrates limited capability in accurately identifying dispersed defects, often leading to missed or false detections. In contrast, as shown in Figs. 10h–i and 11j–l, the proposed SML-YOLO algorithm enables accurate identification and precise annotation of small-scale defects, including damage and flashover, effectively improving detection reliability and overall performance. Further analysis of Figs. 10d and 11e reveals that when the imaging distance increases, the relative pixel proportion of target objects within the image decreases notably. Under such conditions, the original YOLO11n algorithm can effectively detect large-scale objects (e.g., insulators) but tends to miss small-scale damage defects. As depicted in Figs. 9j and 10g, the improved SML-YOLO algorithm accurately identifies small-scale damage defects even in complex background environments, substantially improving detection confidence and demonstrating superior small-object detection capability. The comparative results in Figs. 10 and 11 further validate these findings: compared with YOLO11n, the proposed SML-YOLO algorithm reduces the detrimental influence of complex environmental factors on detection precision and significantly strengthens the identification capability for small defects. In addition, this performance gain exhibits strong consistency across diverse imaging distances, observation angles, and environmental conditions. In summary, SML-YOLO exhibits outstanding performance in insulator defect detection tasks. Heat map comparison chart Heatmaps primarily serve to enhance the interpretability of deep learning models by revealing the rationale behind their decision-making processes. In such visualizations, color intensity reflects the regions that contribute most to the prediction of a specific class. Figure 12 illustrates several heatmaps generated from the detection results of the proposed model. To validate the improved algorithm’s capability in accurately detecting insulator damage and flashover defects, a comparison was conducted between the visualization heatmaps generated by the YOLO11n model and those produced by the proposed SML-YOLO algorithm, as illustrated in Fig. 9. Specifically, Fig. 12a depicts the original input image, Fig. 12b shows the visualization heatmap obtained after processing with the YOLO11n algorithm, and Fig. 12c presents the heatmap produced by the SML-YOLO model. In these class activation maps, warm colors indicate regions where the model focuses most intensely during object detection, while cool colors represent areas with minimal influence on the final detection outcome. As shown, SML-YOLO demonstrates a more concentrated and accurate attention distribution over the defect regions, confirming its superior interpretability and enhanced detection capability. Fig. 12Visualization heat map of prediction results. Visualization heat map of prediction results. Fig. 13LayerCAM-based heatmap visualization of prediction results. LayerCAM-based heatmap visualization of prediction results. An analysis of the visualization heatmap in Fig. 12b reveals that the YOLO11n algorithm still exhibits certain shortcomings in insulator defect detection. When recognizing insulator objects, the model shows no substantial increase in attention to defect regions compared with background areas. This suggests inadequate learning of object-specific representations during training, which hampers its ability to effectively differentiate between objects and background during detection. In contrast, the visualization heatmap in Fig. 12c demonstrates that the SML-YOLO algorithm exhibits a markedly stronger focus on defect regions while significantly suppressing attention to background areas. This behavior suggests that the model is able to more effectively discriminate between insulator objects and their surroundings. The primary attention regions of SML-YOLO are concentrated on areas associated with self-explosion, structural damage, and flashover defects, indicating enhanced feature localization capability. By comparing Figs. 11c and 12b, it is evident that SML-YOLO provides detection results with higher confidence and precision, more accurately identifying the regions of insulator damage and flashover defects within the image. Comparative experiments on generalizability To evaluate the cross-domain generalization capability of the proposed model, experiments are conducted on two publicly available external datasets, CPLID44 and SFID45, which exhibit different data distributions compared to the training dataset. Specifically, CPLID consists of real-world transmission line inspection images with complex backgrounds and varying object scales, while SFID focuses on foggy and low-visibility scenarios, where defects are generally larger but affected by environmental degradation. These datasets differ from the training data in terms of environmental conditions, background complexity, and object scale, thereby providing a challenging cross-domain evaluation setting. The generalization results are summarized in Table 9. It is worth noting that the objects in the SFID dataset are relatively large and do not include small-scale targets, making the detection task comparatively less challenging. Nevertheless, the proposed SML-YOLO consistently outperforms all competing methods across both datasets. In particular, on the CPLID dataset, SML-YOLO achieves an mAP@50 of 96.6%, outperforming YOLO11 by 0.9%. On the SFID dataset, it reaches an mAP@50 of 98.9%, exceeding YOLO11 by 0.7%. These consistent improvements demonstrate the robustness and adaptability of the proposed model under different environmental and imaging conditions. Although the training dataset includes synthetically generated weather conditions to enhance robustness, the evaluation on CPLID and SFID is conducted without additional augmentation or fine-tuning. The stable performance across these datasets suggests that the proposed model does not rely excessively on synthetic patterns, but instead learns generalizable feature representations. Nevertheless, it should be acknowledged that weather-based augmentation may still introduce potential bias. Further validation on fully real-world datasets without synthetic augmentation will be considered in future work to more rigorously eliminate this concern. Overall, the results indicate that the augmentation strategy improves robustness without significantly compromising generalization capability. Table 9Comparative experiment on CPLID and SFID.ModelCPLIDSFIDParamsGFLOPs P R mAP@50 P R mAP@50YOLOv585.788.392.191.794.598.62.195.9YOLOv790.587.690.397.594.497.36.0112.3YOLOv894.493.595.496.996.198.43.018.1YOLO1194.193.795.796.496.398.22.586.5YOLO1292.594.894.795.497.897.72.525.8YOLO1395.192.295.197.995.998.32.446.1Ours94.393.196.696.794.998.92.046.3 Table 9Comparative experiment on CPLID and SFID.ModelCPLIDSFIDParamsGFLOPs P R mAP@50 P R mAP@50YOLOv585.788.392.191.794.598.62.195.9YOLOv790.587.690.397.594.497.36.0112.3YOLOv894.493.595.496.996.198.43.018.1YOLO1194.193.795.796.496.398.22.586.5YOLO1292.594.894.795.497.897.72.525.8YOLO1395.192.295.197.995.998.32.446.1Ours94.393.196.696.794.998.92.046.3 Comparative experiment on CPLID and SFID. Interpretability and failure case analysis To provide a more rigorous understanding of model behavior, we further analyze the interpretability of the proposed SML-YOLO using gradient-based visualization techniques. Specifically, Grad-CAM and LayerCAM are employed to examine the spatial attention distribution of the model. Grad-CAM is widely used for visual explanation; however, it tends to produce relatively coarse activation maps due to its reliance on high-level feature maps and global gradient information. Consequently, the highlighted regions may span the entire object or even surrounding context, rather than precisely localizing small defect areas. This phenomenon explains why activations sometimes appear distributed over large regions instead of focusing strictly on defect locations. It should be noted that such activations cannot be solely attributed to visualization limitations, as they may also reflect biases in the learned feature representations. To achieve more precise localization, LayerCAM is additionally introduced46. Compared with Grad-CAM, LayerCAM utilizes pixel-wise gradient information from multiple layers, enabling finer spatial resolution and more accurate identification of salient regions. As illustrated in Fig. 12, Grad-CAM visualizations for both YOLO11 and SML-YOLO exhibit a certain degree of background activation. These failure cases can be broadly categorized into three types: (1) structural edge bias, where prominent contour features (e.g., insulator boundaries) are incorrectly emphasized; (2) contextual co-occurrence bias, where background elements frequently associated with defects are mistakenly highlighted; and (3) insufficient response to small or low-contrast defects. Representative failure cases are visualized in Fig. 12 to support the above analysis. As illustrated in Fig. 13, compared with YOLO11, the proposed SML-YOLO exhibits noticeably fewer irrelevant activations, indicating improved feature discrimination capability. Furthermore, LayerCAM results show that the attention of SML-YOLO is primarily concentrated on defect regions, while still revealing minor residual bias in more challenging scenarios. These observations suggest that, although the proposed model enhances defect-oriented feature learning, it remains partially sensitive to contextual bias and structural distractions. Overall, the combined use of Grad-CAM and LayerCAM provides complementary insights, demonstrating that SML-YOLO achieves more accurate and robust feature localization than the baseline model. Conclusions In this study, a lightweight and high-accuracy insulator defect detection model, SML-YOLO, is proposed to address the challenges of small-object detection and complex background interference under adverse weather conditions. By integrating the StarNet-nano backbone, MMFPN feature fusion neck, and LSCD detection head, the proposed model achieves an effective balance between detection accuracy and computational efficiency. Experimental results demonstrate that SML-YOLO consistently outperforms the baseline YOLO11 and other lightweight YOLO-based models in terms of mAP, inference speed, and model compactness, while maintaining strong robustness across diverse environmental scenarios. These results confirm the effectiveness of the proposed architecture in enhancing small-defect representation, improving feature discrimination, and enabling real-time deployment on resource-constrained platforms such as UAV-based inspection systems. Furthermore, interpretability analysis based on Grad-CAM and LayerCAM reveals that the proposed model is capable of focusing more accurately on defect regions, with reduced irrelevant background activations compared to the baseline. However, residual activation in non-defect regions is still observed in certain challenging cases. Rather than being solely caused by texture similarity, these phenomena are more closely related to structural edge bias, contextual co-occurrence bias, and limited sensitivity to small or low-contrast defects, as discussed in the failure case analysis. These observations indicate that, while SML-YOLO improves defect-oriented feature learning, it still exhibits partial sensitivity to contextual interference and structural distractions. In future work, a more systematic failure case analysis will be conducted to further investigate these limitations. Additionally, more discriminative attention mechanisms and spatial regularization strategies will be explored to suppress irrelevant activations. The framework will also be extended to broader power equipment inspection tasks, incorporating multimodal data (e.g., infrared or depth information) to enhance robustness. Finally, improving cross-domain generalization under unseen environmental conditions remains an important direction for further research.