Early research in insulator defect detection predominantly employed handcrafted feature extraction and classical machine learning classification techniques (Yuan et al., 2021; Zhou et al., 2022; Liu et al., 2021; Souza et al., 2023). These methodologies typically involve multi-stage processing pipelines incorporating image preprocessing, region segmentation, feature engineering, and supervised classification. Representative approaches utilized edge detection operators, morphological transformations, and texture descriptors, including Local Binary Patterns and Histogram of Oriented Gradients, to characterize insulator appearances and defect signatures. While these classical methods achieved moderate success under controlled imaging conditions, they exhibited fundamental limitations in generalization capability, requiring extensive manual feature engineering and demonstrating brittleness when confronted with variable environmental conditions, complex backgrounds, and diverse defect morphologies encountered in operational transmission line inspection scenarios.

The introduction of region-based convolutional neural network architectures marked a paradigmatic shift in object detection methodology, leveraging deep learning to automatically learn hierarchical feature representations from training data (Girshick et al., 2014; Girshick, 2015; Ren et al., 2015). The seminal R-CNN framework established the foundational paradigm of region proposal generation, followed by CNN-based classification and bounding box regression. Subsequent refinements, including Fast R-CNN and Faster R-CNN, progressively improved computational efficiency through shared convolutional feature extraction and integrated region proposal networks. Recent applications of these architectures to insulator defect detection have demonstrated promising results (Shuang et al., 2023; Ou et al., 2023; Wang et al., 2021), with researchers reporting substantial improvements over classical approaches. However, the two-stage detection paradigm inherent to R-CNN variants introduces computational complexity and latency that constrain real-time deployment feasibility, particularly for resource-constrained UAV platforms requiring onboard processing capabilities.

The emergence of single-stage object detection architectures, particularly the YOLO family, revolutionized real-time visual recognition by formulating detection as a unified regression problem (Redmon et al., 2016; Liu et al., 2016). The original YOLO architecture introduced the concept of dividing input images into spatial grids and directly predicting bounding boxes and class probabilities from full images in a single forward pass, achieving unprecedented inference speeds while maintaining competitive accuracy. Subsequent iterations, including YOLOv3 (Redmon and Farhadi, 2018), YOLOv4 (Bochkovskiy et al., 2020), and YOLOv8, progressively enhanced detection performance through architectural refinements including residual connections, spatial pyramid pooling, and path aggregation networks. Contemporary research has extensively explored YOLO adaptations for insulator defect detection (Hu Z. et al., 2025; Liu Q. et al., 2025; Shen et al., 2025; Lu et al., 2025; Zhang Y. et al., 2025), with modifications targeting improved small object detection, enhanced feature pyramid architectures, and optimized loss functions. Despite these advances, existing YOLO-based frameworks exhibit limited capacity for capturing global context and long-range dependencies, constraining performance in complex transmission line environments characterized by severe occlusion, scale variation, and background interference.

The introduction of attention mechanisms and transformer architectures has catalyzed significant advances in computer vision, enabling models to selectively focus on salient image regions while capturing long-range spatial relationships (Woo et al., 2018). The Convolutional Block Attention Module demonstrated the effectiveness of channel and spatial attention for enhancing CNN feature representations. More recently, Vision Transformers have achieved state-of-the-art performance across diverse visual recognition benchmarks by modeling images as sequences of patches processed through multi-head self-attention mechanisms. Hybrid architectures combining convolutional feature extraction with transformer-based context modeling have emerged as particularly promising (Li C. et al., 2025; Xu et al., 2025; Wang et al., 2025), leveraging the complementary strengths of local feature learning and global dependency capture. However, the application of transformer-enhanced architectures to insulator defect detection remains limited (Li J. et al., 2025; Liu J. et al., 2025), representing a significant research opportunity for improving detection robustness under challenging operational conditions (Huang et al., 2022).

Despite notable progress in automated insulator defect detection, several critical limitations continue to hinder the practical deployment of existing methodologies (Huang et al., 2022). First, conventional YOLO-based detectors exhibit reduced effectiveness when handling small defects and extreme scale variations typical of aerial inspection imagery (Zhang Q. et al., 2025; Ma et al., 2025; Yu et al., 2025; Chen et al., 2017), largely due to restricted receptive fields and rigid multi-scale fusion strategies. Second, current approaches insufficiently capture global context and long-range spatial dependencies, limiting their ability to distinguish defect features from visually similar background elements in cluttered transmission line environments (Wei and Wei, 2025; Hu M. et al., 2025). Third, feature fusion mechanisms often rely on fixed weighting schemes that fail to adaptively emphasize informative features, thereby constraining representational flexibility. Finally, standard loss functions suffer from slow convergence and sensitivity to bounding box geometry variations, which negatively impact training efficiency and localization accuracy. While prior studies have introduced attention mechanisms and multi-scale fusion strategies, these methods either lack adaptive weighting or fail to adequately capture global dependencies in complex aerial imagery. To overcome these shortcomings, we propose the TE-YOLOv8 framework, which systematically integrates several targeted architectural innovations: the Global Convolution (GConv) module for efficient large receptive fields, the C3f-Global Pooling Fusion (C3f-GPF) module for enhanced feature recalibration, the Multiscale Information Fusion (MSIF) module with learnable fusion weights, and the Weighted Feature Information Fusion (WFIF) module for dynamic channel prioritization. Collectively, these components directly address insulator-specific challenges such as elongated geometries, extreme scale variation, and background interference, thereby advancing the robustness and accuracy of defect detection in high-voltage transmission systems.

The YOLOv8-based detection algorithm, while highly effective in object detection tasks, often faces challenges when applied to the detection of small, complex defects in transmission lines. YOLOv8, building upon the advancements of YOLOv4, employs a strong backbone network for feature extraction and a decoupled head scheme for bounding box regression and classification. However, despite these advancements, YOLOv8 uses standard convolution layers in its backbone network, which can sometimes fail to capture fine-grained spatial and contextual features essential for detecting subtle and small defects. This limitation is especially significant in the context of insulator defect (ID) detection, where the defect features often overlap with or are obscured by background information, making them difficult to detect using traditional convolutional techniques. Figure 2 illustrates the baseline YOLOv8 architecture.

To address these limitations, we propose a new hybrid algorithm, TE-YOLOv8, specifically designed for high-precision insulator defect detection in power transmission lines. As illustrated in Figure 3, TE-YOLOv8 integrates several advanced modules to enhance feature extraction, multi-scale detection, and contextual understanding, ensuring better performance in challenging detection scenarios. The backbone network is enhanced with the introduction of the Global Convolution (GConv) module, which replaces traditional convolution layers at critical points. This modification allows the network to capture broader spatial contexts, improving the network's ability to distinguish defect features from background noise. We also modified the C3 module by incorporating the C3f-Global Pooling Fusion (C3-GPF) module. This enhancement strengthens the network's discriminative ability by recalibrating features through global pooling operations. As a result, the network is better able to focus on important defect features while minimizing the impact of irrelevant background information. The Multiscale Information Fusion (MSIF) module is incorporated to replace the SPPF module, enhancing the network's capability to detect objects across multiple scales, which is particularly important in detecting small or multi-scale defects in transmission lines. Additionally, the Weighted Feature Information Fusion (WFIF) module is integrated to replace the standard Concat module, enabling more precise processing of critical defect-related features through learned attention weights.

This allows the model to dynamically prioritize the most informative feature channels and suppress less relevant background noise. To further enhance localization precision and accelerate convergence, we introduce the SCYLLA-IoU (SIoU) loss function, which replaces the CIoU loss function traditionally used in YOLOv8. This modification speeds up the model's training and improves its ability to accurately localize insulator defects. The TE-YOLOv8 framework integrates these advanced modules to create a comprehensive strategy that combines defect and background features, making it particularly well-suited for detecting small and complex targets in power transmission lines. These improvements significantly enhance the network's ability to meet the recognition requirements for detecting insulator defects under varying operational conditions.

Conventional convolution operations are spatial-agnostic and channel-specific, which makes them efficient but limits their ability to capture extended spatial dependencies. Small kernels process only local neighborhoods, restricting the network's capacity to disambiguate defect features in cluttered backgrounds. While large-kernel convolutions can expand the receptive field, their quadratic growth in parameters and computation renders them impractical for real-time deployment. Recent alternatives such as involution (Li et al., 2021) introduce spatial-specific and channel-agnostic properties, dynamically generating kernels at each spatial location to better capture fine-grained variations. However, involution alone remains insufficient for modeling global continuous information across the entire image, which is critical for coherent defect detection in complex aerial scenes.

To overcome these limitations, we propose the Global Convolution (GConv) module, which integrates spatial and channel information to capture global features while maintaining computational efficiency (Elfwing et al., 2018). As illustrated in Figure 4, GConv employs an asymmetric kernel decomposition strategy, factorizing large-kernel convolutions into sequential horizontal and vertical one-dimensional operations. This design preserves the effective receptive field of a large kernel while reducing complexity by approximately 3.5 × for typical kernel sizes (e.g., k = 7). Unlike standard convolution, which risks redundancy across channels, or involution, which focuses primarily on localized variations, GConv provides a balanced mechanism that retains global continuous information, minimizes channel redundancy, and enhances feature discrimination.

Mathematically, the GConv operation is defined as:

where Fin ε RC × H × W denotes the input feature map with C channels, height H, and width W, Wh ε RC × 1 × k represents the horizontal convolution kernel, Wv ε RC × k × 1 represents the vertical convolution kernel with kernel size k, * denotes the convolution operation, and σ represents the activation function (Appendix A).

The computational complexity reduction achieved through kernel decomposition can be quantified by comparing the number of multiply-accumulate operations required. For a standard convolution with kernel size k × k and C input and output channels, the computational cost is:

In contrast, the decomposed GConv operation requires:

The computational efficiency ratio, therefore, becomes:

Detecting insulator defects (IDs) in power transmission lines requires precise localization of target features that often resemble background structures. While the standard C3 module within the Cross Stage Partial (CSP) framework is effective for hierarchical feature extraction, it struggles to emphasize defect-specific information under complex backgrounds. To overcome these limitations, and inspired by ResNet (He et al., 2016) and CBAM (Woo et al., 2018), we propose the C3-Global Pooling Fusion (C3-GPF) module, illustrated in Figure 5.

The C3-GPF module enhances the discriminative capacity of feature representations by integrating global pooling operations into the CSP bottleneck structure. Unlike the conventional C3 block, which relies solely on local convolutional operations, C3-GPF aggregates spatial statistics across the entire feature map, allowing each position to access global context. This recalibration strengthens the network's ability to distinguish defect features from visually similar background elements, thereby improving detection precision. Structurally, the input feature 1 × 1 is divided into two branches. The first branch passes through a 1 × 1 convolution followed by the Bottleneck Enhanced X (BEX) module, while the second branch undergoes only a 1 × 1 convolution. The outputs of both branches are concatenated and processed through a final 1 × 1 convolution to produce the output feature y. The BEX module itself consists of 1 × 1 and 3 × 3 convolutions combined with the Global Pooling Fusion (GPF) operation, which is implemented in two variants: BE1 and BE2 (Figure 6).

The mathematical formulation of the C3-GPF module is expressed as:

where i indexes the bottleneck stages, and the Global Pooling Fusion operation is defined as:

Here, GAP denotes global average pooling that computes the spatial average across feature maps:

The global statistics are then transformed through a convolutional layer and sigmoid activation σ to produce channel-wise modulation weights, which are applied to the feature map through element-wise multiplication ?. This mechanism enables adaptive feature recalibration based on global context, enhancing the network's sensitivity to defect signatures while suppressing irrelevant background activations.

Detecting insulators and defects in aerial imagery requires effective integration of features across multiple spatial scales, as targets often vary widely in size and appearance. Conventional feature pyramid networks typically employ fixed fusion strategies, which fail to adapt to input characteristics or optimally weight contributions from different scales (Yu and Koltun, 2015). This limitation reduces their ability to handle extreme scale variation in complex inspection environments. To address this, we propose the Multiscale Information Fusion (MSIF) module, which implements a bidirectional feature pyramid architecture with learnable fusion weights and cross-scale feature interactions. The MSIF design consists of bottom-up and top-down pathways connected through lateral links with adaptive weighting. The bottom-up pathway aggregates features from fine to coarse scales:

where Plbu denotes the bottom-up feature at scale level ℓ, F_down represents a down-sampling operation, Flbackbone indicates backbone features at level ℓ, and are learnable lateral connection weights.

The top-down pathway propagates semantic information from coarse to fine scales:

where f_up performs up-sampling, and wℓtd are learnable top-down fusion weights. The final multi-scale feature representation results from a weighted combination of bottom-up and top-down features:

The learnable weights are constrained to sum to unity through SoftMax normalization:

where αl(k) are unconstrained learnable parameters and k indexes fusion pathways. This adaptive fusion strategy enables the network to dynamically adjust feature contributions based on input characteristics and scale-specific information content, improving detection performance across the wide range of object sizes encountered in transmission line inspection applications.

Insulator defect detection often involves discriminative features that occupy only a small subset of the total feature space, while background clutter generates substantial uninformative activations (Lin et al., 2017). Conventional fusion strategies treat all channels equally, which dilutes the importance of defect-specific signals. To overcome this limitation, we introduce the Weighted Feature Information Fusion (WFIF) module, which implements channel-wise attention to selectively emphasize informative channels and suppress irrelevant ones (Tan et al., 2020). The WFIF operation begins by computing global channel statistics using both average pooling and max pooling:

where GMP denotes global max pooling:

The global statistics are processed through a shared multi-layer perceptron to generate channel attention weights:

The MLP employs a bottleneck architecture with a reduction ratio r to constrain parameters:

where W1 ε RC/r × C and W2 ε RC × C/r are learnable weight matrices, and δ represents the ReLU activation function.

The final recalibrated feature map results from channel-wise multiplication:

This attention mechanism enables dynamic feature channel reweighting that adapts to input content, enhancing the network's focus on discriminative defect signatures while attenuating background interference.

The transformer-enhanced neck architecture represents a fundamental restructuring of the feature fusion network through integration of transformer encoder modules that enable global receptive fields and superior feature interaction capabilities (Vaswani et al., 2017). Traditional neck networks employ purely convolutional operations that process local neighborhoods, limiting their ability to capture long-range dependencies and global context essential for robust detection in complex scenes (Shaikh et al., 2025). Our transformer-enhanced design, illustrated within Figure 3, replaces selected convolutional layers in the neck network with transformer encoder blocks. Each transformer encoder processes the feature map as a sequence of spatial tokens and applies multi-head self-attention to model global dependencies:

where Q, K, and V denote query, key, and value matrices derived from input features through learned linear projections, and dk represents the key dimension. The multi-head attention mechanism employs parallel attention operations with different learned projections:

where each attention head is computed as:

The transformer encoder incorporates feed-forward networks and residual connections:

where the feed-forward network is defined as:

This transformer integration enables each spatial position to attend to all other positions in the feature map, capturing global context and long-range dependencies that are critical for disambiguating defect features from visually similar background elements in cluttered transmission line environments.

Accurate bounding box regression is critical for insulator defect detection, where small localization errors can significantly impact detection reliability (Gevorgyan, 2022). To improve localization precision and accelerate training convergence, we replace the conventional Complete Intersection over Union (CIoU) loss with the SCYLLA-IoU loss function. Unlike CIoU, which primarily considers centroid distance and aspect ratio, SCYLLA-IoU introduces four complementary components—angle cost, distance cost, shape cost, and IoU cost—to provide more comprehensive supervision for bounding box optimization. The SIoU loss is formulated as:

The angle cost term penalizes orientation differences between predicted and ground truth boxes:

where c_h represents the height difference between box centers, and σ is the Euclidean distance between centers:

The distance cost captures the center point deviation:

The shape cost addresses aspect ratio differences:

The complete loss function for the detection task combines classification, objectless, and localization terms:

By jointly modeling orientation, centroid alignment, aspect ratio, and overlap, SCYLLA-IoU provides richer geometric constraints than CIoU. This results in faster convergence during training and improved bounding box localization accuracy, particularly for elongated and irregular defect structures in aerial inspection imagery.

The training and inference procedures for TE-YOLOv8 are detailed in Algorithm 1, which outlines the training process, where the model receives a dataset, learning rate, batch size, and the number of epochs as input. The model's parameters are initialized randomly, and for each epoch, data augmentation is applied to the images. The images pass through the enhanced YOLOv8 architecture, which includes GConv, C3-GPF, MSIF, and WFIF modules to improve feature extraction and multi-scale fusion. Predictions are generated, loss is computed, and gradients are backpropagated to update model parameters. The model is evaluated on a validation set after each epoch, with the learning rate adjusted as needed, and the optimized model parameters θ ^* are returned at the end.

Algorithm 2 describes the inference procedure, where a test image, the trained model, confidence threshold, and NMS threshold are input. The test image is preprocessed, passed through the trained model, and predictions are made. These predictions are filtered by confidence, and Non-Maximum Suppression is applied to remove redundant boxes. The final bounding boxes are mapped back to the original image coordinates, and the refined detection results are returned. These procedures ensure TE-YOLOv8 efficiently detects defects in transmission line images with high accuracy and real-time performance.

The experimental evaluation employs two publicly available insulator defect datasets to ensure comprehensive assessment, cross-dataset validation, and reproducible comparisons. Both datasets provide diverse and challenging scenarios representative of real-world transmission line inspection conditions.

Extensive comparative experiments were conducted to evaluate TE-YOLOv8 against a comprehensive set of state-of-the-art object detection algorithms, including both two-stage and single-stage architectures. Tables 6, 7 present quantitative results across both datasets, showing the superiority of TE-YOLOv8 in multiple performance metrics.

Moreover, on the IDID dataset, TE-YOLOv8 outperformed all other detection algorithms, achieving the highest mean average precision (mAP) of 94.2%. This represents a significant improvement over baseline YOLOv8 (4.9%), YOLOv7 (4.3%), YOLOv8m (3.3%), YOLOv8-IDX (2.3%), and TE-YOLOv8 (1.9%). The model also demonstrated a precision of 95.3% and a recall of 93.1%, highlighting its strong balance between detection accuracy and completeness. These results validate the effectiveness of integrating transformer-based attention mechanisms and advanced feature fusion modules, significantly enhancing the ability of TE-YOLOv8 to detect insulator defects with improved precision and recall.

Similarly, on the CPLID dataset, TE-YOLOv8 achieved an mAP of 93.8%, outperforming YOLOv8 (5.1%), YOLOv7 (4.5%), YOLOv8m (3.4%), YOLOv8-IDX (2.5%), and TE-YOLOv8 (2%). The model's precision of 94.7% and recall of 92.9% further emphasize its effectiveness in detecting defects with minimal false positives while maintaining high detection completeness. The consistent improvements in both precision and recall across these datasets highlight the superior performance of TE-YOLOv8, underlining the importance of transformer-based attention mechanisms and feature fusion in improving overall detection performance. These advancements make TE-YOLOv8 a promising solution for real-time defect detection in complex environments.

Figure 7 illustrates successful detections on the IDID dataset, while Figure 8 shows successful detections on the CPLID dataset using TE-YOLOv8. Additionally, Figures 9, 10 track the progression of performance metrics such as training loss, validation loss, precision, recall, and mAP scores over the course of training. As the number of training epochs increases, both training and validation losses decrease, indicating that the model is improving. Concurrently, metrics like precision, recall, mAP@0.5, and mAP@0.5:0.95 increase, demonstrating enhanced detection capabilities. These results highlight the model's ability to effectively learn and improve over time.

Moreover, Figures 11, 12 present confusion matrices comparing the performance of TE-YOLOv8 against traditional detection models. The results clearly demonstrate that our method outperforms conventional models, underscoring its potential for further advancements in defect detection research.

Figures 13, 14 present representative failure cases of TE-YOLOv8 evaluated on the IDID and CPLID datasets, respectively. As shown in Figure 13 (IDID), the model occasionally produced false negatives, where insulator defects were missed under severe occlusion or when cracks appeared faint, low-contrast, or partially blended into the surface. False positives were also observed, particularly when background structures such as clamps, stains, or shadows resembled defect patterns and were incorrectly classified as faults. Similarly, Figure 14 (CPLID) illustrates failure cases in which defects were overlooked due to extreme scale variation (false negatives), as well as instances where normal insulators were incorrectly flagged as defective. These false positives often arose from background clutter, unusual lighting conditions, or reflections that created visual artifacts mimicking real damage. In both datasets, mislocalized bounding boxes were also observed, especially for elongated insulators or small defect regions.

These examples highlight the remaining challenges for practical deployment of TE-YOLOv8 in real-world inspection scenarios. They emphasize the need for further refinement of the detection framework, particularly in handling low-contrast defects, occlusion, and background interference. Future work will focus on domain adaptation to improve generalization across diverse environments and semi-supervised learning strategies to enhance robustness when annotated data are limited.

Comprehensive ablation experiments were conducted to quantify the individual and cumulative contributions of the proposed modules in TE-YOLOv8. We progressively integrated the Global Convolution (GConv), C2f-Global Pooling Fusion (C3-GPF), Multiscale Information Fusion (MSIF), Weighted Feature Information Fusion (WFIF), and the transformer-enhanced neck, starting from the YOLOv8 baseline, and measured changes in precision, recall, F1-score, mAP, parameters, and FPS. The results on IDID and CPLID (Tables 4, 5) demonstrate consistent, stepwise gains in detection accuracy with modest computational overhead, confirming that each component contributes meaningfully to performance.

Despite its promising performance, several limitations must be addressed. First, the current framework is focused exclusively on visible defects in optical imagery. Incorporating multi-modal fusion, such as infrared thermography or ultraviolet corona imaging, could help detect internal degradation and electrical tracking. Second, current datasets employ coarse classification, and finer taxonomies could enable more detailed severity assessments and failure mechanism identification. Third, few-shot learning techniques or synthetic data augmentation could address the challenge of limited training examples for rare defect categories. Finally, integrating the framework with active vision systems for adaptive image capture upon defect detection could improve inspection efficiency and reliability.