Two-Stage Deep Learning Framework Improves Aerial Insulator Defect Detection Accuracy

In an advance for autonomous power grid inspection, computer vision researchers Sultan Daud Khan, Atif Naseer, Habib Ullah, and Mohib Ullah have engineered a coarse-to-fine-grained two-stage deep learning framework that enhances insulator defect detection in aerial drone imagery. Published in Frontiers in Artificial Intelligence, the study demonstrates that decoupling spatial localization from fine-grained defect classification delivers superior diagnostic accuracy over traditional single-stage detectors and standard region-based convolutional frameworks. Evaluated on the benchmark Insulator Defect Image Dataset, the proposed architecture attained an overall precision of 0.92, recall of 0.94, F1-score of 0.93, and mAP@50 of 0.93. By pairing an initial object detector with an isolated patch-level classifier governed by a targeted contrastive loss function, the research establishes a robust method for mitigating visual ambiguity in complex aerial inspection environments.

Electrical insulators serve as fundamental components across high-voltage transmission and distribution systems, providing mechanical support for conductors while maintaining electrical isolation from supporting towers. Because utility networks span vast distances, these components remain continuously exposed to severe outdoor operational environments, including extreme weather, cyclic mechanical stresses, dust accumulation, and pollution. Over time, environmental degradation triggers surface flaws, including glaze flashover damage, broken shells, and abnormal operating conditions. Left undetected, minor abnormalities degrade grid performance, cause electrical leakage currents, increase flashover hazards, and precipitate power outages. Early detection during routine surveillance remains crucial to preempting physical line failures and avoiding the substantial maintenance costs associated with reactive repairs.

Historically, power utilities have relied on manual visual inspections conducted by field experts traversing transmission corridors or climbing elevated tower structures. This conventional inspection paradigm presents severe operational shortcomings: it is labor-intensive, time-consuming, expensive, and fundamentally reactive, often initiated only after visible power disruptions occur. Furthermore, dispatching maintenance personnel near high-voltage equipment exposes workers to serious safety risks. Visual examination frequently misses minor defects, such as subtle glaze flashover tracking or slight ceramic chipping, because fine-grained flaws appear minuscule from afar. Although deploying unmanned aerial vehicles equipped with high-resolution digital cameras has automated aerial image capture, automated computer vision algorithms struggle to reliably interpret aerial data due to complex natural backgrounds, changing viewpoints, illumination variations, and tiny defect footprints.

To parse the growing volume of drone imagery, prior engineering efforts explored various automated deep learning detection architectures, ranging from multi-stage region proposal networks like Faster R-CNN to fast single-stage detectors such as YOLOv4, YOLOv5, and YOLO11 models. Despite achieving real-time inference speeds, standard single-stage detection models simultaneously predict bounding box coordinates and semantic class probabilities in a single forward pass, which severely constrains their ability to disentangle subtle class confusions between visually identical components. When defect targets occupy only a tiny fraction of an image, single-stage backbones often merge local structural defects with background noise. Conversely, traditional two-stage frameworks employing region proposal networks generate candidate proposals across broad receptive fields, which frequently miss tiny component defects and demand substantial memory and compute cycles, rendering them impractical for edge computing boards mounted on UAVs.

To overcome these limitations, the authors developed a coarse-to-fine-grained architecture that separates candidate localization from fine-grained defect classification. In the primary stage, a YOLOv8m detector processes 960 by 960 pixel images to predict coarse bounding boxes, preliminary class labels, and confidence scores for potential insulator assemblies. The researchers selected YOLOv8m because its multi-scale architecture enables reliable spatial modeling across diverse aerial viewpoints, distances, and image resolutions while maintaining computational efficiency. The first-stage network is trained using bounding box regression loss, distribution focal loss addressing boundary ambiguity around object edges, and preliminary classification loss. To eliminate redundant, overlapping detections generated during inference, non-maximum suppression is applied at a specified intersection-over-union threshold before candidate regions are forwarded to the patch extraction module.

Following suppression, bounding box coordinates serve as boundaries to crop image patches directly from the input frames. Rather than cropping strictly along predicted box edges, the framework incorporates a contextual padding mechanism parameterized by an expansion ratio of alpha equals 0.10. This deliberate padding adds a ten percent contextual margin around each detected region, ensuring that peripheral physical features, disk rims, and surrounding mounting interfaces are preserved within the cropped patch. Following padding, the localized image patches are resized to match the uniform input dimensions of the secondary classifier. Decoupling spatial localization in this manner removes the vast majority of irrelevant background clutter—such as vegetation, terrain, sky, and transmission tower framework—allowing subsequent neural layers to concentrate computational capacity entirely on discriminative object-level details.

In the secondary stage, the cropped and padded patches enter a dedicated patch-level classifier constructed around a lightweight MobileNetV3 backbone. Selected specifically for its balance of high classification accuracy and low computational complexity, MobileNetV3 functions as an efficient feature extractor suitable for deployment on hardware-constrained UAVs. The architecture comprises a feature extractor, an embedding projection module, and a classification head. Input patches pass through the feature extractor to produce hierarchical representations, which the projection module transforms into L2-normalized embedding vectors in a d-dimensional feature space. These normalized vectors are subsequently provided to a linear classification layer that computes class logits and yields a softmax probability distribution across the four distinct target classes. Because this classifier operates exclusively on localized patches, it captures minute surface patterns that are typically lost during full-image convolutional sweeps.

To resolve persistent misclassifications between visually similar insulator conditions, the authors introduced an objective termed Hard-Pair Contrastive Classification Loss. Conventional cross-entropy loss functions treat all target categories equally during backpropagation, penalizing classification errors uniformly regardless of whether confusion occurs between distinct objects or visually similar classes. In aerial insulator inspection, this uniform penalty causes frequent false alarms between normal undamaged units and intact components that lack flashover marks but share identical geometric contours. To enforce stricter class boundaries, the proposed loss function augments standard cross-entropy loss with an explicit hard-pair contrastive regularization term in the embedding space, weighted by a coefficient of lambda equals 0.2. Operating exclusively on designated hard classes, this metric learning term minimizes cosine distance between positive sample pairs belonging to the same category, while enforcing a contrastive distance margin of mu equals 0.5 to repel negative pairs.

The end-to-end inference workflow integrates both stages through a deterministic decision hierarchy and confidence aggregation rule. When an aerial image is evaluated, the first-stage YOLOv8m detector establishes spatial bounding boxes and generates preliminary class estimates. The cropped patches corresponding to these boxes are then independently classified by the MobileNetV3 network. If a categorical discrepancy arises between the first-stage proposal and the second-stage classification, the patch-level prediction takes precedence because the second stage is specifically optimized to resolve fine-grained class confusion. The final detection confidence score is computed by multiplying the initial YOLOv8m detection confidence score by the MobileNetV3 patch classifier’s predicted class probability. This multiplicative formulation ensures that the framework incorporates both the localization reliability of the primary detector and the fine-grained classification certainty of the dedicated patch model.

Experimental evaluation was conducted using the publicly available Insulator Defect Image Dataset, an aerial collection comprising 1,600 drone-captured transmission line images distributed across four operational categories. The dataset includes 1,532 images containing 13,060 normal instances labeled No_issues, 824 images featuring 2,564 instances of glaze_Flashover_damage, 812 images with 1,180 instances of shell_Broken, and an extreme minority class of notbroken-notflashed represented in only 40 images with 304 instances. To prevent data contamination, the authors partitioned the original images into mutually exclusive 70 percent training and 30 percent testing subsets at the source image level prior to data augmentation. Targeted image augmentation was restricted strictly to the training split to upsample minority classes without permitting augmented duplicates to leak into the test set.

Implemented in Python 3.11 with PyTorch 2.11.0, Ultralytics 8.4.26, and CUDA 12.8 on an NVIDIA GeForce RTX 2070 GPU, the proposed framework demonstrated substantial performance gains during extensive 100-epoch training trials. Against single-stage baselines, the proposed system outperformed YOLO11n (precision 0.83, recall 0.88, F1-score 0.85, mAP@50 0.89), YOLOv8n (F1-score 0.87, mAP@50 0.87), YOLOv8s (F1-score 0.89, mAP@50 0.90), YOLO11s (F1-score 0.88, mAP@50 0.91), and YOLOv8m (precision 0.88, recall 0.91, F1-score 0.90, mAP@50 0.91). Experimental modifications to YOLO11n incorporating attention modules (DCS-Attn) degraded performance to an F1-score of 0.68, demonstrating that inserting attention layers without structural decoupling increases complexity without improving discriminative power. Furthermore, when compared with established two-stage baselines, the proposed framework surpassed Faster R-CNN (F1-score 0.73, mAP@50 0.70), Cascade R-CNN (F1-score 0.88, mAP@50 0.88), Mask R-CNN (F1-score 0.88, mAP@50 0.89), Libra R-CNN (F1-score 0.90, mAP@50 0.90), and Double-Head R-CNN (F1-score 0.90, mAP@50 0.91).

Category-wise analysis revealed that the framework achieved its most notable breakthrough on the challenging notbroken-notflashed category, elevating classification precision to 0.86, recall to 0.93, and F1-score to 0.89, compared to standard YOLO models whose precision stagnated between 0.53 and 0.61. Architectural ablation experiments confirmed that while pairing YOLO11n with a heavier ResNet-50 second-stage backbone produced an mAP@50 of 0.94, it added 25.56 million parameters and dropped speed to 18 frames per second. By contrast, MobileNetV3 required only 2.54 million parameters and operated at 30 frames per second during patch classification, balancing accuracy and efficiency. The complete two-stage system contains 28.40 million parameters and achieves an end-to-end inference rate of 8.3 frames per second. Grad-CAM heatmaps confirmed that spatial activations concentrate tightly on defect contours rather than background terrain. Finally, the authors noted primary limitations: the pipeline cannot recover defects overlooked by the initial detector, was evaluated on a single dataset, and requires further latency optimization for edge deployment.

Subject of Research: Two-stage deep learning framework for power line insulator defect detection in aerial drone imagery

Article Title: Coarse-to-fine-grained insulator defect detection using two-stage deep learning framework

Article References: Khan, S. D., Naseer, A., Ullah, H., & Ullah, M. (2026). Coarse-to-fine-grained insulator defect detection using two-stage deep learning framework. Frontiers in Artificial Intelligence, 9, Article 1947204. https://doi.org/10.3389/frai.2026.1947204

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1947204

Keywords: deep learning, insulator defect detection, unmanned aerial vehicles, power transmission systems, YOLOv8m, MobileNetV3, hard-pair contrastive classification loss, computer vision, aerial inspection, object detection, fine-grained image classification, Insulator Defect Image Dataset, glaze flashover damage, broken shells, non-maximum suppression, Grad-CAM, power grid maintenance, contrastive loss