30 June 2025 in IEEE Transactions on Geoscience and Remote Sensing
With the advancement of high-resolution aerial imaging technology enabled by unmanned aerial vehicles (UAVs), insulator defect detection based on images has emerged as a key approach for intelligent inspection of overhead transmission lines. However, the insulator defective region detection (IDRD) task based on such imagery continues to face several critical challenges, including complex background interference, difficulty in detecting small-scale defects, and class imbalance in sample distribution. To tackle these challenges, we propose a state space model guided deep feature perception network for insulator defect detection (SGFP-YOLO). Specifically, the model introduces a Bi-directional Feature Enhancement Module (BFEM) and a Dual-context Feature Refinement Module (DFRM), both guided by a state-space framework, to enhance global feature and capture local detail features. Additionally, to address the hard-easy sample imbalance problem in the detection process, we introduce Focused Balance Loss, improving the performance of model in both classification and regression tasks. Experiments are conducted on the self-constructed TLID dataset and the publicly available IDID dataset. The results show that SGFP-YOLO outperforms existing advanced models in several metrics, including mAP@0.5 and F1-score, especially in complex background and small-scale defect detection, providing an efficient and accurate solution for intelligent detection of insulators in transmission line.
25 August 2025 in IEEE Transactions on Industrial Informatics
Intelligent detection of electric power fittings is essential for monitoring conditions and ensuring the stability of transmission and distribution lines. However, the lack of real-world aligned datasets has hindered progress. We developed an electric power fitting dataset that includes challenges, such as long-tailed distribution, multimorphology, multiscale, small targets, and dense occlusion. It contains 1560 unmanned aerial vehicle (UAV)-captured images of 18 types of electric power fittings, annotated with over 13 500 bounding boxes. Compared to existing datasets, ours offers significant advantages in categories, diversity, and complexity. We further propose a multimorphology contrastive learning approach to ensure consistent feature representation and address morphological diversity. By minimizing intraclass variation and maximizing interclass separability, our method achieves more discriminative feature representations. Experimental results indicate that our method achieves an average precision of 86.2% for multifitting detection—improving upon the current state-of-the-art by 2.6%. Meanwhile, cross-dataset testing further verifies the broad applicability of our approach.