@article{Naya2026, 
author = {Christopher D. Naya and Elvis Twumasi and Eliel Keelson and Abdul-Majid Issah Malori},
title = {A lightweight frequency-aware enrichment module (FAEM) architecture for tiny insulator defect detection},
year = {2026},
journal = {AIMS Electronics and Electrical Engineering},
volume = {10},
number = {2},
pages = {205-239},
keywords = {tiny object detection, UAV inspection, insulator defects, lightweight deep learning, frequency-domain features, Fast Fourier Transform (FFT), computer vision},
url = {https://www.sciopen.com/article/10.3934/electreng.2026009},
doi = {10.3934/electreng.2026009},
abstract = {Real-time detection of tiny insulator defects is critical for power grid reliability. However, it remains a major challenge for Unmanned Aerial Vehicle (UAV) based inspection systems due to the trade-off between model accuracy and computational efficiency. To address this challenge, we proposed a novel lightweight object detector that incorporates a Frequency-Aware Enrichment Module (FAEM) into a You Only Look Once (YOLO) architecture for tiny insulator defect detection. FAEM introduces a learnable frequency-domain filtering pipeline operating on the Fourier magnitude spectrum while preserving phase information. It selectively amplifies high-frequency textural signatures of small defects (e.g., cracks and pollution flashovers) often missed by conventional detectors (YOLOv5n, YOLOv8s, and LiteYOLO-ID (2024)). The proposed approach was evaluated on three public datasets: IDID-Plus (~1, 600 images), PLIDD (~4, 927 images), and SFID (~320 images). The IDID-Plus and PLIDD datasets were used for training, validation, and testing with a 70/15/15 split, while SFID was employed exclusively for benchmarking robustness under degraded visibility. The performance was measured using Mean Average Precision (mAP) derived from the area under the precision recall curve. All experiments were conducted in a reproducible Google Colab Pro environment using an NVIDIA Tesla T4 GPU and PyTorch 2.6.0. The experimental results demonstrated that FAEM-YOLO achieved 90.7% mAP@0.5 and 50.2% mAP@0.5:0.95 on IDID-Plus, significantly outperforming LiteYOLO-ID (2024). On PLIDD, the model attained 73.9% mAP@0.5 and 40.5% mAP@0.5:0.95, surpassing the YOLOv8n baseline by 3.9 points on the stricter metric while using 20% fewer parameters (4.2M vs. 5.3M). Furthermore, FAEM-YOLO recorded 99.2% mAP@0.5 and 84.4% mAP@0.5:0.95 on the SFID dataset. These results underscore that frequency-domain enrichment generalizes well on different dataset; hence, it is an efficient strategy for visual inspection in UAV-based power line monitoring.}
}