LPC-Det: Attention-based lightweight object detector for power line component detection in UAV images

Abstract

Lacking timely maintenance of power line infrastructures is a prime cause of power shortages and large-scale blackouts. The current manual inspection method used in power line monitoring is time-consuming, less accurate, expensive, and prone to human error. Thus, there is a requirement for intelligent monitoring of power line infrastructure. Recent advancements in Unmanned Aerial Vehicles (UAVs) and deep learning have opened the area of intelligent power line infrastructure monitoring. However, the diversity of the UAV dataset can hurt the detection accuracy of lightweight object detectors, while the heavier one has a high computational cost. Thus, achieving a suitable trade-off between computational cost and detection accuracy is challenging. To this end, this work presents a lightweight and robust object detector named LPC-Det for power line component detection. The proposed LPC-Det, built on top of the YOLOv7 object detector, uses parameter-efficient attention modules to enhance the detection accuracy without much enhancement in the computation time. We also introduce a custom in-house power line dataset captured using UAV at different power line infrastructure sites in India. The dataset contains 10,968 power line images labeled into five types of components and aims to highlight diversity in power line infrastructure. Evaluated on the newly introduced dataset, the proposed LPC-Det using 640 × 640 input images achieved a remarkable baseline mAP@50 of 90.30%, a 1.7% improvement over the baseline YOLOv7. To further validate the efficacy of the proposed LPC-Det model, we trained and tested it on five public benchmark power line datasets. The proposed model consistently achieved a better mAP on all these datasets with slightly increased model size and parameters, GFLOPs, and inference time than the baseline YOLOv7 object detector.