摘要
电力线自动巡检需要快速且准确的目标检测算法。传统算法如DPM由于精度不高难以满足要求,基于深度学习的两步法如Faster R-CNN虽然可以达到高精度要求,但是检测速度仍远不及实时标准。基于深度学习的一步法如YOLO、SSD同时实现了高精度和高速度。针对自动巡检的精度和速度需求,提出了基于YOLOv3的电力线多种关键部件的实时检测方法。构建了包含5种电力线关键部件的数据集,并在该数据集上应用YOLOv3实现了高精度实时检测。结果显示,本文所采用的方法在检测精度评价指标mAP上达到90.8%,检测速度达到57.658fps。相比采用Faster R-CNN的两步法mAP提高约3%,检测速度提高了约7倍,说明基于YOLOv3的电力线关键目标检测方法能够很好地满足电力线自动巡检的精度和速度要求。
Fast and accurate object detection algorithms are required to realize automatic inspection on power transmission line.Traditional algorithms like DPM,however,cannot meet the requirements due to their poor accuracy.Deep Learning two-stage algorithms like Faster R-CNN meet the high accuracy requirement,but they still fall far below real-time benchmark.Deep Learning one-stage algorithms like YOLO and SSD achieve both high accuracy and speed.We presented a key components detection method based on YOLOv3 for automatic inspection on power transmission line.Then we trained and tested YOLOv3 based method and Faster R-CNN based method on power transmission line dataset which contains 5 kinds of key components.According to the experimental result,YOLOv3 achieves 90.8%on detection accuracy metrics mAP,which is about 3%higher than Faster R-CNN.YOLOv3 is also about 7×faster(57.658 fps)than Faster R-CNN.The result shows that YOLOv3 based power transmission line key object detection method meets the requirements of detection accuracy and speed.
作者
董召杰
Dong Zhaojie(Dingxin Information Technology CO.,LTD,Guangzhou 510623,China)
出处
《电子测量技术》
2019年第23期173-178,共6页
Electronic Measurement Technology
基金
中国南方电网有限责任公司科技项目(090000KK52170124)资助。