摘要
为了改善接触网管帽这类小尺度部件在故障检测过程中定位困难的情况,提出一种基于改进Faster R-CNN的接触网管帽目标定位算法。通过K均值聚类算法(K-means)对region proposal network(RPN)层中生成anchor boxes的比例及面积进行改进,所提算法在定位接触网管帽这类小部件上具有较好的表现。并通过比较VGG16、resnet50、resnet101、resnet152等4种特征提取网络在原始及改进的Faster R-CNN上定位管帽的准确率、召回率、准确率和召回率的调和平均F_(1)、单张检测时间等指标来选择最优特征提取网络。实验结果表明,基于resnet50的改进Faster R-CNN深度网络模型在接触网管帽定位中具有明显的优势,召回率为89.78%,定位准确率可以达到83.16%,F_(1)值为86.34%,单张检测时间为0.283 s。
In order to improve the difficulty of locating small-scale components such as contact network pipe cap in the process of fault detection, a contact network pipe cap target location algorithm based on improved Faster R-CNN is proposed. The proportion and area of anchor boxes generated in the region proposal network(RPN) layer are improved by K-means clustering algorithm(K-means), the proposed algorithm has good performance in locating small components such as contact network pipe caps. The optimal feature extraction network is selected by comparing the accuracy, recall, accuracy, harmonic average of accuracy and recall F_(1), and single sheet detection time of VGG16, resnet50, resnet101, and resnet152 feature extraction networks on the original and improved Faster R-CNN. The experimental results show that the improved Faster R-CNN deep network model based on resnet50 has obvious advantages in contact network pipe cap locating, the recall rate is 89.78%, the locating accuracy can reach 83.16%, the F_(1)value is 86.34%, and the single detection time is 0.283 s.
作者
顾桂梅
陈充
余晓宁
张存俊
仝甄
梅小芸
Gu Guimei;Chen Chong;Yu Xiaoning;Zhang Cunjun;Tong Zhen;Mei Xiaoyun(School of electrical engineering and automation,Lanzhou Jiaotong University,Lanzhou,Gansu 730070,China;China Railway Lanzhou Bureau Group Co.,Ltd.,Lanzhou,Gansu 730030,China;Qingyang Power Supply Company of State Grid Gansu Electric Power Company,Qingyang,Gansu 745000,China)
出处
《激光与光电子学进展》
CSCD
北大核心
2022年第4期132-142,共11页
Laser & Optoelectronics Progress
基金
甘肃省自然科学基金(20JR10RA216)
甘肃省科技计划(20JR10RA216)。
