摘要
为实现在海量线路巡检图像中对绝缘子的自动识别,提出一种基于对抗自学习的跨域绝缘子识别方法。该方法由对抗学习和自训练学习2个阶段组成。在对抗学习阶段,通过特征提取器和分类器之间的对抗学习,使模型分别获得对玻璃绝缘子和复合绝缘子具有鲁棒性的分类特征。在自训练学习阶段,首先,采用有标签的玻璃绝缘子样本对模型进行预训练;然后,将无标签的复合绝缘子样本输入网络,并选择置信度高的样本赋予软标签对模型进行再次训练,使模型最终获得在不同域上的泛化能力。与现有方法相比,该文方法采用分属不同材质的绝缘子样本对深度神经网络进行2个阶段的训练,在有效降低模型训练过程中样本标注量的同时,解决了跨域识别不同材质的绝缘子的问题。
In order to realize the automatic recognition of insulators in inspection images of massive lines,a cross-domain insulator recognition method based on adversarial self-learning is proposed.The method consists of adversarial learning and self-training learning two stages.In the adversarial learning stage,through adversarial learning between the feature extractor and the classifier,the model obtains robust classification features of glass insulators and composite insulators.In the self-training learning stage,firstly,the model is pre-trained with labeled glass insulator samples;then,the unlabeled composite insulator samples are input into the pre-trained model,and the sample with high confidence is selected to give the soft label to retrain the model,so that the model obtains generalization ability in different domains.Compared with the existing methods,the proposed method uses insulator samples of different materials to train the deep neural network in two stages,which effectively reduces the amount of sample annotations in the model training phase and solves the problem of cross-domain identification of different materials.
作者
李凡
高瞻
王红斌
李爽
庞健
徐开雄
余正涛
Li Fan;Gao Zhan;Wang Hongbin;Li Shuang;Pang Jian;Xu Kaixiong;Yu Zhengtao(Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650500,China;Yunnan Key Laboratory of Artificial Intelligence,Kunming University of Science and Technology,Kunming 650500,China)
出处
《南京理工大学学报》
EI
CAS
CSCD
北大核心
2020年第6期651-659,共9页
Journal of Nanjing University of Science and Technology
基金
国家重点研发计划(2018YFC0830105,2018YFC0830100)
云南电网公司科技项目(YNKJXM20190729)。
关键词
对抗学习
绝缘子
线路巡检
图像
自训练学习
特征提取器
玻璃绝缘子
复合绝缘子
adversarial learning
insulators
lines inspection
images
self-training learning
feature extractor
glass insulators
composite insulators
