摘要
电力设备外表面不规则缺陷具有特征不明显、形态多变的特点,当前常规端到端的图像识别算法表现出特征提取能力不足、泛化性差等问题。为此,提出了一种基于域适应网络的设备外表面不规则缺陷图像检测模型。该模型首先构建了包含特征生成器和分类器的域适应架构,以增强模型的泛化能力;然后通过添加纹理提取支路、辅助损失支路的方式增强特征生成器对纹理信息的提取能力;最后通过模型的对抗学习,实现在目标域上的准确识别。测试结果表明,所提方法能在角度、光照差异较大的目标域锈蚀和渗漏油图像中依然保持较高的识别精度,其中针对漏油、锈蚀隐患交并比指标分别达到了89%、85%。所提模型可为设备缺陷检测提供参考。
Irregular power defects on outer surface of power equipment have the characteristics of inconspicuous features and changeable shapes.Conventional image recognition algorithms show insufficient feature extraction ability and poor generalization.In this paper,a model for detecting the power irregular defects on outer surface of power equipment based on domain adaptation was proposed.Firstly,a domain adaptation architecture including feature generator and classifier is constructed to enhance the generalization ability of the model.Secondly,the ability of feature generator to extract texture information is enhanced by adding texture extraction branches and auxiliary loss branches.Finally,a high-precision detection model is obtained by the adversarial learning between generator and classifier.The experiment results show that the method proposed in this paper can still maintain high recognition accuracy in complex environments.The index of the intersection over union of oil leakage and corrosion defects can reach 89%and 85%,respectively.The model proposed in this paper can provide reference for equipment defect detection.
作者
张迎晨
王波
马富齐
罗鹏
张嘉鑫
李怡凡
ZHANG Yingchen;WANG Bo;Ma Fuqi;LUO Peng;ZHANG Jiaxin;LI Yifan(School of Electrical and Automation,Wuhan University,Wuhan 430072,China)
出处
《高电压技术》
EI
CAS
CSCD
北大核心
2022年第11期4516-4526,共11页
High Voltage Engineering
基金
贵州省科技计划(基于多源视觉大数据感知的输电设备状态智能诊断及通用平台)([2020]2Y039)。
关键词
不规则缺陷
域适应
电力深度视觉
纹理特征
缺陷识别
卷积神经网络
irregular defects
domain adaptation
power depth vision
texture features
defect recognition
convolutional neural networks
