摘要
为准确定位航拍红外图像中光伏组件故障的区域位置,在单阶段多框检测(Single Shot Multibox Detector,SSD)模型的基础上设计了ResNet18_FPN_DN_SSD模型。首先用ResNet18替代SSD模型的基础网络VGG16,以提高故障特征的提取能力;然后引入DR loss,针对目标样本类别失衡及负样本过多的问题进行优化改善;最后在非极大值抑制(Non-Maximum Suppression,NMS)基础上做加权处理,使分类置信度高的边框充分利用周围对象的信息,提高预测框的分类置信度与定位准确率。实验表明:所提出的模型对图像中故障目标的检测效果,在定位精度、分类置信度和m AP上均优于传统SSD模型。
For accurate positioning of photovoltaic component failures in aerial infrared image area location,the ResNet18_FPN_DN_SSD model is designed on the basis of single stage box testing( Single Shot Multibox Detector,SSD) model. Firstly,Resnet18 was used to replace the basic network VGG16 of the SSD model to improve the ability of fault feature extraction. Then,DR loss is introduced to optimize and improve the imbalance of target sample categories and excessive negative samples. Finally,non-maximum suppression( NMS) was applied to improve the classification confidence and location accuracy of the prediction box by making full use of the information of the surrounding objects. Experimental results show that the proposed model is superior to the traditional SSD model in location accuracy,classification confidence and m AP.
作者
姜萍
王天亮
栾艳军
JIANG Ping;WANG Tianliang;LUAN Yanjun(College of Electronic Information Engineering,Hebei University,Baoding Hebei 071000,China)
出处
《激光杂志》
CAS
北大核心
2022年第5期43-48,共6页
Laser Journal
基金
河北省自然科学基金(No.A2020201021)。
