基于CenterNet的半监督起落架自动标注  

Automatic marking of semi-supervised landing gears based on CenterNet

作　　者：方伟 汤淼 闫文君 张婷婷 FANG Wei;TANG Miao;YAN Wenjun;ZHANG Tingting(Naval Aviation University,Yantai 264001,China)

机构地区：[1]海军航空大学,山东烟台264001

出　　处：《兵器装备工程学报》2023年第4期239-244,共6页Journal of Ordnance Equipment Engineering

基　　金：国家自然基金项目(91538201);泰山学者工程专项经费基金项目(ts201511020);信息系统安全技术重点实验室基金项目(6142111190404)。

摘　　要：针对飞机起落架标注人工标注费时费力问题,提出了将CenterNet目标检测模型与半监督学习结合起来对飞机起落架进行自动标注。该方法在CenterNet的主干特征网络ResNet50基础上嵌入通道注意力机制并对其有效性进行了验证,结合半监督学习,用标记样本训练的模型对未标记样本进行标注并对得到的问题样本进行人工修正后叠加进原标记样本组成新的数据集继续训练,最终生成性能较好、能够自动标注的目标检测模型。实验结果表明,模型经过5次迭代训练后,得到标注模型的精确率达到95.29%,平均准确率达到92.16%,对飞机起落架的定位能够满足标注要求。Aiming at the time-consuming and laborious problem of manual tagging of aircraft landing gears,this paper proposes an automatic tagging of aircraft landing gears by combining the target detection model of CenterNet with semi-supervised learning.Based on ResNet50,the backbone feature network of CenterNet,this method embeds the channel attention mechanism and verifies its effectiveness.Then,combined with semi-supervised learning,unlabeled samples are marked with the model of labeled sample training,and the obtained problem samples are manually corrected and superimposed into the original labeled samples to form a new dataset for further training.Finally,a target detection model with good performance and automatic labeling is generated.The experimental results show that,after five iterative training for the model,the precision of the annotation model is 95.29%,and the average accuracy is 92.16%,which meets the annotation requirements for the positioning of the aircraft landing gear.

关 键 词：图像自动标注 CenterNet 通道注意力机制 半监督学习 目标检测模型 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]