摘要
针对地面非合作特种车辆目标红外数据获取难度大、成本高,深度学习网络小样本数据条件下易于出现过拟合、网络泛化能力差等问题,本文以地面车辆红外数据为对象,提出了一种基于几何-特征空间变换的数据增强方法。首先,通过高清红外设备构建了原始地面车辆红外数据集;在此基础上,利用金字塔生成对抗网络(SinGAN)的空间特征重构机制,联合几何空间变换,对原始车辆红外数据进行了增广,并建立了地面目标红外数据集Infrared-VOC;最后,利用几种不同深度学习目标检测模型对增强后的红外数据集进行测试,验证了几何-特征空间联合变换方法数据增强的有效性,为地面非合作特种车辆红外数据增强提供了新方法。
In order to deal with the difficulty and excessive cost in acquiring infrared data of ground non-cooperative special vehicle targets,and solve the problems of over-fitting and poor generalization ability for small sample data sets,a data-augmented method based on geometric-feature space transformation is proposed for ground vehicle infrared data.Firstly,the original ground vehicle infrared data set is constructed by high-definition infrared equipment.Then,in conjunction with the geometric-feature space transformation method,the reconstruction mechanism of the SinGAN neural network is leveraged to augment the infrared data sets and build the Infrared-VOC data sets.Finally,a variety of the target detection models is employed to validate the performance of the augmented infrared data sets.The effectiveness of the geometric-feature space transformation for data augmentation is verified by several benchmark test cases,which provides a new method for ground non-cooperative special vehicle infrared data augmentation.
作者
赵晓枫
夏玉婷
徐叶斌
牛家辉
张文文
ZHAO Xiao-feng;XIA Yu-ting;XU Ye-bin;NIU Jia-hui;ZHANG Wen-wen(Rocket Force Engineering University,Xi′an 710025,China)
出处
《激光与红外》
CAS
CSCD
北大核心
2023年第7期1117-1124,共8页
Laser & Infrared
关键词
地面目标
红外数据集
数据增强
几何-特征空间变换
ground target
infrared datasets
data augmentation
geometric-feature space transformation
