摘要
针对空中红外目标样本数目不足、细粒度分类精度低等问题,提出一种基于元学习的少样本红外空中目标分类的方法。该方法以元学习为基础,结合多尺度特征融合,在减少计算量的同时有效提取不同分类任务之间的共性,再利用微调策略实现对不同任务的分类。实验证明,此方法在提升mini-ImageNet数据集分类精度的同时可减少约70%的计算量,对仅有少量样本的红外空中目标细粒度分类准确率可达到92.74%。
Aiming at the problem of insufficient samples of infrared aircrafts and low accuracy of fine-grained classifica‐tion,a method of infrared aircraft few-shot classification based on meta learning is proposed.Based on meta learning and combined with multi-scale feature fusion,this method can effectively extract commonness among different classifi‐cation tasks while reducing computation,and then classify different tasks with fine-tuning.The experiments proved that this method could improve the classification accuracy of mini-ImageNet dataset while reducing the calculation amount by about 70%.The accuracy of fine-grained classification for infrared aircrafts with few samples reached 92.74%.
作者
陈瑞敏
刘士建
苗壮
李范鸣
CHEN Rui-Min;LIU Shi-Jian;MIAO Zhuang;LI Fan-Ming(Key Laboratory of Infrared System Detection and Imaging Technology,Chinese Academy of Sciences,Shanghai 200083,China;University of Chinese Academy of Sciences,Beijing 100049,China;Shanghai Institute of Technical Physics,Chinese Academy of Sciences,Shanghai 200083,China)
出处
《红外与毫米波学报》
SCIE
EI
CAS
CSCD
北大核心
2021年第4期554-560,共7页
Journal of Infrared and Millimeter Waves
基金
上海市现场物证重点实验室基金(2017xcwzk08)。
关键词
红外图像
细粒度分类
少样本学习
元学习
多尺度特征融合
微调
infrared image
fine-grained classification
few-shot learning
meta learning
multi-scale feature fusion
fine-tuning
