基于元学习的少样本红外空中目标分类方法

Infrared aircraft few-shot classification method based on meta learning

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%.