基于深层-浅层双流学习图模型的无监督少样本红外空中目标识别网络

An unsupervised few-shot infrared aerial object recognition network based on deep-shallow learning graph model

在军事空中目标识别领域,由于样本数量缺失,现有人工智能算法无法完成准确识别。文章利用已有足量辅助域图像辅助少样本应用域进行跨域目标识别,解决因标签缺失与样本稀疏导致的识别模型泛化能力不强及性能不佳问题。文章提出一种基于深层-浅层双流学习图模型(D-SLGM)的跨域目标识别算法。首先,提出一种深层-浅层双流特征提取算法,解决无监督少样本条件下特征表示困难的问题;同时,提出一种基于图模型的特征融合算法,实现特征间高精度融合;基于融合后的特征训练识别模型,提升算法的泛化能力。使用自建空中目标数据集,设计三种应用场景。实验结果表明,D-SLGM平均识别准确率均值达到78.2%,优于对比方法,在实际空中目标识别应用中具有较大潜力。

In the field of military aerial object recognition,due to the lack of samples,current artificial intelligence algorithms cannot perform well.This paper uses the existing sufficient auxiliary domain images to assist the application domain with few samples for cross-domain object recognition and solves the problem of weak generalization ability and poor performance of the recognition model caused by missing labels and sparse samples.A cross-domain object recognition algorithm named Deep-Shallow Learning Graph Model(D-SLGM)is proposed.Firstly,a deep-shallow two-stream feature extraction algorithm is proposed to solve the problem of feature representation under unsupervised few-shot conditions.At the same time,a feature fusion algorithm based on graph model is proposed to realize high precision fusion between features.Then,a recognition model is trained based on the fused features,the generalization ability of the algorithm is improved.The self-built aerial object dataset is adopted with three application scenarios.The experimental results show that the mean average recognition accuracy of D-SLGM reaches 78.2%,which is better than those of the comparison methods.D-SLGM has great potential in actual aerial object recognition applications.