针对地对空红外空中目标识别任务中数据量严重不足的问题,提出一种基于改进关系网络的小样本红外空中目标分类方法。该方法将关系网络模型、多尺度特征融合方法及元学习训练策略相结合,首先构造多尺度特征提取模块提取输入图像的特征向量,然后将支撑样本和预测样本的特征向量输入到关系模块中,根据关系值得到预测样本的类别标签。mini-ImageNet数据集上的实验结果表明:所提模型的分类精度显著高于其他经典的小样本学习模型。Infra-aircraft dataset上的实验结果表明:所提方法在仅有个位数样本的情况下,可完成多种机型的地对空红外图像分类任务。

To resolve the problem that the available data on the ground-to-air infrared aircraft identification task is considerably scarce, the small samples infrared aircraft identification classification method is proposed on the basis of an improved relation network. This method combines the relation network model and the multi-scale feature fused method with the meta learning training strategy. First, a multi-scale feature extraction module is constructed to extract the feature tensors of input images. Then, the feature tensors of support samples and test samples are inputted into the relation module, and the category labels corresponding to test samples are predicted based on the relation value. The results of the proposed model on the mini-ImageNet dataset show that the classification accuracy of the proposed model is significantly higher than those of other conventional learning models using small samples. The experimental results based on the Infra-aircraft dataset verify that the proposed model can realize the ground-to-air infrared image classification task of various aircraft types even when the number of samples is limited.