克服深度学习训练数据不足问题的主流方法是深度迁移学习,当目标域训练样本缺失不严重且目标域与源域样本空间相差不大时,该方法的性能几乎不亚于在充足训练样本集上训练得到的深度模型;但当训练样本数很少甚至只有几个时,该方法会由于过拟合导致模型泛化能力太差而失效。在非合作目标雷达像识别中,往往难以获取充足的训练图像 样本。针对这一问题,文中将单样本学习(OSL)引入训练样本极少情况下的雷达像识别中,提出了一种特征迁移和原型网络相结合的小样本雷达像识别新方法。文中详细描述了该方法原理及流程,并采用MSTAR数据集对所提方法进行了实验验证。实验结果表明,在目标域训练样本数量极少的情况下,所提方法的识别性能优于传统的深度迁移学习方法。

Deep transfer learning is the widely prevalent method for training deep neural networks with insufficient training data. On condition that the training samples are not seriously insufficient and the target domain is similar to the source domain, this method can achieve the similar performance as normal deep learning method. But the recognition performance of transfer leaning method will greatly decrease while only several samples or only one sample are available. In the field of non-cooperative target recognition, it is difficult to obtain sufficient training samples which limits its application. To deal with this problem, this paper introduces oneshot learning (OSL) into non-cooperative radar image recognition, and proposes a one-shot recognition method based on feature transfer and prototypical networks. The principle and procedures of this method are given in this paper. Experiments on MSTAR dataset show that the performance of our method is superior to deep transfer learning when the training samples in the target domain are seriously deficient.

TN957.52

国家自然科学基金委资助项目(61921001)