迁移学习用于电磁目标识别

Transfer learning for electromagnetic target recognition

摘要:

迁移学习技术可以利用经验信息辅助当前任务，已在计算机视觉和语音识别领域得到广泛应用，但在电磁领域还没有取得明显的成就。电磁环境变化速度快，源数据或分类器模型在新环境中性能会显著下降，重新训练不仅需要大量的数据且费时费力。迁移学习技术与电磁目标识别任务十分相关，本文采用实测电磁目标数据集，探索迁移学习在解决电磁目标小样本问题中的几种应用，包括同类目标迁移和异类目标迁移。实验结果表明，通过将预训练模型迁移到目标域小样本识别任务，当目标域为同类源且标记样本只有20个情况下，相较于非迁移模型，验证准确率提高25%，并且大大缩短了目标域训练时间；当目标域为异类源时，也能够在保证识别准确率的同时使训练时间少于源域的1/5。

Abstract:

Transfer learning technology can use experience information to assist current tasks. It has been widely used in the fields of computer vision and speech recognition, whereas it has not made obvious achievements in the electromagnetic field. The electromagnetic environment changes quickly, and the performance of the source data or the classifier model in the new environment will be significantly degraded. Retraining not only requires a lot of data but also takes time and effort. Transfer learning technology is very related to the task of electromagnetic target recognition. Based on the measured electromagnetic target data set, this paper explores several applications of transfer learning in solving the problem of small samples of electromagnetic targets, including the transfer of similar targets and the transfer of heterogeneous targets. Experimental results show that by migrating the pre-training model to the target domain small sample recognition task, when the target domain is a similar source and there are only 20 labeled samples, the verification accuracy is increased by 25% compared with the non-transfer model and the training time is greatly shortened; when the target domain is a heterogeneous source, the training time can be less than 1/5 that of the source domain while ensuring the recognition accuracy.