迁移学习在天基红外目标识别中的应用

Application of transfer learning in space-based infrared target recognition

星载红外传感器对飞行的火箭进行识别时,因为观测数据有限,一般属于小样本甚至单样本学习的分类问题。本文建立了一种以一维全卷积为主体结构的孪生神经网络,将多分类问题转化为比较相似度的验证问题;并利用UCR时间序列数据集的预训练权重,对孪生神经网络的卷积特征提取部分进行知识迁移,最后使用公开文献中火箭红外辐射强度序列数据对网络进行微调,形成了一个能够比较两型火箭相似度的迁移学习网络。实验结果表明,本文建立的模型能够从其他数据集中学习到有利于时间序列相似性度量的信息,训练过程也具备可行性,在单样本情况下能较好地实现对火箭的识别。

When the space-based infrared sensor is used to identify the type of flying rocket,because of the limited observation data,it generally belongs to a classification problem of small-sample learning or even one-shot learning.In this paper,the main structure of Siamese Neural Network with one-dimensional full convolution was established,which transformed the multi-classification problem into a verification problem of comparative similarity.With the pre-training weights of UCR time series data set,the knowledge transfer of convolution feature extraction part of the Siamese Neural Network was carried out.Finally,the observed time series of rocket infrared radiation intensity were used to fine-tune the network,forming a transfer learning network which can compare the similarity between two types of rockets.The experimental results show that the model in this paper can learn the knowledge which is conducive to the similarity measurement of time series from other data sets,and presents the feasibility of the training process,which can better realize the recognition of rocket types in the case of one-shot sample.