| Radar target recognition based on High Resolution Response Profile(HRRP)has attracted much attention in RADAR-related research.At present,deep learning e.g.convolutional neural networks(CNN)has achieved state-of-the-art performances compared with traditional methods in HRRP recognition.This is due to the ability of addressing target-aspect and amplitude-scale sensitivity in HRRP applications.However,deep learning requires large-scale annotated data for training the network,besides,most deep learning methods rely on the assumptions that the training data and testing data are following the same distribution.In the real-world applications,it is challenge to acquire large-scale annotated HRRP data.In particular,it is extremely tough to collect sufficient data from the enemy to train the network.All of them will impede the application of deep learning in HRRP recognition.This paper focuses on incomplete HRRP target recognition.It is prerequisite to require a large amount of labeled HRRP data with all the aspect to train the network if the deep learning methods are applied to the HRRP target recognition.However,it is extremely rough in the real-world application.In order to reduce the limitation,we introduces the idea of transfer learning deep learning,and then proposes a method based on deep transfer learning and aspect-completed auxiliary HRRP data.The latter of which is much easier to be acquired in the real-world application.Firstly,this paper studies the HRRP target recognition with some widely-used convolutional neural network(CNN)to solve the aspect-sensibility problem in HRRP recognition.Secondly,with the idea of inductive deep transfer learning,this paper improves the recognition performance of aspect-incomplete HRRP by using auxiliary data sets.Thirdly,by combining the unsupervised domain adaptation methods in deep learning and transductive deep transfer learning,this paper use a regularization term into the deep learning models,and then further improves the recognition performance of the model on aspect-incomplete HRRP data sets.Finally,this paper compares the performance of the deep learning models and the deep transfer models validated on a set of incomplete HRRP data.The proposed model is evaluated under the HRRP of three real targets,and various aspect-incomplete settings are designed through different experimental settings.The situation is to verify the performance of the algorithm and achieve a deep transfer learning algorithm framework that can effectively improve the performance of HRRP recognition. |
