| Synthetic Aperture Radar(SAR),also known as Synthetic Aperture Radar(SAR),is a kind of Radar which synthesizes the real Aperture of the smaller Radar into a larger equivalent Aperture by means of data processing by using the relative motion between the Radar and the target,which can effectively improve its resolution.Therefore,SAR has played an important role in military reconnaissance and other fields since it came into being.In recent years,with the rapid development of deep learning technologies,researchers at home and abroad have conducted a great deal of studies on SAR object recognition task focusing on Convolutional Neural Network(CNN).However,the heavy dependence of Neural Network on large data brings new challenges to SAR interpretation:(1)The insufficiency of labeled SAR data leads to over-fitting of the neural network.(2)The multiparameter sensitivity of SAR targets causes the degradation of model generalization performance.(3)The incompleteness of multi-parameter datasets makes the model feature extraction algorithm invalid under the target recognition task under the complex environment.From the perspective of transfer learning(TL),this paper designs three algorithms aiming at few-shot SAR target recognition,SAR image target domain adaptation and cross-domain SAR target recognition under the missing of heterogeneous data.The main research work of this paper is as follows:(1)Aiming at the over-fitting problem of deep learning models caused by limited SAR image target samples,combined with the concepts of transfer learning,the feasibility of model transferring in SAR image target recognition is verified.Furthermore,a new SAR image target recognition technique based on adversarial generative network and model transfer learning is proposed.The proposed method pretrains the feature extraction module using GAN and target domain data and finetunes the network using source domain data within a multi-class hinge loss.Experiments based on MSTAR data illustrate the proposed method effectively mitigates the over-fitting problem under the condition of few-shot training samples.(2)Aiming at the poor generalization of neural network models caused by the multiple-parameters sensitivity of SAR targets,the classical domain adaptation methods based on metric divergence and admissibility learning are studied combining with the unsupervised Domain adaptation(DA)technology.A SAR image target domain adaptation technique based on class confusion regularization and selective pseudo-labelling generation strategy is proposed.The proposed method directly learns knowledge from target domain data,which enables model’s adaptation in target domain.To suppress the local maximum problem caused by error accumulation,class confusion loss is introduced into the iterative network training process.Domain adaptation experiments among data of different depression angles in MSTAR datasets proves the proposed method outperform other classical domain adaptation methods.(3)Aiming at the problem of the lack of target domain data and the poor adaptability of inductive learning and transductive learning.A task-driven domain adaptation method for cross-domain SAR target recognition is proposed.The proposed method acquires simulated data according to the imaging condition information of source and target domain.Then,the domain discrepancy of measured data is decreased with the optimizing of domain discrepancy metric on simulated data,which both improves the domain robustness and classification ability of the model,and avoids the availability of target domain measured data.Experiments based on MSTAR and SARSIM datasets proves the feasibility of bridging role of simulated data in reducing the domain discrepancy of measured data,and the proposed method significantly improves the cross-domain target recognition performance under different imaging conditions.In this paper,three target recognition algorithms are designed by optimizing model training,designing processing framework and introducing simulation data,and the effectiveness of the proposed algorithm is verified by several groups of experiments.The research in this paper is helpful to improve the performance of target feature extraction,detection and recognition in high-resolution SAR images. |
