| Synthetic Aperture Radar(SAR)is an important remote sensing device with widespread applications in both civilian and military fields.At present,with sufficient labeled samples,traditional and deep learning-based SAR target recognition algorithms have been able to obtain good recognition performance.However,obtaining a large number of labeled samples itself is a difficult task,it is also necessary to quickly and accurately determine the type of target when lacking labeled samples.This dissertation focuses on SAR target recognition tasks when lacking labeled training samples.“Lacking labeled samples”mainly covers the following two situations: First,there are only a small number of labeled training samples used for the recognition task,which is called the small sample SAR target recognition problem in this dissertation,and the target recognition tasks under limited training samples and fewshot training samples are collectively referred to as the small sample SAR target recognition task.Taking the Moving and Stationary Target Acquisition and Recognition(MSTAR)public dataset as an example,limited training samples in this dissertation mean that the number of training samples for each category is 10% of the original complete dataset,and the number of samples for each category is about 20,and few-shot training samples in this dissertation means that the number of labeled training samples for each category is only 5;Second,there are no labeled real training samples,but only labeled simulated samples for the recognition task,which is called the heterogeneous SAR target recognition task.According to the number of real labeled training samples used for training,the research of this dissertation can be divided into the study of limited-sample,few-shot and heterogeneous SAR target recognition methods,and be summarized as the following six parts:(1)The first part of the dissertation explores the SAR target recognition problem under limited training samples based on Gaussian Restricted Boltzmann Machine(GRBM)and the corresponding Deep Belief Networks(DBN).When the labeled training samples are limited,RBM can use unlabeled data to co-train the network by its unique unsupervised training process,to learn the common feature representations of both labeled training data and unlabeled test data,and can avoid the overfitting problem associated with limited training samples to a certain extent.The experimental results show that for the three-type target recognition task of the MSTAR dataset,with only 23 labeled training samples for each category,DBN trained with only 69 labeled samples achieves about 53% recognition rate;When the unlabeled test data and the 69 labeled samples are used to pre-train DBN together in the unsupervised training way,and only labeled training samples are used to fine-tune the network,the recognition accuracy can be increased to about 85%.(2)The second part of the dissertation proposes an Angular Rotation Generative Network(ARGN)for SAR target recognition tasks under limited training samples based on the idea of transfer learning.The feature space of trained ARGN contains complete target information under different aspect angles of the source data,which compensates for the missing target information under certain aspect angles in the target domain,and it is helpful to improve the performance of image generation and target recognition in the target domain.The experimental results show that for the three-type target recognition task of the MSTAR dataset,with only 23 labeled training samples for each category,87% recognition accuracy can be achieved by extracting features of ARGN and using support vector machine classifier.(3)The third part of the dissertation proposes an Attribute-Guided Generative Adversarial Network(AGGAN)and an Improved Episode Training Strategy(I-ETS),which aims to improve the quality of generated images,and then realize the data augmentation for the corresponding recognition task under few-shot training samples.The experimental results show that AGGAN and I-ETS can effectively improve the quality of the generated images,and the generated images can be applied to the few-shot SAR target recognition problem.Ultimately,by combining the generated data with traditional data augmentation methods,68%recognition accuracy can be achieved for the three-type target recognition task of the MSTAR dataset,with only 5 labeled training samples for each category.(4)For the heterogeneous SAR target recognition problem,the fourth part of the dissertation designs a multi-scale feature extraction module and combines it with the deep domain confusion method;furthermore,this chapter adopts the ensembling technique,and a target recognition method called an ensemble multi-scale deep domain adaptation recognition framework is proposed to obtain more stable recognition results.Using the three-type simulated SAR target data provided by the Science and Technology on Electromagnetic Scattering Laboratory,Beijing Institute of Environmental Features,Beijing,China,and the MSTAR dataset,two experimental settings are considered and the proposed method has better recognition performance.Especially with the target variants in the test data,the results of the proposed method are more stable,and the recognition performance of our approach improves by at least 9% compared with the existing representative method A-Conv Nets.(5)The fifth part of the dissertation proposes a heterogeneous SAR target recognition method based on data transformation and label smoothing classification loss function.On the one hand,this chapter improves the Cycle-Consistent Adversarial Networks(Cycle GAN)by adding cross domain constraints,to realize the transformation from simulated data to real data and maintain the label consistency in the transformation process;on the other hand,this chapter combines the multi-scale feature extraction network for recognition and uses label smoothing classification loss function to reduce the influence of transformed data under label noise on the recognition model.The experimental setup is the same as that in the fourth chapter.The experimental results verify that Cross Domain Cycle GAN(CD-Cycle GAN)can improve the data distribution efficiently,and at the same time,maintain the category information in the transformation process;label smoothing classification loss function also contributes to the improvement of the recognition performance,and without using the domain adaptation methods,the method proposed in this chapter can achieve comparable recognition performance with the method proposed in the fourth part.(6)Based on the newly published Synthetic and Measured Paired Labeled Experiment(SAMPLE)dataset,the sixth part of the dissertation proposes a gradual domain adaptation recognition framework with pseudo-label denoising method for the heterogeneous SAR target recognition problem.The method combines the domain adaptation loss with self-training to gradually reduce the feature differences between the simulated data and the real data,so as to improve the recognition performance progressively;in the self-training process,a pseudolabel denoising method is proposed to effectively eliminate some noisy pseudo labels and improve the reliability of pseudo-labeled data.Two common experimental scenarios based on the SAMPLE dataset are considered in this chapter,and the experimental results show that,the proposed method in this chapter achieves an average recognition result of about 96% and97% in the two recognition scenarios,respectively,which is superior compared with some existing methods for the SAMPLE dataset. |
