Research On Synthetic Aperture Radar Target Recognition Based On Transfer Learning For Small Samples

Image recognition technology is the focus of current research,which is helpful to understand the basic technology of artificial intelligence.Thereinto,Synthetic Aperture Radar(SAR)targets are extensive considered in urban monitoring,natural environment investigation,military target reconnaissance and etc.According to the imaging characteristics,there exists the following difficulties concerned to SAR target recognition.One is the effective model design,embodied in the design of objective function and cost function,and the optimization of ultra-high dimensional parameters,etc.under the small samples circumstance.The next is the model generalization,in which the lack of samples leads to the over-fitting limitation and the model generalization ability.The finally is the insight of feature acquisition with powerful discriminative under the case of small samples,emphasizing on intuitive features from raw images and the impact of small samples.In this paper,we introduce the transfer learning approaches for small samples,combined with deep learning framework,for SAR target recognition task.The details are abstracted as follows:(1)Suffering from the independent changes such as over-fitting and angle estimation error caused by insufficient training data,a lightweight SARNet using data amplification technique is proposed in this paper.Centroid method is adopted combined with Clock-wise rotation technique,to obtain the adequate training samples to avoid the over-fitting problem in the course of training.Meanwhile,the proposed network has been validated lightweight and simple,the performance is greatly improved compared with the traditional complex network model,not only does the storage requirement reduce,but also the redundant computation is considerably decent.We also explore the function of various activation functions,studying the influence of different sample numbers on the performance of the deep learning network,and obtains better performance on the public MSTAR dataset.(2)Limited by the challenge of SAR target training samples,a joint regularization and transferred MS-CNN network framework is proposed in this thesis.ROI feature extraction and data amplification are combined to preprocess the regions of interest within SAR images,to enrich the variety of SAR training samples and expand the dataset effectively.Secondly,combining with MS-CNN framework,L~2 regularization strategy is used to extract the features of robust SAR targets.Meanwhile,the dropout strategy is utilized to alleviate the overfitting problem.Finally,transfer learning theory is introduced to improve the SAR image feature recognition performance,and to validate model effectiveness and robustness.The method achieves high recognition accuracy on public datasets MSTAR and established small sample datasets 1/2,1/3,1/4,1/8,which is superior to other CNN advances.(3)For the purpose of solving the problem of small samples in SAR ATR,a novel deep transferred multi-level feature fusion attention network(MFFA-SARNET)with dual optimization loss is presented for the task.Firstly,a multi-level feature attention network(MFFA)is schemed to obtain more discriminative features from SAR images by fusion method,and alleviate the impact of background features through the followed attention module that focuses more on the target features.Secondly,a novel dual optimized loss is incorporated to further optimize the classification network,learning more robust and discriminative power of the features.Thirdly,transfer learning is utilized to validate the variances and small samples classification task.Extensive experiments on public database with three different configurations have consistently demonstrated the effectiveness of our proposed network,significant improvements have been yielded compared to the state-of-the-arts under the small samples conditions.