Research On Synthetic Aperture Radar Image Target Recognition Based On Deep Learning

Synthetic Aperture Radar(SAR)images play an important role in the fields of service earth resource exploration,regional development level assessment and national defense system construction.The continuous breakthrough of deep learning technology has laid a solid foundation for the efficient interpretation of SAR images,and the research of SAR image recognition based on deep learning has entered the fast lane of development.At present,SAR image automatic recognition technology faces the problems of insufficient training samples,complex model structure and weak robustness.Aiming at the difficult problems of the above SAR image recognition tasks,this paper is based on the deep learning framework,combined with transfer learning,self-attention mechanism and knowledge distillation technologies to carry out research.The main research contents related to this article are summarized as follows:(1)Aiming at the problem of severe coherent speckle noise in the background of SAR images and over-fitting of small training sample data,this paper combines the atrous convolution and the Inception module to design transferred TAI-SARNet,which can exponentially amplify the receptive field to help the model extract discriminative feature information.In addition,the network strictly controls the parameter growth,and combines with batch normalization strategy to alleviate the over-fitting phenomenon caused by model parameter redundancy and internal covariate transformation.Finally,in this paper,the related methods of joint transfer learning explore the recognition performance of the transfer of prior knowledge in the optical field,non-optical field,joint optical and non-optical field to the SAR small sample dataset,and obtain good experimental results.(2)Aiming at the problem that it is difficult for deep learning models to effectively extract and use the feature information of small sample data,this paper draws on the ideas of feature reuse and feature fusion,and combines the self-attention mechanism to design a multi-scale feature fusion convolutional neural network(CNN).First,the self-attention ghost module is constructed by combining the self-attention mechanism with the efficient and lightweight ghost module,and the module is used to replace the traditional convolution operation to effectively extract the salient features of the SAR image.Secondly,by constructing an efficient bottleneck unit for network architecture to obtain different levels of feature information of the target image.Finally,the channel shuffling unit and the maximum pooling layer are introduced to construct a multi-scale information branch to promote the full interaction of information.The experimental results show that the multi-scale feature fusion network based on self-attention mechanism has obtained satisfactory recognition results on MSTAR datasets collected under various working conditions,and has also shown good robust performance on self-constructed SAR small sample datasets.(3)Aiming at the related problems of CNN model structure complexity and parameter redundancy,this paper introduces the knowledge distillation technology to redesign the complex model to obtain a simplified simple model,and distill and transfer the rich knowledge obtained through the teacher model.Knowledge distillation breaks through the limitation of the parameter-based transfer learning method on the model structure,and distills the information with rich knowledge by adjusting the temperature parameters,thereby further improving the recognition performance of simple model on related tasks.The relevant results of the experiment show that the simple model combined with knowledge distillation still has the powerful learning ability similar to the complex model,and it also helps to improve its recognition performance on small sample data of SAR.