SAR Automatic Target Recognition Based On Deep Learning

Synthetic aperture radar can work in all-time and all-weather,and the acquired SAR images can provide useful information in various fields.SAR automatic target recognition can automatically interpret the useful information from SAR images,which is also a research hot-spot.In recent years,deep learning has been successfully applied in SAR automatic target recognition.SAR automatic target recognition based on Deep-learning can effectively improve the target recognition rate and the application range of SAR.In this paper,we have deeply studied the framework of convolutional neural networks and autoencoders.Based on these,An improved convolutional neural network and autoencoder network are proposed and applied in SAR automatic target recognition,which achieved a high recognition accuracy rate.The contents of this paper are as follow:(1)A SAR-ATR method based on all convolutional neural network is proposed.In this method,lack of SAR image firstly solved by data augmentation,and then a feasibility analysis of that convolution layer replaces the pooling layer and the full connected layer in the traditional convolutional neural network is given.The experiment shows that all convolutional neural network has more advantage than traditional convolutional neural network.(2)A SAR-ATR method based on improved autoencoder network is proposed.The SAR images are insufficient when deep learning applies to SAR-ATR is presented.Data augmentation can solve this problem,but training deep learning with large-scale data is a time-consuming and labor-intensive issue.Combined convolutional neural network with autoencoder,and an improved autoencoder network is proposed.Replacing fully connected layer with convolutional layer in this improved model.Finally,the experiment shows that the improved autoencod network is superior to the traditional encoding network in SAR-ATR whether as a feature extractor or as a pre-training network.