Target Detection And Classification Of High Resolution SAR Images With Multi-scale Deep Network And Visual Attention Mechanism

With the development of SAR imaging technology,the resolution of SAR images is getting higher and higher,the size of the targets is getting larger and larger,and the shape,texture,and spatial information contained in the targets is becoming more and more abundant.The traditional detection and classification methods of low resolution SAR images are challenged by speed and accuracy.In recent years,deep neural networks have received extensive attention because of far more excellent performance than traditional methods.Designing deep neural networks which are adapted to SAR image processing to improve the performance of SAR image target detection and classification is a very significant research content.Based on traditional SAR image target detection and classification methods and the current rapidly developing deep learning,this paper proposes a high-resolution SAR image target detection and classification method with multi-scale deep network and visual attention mechanism.The main work is as follows:The first is the SAR image target detection method based on the visual attention mechanism.This method constructs a fully convolutional neural network to detect target saliency,and obtains the final target detection result by performing morphological processing on the saliency map.Compared with the traditional target detection methods,this method uses the prior information in the training data to improve the detection accuracy,and improves the computational efficiency through a fully convolutional network architecture.Experiments on the MSTAR dataset show that compared with OS-CFAR,spectral residual and other traditional methods as well as deep learning methods such as CNN and RPN,this method has a fast detection speed and a higher detection accuracy and recall rate.Then a SAR target classification method with DC-Res Net is proposed.This method improves the deep residual network and constructs a deep convolutional neural network model DC-Res Net based on the deformable convolution residual module by using deformable convolution kernels.Compared with the traditional deep residual network,this model extracts more abundant and flexible target features and has better generalization performance.Experiments on the MSTAR datasets under standard operating conditions show that the DC-Res Net model has higher test accuracy than deep neural network models such as CNN and residual network.Finally,a SAR image target classification method with multi-scale deep network is proposed.Aiming at the problem of poor generalization performance of DC-Res Net on MSTAR datasets under extended operating conditions,two kinds of SAR image target classification models based on multi-scale deep network are proposed.The first is MCK-CNN realized by deep fusion of convolution kernels of different scales,and the second is MGA-CNN realized by replacing certain convolution filters of CNN with Curvelet filters.Compared with the traditional deep neural networks,the multi-scale deep network combines features on different scales and improves the representation ability of the model.Experiments on the MSTAR datasets under extended operating conditions show that compared with CNN,residual network,DC-Res Net and other deep neural networks,the multi-scale deep network has better generalization performance.