Radar Sparse Imaging Using Deep Learning

With the rapid development of radar technology and the changing needs in the military and civilian fields,a variety of performance of the radar have been put forward higher requirements.In order to provide higher resolution images,imaging radars use synthetic aperture technology to improve the azimuth resolution and improve the distance resolution by emitting signal with large bandwidth.However,with the restriction of the Nyquist's theorem,there will be a high data sampling rate.And the transmission,storage,and real-time processing of massive data bring great challenges to hardware implementation and algorithm complexity.The electromagnetic scattering characteristic of targets makes it possible to apply compressed sensing theory to SAR imaging.Compressive sensing aims at sparse sub-Nyquist sampling of scene data,studies sparse representation of data,sparse observations,and reconstruction of sub-Nyquist sampling data.We always expect to achieve high reconstruction accuracy from lower complexity.However,in general,the construction and solution of sparse representation dictionaries in complex scenes is difficult,and it is also hard to fully exploit the correlation and sparseness of data.In addition,the design of observation matrix and reconstruction algorithm are another two difficulties.During the past few years,the machine learning based on deep learning has achieved rapid advancements,and has make significant breakthroughs in the fields of speech recognition,text translation,image processing,and automatic driving technology.The deep structure can automatically extract more abstract and sparse features of the data,thus provide the possibility for achieving sparse representation of complex scenes.Based on the above issues,we combine deep learning with radar signal processing techniques to explore the feasibility of deep learning in radar sparse imaging applications in three ways including sparse representations,sparse observations,and sparse reconstructions of data.The proposed methods have a certainly theoretical and practical signification.The main work in this paper is as follows:Firstly,we designed a SAR imaging method based on one-dimensional deep autoencoder network.The deep autoencoder network has excellent ability in data sparse representation and reconstruction,therefore,it can be used for sparse representation,sparse observation and reconstruction of data.The front end of the network is the layer of sparse representation and observation,and the back end of the network is the layer of reconstruction.Then the network is optimized by SGD and other optimization algorithms.After training,we could obtain a one-dimensional deep autoencoder network for signal reconstruction.Then the reconstructed signals are feed into traditional imaging models to obtain the final reconstructed image.The effectiveness of the method is verified by experimental simulation.Secondly,we proposed a SAR imaging method based on convolution neural network.Aimed at solving the problem of two-dimensional radar imaging and fully excavating the correlation and sparsity of two-dimensional echo data,we designed a two-dimensional deep autoencoder network based on convolutional neural network.Considering that the echo signals received by SAR are plurality,the real part and imaginary part are treated as two different channels in the net.Then input data are encoded and decoded by layer-wised convolution,pooling and nonlinear feature mapping operation.The Adam optimization method is utilized to minimize the error between the reconstructed data and the original data.Adam optimization method to optimize the error function of the reconstructed data and the original data.Once the network is optimized,we can obtain a stable and robust two-dimensional signal reconstruction network.At last we can achieve imaging by using traditional imaging method to reconstructed signal.Finally,we studied a SAR imaging method based on Generative Adversarial Networks.Different from the idea of sparse signal reconstruction,pointing at the sparse observation and reconstruction of radar echo signal,we analyzed the feasibility of the application of generative adversarial networks on SAR sparse imaging and constructed an end to end SAR sparse imaging network.The output of the network is a reconstructed SAR image,which avoids the design of measurement matrix and the iterative operation of CS reconstruction algorithm in the traditional CS imaging.In addition,this method greatly reduces the complexity of the reconstruction algorithm and improves the efficiency of real-time imaging,and we verified it by simulation experiments.