SAR Target Classification Based On Complex Full Convolutional Neural Network And Convolutional Auto Encoder

Synthetic aperture radar(SAR)has long-range,all-day and all-weather working capabilities,and plays a very important role in the military and civilian fields.SAR Automatic Target Recognition(SAR-ATR)has always been one of the research hotspots in the field of SAR.It can use a computer to automatically recognize a target and determine its category.In recent years,Convolution Neural Network(CNN),as a typical deep learning technology,has been widely used in SAR-ATR,and has achieved remarkable results.At present,most CNN-based SAR-ATR methods only use the SAR image amplitude data and lose the phase data.In fact,SAR phase data also contains rich target information.Therefore,it is extremely urgent to study the SAR-ATR method based on the complex-valued convolutional neural network(CCNN)network.This paper mainly proposes three improved complex convolutional neural networks,and on this basis,introduces an improved complex convolutional auto encoder(ICCAE)to carry out unsupervised pre-training on one of them,the specific content as follows:(1)This paper proposes three improved complex convolutional neural network structures,namely Complex-Valued Full Convolution Neural Network(CFCNN),Real and Imagery Two-Channel Complex Full Convolutional Neural Network(RIDC-FCNN),Amplitude and Phase Dual-Channel Full Convolutional Neural Network(APDC-FCNN).Then,the back propagation formulas of the three networks are derived,and the computational complexity of CFCNN is analyzed.The experimental results of the MSTAR data set show that CFCNN has a higher recognition rate than CNN and CCNN.In addition,RIDC-FCNN has a higher recognition rate than CFCNN and APDC-FCNN.(2)This paper proposes a SAR target classification method combining RIDC-FCNN and ICCAE.This method first performs unsupervised pre-training of ICCAE,then initializes some parameters of ICCAE to RIDC-FCNN,and finally fine-tunes RIDC-FCNN.In addition,the back propagation formula based on RIDC-FCNN and ICCAE is derived.The experimental results of the MSTAR data set and the fully polarized Flevoland data set show that the method based on RIDC-FCNN and ICCAE can not only obtain a higher recognition rate with a small number of data sets,but also have a stronger Anti-noise capability than the method based directly on RIDC-FCNN.