SAR ATR Using Complex-Valued CNN

Abstract

In the past years, many classification algorithms have proven to be effective and efficient in Synthetic Aperture Radar (SAR) Automatic Target Recognition (ATR). Among them, classification algorithms based on convolutional neural network (CNN) have attracted extensive interest because of high accuracy. However, many CNN-based SAR ATR algorithms just use the amplitude information of SAR data, ignoring the phase information. In this study, a complex-valued CNN is proposed to address this issue. We introduce the complex-valued operations in the calculation of the network, so the complex-valued CNN can extract features in the complex-valued field. It is beneficial to utilize both the amplitude and phase information of the SAR data. And our complex-valued CNN utilizes the Squeeze and Excitation Module (SE Module) to apply weighting factors for the feature maps in the same layer. In ten military targets of MSTAR classification experiment, our model achieves a recognition rate of 98.97%. The accuracy of our complex-valued CNN is superior to many real-valued CNN algorithms in the MSTAR classification experiment.