Abstract:

Sufficient synthetic aperture radar (SAR) template data is the key to achieve excellent recognition performance of target recognition algorithms (especially intelligence target recognition algorithms based on deep learning). It is unrealistic to obtain sufficient SAR data from the actual measurements, so SAR simulation based on electromagnetic scattering modeling has become an effective way to obtain sufficient samples. Simulated SAR image and measured SAR image are non-homologous data. Due to the fact that the target geometric model of SAR simulation is not inevitable consistent with the real object, the SAR sensor model in SAR simulation may be different from the actual sensor performance, the background environment of the object is also inevitably different from that of SAR simulation, and the error of electromagnetic modeling method itself, etc., difference is a inevitable existing between the simulated and measured SAR images, which will affect the recognition performance. To address this problem, this paper first adopts a SAR simulation method based on high frequency asymptotic technique and discrete ray tracing technique to obtain SAR simulation images of ground vehicle targets, and then uses the convolutional neural network (CNN) method and the linear/nonlinear feature transformation method to realize the comparative analysis of non-homologous SAR target identification performance of MSTAR measured data. The experimental results show that the direct use of SAR simulation data cannot achieve the ideal recognition performance of the measured SAR data, while the feature transformation based on linear/nonlinear can improve the identification performance of non-homologous SAR target recognition, and to some extent, alleviate the poor recognition performance caused by the difference between SAR simulation data and measured data.

Key words: synthetic aperture radar (SAR) simulation, feature transformation, non-homologous SAR target recognition