Target Identification of Low-Resolution SAR Image via Super-Resolution

Abstract

Resolution of synthetic aperture radar (SAR) is limited by hardware specifications of the radar and operational mode of the SAR system. The resolution is an important factor affecting SAR automatic target recognition (ATR) and low-resolution SAR images provide poor ATR performance. To solve this problem, a framework for target identification for low-resolution SAR is proposed in this thesis. We combine a spectral estimation-based super-resolution technique with deep neural network (DNN)-based ATR frameworks. The experimental results under SOCs (standard operation conditions) and EOCs (extended operation conditions) using various DNN backbones show significant improvement in identification performance for low-resolution image by the proposed framework. For the target class with significantly improved accuracy, we confirm that the unresolved scattering point in the low-resolution is successfully resolved through the method, resulting in a difference between the two classes.