Improving Target Pose Estimation from Synthetic Aperture Radar Imagery

Abstract

Automatic target recognition (ATR) from synthetic aperture radar (SAR) image has been studied for many years. In recent years, many convolution neural networks (CNNs) for SAR ATR based on deep learning have been proposed, but most of them classify target classes from fixed size target chip extracted from SAR image. On the other hand, we proposed CNN for end-to-end SAR ATR that input arbitrary size SAR image of multiple targets and outputs the position, class, and pose of each detected target as SAR ATR image. Among them, the evaluation metrics of the pose class of the CNN is lower than the evaluation metrics of the target classes. In this paper, we propose a method to improve posture estimation performance of end-to-end SAR ATR.