Automatic Target Recognition from SAR Images Using Capsule Networks

Abstract

Synthetic Aperture Radar (SAR) imagery has become popular in the past few decades owing to its operability under difficult weather conditions. In this paper, we introduce a deep learning architecture using a Capsule Network (CapsNet) for automatic target recognition (ATR) of images of targets captured using an X-band SAR sensor. The architecture consists of a single convolutional layer, followed by two Capsule layers, and a decoder network at the end. Since, traditional Convolutional Neural Networks (CNNs) often require a significant number of training images, their performance is limited for small number of training examples. Unlike CNNs, Capsule Networks encapsulate the instantiation parameters of an object within an image, thus, they do not require a large number of training samples. In addition, Capsule Networks are view-point invariant. For the evaluation of the proposed method, we have used the MSTAR database, containing SAR images of 10-classes of military vehicles. We have achieved 98.14% overall classification accuracy on this dataset.