Convolutional neural networks for SAR image segmentation

Abstract

Segmentation of Synthetic Aperture Radar (SAR) images has several uses, but it is a difficult task due to a number of properties related to SAR images. In this article we show how Convolutional Neural Networks (CNNs) can easily be trained for SAR image segmentation with good results. Besides this contribution we also suggest a new way to do pixel wise annotation of SAR images that replaces a human expert manual segmentation process, which is both slow and troublesome. Our method for annotation relies on 3D CAD models of objects and scene, and converts these to labels for all pixels in a SAR image. Our algorithms are evaluated on the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset which was released by the Defence Advanced Research Projects Agency during the 1990s. The method is not restricted to the type of targets imaged in MSTAR but can easily be extended to any SAR data where prior information about scene geometries can be estimated.