A Multi-View CNN for SAR ATR

This paper describes an approach for exploiting 3D information to improve synthetic aperture radar (SAR) automatic target recognition (ATR) performance. Historically, SAR ATR is performed using a single look, where an individual SAR image formed over a particular azimuth dwell is used as input to a classification algorithm. Recent data collection and processing developments enable techniques which exploit information from multiple looks simultaneously. Using more information from distinct geometries promises to improve both algorithm robustness and overall classification performance. Our approach to using this additional information is to adopt techniques for 3D image recognition from the natural image literature and adapt them to the SAR modality. We find SAR ATR performance can be improved substantially using a multi-view network employing a backbone architecture known to perform well on SAR ATR and enhanced with SAR-specific preprocessing and augmentation.