An energy-efficient embedded implementation for target recognition in SAR imageries

Abstract

This paper presents an energy-efficient deep learning model design, training and implementation method for the synthetic aperture radar (SAR) image classification application on a neuromorphic processor. The proposed approach adopts emerging neuromorphic computing models and hardware to achieve significant improvement in computational energy efficiency over deep learning algorithms on conventional embedded processors. A deep convolutional neural network (DCNN) is designed specifically for implementing image classification on the TrueNorth neurosynaptic processor. We have explored the DCNN model design parameters to obtain a comprehensive solution set in the energy-performance trade-off space. Using a SAR image classification dataset, evaluation results show that the proposed design and implementation approach achieves at least 20X reduction in energy-per-image-classification over one of today's most energy-efficient conventional embedded processors. while achieving a classification accuracy of 95% and a processing throughput of 1,000 images per second.