Deep Cnn For Extraction Of Sidelobes From Sar Imagery In Spectrally Restricted Environment

This paper explores the critical problem of spectral estimation and recovery via deep convolutional neural networks (CNNs). This problem often arises in ultra-wideband (UWB) synthetic aperture radar (SAR) applications where the wide available bandwidth dictates the final SAR image resolution and quality. Our simple robust solution here also complements well the recently developed stepped-frequency radars built by the U.S. Army Research Laboratory, allowing the system to bypass prohibited frequency bands specified by local frequency management agencies and congested frequency bands that are deemed to contain too much interference. Extensive experimental results confirm the validity as well as the promising prospect of the proposed approach - our CNN-based generative network can still robustly recover SAR images when up to 90% of the operating spectrum is missing.