Generative adversarial networks for recovering missing spectral information

• Main Authors: Tran, Dung N., Tran, Trac D., Nguyen, Lam
• Format: Conference Proceeding
• Language: English
• Subjects: Frequency-domain analysis; Gallium nitride; generative adversarial network; Generators; missing frequency bands; Radar imaging; Spectral recovery; Synthetic aperture radar; Ultra wideband radar; ultra-wideband
• ISSN: 2375-5318
• DOI: 10.1109/RADAR.2018.8378737

Ultra-wideband (UWB) radar systems nowadays typically operate in the low-frequency spectrum to achieve pene­tration capability. However, this spectrum is also shared by many other communication systems, leading to tremendous interference in a large number of frequency bands. Although avoiding these frequency bands is a natural solution, the resulting frequency gaps not only lower the signal-to-noise ratio in the received radar signals but the fragmented radar spectrum also creates severe sidelobes, which can destroy the original features of the targets of interest. In this paper, we propose to recover this missing spectral information via a generative adversarial network, called SARGAN, which learns the relationship between original and missing-frequency-band signals by observing numerous possible training pairs in a clever way. Initial results shows that this approach is promising in tackling this challenging problem: our recovered signals achieve on average more than 18 dB gain in the signal-to-noise ratio when up to 90% of the operating spectrum is missing. Moreover, the proposed SARGAN accomplishes this performance level without any prior knowledge of the locations of the missing frequency bands.