Transfer Learning for RCS Reduction of Antenna Arrays Across Arbitrary Frequency Bands

Recent studies reveal that deep learning networks can reduce the radar cross section (RCS) of antenna arrays. However, the existing deep learning networks are all for a fixed frequency band. In this work, a transfer learning feature derivation network (TL-FDN) is proposed to reduce RCS across arbitrary frequency bands. Firstly, a feature derivation network (FDN) is initially trained on a source frequency band with readily available calibrated data. Secondly, the pre-trained FDN is reused for a data-insufficient target frequency band. Numerical simulations demonstrate that TL-FDN can improve accuracy by more than 15.64% when trained with insufficient data. Experimental results show that our TL-FDN provides effective RCS reduction of more than 6.66 dB across new frequency bands.