SAR Target Recognition with Auxiliary Data Generated by Infrared-to-SAR Translation Models

Joint automatic target recognition (ATR) in multimodal aerial-view images by exploiting the complementary information collected by heterogeneous sensors plays a key role in supporting military intelligence, surveillance and reconnaissance. Compared to joint ATR in pixel/time-aligned electro-optical (EO) and infrared (IR) images, cross-modality ATR based on synthetic aperture radar (SAR) and EO/IR images is more technically challenging due to the unique data acquirement and imaging mechanism of SAR. In this work, the problem of ATR in unpaired SAR and IR images collected by sensors mounted on drone platforms is considered. To compensate for the time delay between SAR data collection and imaging and tackle potential target movements, an IR-to-SAR (I2S) module is proposed, which is compatible with various generative adversarial networks (GANs). Experiment results show that the synthetic SAR images generated by the I2S module based on the IR images are with high-fidelity and could be used in substitute of real SAR data in network training for SAR ATR.