Multi-Target ISAR Image Prediction and Background Decoupling Using Deep Learning

This paper proposes a dual-channel ISAR prediction network based on U-Net for image-to-image tasks. This network takes the top-view optical images as inputs to predict ISAR images of multiple targets in complex backgrounds, and can also mitigate the electromagnetic coupling effects caused by complex backgrounds or other interfering targets. Our method focuses on decoupling the strong interference between complex backgrounds and multiple targets using only ISAR images or optical images. The proposed end-to-end prediction method bypasses meshing, matrix-solving and post-processing procedures of conventional simulation approaches, significantly reducing computational resource requirements. Two numerical experiments demonstrated the proposed method can predict and decouple ISAR of multiple targets, while maintaining acceptable accuracy. We expect this approach could provide a foundation for solving practical non-cooperative target electromagnetic imaging in the future.