Classification of ISAR Images Generated from Model-based Approaches Using Deep Learning

The classification of small, low-observable airborne targets, such as drones and birds, poses significant challenges due to their low detection rates. Conventional vision sensorbased approaches often suffer from reduced performance in low-visibility environments or adverse weather conditions. Additionally, the integration of infrared sensors alongside camera sensors increases hardware complexity and cost, rendering such solutions inefficient. To address these limitations, we propose a method that leverages deep learning and inverse synthetic aperture radar (ISAR) imaging for accurate target classification, using only radar sensors. Our proposed deep learning-based ISAR image classifier comprises two key components: simulated ISAR image generation and deep learning-based classification. We construct simulated ISAR datasets using point scatter (PS) modeling for quadcopter drones, hexacopter drones, and aircraft, and three-dimension (3D) mesh modeling for birds, unmanned aerial vehicles, and quadcopter drones. The two datasets based on PS and 3D mesh modeling are used to train a proposed deep learning classifier. The proposed classifier can achieve a classification accuracy of $\mathbf{9 8 \%}$ on the PS-based dataset and $\mathbf{9 6 \%}$ on the 3D mesh-based dataset, where scattering was calculated using the physical optics method.