Physical-aware Radar Image Synthesis with Projective Network

This paper proposed a new network module named as projection network, which explicitly combined radar's projection process with trainable network. It assumes that each 2D radar cross section (RCS) map is a projection of a 3D RCS map. And it models the projection mechanism as a differentiable layer, so that it can be integrated with other neural network layers, such as convolutional and pooling layers. The proposed model is consistent with radar projection process, hence effects such as layover is considered. It is designed and used specifically for radar applications. This paper applied the proposed network on radar image synthesis, and the simulation results showed great potential of projective network.