A New Image Simulation Technique for Deep-Learning-Based Radar Target Recognition

Abstract

Radar target recognition via deep learning has been an active research area recently. However, this family of methods depends on the quality of radar images and the number of training samples. Given limited training samples of poor quality, these methods will cause severe overfitting. To solve this problem, this paper proposes a new method combining the physical model and deep learning. A new radar measurement simulation technique is presented. The radar measurements are first modeled by the attributed scattering center model. The target is reconstructed by the estimated model parameters. The corresponding residual is formed simultaneously. The reconstructed target and the residual are then frozen. A mask with random shape is imposed on the frequency data. The masked components are reset accordingly. The resulting data are combined and transformed from the frequency domain into the image domain. The new radar measurement can be then generated via the re-imaging process of target. We aim to simulate the unforeseen disturbances during data collection. The simulated radar measurements are used to improve the learning efficiency of deep models under limited sample environments. Multiple comparative studies are performed to demonstrate the advantages of the proposed method.