ESFFM: An Electromagnetic Scattering Feature Fusion Model for SAR Automatic Target Recognition

In recent years, deep learning has been widely used in automatic target recognition (ATR) task of synthetic aperture radar (SAR). However, the visual feature and the electromagnetic scattering feature of the targets are not effectively combined, which limits the recognition accuracy and stability in SAR ATR tasks. Aiming at this problem, an electromagnetic scattering feature fusion model (ESFFM) is proposed for SAR ATR in this paper, which incorporates the electromagnetic scattering feature with the visual feature into an end-to-end deep learning framework. First, based on the idea of knowledge distillation, an attribute scattering center (ASC) reconstruction branch is designed, which can extract the electromagnetic scattering feature through a convolution network. Subsequently, a feature-fusion module is designed, which assigns different weights for the ASC maps in fusing the electromagnetic scattering feature with the visual feature. Finally, the whole model is jointly trained with the ASC reconstruction loss and the object recognition loss, which shares the idea of multi-task learning. Experiments on the moving and stationary target acquisition and recognition (MSTAR) dataset demonstrate that the proposed model achieves superior recognition performance compared to the state-of-the-art methods under different operational conditions.