SAR Target Recognition based on Lightweight Parallel Multi-channel CNN

Compared with the field of computer vision, deep learning (DL)-based SAR automatic target recognition (ATR) faces more technical challenges, which include the scarcity of annotated SAR imagery that could be used for network training, the dramatically-varying SAR image features of the same targets obtained from different angles, and the far-from-ideal imaging quality due to artifacts such as speckle noise, image defocusing, and multipath effects. As a result, deep convolutional neural networks (CNN) designed for SAR ATR are more prone to overfitting problems and often features poor cross-dataset generalizability. To encounter these problems, we leverage the proprietary NUAA MiniSAR ground military vehicle dataset and propose a prior-knowledge aided (KA) lightweight multichannel CNN architecture, the IncepResNet, which combines the essential design principles of multiple classic neural network models. Experiment results show that the proposed IncepResNet exhibits a target classification performance much better than the existing CNN models at a strikingly low computational cost by jointly leveraging novel SAR image preprocessing techniques and the KA multiview deep feature fusion method.