A Comprehensive Survey on SAR ATR in Deep-Learning Era

Due to the advantages of Synthetic Aperture Radar (SAR), study Automatic Target Recognition (ATR) has become a hot topic. Deep learning, especially in case Convolutional Neural Network (CNN), works an end-to-end way and powerful feature-extracting abilities. Thus, researchers SAR ATR also seek solutions from deep learning. We review related algorithms with regard this paper. firstly introduce commonly used datasets evaluation metrics. Then, we before They are template-matching-, machine-learning- model-based methods. After that, mainly methods deep-learning era (after 2017); those core The non-CNNs CNNs, that is, ATR, summarized at beginning. found tend design specialized CNN for ATR. solve problem raised by limited samples reviewed. data augmentation, Generative Adversarial Networks (GAN), electromagnetic simulation, transfer few-shot semi-supervised metric leaning domain knowledge. imbalance problem, real-time recognition, polarimetric SAR, complex adversarial attack principles problems them introduced. Finally, future directions conducted. In part, point out dataset, architecture designing, knowledge-driven, explainable should be considered future. This paper gives readers quick overview current state field.