Two Stage Fine-Tuning Prototypical Network for Few Shot SAR Aircraft Recognition

Over past few years, the development of deep learning has greatly facilitated the SAR target recognition tasks. However, most deep learning-based SAR aircraft recognition methods suffer from a drawback of heavy reliance on large-scale labeled training data, making it difficult to apply them to actual remote sensing tasks. In term of these issues, we propose an improved prototypical network based on two stage fine-tuning for few shot SAR aircraft recognition. In this method, firstly, the deep learning image encoder is trained on large scale open source datasets supervised. This gives the encoder the ability of recognizing basic geometry. Secondly, based on the contrastive learning, the encoder is trained on SAR aircraft data in a self-supervised manner. This helps the encoder learn the ability to encode the semantics of scatterers in SAR image. Finally, the encoder is further fine-tuned on the specific SAR aircraft few shot recognition dataset in the way of prototype network. The comparative experiments shows that, the proposed method outperforms other state-of-the-art methods on public SAR aircraft dataset. Especially in the 1-shot case, the proposed method achieves a recognition accuracy of 81.34%. Simulation results show that the proposed method has the potential to be applied to real-world scenarios.