Few-shot ship classification based on metric learning

In recent years, deep learning has become very popular and its application fields have been increasing, but it relies heavily on large number of labeled data. Therefore, it is necessary to find a few-shot learning method which can obtain a good training model using few samples. In this paper, a few-shot classification method based on MSFR is introduced for the ships classification task for the first time. In addition, we made a dataset of ships for the few-shot classification task, which we called FSCD. FSCD contains nine categories and 1500 samples. We used two methods of measuring learning called ProtoNet and MSFR, and a non-measuring method MAML for comparison. A large number of experiments have been implemented to prove that the performance of our proposed MSFR method on the ship dataset can reach 61% in 1-shot and 77.5% in 5-shot, which is better than the MAML and ProtoNet. In addition, we explore the effects of different network depths and different epochs on network performance in the ship dataset. As a few-shot ship classification study, this work opens up a new way of thinking and lays the foundation for further research.