Small sample radar target recognition based on metric learning

Radar maritime target recognition, not affected by weather and illumination, plays a vital role in sea state detection. But the development of radar ship target recognition has been obstructed due to complicated sea conditions and difficult data acquisition. Traditional recognition methods are difficult to extract robust and highly discriminative features. CNNs is widely used in radar target recognition because of its self-learning. But CNNs has low learning efficiency and poor classification performance under small sample conditions. In this paper, prototype-based metric learning method(PML) is proposed. Specifically, we sample two subsets from original training data as a support set and a query set. The mean of a class's support vectors is calculated to get its centroid in the embedding space, which is called prototype. We find the nearest category prototype for embedded query points to make a classification. It is because our model is more convenient for extracting highly discriminative features and easy to train that it has higher learning efficiency. The experiments is based on Open-SARShip classification dataset in TOPSAR data of the Sentinel-1 satellites for algorithm verification. Experimental results show that recognition accuracy of our model is significantly higher than those achieved by CNNs and traditional radar target recognition models, especially in the limited-data regime.