A Study on the Improvement of Fine-grained Ship Classification through Data Augmentation Using Generative Adversarial Networks

Identification of the type of ship is an important issue in maritime surveillance. However, unlike the land environment where data sets are easy to build, it is difficult to build large amounts of marine environment data where it is difficult to collect images. In this situation, the use of large-scale data obtained in the surveillance and defense fields is essential for research, but the use of data for private research is impossible due to security issues. In this paper, a ship dataset free from security problems was constructed through data augmentation using GANs. We conduct an experiment on improving fine-grained ship classification performance through the use of a small amount of real ship images and augmented data, and try to show that the augmented data is useful for ship classification network training.