Data Augmentation and Transfer Learning for Limited Dataset Ship Classification

Abstract

Fine-grained classification consists of learning and understanding the subtle details between visually similar classes, which is a difficult task even for a human expert trained in a corresponding scientific field. Similar performances can be achieved by deep learning algorithms, but this requires a great amount of data in the learning phase. Obtaining data samples and manual data labeling can be time-consuming and expensive. This is why it can be difficult to acquire the required amount of data in real conditions in many areas of application, so in the context of a limited dataset it is necessary to use other techniques, such as data augmentation and transfer learning. In this we paper we study the problem of fine-grained ship type classification with a dataset size which does not allow learning network from scratch. We will show that good classification accuracy can be achieved by artificially creating additional learning examples and by using pretrained models which allow a transfer of knowledge between related source and target domains. In this, the source and target domain can differ in their entirety.