Application of Transfer Learning for Fine-Grained Vessel Classification Using a Limited Dataset

The automatic classification of maritime vessel type from low resolution images is a significant challenge and continues to attract increasing interest because of its importance to maritime surveillance. Convolutional neural networks are the method of choice for supervised image classification, but they require a large number of annotated samples, which prevents many superior models being applied to problems with a limited number of training samples. One possible solution is transfer learning where pre-trained models are used on entirely new predictive modeling, transferring knowledge between related source and target domains. Our experimental results demonstrate that a combination of data augmentation and transfer learning leads to a better performance in the presence of small training dataset, even in the a fine-grained classification context.