A Viewpoint Adaptation Ensemble Contrastive Learning framework for vessel type recognition with limited data

Abstract

Unmanned Aerial Vehicle (UAV)-based systems are gaining increasing attention in the maritime industry, but one of their major challenges is accurately identifying vessel types from bird-view images captured by UAV. The use of computer vision and deep learning technologies in image recognition requires large amounts of annotated images for model training, but collecting and manually annotating these images is a costly and time-consuming task. To overcome these challenges, we propose a novel Viewpoint Adaptation Ensemble Contrastive Learning (VAECL) framework. With the VAECL framework, first, an improved deep generative model (DGM) is constructed to learn the distribution of the limited vessel image data and to generate more images for data augmentation. Second, transfer learning using a pre-trained Inception V3 network is then presented for vessel viewpoint transfer and adaptation learning. Third, a contrastive learning paradigm is adopted by formulating a new loss function to obtain more contrastive feature representation via pre-training. Finally, an ensemble learning algorithm is highlighted to further improve the performance of vessel type recognition. Extensive experiments on a newly established dataset reveal the following encouraging findings: (1) VAECL can achieve a high accuracy of 92.96% with proper parameters; (2) DGM-based image augmentation improves the accuracy by at least 14.68%, and it performs better than the traditional data augmentation techniques; (3) viewpoint transfer learning surges the accuracy by as high as 18.50%; (4) contrastive learning increases performance by at least 2.79% in terms of accuracy; (5) ensemble learning enhances the accuracy by as high as 1.03%.