Ocean Exploration with Artificial Intelligence
Large and diverse data is crucial to train object detection systems properly and achieve satisfactory prediction performance. However, in some areas, such as ma rine science, gathering sufficient data is challenging and sometimes even infeasible. Working with limited data can result in overfitting and poor performance. Further more, underwater images suffer from various problems, like varying quality, which have to be considered. Therefore, alternative means need to be used to increase and enhance the data to facilitate marine scientists’ work. In this thesis, we explore building a more robust system to improve the detec tion accuracy for deepwater corals and analyze underwater movies under different conditions. We experiment with several Generative Adversarial Networks (GANs) to enhance and increase the training data. Our final system comprises two steps: Image Augmentation using StyleGAN2 combined with the augmentation strategy DiffAugment, and Object Detection using YOLOv4. The results indicate that generating realistic synthetic data combined with an ad vanced detector could provide marine scientists with the tool they need to extract species occurrence information from underwater movies. Our proposed system shows increased performance in different domains compared to prior work and the potential to overcome the limited data issue.