Deep Learning @ UGA

Semantic Deep Learning of Underwater Ecology

A fundamental issue limiting ecological studies in marine environments, such as coral reefs, is the difficulty of generating accurate and repeatable maps of the underlying ecosystems. We examine two major deep learning methods to assist with this task, patch-based convolutional neural network approaches and fully convolutional network models. We explore 2D and 3D semantic map creation using these methods on image data extracted from underwater video. For our patch-based CNN approaches we use individual point-wise ground truth annotations. For our fully convolutional networks we develop a tool for the fast creation of ground truth image segmentations.

Andrew King is a machine learning scientist currently completing a master’s degree in Artificial Intelligence at the University of Georgia. His research as a master’s student has focused on deep learning models for semantic segmentation. He is the author of Deep Segments, Scopi, and a number of other machine learning and computer vision based applications.