An Evaluation of Object Detection Algorithms Applied to Oceanographic Midwater Imagery

Charlotte A. DeBossu, University of Rhode Island


Object detection algorithms have emerged as powerful tools in various domains, yet their application to oceanographic images remains in its early stages. Particularly, the use of stereo camera systems for capturing images of macroscopic marine organisms is still considered novel imagery. Therefore, it is critical to understand how well deep learning algorithms perform in classifying marine organisms in these types of images. This thesis aims to contribute to the nascent literature on this task in the context of stereo camera imagery. The study evaluates the classification performance of two prominent deep learning models, YOLOv5 and ResNet50, when applied to images of macroscopic marine organisms captured in a low-light environment. The findings in this thesis aim to provide insights into the effectiveness of deep learning algorithms when applied to images of this nature, and might serve as a resource for researchers seeking to implement a deep learning model for their specific use case.