PUB - Publikationen an der Universität Bielefeld

This thesis explores methods in machine learning and computer vision to extract meaningful semantic information from images, with a primary focus on marine images, while addressing challenges associated with limited prior knowledge and annotation constraints. Traditional deep learning approaches in computer vision—such as image classification, object detection, and semantic segmentation—have achieved impressive results but depend heavily on large datasets of labeled images and predefined class information. However, marine imaging often lacks these resources, as species in remote marine environments are not always known in advance, and annotated images are scarce due to limited access to trained experts. This work develops and evaluates novel techniques to address these gaps, aiming to reduce the need for domain specific labeled training data and the effort required to obtain it.