Image segmentation is a fundamental task in computer vision, with applications ranging from medical diagnostics to archaeological research. Although deep learning methods have achieved impressive results in recent years, researchers in specialised scientific domains often face barriers due to the high cost and effort of generating reliable annotations, as well as the opacity of the models themselves.
In this thesis, we address these barriers in two areas where annotation is particularly demanding: cell microscopy and archaeological remote sensing. We first analyse how different annotation errors affect cell segmentation performance. Building on these insights, we introduce a method that upgrades low-quality annotations, allowing researchers to train accurate models while reducing expert workload. We then develop a few-shot learning approach for cell segmentation, showing that competitive results emerge with as few as five annotated examples. Finally, we apply...
Show moreImage segmentation is a fundamental task in computer vision, with applications ranging from medical diagnostics to archaeological research. Although deep learning methods have achieved impressive results in recent years, researchers in specialised scientific domains often face barriers due to the high cost and effort of generating reliable annotations, as well as the opacity of the models themselves.
In this thesis, we address these barriers in two areas where annotation is particularly demanding: cell microscopy and archaeological remote sensing. We first analyse how different annotation errors affect cell segmentation performance. Building on these insights, we introduce a method that upgrades low-quality annotations, allowing researchers to train accurate models while reducing expert workload. We then develop a few-shot learning approach for cell segmentation, showing that competitive results emerge with as few as five annotated examples. Finally, we apply explainability techniques in the archaeological case study, using activation maps both to improve trust in model predictions and to generate semi-automated annotations.
Together, these contributions demonstrate how annotation-efficient and more interpretable deep learning methods can support the adoption of computer vision techniques in domains where data is limited and expert time is scarce.
Show less
