Reducing manual annotation load for CNN-based image segmentation
(2024)
Files
PhD_thesis_NanaYang.pdf
Open Access - Adobe PDF
- 2.68 MB
Details
- Authors
- Supervisors
- Christophe , De Vleeschouwer
- Abstract
- This thesis explores methods to achieve semantic segmentation with minimal manual annotation, addressing the challenge of obtaining extensive labeled datasets. Three main approaches are investigated: leveraging single-class samples, using pseudo-labels, and applying graph-cut methods with prior knowledge. First, when single-class samples are available, a multi-class model can be trained using a simple background/foreground segmentation approach, reducing annotation effort. For example, pure pollen images are easier to label than mixed pollen images. A universal segmentation model trained on these pure samples can automate the labeling of larger datasets, improving segmentation of mixed pollen images. Second, pseudo-labels generated from large unlabeled datasets can enhance model performance, even when they contain errors. Combining traditional and deep learning techniques with diverse pseudo-label generation strategies further improves segmentation accuracy. Third, graph-cut methods leverage prior information about object topology to create rough segmentations that are refined by non-experts. In medical imaging, this reduces the need for expert manual annotations. By training a CNN with these refined segmentations, better models can be developed. Overall, these approaches reduce reliance on extensive labeled data and minimize annotation costs, with applications across various fields.
- Affiliations
Citations
Nana, Y. (2024). Reducing manual annotation load for CNN-based image segmentation. https://hdl.handle.net/2078.5/234476