dc.description.abstract | Deep learning and computer vision are two thriving research areas within machine learning.
In recent years, as the available computing power has grown, it has led to the possibility
of combining the approaches, achieving state-of-the-art results. An area of research that
has greatly benefited from this development is building detection. Although the algorithms
produce satisfactory results, there are still many limitations. One significant problem is the
quality and edge sharpness of the segmentation masks, which are not up to the standard
required by the mapping industry. The predicted mask boundaries need to be sharper and
more precise to have practical use in map production.
This thesis introduces a novel Data Enhancement Technique (DET) to improve the boundary
quality of segmentation masks. DET has two approaches, Seg-DET, which uses a segmentation
network, and Edge-DET, which uses an edge-detection network. Both techniques
highlight buildings, creating a better input foundation for a secondary segmentation model.
Additionally, we introduce ABL(RMI), a new compounding loss consisting of Region Mutual
Information Loss (RMI), Lovasz-Softmax Loss (Lovasz), and Active Boundary Loss (ABL).
The combination of loss functions in ABL(RMI) is optimized to enhance and improve mask
boundaries.
This thesis empirically shows that DET can successfully improve segmentation boundaries,
but the practical results suggest that further refinement is needed. Additionally, the results
show improvements when using the new compounding loss ABL(RMI) compared to its
predecessor, ABL(CE) which substitutes RMI with Cross-Entropy loss(CE). | |