Weakly-supervised Semantic Segmentation in the Multi-class Setting across Different Image Domains

Weakly-supervised semantic segmentation from image-level annotations has been proposed for segmenting natural scene images and this performs well compared to full supervision with pixel-level annotations, which are harder and slower to obtain. However, this is a challenging problem because image labels lack location-specific information to learn object extents. Weakly-supervised semantic segmentation is unexplored in other image domains, such as histopathology and satellite imagery and it is unknown whether different methods are needed for these images. This thesis proposes a novel weakly-supervised semantic segmentation algorithm for histopathology images and explores its performance compared to state-of-the-art methods developed for natural scene images. The compatibility of each method for different image domains is analyzed and this thesis analyzes the effect of classification network choice, the self-supervised learning approach, and class co-occurrence on segmentation performance.M.A.S