Exploring Hierarchy Simplification for Non-Significant Region Removal

Abstract

Image segmentation is a classic subject in the field of digital image processing, and it can be used to solve a large variety of problems or serve as preprocessing for other methods of image analysis. Hierarchical image segmentation methods provide a multiscale representation, therefore they produce a nested set of image segmentations in which a result at a given level can be produced by merging regions of the segmentation at its previous level. However, a hierarchical representation may produce small components at its coarser levels, leading to oversegmentations on such scales. To solve this problem, we explore strategies to simplify hierarchies in order to remove nonsignificant regions, in terms of area, while trying to preserve the hierarchical structure. We evaluate the proposed simplification strategies with different hierarchical segmentation methods on the Pascal Context dataset by using precision-recall measures and fragmentation curves, along with a qualitative assessment showing that the simplification of hierarchies can lead to visually better image segmentations.

Higra: Hierarchical Graph Analysis

“…• manipulate and explore hierarchies: simplification [26,27], accumulators, cluster extraction, various attributes [28] (size, volume, dynamics, perimeter, compactness, moments, etc. ), horizontal and non-horizontal cuts, alignment of hierarchies [29];…”