Combining regular decimation and dual graph contraction for hierarchical image segmentation

The Bounded Irregular Pyramid (BIP) is a hierarchical structure for image representation whose aim is to combine concepts from regular and irregular pyramids. The data structure is a combination of the simplest regular and irregular structures: the 2 × 2/4 regular one and the simple graph representation. However, simple graphs only take into account adjacency relationships, being unable to correctly encode the topology of the image. This paper proposes a new version of the BIP, where the regular decimation process is now merged with a stochastic graph decimation strategy. Experiments demonstrate that this new irregular pyramid is able to provide qualitative good segmentation results and to preserve the topology of the input image at higher levels of its hierarchy