GraphBPT: An efficient hierarchical data structure for image representation and probabilistic inference

Abstract

This paper presents GraphBPT, a tool for hierarchical representation of images based on binary partition trees. It relies on a new BPT construction algorithm that have interesting tuning properties. Besides, access to image pixels from the tree is achieved efficiently with data compression techniques, and a textual representation of BPT is also provided for interoperability. Finally, we illustrate how the proposed tool takes benefit from probabilistic inference techniques by empowering the BPT with its equivalent factor graph. The relevance of GraphBPT is illustrated in the context of image segmentation.