Semantic Segmentation Tree for image content representation

Abstract

In this paper, we present a novel algorithm for representing image content by constructing a hierarchy of semantic image regions, called a Semantic Segmentation Tree (SSeg-tree). First, the hill-manipulation algorithm divides an image into several visually coherent segments (small regions), which form the leaves of the SSeg-tree. Then, the method groups these segments based on well-defined spatio-visual grouping rules to produce bigger and more semantic regions, which form the intermediate nodes of the SSeg-tree. The SSeg-tree is a region-based description of the image semantic content that could be useful in many applications such as CBIR and filtering unwanted objects.