Hierarchical Regions for Image Segmentation

Image segmentation is one of the key problems in computer vision. Gibbs Random Fields (GRFs), which produce elegant models, but which have very poor computational speed have been widely applied to image segmentation. In this paper, we propose a hierarchical region-based approach to the GRF. In contrast to block-based hierarchies usually constructed for GRFs, the irregular region-based approach is a far more natural model in segmenting real images. By deliberately oversegmenting at the finer scales, the method proceeds conservatively by avoiding the construction of regions which straddle a region boundary. In addition to the expected benefit of computational speed and preserved modelling elegance, our approach does not require a stopping criterion, common in iterated segmentation methods, since the hierarchy seeks the unique minimum of the original GRF model.