A class hierarchy approach for the segmentation of natural scenes

Scene segmentation techniques normally model scene content in terms of a set of mutually-exclusive classes. This model is used both by local feature classifiers and by later stages of processing such as relaxation labelling. This paper generalises these schemes to a hierarchical class model where the labelling of each pixel is the result of a path down the class hierarchy. The benefits of the hierarchical scheme are discussed, and an efficient relaxation labelling algorithm for the class hierarchy is given. The technique is demonstrated on synthetic textured images with added noise and a comparison in terms of performance and complexity is made with the traditional model