From Primitive Regions to Shape Partition

Abstract

The work presented in this paper has been developed within the framework of a novel hierarchical method to decompose a figure into a connected set of possibly overlapping components. The method is based on the exploitation of the shape information carried by the skeleton to initially decompose the figure into a number of pattern subsets representing the primitive regions composing the figure. Once the initial decomposition has been obtained, it is exploited to obtain the highest level of our decomposition. This paper addresses specifically the problem of merging the primitive pattern subsets found within the figure to get the minimal number of components. It is assumed that the skeletal pixels are associated with a label specifying the local thickness of the figure, and it is shown that the trend of the label along the skeleton branches allows both to identify the primitive regions within the figure and to merge them into the figure components. Experimental results show that the decomposition provided by the algorithm is very natural in both the number and the type of the components