A minimum spanning tree approach to line image analysis

We propose a graph theoretic approach for extracting the skeleton of a binary line image. Unlike other thinning methods, emphasis is placed on the preservation of the topology of both the foreground and the background. Such conditions guarantee a relevant resulting structure that can be used as input for pattern recognition. Using the underlying graph structure, we can readily formulate this problem as an optimisation problem. Local information such as centrality is given by a distance transform operation. Global information such as location of a branch end is given via a minimum weighted spanning tree which spans all foreground pixels. The resulting structure is then characterised as a union of central paths between end points with their adjacency inter-relationships. Other image characteristics (e.g. width and length of the branches) are also provided. Computational results applied on real images illustrate the noise insensitivity of this method