On the determination of inconsistent edges in graph-based segmentation algorithms

In this paper we introduce a new method for decomposing a minimum spanning tree (MST) into N connected components (representing N textures) for graph-based texture segmentation. Currently, ad hoc methods are used to perform this decomposition. Under our approach, the local maxima of the histogram of MST edge weights are used to define the centers of textures. A collection of thresholds between these maxima is then selected to optimize a square error criterion. These thresholds directly define the decomposition of the MST. We demonstrate the performance of our approach on synthetic images and high-resolution infrared imagery.