Hierarchical image segmentation by multi-dimensional clustering and orientation-adaptive boundary refinement

摘要

In this paper we present a new multi-dimensional segmentation algorithm. We propose an orientation-adaptive boundary estimation process, embedded in a multiresolution pyramidal structure, that allows the use of different clustering procedures without spatial connectivity constraints. The presence of noise in the feature space, mainly produced by modeling errors, causes a class-overlap which can be reduced in a multiresolution pyramid. At the coarsest resolution level, the separation between the different classes is increased and the within-class variance reduced. Thus, at this level, the classes can be obtained with different multi-dimensional clustering algorithms without connectivity constraints. Small and scattered classes as well as isolated class labels are reassigned to their neighborhood by a process which guarantees the spatial connectivity. The resolution is then increased by projecting down the class labels. At each level, the borders are improved by reassigning the boundary pixels to their spatially closest class. However, the class-uncertainty astride the borders has first to be reduced, and we propose to do this by means of orientation-adaptive butterfly-shaped filters. This refinement process further eliminates spatially misclassified pixels produced by the unconstrained clustering. Experimental results show that similarly accurate boundaries are obtained with different clustering algorithms for various test images.