(A) region relaxation for range image segmentation and hexagonal edge relaxation

Abstract

In this thesis two algorithm are proposed. One is a reliable range image segmentation algorithm and the other is a hexagonal edge relaxation. The range image segmetation algorithm cosists of a region relaxation stage and a stepwise optimal region merging stage. The region relaxation is based on the iterative procedures of labeling, connected component analysis and function approximation. The relaxed image is represented by the region adjacency graph(RAG) and merged by means of stepwise optimal approach[LIM 88]. These stages are all based on the polynormial function approximation. The proposed range image segmentation algorithm is applied to various range imagessuch as ERIM, synthetic, or structured light range finder, because of the general smoothness assumption that does not depend on thchniques for range finding. To acquire more accurate boundary information, we devise hexagonal edge relaxation. It has been found that hexagonal spatial sampling yields smaller quantization errors than square sampling. The hexagonal grids remarkably improve the understanding of connectivity. We exploited a new hexagonal edge relaxation algorithm[CHO 92]. In this method, the vertex types of an edge are simple and edge classifications according to the paring of vertex types are reasonable, therefore the overall enhancement of edges is more reliable than Prager``s square edge relaxation method[PRAG 80]. This hexagonal edge relaxtion method is extended to find converging one. The vertex types are classified by extended neighbors than hexagonal edge relaxation. For each edge, vertex types are defined by the dictionary according to the edge presence in the neighbors and this relaxation algorithm is experimentally shown to converge.