Relaxation labeling algorithm for information integration and its convergence

摘要

Relaxation labeling is an iterative and parallel processing scheme, which has been used to implement information integration. In the framework of relaxation labeling, two sources of given information is encoded in a form of constraints and that of certainty measures of the labels, respectively. With this formulation, the information integration is implemented by the relaxation operator by forcing neighborhood constraint satisfaction. Various versions of relaxation labeling algorithm have been proposed by other researchers and their applications to image processing, scene analysis, object identification have produced satisfactory and partially satisfactory results in the past. As an iterative algorithm, the convergence properties of relaxation labeling have been investigated by some researchers. A new relaxation labeling algorithm is developed and presented in this paper. The motivation as well as the performance of the new algorithm is discussed. Its local and global convergence properties are established analytically and numerically. It is also demonstrated that the convergence analysis of the new algorithm can be easily generalized.