Learning the Graph Edit Distance Parameters for Point-Set Image Registration

Abstract

Alignment of point sets is frequently used in pattern recognition when objects are represented by sets of coordinate points. The idea behind this problem is to be able to compare two objects regardless of the effect of a given transformation on their coordinate data. This paper presents a method to align point sets based on the graph edit distance. The main idea is to learn the edit costs (in a learning step) and then apply graph edit distance (in a pattern recognition step) with the learned edit costs. Thus, the edit cost would have to incorporate the transformation parameters. In the experimental section, we show that the method is competitive if the graph edit distance parameters are automatically learned considering the learning set. These parameters are the insertion and deletion costs and also the weights on the substitution costs.