Genetic Approximate Matching of Attributed Relational Graphs

Image segmentation algorithms identify meaningful spatial entities for content-based image retrieval. One or several visual features are extracted for each entity. Based on the feature vectors of the spatial entities and their mutual relationships, attributed relational graphs (ARG) can effectively model entire images. The image retrieval process in an ARG context requires efficient methods to compare these graph models. Each comparison involves the resolution of a general inexact (sub-)graph matching problem. Inexact graph matching already has a long tradition in the domain of pattern recognition (Conte et al., 2004). The variety of existing methods can be classified as either exhaustive state-space search approaches, which guarantee the optimal solution, or approximate methods. The last ones trade the global optimality in for a complexity reduction by accepting sub-optimal solutions. They are usually based on the optimization of an objective function. Although our primary research interest lies in content-based image retrieval, this paper focuses on the general comparison of the two classes. We compare a state-of-the-art exhaustive tree search algorithm (Berretti et al., 2001) and a prototype based on a genetic approach. Both methods are universally applicable, i.e. they do not impose neither any constraints nor any preprocessing steps on the graphs. In the following, we summarize both approaches, before presenting the experimental results and drawing conclusions.