A new algorithm for image similarity measure and graph edit distance

For the difficulty of reasonably defining cost functions for edit operations in Graph Edit Distance (GED), a purely structural method for GED computing is proposed which is thoroughly independent of cost functions. It makes use of Edge Direction Histogram (EDH) to characterize the structure of graphs, and estimates the dissimilarity of graphs by computing the distance of EDHs without defining cost functions. Earth Mover's Distance (EMD) examines histograms' distance exactly and is robust with respect to the diversification of histograms caused by rotating a graph in the same plane, so GED is measured by the EMD of EDHs in this paper. Images are represented by graphs and this new distance measure is used for examination of image similarity. Experimental results demonstrate that the proposed method is effective for classifying and clustering images. Compared with the GED from spectral seriation, the proposed method can capture the structure difference of graphs better.