Segmentation of Similar Images Using Graph Matching and Community Detection

In this paper, we propose a new method to segment sets of similar images using graphmatching and community detection algorithms. The images in a database are represented by Attributed Relational Graphs, allowing the analysis of structural and relational information of the regions (objects) inside them. The method gathers such information by matching all images to each other and stores them in a single graph, called Match Graph. From it, we can check the obtained pairwise matchings for all images of the database and which objects relate to each other. Then, with the interactive segmentation from one single image from the dataset (e.g. the first one) we can observe these relationships between them through a color label, thus leading to the automatic segmentation of all images. We show an important biological application on butterfly wings images and a case using images taken by a digital camera to demonstrate its effectiveness.