Image classification based on web community structure model

This paper describes a network-theoretic approach for clustering pixels in a remote-sensing image and has received relatively satisfactory results. This approach is divided into two stages-the local classification stage and the global clustering stage. During local classification, the original image is partitioned into small blocks which are converted into local network graphs respectively. Then, the method of optimal modularity based on the modularity matrix is exploited to detect communities(classes) in each block. This paper also performs class merging to those classes that we think they are geometrically dispersed. All the classes left after merging are considered as nodes in the global network graph, and the main color is computed for each class. Finally, the community detection method mentioned above is employed again to achieve the final classification results.