Self-Organizing Map Based Multiscale Spectral Clustering for Image Segmentation

Spectral clustering receives wide attention in recent years since its efficiency in image segmentation and irregular data clustering. However, the applications of it in large scale data processing, such as web data categorization and image segmentation, are greatly restricted because of its O(n³) computational complexity. To address this problem, we propose a new effective scheme to greatly decrease the complexity while keep the clustering quality. The scheme adopts the Self-Organization Map(SOM) to encode the original data and then groups the obtained prototypes using multiscale spectral clustering proposed by us. We analyze and compare the performance of our approach with NJW and find that ours has less time consumption. Furthermore, we carry out an experiment on color image segmentation and results show that our approach behaves better than Kmeans algorithm.