Unsupervised Image Segmentation via Graph-Based Community Detection

Abstract

Community detection arises in a variety of fields ranging from mathematics, physics, biology, computer science, and the social sciences, among many others. It differs from the classical problem of graph partitioning in that the number and size of the groups into which the network is divided are not specified by the user. This makes community identification algorithms ideally suited to deal with image segmentation from a non-supervised perspective. In this chapter we present an unsupervised framework that automatically identifies semantic objects in images by formulating the general problem of semantic segmentation as community detection problem in graphs. The framework broadly follows a four-step procedure. First, we perform an over-segmentation of the original image using the well-known statistical region merging (SRM) algorithm which presents the advantage of not requiring any quantization or color space transformations. Second, we compute the feature descriptors of the resulting segmented regions. For encoding color and other textural information, each region is described by a hybrid descriptor based on color histograms and covariance matrix descriptor. Third, from the set of descriptors we construct different weighted graphs using various graph construction algorithms. Finally, the resulting graphs are then divided into groups or communities using a community detection algorithm based on spectral modularity maximization. This algorithm makes use of the eigenspectrum of matrices such as the graph Laplacian matrix and the modularity matrix which are more likely to reveal the community structure of the graph. Experiments conducted on large orthophotos depicting several zones in the region of Belfort city situated on the northeastern of France provide promising results. The proposed framework can be used by semiautomatic approaches to handle the challenging problems of scene parsing.