Graph Theory and Mean Shift Segmentation Based Classification of Building Facades

Abstract

Automatic classification of building facades has become a field of interest for researchers, since it can provide important solutions to many remote sensing research field. Urban scene understanding, mobile building recognition systems can be counted as the most important application fields. Classification results can be also used to increase the level of details in threedimensional city models which are generated from aerial or satellite imagery. Characterization of stable regions in facades is also necessary for robust indexation and image retrieval. Although in the literature many different facade classification techniques are proposed, because of the requirements of real-time applications robust and fast methods are still needed. Therefore in this study we propose a mean shift segmentation and graph theory based fast facade classification method. For this purpose, first mean shift segmentation is applied to color image and the connected pixels having similar color properties are labeled as one segment. Then, we benefit from graph theory to classify the segments as window, wall, and sky. The mass centers of the detected segments form the nodes of graph, and the neighborhood information is obtained through Delaunay triangulation. Next, a graph-cut is applied through graph of the image. Finally, we use a probabilistic verification step to compensate shortcomings of graph-cut step. Proposed algorithm is tested on a facade data set which is composed by many different kind of optical facade images which are taken from different places.