Une méthode de réduction exacte pour la segmentation par graph cuts

From now on, graph cuts are a standard in the computer vision community. However, their huge memory consumption remains a challenging problem since underlying graphs contains billions of nodes and even more edges. Except some exact methods [14, 10, 5], the heuristics in the literature can only obtain an approached solution [12, 8]. First, we present a new strategy for reducing exactly these graphs : the graph is built by adding nodes which satisfy locally a given condition and corresponds in a narrow band around the segmented object edges. The experiments presented for segmenting gray-levels and color images highlight low memory usage and show a low distance between segmentations. We also present an application of this method for segmenting lung tumors in CT images.