Multilevel Graph Cuts Based Image Segmentation

This work deals with the graph cuts based image segmentation of gray scale, color and texture images. Multilevel graph partitioning approach is used along with the normalized cuts framework. From the input image, an optimized graph is constructed using intensity, color and texture profiles of the image simultaneously. Based on nature of the image, a fuzzy rule based system is designed for weighted average of image features during graph development. Multilevel graph cuts algorithm is then applied to this graph in which graph is first coarsened to multiple levels in order to reduce the size of graph, coarsened graph is iteratively bi-partitioned through normalized cuts framework to get optimum partitions and then these partitions are projected back towards the original graph during uncoarsening phase to get the segmented image. Berkeley segmentation database is used to test our algorithm and segmentation results are evaluated through probabilistic rand index and global consistency error methods. It is shown that the presented segmentation method provides effective results for most type of images in terms of both accuracy and computational efficiency.