Spectral rounding and image segmentation

摘要: The task of assigning labels to pixels is central computer vision. In automatic segmentation an algorithm assigns a label each pixel where connote shared property across (e.g. color, bounding contour, texture). Recent approaches image have formulated this labeling as partitioning graph derived from the image. We use spectral denote family algorithms that seek by processing eigenstructure associated with graphs. In thesis we analyze current and explain their performance, both practically theoretically, on Normalized Cuts (NCut) criterion. Further, introduce novel methods, rounding, apply them tasks. Edge separators are produced iteratively reweighting edges until disconnects into prescribed number components. At iteration small eigenvectors eigenvalue computed used determine reweighting. way rounding directly produces discrete solutions must map continuous employing heuristic geometric separator k-means). We show compares favorably approximations NCut criterion in natural segmentation. Quantitative evaluations performed multiple databases including Berkeley Segmentation Database. These experiments demonstrate segmentations improved value (obtained using SR-Algorithm) more highly correlated human hand-segmentations.