If you have any problems related to the accessibility of any content (or if you want to request that a specific publication be accessible), please contact us at firstname.lastname@example.org.
Iterative Tensor Voting for Perceptual Grouping of Natural Shapes in Cluttered Background
AuthorLoss, Leandro A.
Computer Science and Engineering
AltmetricsView Usage Statistics
Grouping processes provide reliable pre-processed information to higher level computer vision functions by eliminating irrelevant items and preserving salient, organized structures that are more likely to belong to objects. Applications involving object detection or recognition are particularly benefiting from them. In this dissertation, we consider the problem of grouping pixels into contours and lines. A common characteristic of the images we deal with here is the high heterogeneity of the structures that compose both foreground and background. Besides resulting in complete prohibition of any kind of mid- or high-level modeling of foreground objects, this heterogeneity is also responsible for high amounts of clutter produced by background contents. Inthis context, we developed three general and powerful methods based on iterative tensor voting. In tensor voting, segments are represented as second-order tensors and communicate with each other through a voting scheme that incorporates the Gestalt principles of visual perception. Derived from experiments with human beings, such principles provide low level cues that allow computer algorithms to assess structural agreement among neighboring pixels. By iterating the classical tensor voting, we aim at reinforcing this agreement within real structures and weakening local, fortuitous configurations of clutter elements. Besides exploring the benefits of multi-scale analysis and tunable tensor fields, the approaches introduced in this dissertation are all fundamented onthe conservative removal of background elements and re-voting on the retained ones. We have performed extensive experiments to evaluate the strengths and weaknesses of our approaches using both synthetic and real images from benchmarked datasets. Our experiments are enriched with tests on real, challenging microscopic images of breast cells. Our results and comparisons indicate that the proposed method improves contour and line segmentation considerably, especially under severe background clutter. Results also demonstrate the method's capacity of producing high quality detection of breast cells and segmentation of their membranes.