Computational methods for perceptual organization and object recognition.

摘要

Perceptual organization, the initial organization of an image into meaningful components, is a fundamental topic in computer vision. In this thesis, I focus my research on developing computational models for perceptual organization and its utility in shape-based object recognition from images.;For perceptual organization, I introduce a novel global contour criterion for guiding image segmentation. For this criterion, instead of modeling boundaries by local discontinuity, I propose to use the global derivative of a region-based criterion with respect to the entire contour of a segment. For several existing region-based criteria, including MDL and NCuts, I show theoretically and experimentally that this global derivative criterion is related to the simultaneous contrast effect of visual psychology and provides important complementary information to the original region-based criterion. Incorporating this global discontinuity criterion significantly improves the performance.;For the utility of perceptual organization, I propose to use image segmentations to represent and compare shapes. Matching segmentations based on mutual information allows accurate shape matching between images and achieves efficiency by avoiding the need for computing point-to-point edge correspondences. To address the unreliability of low-level image segmentations, a Bayesian technique is developed to estimate the average matching score of all possible segmentations of the compared images. This Bayesian averaging technique also has implications for perceptual organization. Based upon it, I develop a novel image segmentation algorithm that computes the mean segmentation of an image. This technique gives better consistency and predictability in computing image segmentations. I also introduce a new perspective on structure preserving image smoothing that smooths images according to their global optimal structure, i.e. the mean segmentation.;Promising results are shown in our extensive experiments in object detection, shape-based tracking, image smoothing, and image segmentation.