Image editing in the contour domain.

摘要

This thesis presents a novel, contour-based, image editing system called Interactive Contour Editing (ICE) in which an image may be edited by directly accessing the edge and contour information. The system relies upon a recent algorithm [26] for reconstructig a high-fidelity approximation of an image from its edge representation. Unlike traditional image editors which provide only indirect access to image content through pixels, our contour-based image editor provides direct access to image content, allowing desired changes to be effected more easily.; Essential to the success of this contour based image editing system, is an efficient and accurate contour grouping function. The contour grouping algorithm used in ICE [34] is Bayesian in nature, thus its performance is highly dependent on the models used for the prior distributions and likelihood functions of the various grouping cues. Originally, these models were heuristic in nature and based largely on intuition and common sense. In this thesis I report the results of an empirical study which quantitatively relates three classical Gestalt principles of perceptual grouping to the statistics of human perceived contours in the natural world. From these statistics I derive empirically-grounded prior distributions and likelihood functions for Bayesian contour grouping, and show that these more accurate models lead to a more effective contour grouping algorithm.; Experimental results suggest that, equipped with this statistically accurate grouping algorithm, our novel contour-based image editing system may increase the efficiency of certain classes of image editing operations.