Jack Tumblin: PhD Dissertation


Three Methods For Detail-Preserving Contrast Reduction For Displayed Images


(Jack) John Erwin Tumblin
161 Pages
Directed by Dr. Greg Turk
Georgia Institute of Technology
December, 1999

PDF file(18,179KB)

We are immersed in a world of high contrast scenes that we cannot directly reproduce in displayed images. Night scenes, sunny days and glaring reflections are filled with contrasts measured in ratios of thousands or millions to one, but we use display devices such as CRTs and printers with maximum contrasts measured in ratios of tens or hundreds to one. How can we reduce the large contrasts of a scene sufficiently for display yet still preserve the small contrasts of important scene details and textures made visible by local adaptation processes in human vision?

This dissertation argues that we should first separate the scene into ``large features'' and ``fine details'' and then construct the displayed image by combining compressed large features and preserved fine details. Most previous contrast-reducing methods either avoid this separation and suffer some loss of fine details, or perform separations based on linear bandpass filter decompositions that introduce halo-like artifacts in displayed images. Borrowing from computer vision, physiology and visual psychophysics, this dissertation presents three new display methods for high contrast scenes.

The layering method uses a new sigmoid-shaped function, similar to the response of film or retinal ganglia, to compress only the illumination components of a computer graphics rendering, preserving scene reflectances and transparencies as fine details. The foveal method interactively adjusts the displayed image for best reproduction in a small region centered around the user's direction of gaze. This ``foveal'' region is preserved as fine detail, and large peripheral features are compressed. The LCIS (Low Curvature Image Simplifier) method uses a variant of anisotropic diffusion to separate fine details from large features defined by scene boundaries and smooth shading. Only large, simple, but sharply-bounded scene features are compressed. Example images made from extremely high contrast scenes demonstrates how successfully each method captures visual appearance.

Back to Home Page
Last Updated 06/28/2004