Surface Segmentation from Luminance and Color Differences
Inventors: Donald MacLeod, Geoffrey Boynton and Ione Fine
Potential Uses: Automatic Segmentation of Targets in Satellite Photos, Segmenting Different Tissue Types in Medical Imaging, Use of Color to Segment and Identify "Meaningful Regions" within a Scene.
A Bayesian model used to automatically segment scenes based on the luminance and color statistics of a particular image set.
The invention describes a plausible segmentation model based on luminance and color differences within natural scenes. Color and luminance differences show a striking lack of independence. When two points fall on the same surface, differences in luminance and color between the two points tend to be small. Conversely, when two points fall on different surfaces, the distribution of luminance and color differences tend to be larger. These statistics for color and luminance differences are not easily captured by correlation statistics or independent component analysis. However the Bayesian model of the invention captures these dependencies well. The model (with no free parameters) based on color and luminance pairwise differences, segments images similarly to human observers, both with an image set based on natural scenes and with a very different novel image set containing both natural and man-made environments. The model is based on the statistics governing differences in luminance and color between neighboring pixels in a given scene (such as a picture of a natural scene, a mammogram image or a satellite photo). For pixels of a given separation, differences in luminance and color are not independent, a change in luminance predicts a change in color and vice versa. Differences in luminance and color between nearby pixels can be modeled by assuming that the probability of belonging to the same surface decreases exponentially with the distance between the two pixels. The Bayesian model based on these statistics can be used to automatically segment scenes based on the luminance and color statistics of the particular image set.
Patent Status: U.S. Patent Application published as 2005-0058351
Publications: J. Opt. Soc. Am. A., 20, No. 7, July 2003
License Terms: Non-exclusive and Exclusive by Field of Use Licenses Negotiable
Contact: Michelle Booden, Ph.D., Director of Licensing, 858.453.4100 x1612, email@example.com