Labeled color image segmentation through perceptually relevant chromatic patterns

Resumen

This thesis defines a computerized system to perform Color Image Segmentation, i,e. to split any digitized image into compact regions of homogeneous color pixels, so that subsequent Image Analysis systems can interpret those regions as objects of the scene. The proposed system is intended to be robust in front of color uncertainty sources, i.e. input noise, texture, shading and shadows, highlights, etc. In order to assure the maximum concordance with the human being's perception of color, we have decided to transform the RGB color components (Red-Green-Blue) provided by typical CCD cameras into the Smith's HSI perceptual components (Hue-Saturation-Intensity). The new color components trim down the uneven RGB variations derived from a single chromaticity when it is illuminated with a range of intensities. However, the Smith's HSI space presents some drawbacks. One of them is the annoying amplification of the RGB noise due to non-linear RGB-to-HSI formulae. To deal with such effect, we present a novel study of the intrinsic variability of the HSI components under different illumination level conditions. As a result, we have derived two estimators of the Hue and Saturation deviations. Based on those estimators, we formulate our Hue and Saturation Stability Functions, which express the degree of confidence of the Smith's H-S values of any color pixel. These functions will be used all through our segmentation algorithms to enhance the reliable color information against the unstable color information. The basic idea of our segmentation scheme is to find out a set of relevant chromatic patterns of the image, so that each pixel can be classified (labeled) to the most similar pattern according to its H-S values. We propose three methods to find out a fuzzy-like characterization of the relevant chromatic patterns of an image. The simplest method consists in obtaining the Hue and Saturation histograms from a manual selection of pixels belon