A new color representation for non-white illumination conditions: An effective approach to color machine vision.

摘要

Color is an image attribute that has been used extensively in many areas of computer vision such as image segmentation, object recognition, and object tracking. Color values, however, heavily depend on both the illumination intensity and the color of the illuminant. Variations in illumination intensity create shading effects on object surfaces that need to be discounted. Moreover, with non-white illumination, object surfaces exhibit illumination-induced color changes. It therefore becomes important to use color descriptors that are maximally independent of not only the variations of illumination intensity but also the color content of the illumination. Although the RGB color space is one of the most commonly used color representations, it does not provide illumination-invariance. Other color spaces such as the HSI and the normalized RGB provide simple but ineffective mechanisms to cope with variations of illumination. To the best of our knowledge, a color descriptor that is completely free from illumination effects has not yet been reported in the literature.; In this dissertation, we propose a transformation technique that adapts the color space to the color of the illuminant and leads to a color representation that is more independent of illumination than any existing approaches. This color space transformation extends the well-known RGB-to- HSI transformation to the case of non-white illumination in such a way that the saturation is measured as radial distance from the illumination axis and the hue as the polar angle around the same axis. When a color space is constructed in this manner, it becomes possible to characterize object color with illumination-adapted hue and illumination-adapted saturation. Another benefit of the new color space is that the dichromatic plane now acquires a single-parameter characterization. Our experimental results on color constancy, color image segmentation, specularity detection and color object tracking establish the usefulness of the new approach to color representation.