| In an image, color feature is highly discriminative for identifying the scene objects, but it is often difficult to uniquely extract. The challenge is primarily related to variations in the object's color in different images captured under varying illuminations at different viewpoints. A number of methods have been developed to find true object color, which is independent of illuminations and viewpoints. In this dissertation, we will address the problems of color variations in images and introduce a novel approach for simultaneously estimating the diffuse and specular reflections from a collection of images based on the dichromatic reflection model. By minimizing a cost function, the observed color in a scene can be decomposed into diffuse reflection (modeling the object color), specular reflection (modeling the illumination) and geometry parameters of the dichromatic reflection model. We refer to the resulting color of diffuse reflection as true object color. The estimated parameters of each pixel in the set of images are sequentially obtained by Gauss Seidel iterative approach [60]. The robustness of our method is demonstrated using a set of home-generated and standard image sets. We also demonstrate that our proposed method can be used to estimate the dichromatic reflection model of an object from a single image. The estimated true object color allows for excellent segmentation performance. |
