ABSTRACT
INTRODUCTION The perception of color has been a subject of research in many disciplines, including image understanding. One ap proach to color-image understanding has mainly involved an application of algorithms originally developed for achromat ic images to three-dimensional color space. These include edge detection,1 clustering,2-4 region splitting,5 and region growing.6*7 These schemes regard color as a random vari able that is analyzed statistically without regard to any mod el of specific physical processes of color generation. A sec ond approach involves the explicit use of knowledge about the generation of color. One example is the work of Sloan,8 which shows that simple statements can be made about outdoor shadows and object colors. Similarly, Rubin and Richards910 demonstrated that by analyzing processes that cause changes in color, one can distinguish material changes from shadow boundaries in some situations. Finally, Shafer11 presents a theoretical analysis of highlights and object-color reflection that provides a way to remove high lights from portions of images.