Color Constant Descriptors Combining Image Derivative Structures

Abstract

Color constant image description is a fundamental problem in many computer vision applications. In this paper, the diagonal-offset model is adopted as reflectance model to get color constant image descriptors. This model makes the descriptors much more robust, and also fits the real world images very well. By introducing 3D moment invariants, this paper contributes to give an illumination independent descriptor generation framework. In detail, 0-, 1- and even higher order color constant descriptors can be generated from such framework. These descriptors can characterize n-order derivative image information. Furthermore, the combination thereof can characterize not only original image but also n-order edge image color information. The experiments on real image databases show that all these descriptors are robust to illumination variation and affine transformation, and perform very well for object recognition under various situations.