An important class of color constant im- age descriptors is based on image derivatives. However, these derivative-based descriptors are sensitive to blurring changes of image, and also lose much color information in uniform regions. Aiming at overcoming these drawbacks, the contributions of this paper are in three-folds: First, by introducing 3D moment invariants, we propose a color constant descriptor in RGB color space, which can char- acterize original image color content. Second, we de¯ne a 2D edge-based color space based on the diagonal-o®set re°ectance model. A novel color constant descriptor in the proposed edge-based color space is presented by using mo- ment invariants. This descriptor can characterize edge im- age with nearly no sensitivity to image blurring. Finally, we combine the aforementioned descriptors to construct the third color constant descriptor, which can describe not only original image color content but also the edge image color feature. The experiments on scene recognition show that all these descriptors have satisfying performance un- der various light sources. Furthermore, all these proposed descriptors are insensitive to image a±ne transformations and blurring changes.
DownLoad: