We present a multi-scale approach that uses Laplacian eigenvectorsto extract globally significant contours from an image. The
input images are mapped into the Laplacian space by using Laplacian
eigenvectors. This mapping causes globally significant pixels
along the contours to expand in the Laplacian space. The measure
of the expansion is used to compute the Global Contours. We apply
our scheme to real color images and compare it with several other
methods that compute image and color saliency. The contours calculated
by our method reflect global properties of the image and are
complementary to classic center-surround image saliency methods.
We believe that hybrid image saliency algorithms that combine our
method of Global Contours with center-surround image saliency
algorithms will be able to better characterize the most important
regions of images than those from just using contours calculated
using bottom-up approaches.
| 