Using contour information for image segmentation

This paper proposes an algorithm for image segmentation that improves the graph-based segmentation algorithm by exploiting contour information. The graph-based image segmentation [9] is a fast and efficient method of generating a set of segments from an image. However, its drawback is neglecting the contour information of pixels. Contour can provide significant cues to facilitate the efficient segmentation. We propose an improved weight function that incorporates contour feature into the dissimilarity measure of pixels. We performed experiments on the Berkeley image dataset. Our proposed approach attains significant performance. The experimental results show that our proposed approach is comparable to or even outperforms some state-of-the-art algorithms. In term of global consistency error, our method gives better result while other measures including Probabilistic Rand Index, Variation of Information, and Boundary Displacement Error are close to the best result given by state-of-the-art algorithms.