A robust unsupervised algorithm of contour extraction

Contour extraction is key to automatic object detection. Affected by illumination, clutter and occlusion, extracting contours from real images without templates is a difficult task. In order to improve robustness of unsupervised algorithms of contour extraction, a spectral graph based algorithm searching in two directions is presented. Compared with homogeneous methods, our algorithm is more robust to noise. The experimental results on Berkley image base also show efficiency of our method.