Multi-scale Approach to Salient Contour Extraction

Contour detection is an important and difficult task for object segmentation in computer vision mainly because contours often occur in the presence of background texture. A previously reported scheme, based on the model of cortical cells in primates, analyses the local energy output derived at a single selected scale [7]. Selection of appropriate scale for image analysis is a difficult problem for which the biological visual system has evolved some solutions. We examine the role that scale plays in image analysis and present a scheme for contour detection based on combination of excitatory and inhibitory responses of a multi-scale, log-Gabor filter bank. The proposed scheme is a thus a generalization of the approach in [7]. The resulting operator has been tested on a wide range of images and found to successfully extract salient contours. The results indicate that the proposed operator responds strongly to isolated lines/edges and weakly to edges that belong to a background texture.