Extrapolative Spatial Models for Detecting Perceptual Boundaries in Colour Images

Abstract

We present a novel approach for automatically building image-specific extrapolative spatial models of non-boundary local energy that can be used to perform local statistical tests to detect perceptual boundaries. The non-boundary model consists of statistics of local energy that are spatially extrapolated by a non-boundary confidence map and a scale-adaptive normalised filtering algorithm. We exploit the flexibility of steerable filters to both extract oriented local energy and to provide local statistics of the energy distribution in the orientation-domain to compute the non-boundary confidence map. Finally, we apply our local thresholding technique separately to the three channels of colour images and adopt a max operator to combine the results. We provide a qualitative and quantitative comparison on real images from a hand-segmented natural image database against the best combination of the most widely cited colour edge detectors and automatic global thresholding methods.