Edge detection and skeletonization using quantized localized phase

Abstract

Localized phase conveys much more information about image structure than magnitude does. For example, phase is used in image reconstruction, edge detection and analysis of textures. We take advantage of the fact that phase is more important on edges and contours than it is in smooth, edge-free regions. Based on this observation, we detect edges as follows: we first calculate the local spatial-frequency transform. We then reconstruct the image, using the magnitude and the quantized phase. The effect of phase quantization on the reconstruction error is negligible in smooth areas, while it is very significant around edges. The reconstruction error provides therefore an excellent map of the edges and a skeleton of the image in the sense of primal sketch.