Edge and Ridge Detection with Color Profiles

This paper proposes a method for a subpixel accuracy detection of edges, ridges, and other curvilinear visual patterns, based on identification of their “extended color profiles”. These profiles accurately capture the image color in an orthogonal direction to the curve, in a range of a few pixels. It is well known that various types of “curvilinear visual patterns”, and not only edges and ridges, appear in abundance in high resolution images of the real world. We believe that the most coherent way to adequately capture and represent such patterns is to explicitly identify their shape in the direction transversal to the curve, i.e. to capture their extended color profiles (and not only their blur factor). Also for edges and ridges we argue that their blur factor, as appears in the existing methods, is insufficient in order to reconstruct them in a visually accurate way. To this extent we suggest an experiment visualizing “true” color profiles of edges and ridges. Its results present pretty complicated color profiles of certain edges and ridges in natural high-resolution images. The visual effect of these profiles is easily distinguishable on the image. We describe “shape curves” the main element of our “output format” in which we represent detected edges, ridges, and other curvilinear patterns. A shape curve is defined by a central line, which is a spline curve of the third or second order, equipped with extended color profiles along it. We describe an algorithm starting with the initial detection of edges and ridges, then constructing on this base empiric model elements (EME’s) accurately capturing the image color in a vicinity of the edge/ridge. Finally the shape curves with their color profiles are identified through the chains of the EME’s. We also describe construction of aggregated shape curves with aggregated color profiles. They capture bundles of adjacent shape curves.