A Gaussian derivative operator for authentic edge detection and accurate edge localization

One of the nice properties of the Gaussian scale space map is its well beha vedness. This rather well-behaved nature is somewhat deceptive, however, as portions of the map may not have any direct relationship to the features i n the unfiltered image.(4) It has been shown that not all zero-crossing sur face patches can be associated with intensity changes in the unfiltered ima ge. Zero-crossings give rise to both authentic and phantom scale map contou rs. Recently, we proposed an edge enhancement operator, the LWF, which is a weighted combination of the Gaussian and its second derivative.(6) In this paper, we prove analytically and demonstrate experimentally that the LWF p roduces the authentic scale map contours only. We also show that the LWF ha s excellent edge localization (i.e. the points marked by the operator is ve ry close to center of the true edge). A performance comparison between the Laplacian of Gaussian and LWF operators with respect to the localization pr operty is also presented.