Edge detection may be seen as a goal to be reached, expressed in terms
of the minimization of a cost function f(cost). The present approach
expresses f(cost) in terms of an edge detection error with respect to
an ideal reference which is not explicitly known. This error combines
over- and under-detection errors and is estimated on the basis of a cl
assification criterion for detected edge or non-edge points. The edge
map is obtained by optimization of the estimated error with respect to
the parameters of an edge detection operator. (C) 1997 Elsevier Scien
ce B.V.