A new quality metric for evaluating edges detected by digital image pr
ocessing algorithms is presented. The metric is a weighted sum of meas
ures of edge continuity, smoothness, thinness, localization, detection
. and noisiness. Through a training process, we can design weights tha
t optimize the metric for different users and applications. We have us
ed the metric to compare the results of ten edge detectors when applie
d tc edges degraded by varying degrees of blur and varying degrees and
types of noise. As expected, the more optimum Laplacian-of-Gaussians
(LoG) filter and Haralick's second derivative method out-perform the s
impler gradient detectors. At high SNR, Haralick's method is the best
choice, although it exhibits a sudden drop in performance at lower SNR
s. The LoG filter's performance degrades almost linearly with SNR and
maintains a reasonably high level at lower SNRs. The same relative per
formances are observed as blur is varied. For most of the detectors te
sted, performance drops with increasing noise correlation. Noise corre
lated in the same direction as the edge is the most destructive of the
noise types tested.