Steger Carsten
(a) Color image
(d) Gray image (e) Gradient magnitude (f) Gray value edges
Figure 17: Comparison between color and gray value edge detection. (a) Color image of the Olympic Stadium in Munich.
(b) Color gradient magnitude in the central part of (a); high gradients are displayed dark. (c) Extracted color edges.
(d) Gray value image corresponding to (a). (e) Gray value gradient magnitude. (f) Extracted gray value edges.
(a) Color image (b) Extracted lines (f) Detailed view
Figure 18: (a) Color image of cables. (b) Extracted color lines; lines are shown in white, while their width is shown in
black. (c) Detailed view of the left upper part of (b).
with equal polarity. With this, it is shown that the line position is generally biased if the line exhibits different lateral
contrast, while the line width is always biased. From the scale-space analysis an efficient method to remove the bias
from the line position and width is derived. Additionally, a subpixel accurate edge extraction algorithm is proposed by
regarding edges as bright lines in the gradient image. A performance analysis (Steger, 1998c, Steger, 1998a) shows that
edges and lines can be extracted with better than 1/25 pixel accuracy under good conditions.
The behavior of the line and edge detectors in junction areas is analyzed. For the edge detector, the two traditional
definitions as zero crossings of the second derivative in the direction of the gradient and the zero crossings of the Laplacian
are also considered. This analysis shows the reasons why junctions are often extracted incompletely. From this analysis,
a new and efficient method to complete missing junctions is derived.
The proposed line and edge extraction algorithms have been extended to handle multispectral images. Unfortunately, the
line position and width correction seems to be an elusive goal for multispectral lines. Nevertheless, for lines with low
asymmetry in all bands, the line position and width will be fairly accurate.
154 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.