Stephan Scholze
2 EDGE DETECTION IN COLOR IMAGES
Edge detection in color images is not a widely studied subject. In most practical applications, edge extraction is performed
on the corresponding grey-level intensity image using "standard" edge-detectors such as the Canny-operator (Canny, 1986)
or adaptions as the Deriche-operator (Deriche, 1987). These operators are based on the norm of the intensity gradient in
each pixel. Denoting the smoothed grey-level intensity at pixel (x, y) with z(z, y) one would formally compute
: iN Di”
IVi(z, y)|| = (2) +5) (I)
In case of a color image with red, green and blue bands, the image can be represented as a vector valued function
I c R? 5 R? where I(z, y) — (r(z, y), g(z, y), b(z, y)) represents the RGB-values of the pixel (x, y). We compared
the following combinations of the derivatives in the individual bands for replacing || Vz(z, y) ||.
2. The Maximum
We consider the vector valued function I(z, y) — (r(z; y), g(z, y), b(z, y)), representing the multispectral image at pixel
(z, y). To replace the norm of the intensity gradient || Vi(z, y)|| from the single band case, we choose the maximum of
the norm of the gradients in the individual color bands at each pixel position:
I| Vi(z, y)]| — max t] Vr(z, y)ll, lIVoGr, v)]l, || Vb(z, v)]H ()
2.2 The Spatial Gradient
In case of a scalar field the direction and magnitude of its strongest change are given by the gradient of the field. This idea
can be extended to vector fields (Lee, 1991) and was recently used for edge detection in color images (Zafiropoulos and
Schenk, 1999). Again we consider the vector valued image function I(z, y) — (r(z, y), g(z, y), b(x, y)) and an arbitrary
direction n — (cos à, sin 9) which is defined by the angle ¢ in the image plane. The objective is to find the direction of
strongest change in I(z, y) at point (z, y).
First we compute the directional derivative of I(x, y) with respect to n:
or Or
ox Oy
9r = S S n=Jn (3)
on = 2
ox oy
which is equivalent with forming the scalar product of the Jacobian J of the image function I with n. As measure of
magnitude of change of I(z, y) as a function of n one usually chooses the square of the norm of Jn:
x In]? = n‘3J" In (4)
With this choice, note that the n^ J7 Jn is equivalent with the Rayleigh-quotient of D — J7J. Thus the direction oí
the largest change of I(z, y) at (x, y) is given by max(/?) which is in turn given by the largest eigenvector of D. Since
D is real, the eigenvalues of D are the squares of the singular values of J. Clearly, computing the spatial gradient is
computationally expensive, since we have to perform a singular value decomposition at each pixel position.
2.3 Comparison of both combination schemes
After edge detection and edge linking, straight line segments are fitted to the edges. A very tight threshold is used for li
fitting so that curves are not piecewise linearly approximated. Moreover, only line segments above a minimum length (D
pixels) will be considered for matching. For computational efficiency during further processing steps, the line segments
are stored in an R-Tree (Guttman, 1981) data-structure. This allows efficient queries for adjacent lines, which will be
exploited later. Additionally the straight lines are also stored in raster format, i.e. a pixel indexes a line segment if there
one at that location.
Since we want to investigate polygonal shapes, only the fitted straight line segments will be used all through the rest of
the paper. We will not refer to the underlying unfitted data. Thus, in the following the terms ’edge’ and ’line segment
will be used interchangeably, both referring to the fitted straight line segments.
In Figure (1) the results of both edge detection schemes are compared with the result obtained by performing the edge
detection only on the grey-level intensity image. At first glance, the three methods perform similarly well. The number of
816 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.
extrac
case.
the gr
The Ir
line sé
the 10!
the m:
Takin
maxin
In cas
by ad
edge-1
Figure
gradie
pixels
respec
à LI
Suppo
previo
seconc
the giv
match
based
31 1