AG 1) ——
N
N
JA W
DC
SS
NS
Sum
SS
Ns
SS
N
X) 0,11) ——
Nr
| LA
1 #3 G
0.6
Figure 1: Scale-space behaviour of the ideal line f, when convolved with the derivatives of Gaussian kernels for x € [—3.3] and
e € [0.2, 2]
SNO. ee
(xo. 1.1) —— FAX 051,1) ———
Figure 2: Scale-space behaviour of the bar-shaped line f, when convolved with the derivatives of Gaussian.kernels for x € [—-3. 3] and
c € [0.2, 2]
The line detection algorithm will be developed for this type of
profile, but the implications of applying it to bar-shaped lines will
be considered later on.
2.2 Detection of Lines in 1D
In order to detect lines with a profile given by (2) in an image =(x)
without noise, it is sufficient to determine the points where ='(x)
vanishes. However, it is usually convenient to select only salient
lines. À useful criterion for salient lines is the magnitude of the
second derivative z" (x) in the point where z'(z) = 0. Bright
lines on a dark background will have z” (x) « 0 while dark lines
on a bright background will have z" (x) 3» 0.
Real images will contain a significant amount of noise. There-
fore, the scheme described above is not sufficient. In this case, the
first and second derivatives of z(x) should be estimated by con-
volving the image with the derivatives of the Gaussian smoothing
kernel
2
1 re
——
— 3243 , (3)
Varo
The responses, i.e., the estimated derivatives, will be;
ga Xo
ral wi BY =" arr). f) (4)
"o. Gu h) = iG) * for) (5)
ur, G,w,h) seu u)x (n) (6)
These equations are given in greater detail in (Steger, 1996).
Equations (4)-(6) give a complete scale-space description of
how the ideal line profile f, will look like when it is convolved
with the derivatives of Gaussian kernels. Figure | shows the
responses for an ideal line with w = | and h = 1 (i.e., a bright
line on a dark background) for x € [-3,3] and e € [0.2,2]. As
can be seen from this figure, 7, (x, 6, w, h) = 0 & x = O forall
o. Furthermore, r7 (zx, c, w, h) takes on its maximum negative
value at x = O for all c. Hence it is possible to determine the
precise location of the line for all c. Furthermore, it can be seen
that because of the smoothing the ideal line will be flattened out
as o increases. This means that if large values for o are used, the
threshold to select salient lines will have to be set to an accordingly
smaller value. Section 4 will give an example of how this can be
used in practice to select appropriate thresholds.
For a bar profile without noise no simple criterion that depends
only on z'(z) and z"(x) can be given since z'(x) and z"(x)
822
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
vanish in the interval [—w, w]. However, if the bar profile is
convolved with the derivatives of the Gaussian kernel, a smooth
function is obtained in each case. The responses will be:
Tor, c, w, hy uss h (os (x +w)— d(x — w)) (7)
ry, (ix, Gu hy = h(ge (x + w) — ge(x — w)) (8)
ri (2,0, w, him A (go (x + w) — go (a — w)) 4i n9)
Figure 2 shows the scale-space behaviour of a bar profile with
w = | and h = | when it is convolved with the derivatives of a
Gaussian. It can be seen that the bar profile gradually becomes
"round" at its corners. The first derivative will vanish only at
x = O for all à > O because of the infinite support of go (x).
However, the second derivative ry (x, c, w, h) will not take on
its maximum negative value for small o. In fact, for o. « 0.2
it will be approximately zero. Furthermore, there will be two
distinct minima in the interval [—w, w]. It is, however, desirable
for rj (z, c, w, h) to exhibit a clearly defined minimum at z = 0.
After some lengthy calculations it can be shown that
c 2w/V3 (10)
has to hold for this. Furthermore, it can be shown that
ry (2, 0, w, h) will have its maximum negative response in scale-
space for c. — w/v/3. This means that the same scheme as
described above can be used to detect bar-shaped lines as well.
However, the restriction on « must be observed. The same anal-
ysis could be carried out for other types of lines as well, eg,
roof-shaped lines. However, it is expected that no fundamentally
different results will be obtained. For all o above a certain value
that depends on the line type the responses will show the desired
behaviour of z'(0) — 0 and z"(0) « 0 with ="(x) having a
distinct minimum.
The discussion so far has assumed that lines have the same
contrast on both sides of the line. This is rarely true in real
images, however. For simplicity, only asymetrical bar-shaped
lines
0, x <—w
fol) = L, soil St (11)
had uw
are considered (^ € [0, 1]). The corresponding responses will be:
ra(x,0,w,h) = óc(x-w)-—(h-—1)és(x — w) (12)
ra(x,o,w, h) Jo (x + w) — (h — 1)go (x — w) (13)
Jo (x + w) — (h — 1)g! (2 — w).(14)
ra(x,0,w,h)
Fi