Professional Documents
Culture Documents
2
p
, and
MRC 2:294, for the overall performance index
of RKMRC 3:244.
On the other hand, using G
3
, we could construct
the dierential operator D
3
x 1=2uG
3
x
G
3
x (Fig. 2), which is very similar to Sarkar
and Boyers lter (Sarkar and Boyer, 1991), al-
though not quite as good a performer, yielding
a maximum RKMRC value of 1.98 (i.e., RK
0:703 and MRC 2:812, for / 1:244). The
Fig. 2. Filters D
2
( ) and D
3
() for the same value of u.
Fig. 1. (a) Greens function G
2
and (b) Greens function G
3
.
1756 J.R.A. Torre~aao, M.S. Amaral / Pattern Recognition Letters 23 (2002) 17551759
single-response criterion is generally well satised
by D
3
(0:9 < MRC < 3), but its detection power is
poor (RK < 0:8).
A better step-edge detector can nevertheless be
designed, by combining D
2
and D
3
into a single
dierential operator. This can be achieved thus:
generalizing Eq. (1), we express our estimate for
the derivative of Ix as
I
0
g
x lim
u!0
Ix gu Ix gu
2gu
11
where g is a real parameter. For any nite u, the
larger the g value, the broader will be the range of
x values taken into account in the estimation of the
derivative. It is thus reasonable to expect that a
combination of estimates such as (11), for dierent
values of g, could lead to a better balance in terms
of the conicting requisites of detection accuracy
and robustness against noise. We propose such a
combination under the form
I
0
x
I
0
1
x aI
0
g
x
1 a
lim
u!0
a
21 agu
Ix
_ _
gu
g
a
Ix u
_
Ix
_
gu
g
a
Ix u
__
12
where a is a real parameter.
Estimating Ix u and Ix gu, respectively,
through the Greens functions G
2
and G
3
, we ar-
rive, from (12), at the new dierential operator (see
Fig. 3) Dx 1=2uF x F x, with F x
given by
F x
A
cos /exp
_
a
gu
x
_
cos
c
gu
x
_
/
_
exp
_
b
gu
x
_
g
a
C/ sin
x
u
_ _
exp
_
x
u
_
13
where a, b, c, / and A are the same as in (9), and
C/ 2=uA.
Since A and u are just amplitude and scale
factors, there are three free parameters in Dx
Fig. 3. Filter D.
Fig. 4. Performance indices RK and MRC, as functions of g, for lter D. The various curves correspond to a going from 0.1 to 0.33, in
increments of 0.0165, from left to right on the plots.
J.R.A. Torre~aao, M.S. Amaral / Pattern Recognition Letters 23 (2002) 17551759 1757
which can aect its performance: a, g and /. To
simplify our analysis, we xed / 0, and varied a
and g. The curves of the performance indices RK
and MRC are shown in Fig. 4. For a 1=9 and
g 2:808, the best overall gure of RKMRC
3:457 was found, with RK 1:583 and MRC
2:184. This means a better result than the
RKMRC of 3.388 achieved with the analytical-
form lter of Sarkar and Boyer (1991).
In Figs. 5 and 6, we illustrate the use of Dx for
edge detection in one- and two-dimensional (2D)
signals. In the rst case, the input is a step function
corrupted by additive gaussian noise; in the second,
the clean Lena image is used, with the 1D lter
Fig. 5. (a) Corrupted step-edge signal and (b) signal enhanced by the operator D.
Fig. 6. (a) Lena image, (b) horizontal edges enhanced by operator D, (c) vertical edges enhanced by D, (d) gradient magnitude ob-
tained through D, and (e) gradient magnitude obtained through Cannys edge detector.
1758 J.R.A. Torre~aao, M.S. Amaral / Pattern Recognition Letters 23 (2002) 17551759
(u 0:013) applied separately for the enhancement
of horizontal (Fig. 6(b)) and vertical (Fig. 6(c))
edges, the magnitude of the gradient showing in
Fig. 6(d). For comparison, we also present (Fig.
6(e)) the gradient results obtained through Cannys
(rst derivative of gaussian) operator (Canny,
1986). The scales of the two lters are chosen so
that they have the same width, calculated as in
(Sarkar and Boyer, 1991). This yields r 0:954u,
where r is the scale of Cannys detector.
3. Concluding remarks
Although we have considered only a 1D treat-
ment of our approach, it should be apparent that it
can be extended to 2D, in the same fashion as
described in (Deriche, 1987; Sarkar and Boyer,
1991). Similarly, a recursive implementation of the
digital version of our lters can be readily envi-
sioned. Moreover, for the enhancement of roof
edges, the same approach could be used, now with
the dierentiated signal as input (i.e., we would
just have to apply our operator twice). It is also
worth noting that, for an innite Taylor-series
expansion of the matching equation, our opera-
tor would become D
1
dx u dx u=2u,
where dx is Diracs delta (Fetter and Walecka,
1980). This can be approximated through a dif-
ference of gaussians, in the limit of innitesimal
variance, thus yielding Macleods edge detector
(Macleod, 1970), which is known to approximate
the rst derivative of the gaussian (Canny, 1986).
Finally, we remark that, being explicitly designed
as dierential operators, our lters should nd
applications in signal processing domains other
than edge detection.
Acknowledgements
Partially supported by CNPq 460015/00-1, and
by Finep-Recope 0626/96-SAGE.
References
Canny, J., 1986. A computational approach to edge detection.
IEEE Trans. PAMI 8 (6), 679698.
Deriche, R., 1987. Using Cannys criteria to derive a recursively
implemented optimal edge detector. Int. J. Comput. Vis.,
167187.
Fetter, A.L., Walecka, J.D., 1980. Theoretical Mechanics of
Particles and Continua. McGraw-Hill, New York.
Macleod, I.D.G., 1970. On nding structure in pictures. In:
Kane, S. (Ed.), Picture Language Machines. Academic,
New York.
Poggio, T., Voorhees, H., Yuille, A., 1985. A regularized
solution to edge detection. Tech. Rep., AIM-833, MIT
Articial Intell. Lab., Cambridge.
Sarkar, S., Boyer, K.L., 1991. On optimal innite impulse
response edge detection lters. IEEE Trans. PAMI 13 (11),
11541171.
Spacek, L., 1986. Edge detection and motion detection. Image
Vis. Comput. 4 (1), 4356.
Torre~aao, J.R.A., 2001. A Greens function approach to shape
from shading. Pattern Recogn. 34 (12), 23672382.
J.R.A. Torre~aao, M.S. Amaral / Pattern Recognition Letters 23 (2002) 17551759 1759