0% found this document useful (0 votes)
106 views5 pages

Binomial Filters

Binomial filters provide a compact approximation of the Gaussian filter using finite impulse response. They have appealing properties for hardware implementation as they do not require multiplications. Key applications include constructing multiscale image representations. Binomial filters are defined by binomial coefficients and correspond to rows of Pascal's triangle. Higher order filters are generated by cascading lower order filters. The frequency response of binomial filters approximates a Gaussian low-pass filter with increasing order.

Uploaded by

verdos
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
106 views5 pages

Binomial Filters

Binomial filters provide a compact approximation of the Gaussian filter using finite impulse response. They have appealing properties for hardware implementation as they do not require multiplications. Key applications include constructing multiscale image representations. Binomial filters are defined by binomial coefficients and correspond to rows of Pascal's triangle. Higher order filters are generated by cascading lower order filters. The frequency response of binomial filters approximates a Gaussian low-pass filter with increasing order.

Uploaded by

verdos
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 5

Overview of Binomial Filters

Konstantinos G. Derpanis
Department of Computer Science and Engineering
York University
[email protected]
March 5, 2005
This note gives an overview of binomial lters. Binomial lters form a compact
rapid nite impulse response (FIR) approximation of the (discretized) Gaussian. An
appealing implementation feature of these lters is that they do not require multipli-
cations and thus are amenable to hardware implementation (e.g., (Aubury & Wayne,
1996; Haddad, 1971; van der Wal & Burt, 1992)). A key application of these lters
in computer vision is in the construction of multiscale image/volume representations
(e.g., (Burt, 1981; Hummel & Lowe, 1986; Crowley, Ri & Piater, 2003; Lindeberg &
Bretzner, 2003)).
The binomial coecients, given by:

N
n

=
N!
(N n)!N!
, where n = 0, . . . , N, (1)
form (when L
1
normalized) a rapid approximation of the Gaussian (normal) distrib-
ution:
G[m] =
1

2
e
m
2
/2
2
, (2)
where m = n N/2, the standard deviation =

N/2, and N (i.e, the order of the


binomial lter) denotes the number of cascaded convolutions (denoted by ) of the
binomial lter,

1 1

, used to generate the binomial lter:


B
N
=

1 1

1 1

1 1

. .. .
N times
. (3)
Notice that the binomial lters are closed under convolution,
B
N
= B
NM
B
M
. (4)
1
20 15 10 5 0 5 10 15 20
0
0.1
0.2
0.3
0.4
0.5
(a) B
2
20 15 10 5 0 5 10 15 20
0
0.1
0.2
0.3
0.4
0.5
(b) B
4
20 15 10 5 0 5 10 15 20
0
0.1
0.2
0.3
0.4
0.5
(c) B
1
6
/2 0 /2
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1

16
(d) B
1
6
Figure 1: Binomial lters. (a)-(c) depict the impulse responses of several binomial
lters. (d) depicts the frequency responses of the binomial lters given in (a)-(c).
Alternatively, the binomial lters correspond to the rows of Pascals triangle:
1
1 1
1 2 1
1 3 3 1
1 4 6 4 1
1 5 10 10 5 1
.
.
.
where the sum of the coecients, given by 2
N
, represent the normalization factor (see
Fig. 1 (a)-(c)).
Two-dimensional binomial lters can be generated by using two one-dimensional
binomial lters in a separable fashion, for example:
B
2
= (1/4)

1 2 1

(1/4)

1
2
1

= (1/16)

1 2 1
2 4 2
1 2 1

. (5)
In one-dimension, the discrete-time Fourier transform (DTFT) of the binomial
lter B
2
= (1/4)

1 2 1

corresponds to a single period of a cosine raised by a


2
constant, in other words, a low-pass lter with no ripples in the stop band:

B
2
[] =

n
B
2
[n] cos[n] =
1
2
+
1
2
cos(). (6)
For higher order binomial lters, the Fourier transform is given by the multiplication
1
of B
2
s spectra with itself N times (see Fig. 1 (d)):

B
N
[u] =

1
2
+
1
2
cos(u)

N
. (7)
Figure 2 illustrates the rapid approximation of the Gaussian by the binomial
lters. For further details on binomial lters and empirical evaluations, see (Burt,
1981; Haddad, 1971; Nicholson & Davis, 1986; Crowley, Ri & Piater, 2003; Jahne,
2005).
1
This is due to the equivalence of spatial convolution and frequency multiplication.
3
0 10 20 30 40 50 60 70 80 90 100
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
Binomial order
E
r
r
o
r
:

s
u
m

o
f

a
b
s
o
l
u
t
e

d
i
f
f
e
r
e
n
c
e
s
Figure 2: Plot showing the sum of the absolute errors between the Gaussian (trun-
cated at 3) and its corresponding binomial lter approximation:

G(x, =

N/2) B
N

.
4
References
Aubury, M. & Wayne, L. (1996). Binomial lters. Journal of VLSI Signal Processing,
12(1), 3550.
Burt, P. (1981). Fast lter transforms for image processing. Computer Graphics,
Image Processing, 16(1), 2051.
Crowley, J., Ri, O. & Piater, J. (2003). Fast computation of characteristic scale using
a half-octave pyramid. In International Conference on Scale-Space Theories in
Computer Vision.
Haddad, R. (1971). A class of orthogonal nonrecursive binomial lters. IEEE Trans-
actions on Audio and Electroacoustics, 19(4), 296304.
Hummel, R. & Lowe, D. (1986). Computing gaussian blur. In International Confer-
ence on Pattern Recognition (pp. 910912).
Jahne, B. (2005). Digital Image Processing, sixth edition. Springer-Verlag.
Lindeberg, T. & Bretzner, L. (2003). Real-time scale selection in hybrid multi-scale
representations. In International Conference on Scale-Space Theories in Com-
puter Vision (pp. 148163).
Nicholson, W. & Davis, K. (1986). The binomial window. In SPIE Image processing
for missle guidance (pp. 910912).
van der Wal, G. & Burt, P. (1992). A VLSI pyramid chip for multiresolution image
analysis. 8(3), 177189.
5

You might also like