Mastering Windows: Improving Reconstruction
Thomas Theul Helwig Hauser Eduard Groller
Institute of Computer Graphics
Vienna UniversityofTechnology
(a) (b) (c)
Figure 1: A CT scan of a head reconstructed with (a) linear interp olation and central di erences with linear interp olation, (b)
Catmull-Rom spline and derivative and (c) Kaiser windowed sinc and cosc of width three with numerically optimal parameters.
dows. The latter two are particularly useful since b oth have Abstract
a parameter to control their shap e, which, on the other hand,
requires to nd appropriate values for these parameters. We
Ideal reconstruction lters, for function or arbitrary deriva-
showhow to derive optimal parameter values for Kaiser and
tive reconstruction, have to b e b ounded in order to b e prac-
Gaussian windows using a Taylor series expansion of the con-
ticable since they are in nite in their spatial extent. This
volution sum. Optimal values for function and rst deriva-
can b e accomplished bymultiplying them with windowing
tive reconstruction for windowwidthsoftwo, three, four and
functions. In this pap er we discuss and assess the quality
ve are presented explicitly.
of commonly used windows and show that most of them are
unsatisfactory in terms of numerical accuracy. The b est p er-
Keywords: ideal reconstruction, windowing, frequency
forming windows are Blackman, Kaiser and Gaussian win-
resp onse, Taylor series expansion
ftheussl,helwig,[email protected]
1 Intro duction
Reconstruction is a fundamental pro cess in volume visualiza-
tion. Mo dalities like CT or MRI scanners provide discrete
data sets of continuous real ob jects, for instance patients
in medical visualization. The function in between sample
points has to b e reconstructed.
It is well known that a band-limited and prop erly sampled
function can be p erfectly reconstructed by convolving the
samples with the ideal function reconstruction lter. Simi-
larly, derivatives of the function can, due to linearity of con-
volution and derivation, directly be reconstructed by con- of the reconstructed function. Goss uses the windowed cosc
volving the samples with the corresp onding derivative of the only on sampling p oints,inbetween some interp olation has
ideal reconstruction lter. The rst derivative of a three- to b e p erformed which is not stated explicitly.
dimensional function, the gradient, can b e interpreted as a Most researchers assess the quality of reconstruction lters
normal to an iso-surface passing through the p ointofinter- in frequency domain, whereas Moller et. al [10 ] prop ose a
est. This gradient is usually used for shading or classi cation purely numerical metho d op erating in spatial domain. We
and therefore the quality of gradient reconstruction strongly will review this metho d in Chapter 4 as it will allowusto
a ects the visual app earance of the nal image. derive optimal parameters, in terms of numerical accuracy,
From signal pro cessing theory we know that the ideal for the Kaiser and Gaussian windows.
function reconstruction lter is the sinc lter.
sin x
3 Windowing ideal reconstruction lters
if x 6=0
x
(1) sinc (x)=
1 if x =0
The windows investigated in this work are
The rst derivative of the sinc lter, called cosc lter,
- The rectangular windoworbox function
cos (x) sinc(x)
if x 6=0
x
1 if jxj
(2) cosc (x)=
(x)= (3)
0 if x =0
0 else
consequently is the ideal rst derivative reconstruction l-
which p erforms pure truncation.
th
ter [1]. In general, the ideal reconstruction lter of the n
th
- The Bartlett window, which is actually just a tent func-
derivative of a function is the n derivative of the sinc func-
tion:
tion.
However, since these lters are in nite in their spatial
jxj
if jxj < 1
(4) Bartlett (x)=
extent they are impracticable. Simple truncation causes
0 else
severe ringing artifacts, as can be seen in Fig. 6, which
shows the Marschner-Lobb data set [6 ] reconstructed with
- The Welch window:
linear interp olation and central di erences and linear inter-
2
p olation (Fig. 6a), the Catmull-Rom spline and derivative
x
jxj < 1
(5) Welch (x)=
(Fig. 6b) and the truncated cosc of width three (Fig. 6c).
0 else
(for high resolution images please refer to the web page of
this pro ject [15 ]). We will investigate and explain the cause
- The Parzen window
of this bad result in frequency domain in Chapter 3 and in
(
2 3
spatial domain in Chapter 4.
4 6jxj +3jxj 0 jxj < 1
1
3
To reduce artifacts originating from truncation the ideal
Parzen (x)=
(2 jxj) 1 jxj < 2
4
reconstruction lters can be multiplied with appropriate
0 else
functions which drop o more smo othly at the edges. Wewill
(6)
discuss some of the more commonly used of these functions,
is a piece-wise cubic approximation of the Gaussian
which are adversely called windows. Weshow that most of
window with extend two. Although its width is not
them are unsuitable for reconstruction, esp ecially for higher
directly adjustable it can, of course, b e scaled to every
order derivative reconstruction, although some, for example
desired extend.
Hamming [5 ] and Lanczos [16 ] windows, are rep orted to b e
commonly used.
- The Hann window (due to Julius van Hann, often
wrongly referred to as Hanning window [17 ], sometimes
just cosine b ell window) and Hamming window, which
2 Previous work
are quite similar, they only di er in the choiceofone
parameter :
Turkowski [16 ] used windowed sinc functions for image re-
x
sampling. He found the Lanczos window sup erior in terms
+(1 )cos( ) jxj <
H (x)= (7)
;
of reduction of aliasing, sharpness, and minimal ringing as
0 else
conclusion of an empirical exp eriment.
Marschner and Lobb [6 ] used a cosine-b ell windowed sinc
1
b eing the Hann windowand =0:54 the with =
2
(often called Hann window) and found it sup erior to the en-
Hamming Window.
tire family of cubic reconstruction lters. According to their
metric there is always a windowed sinc with b etter smo oth-
- The Blackman window, which has one additional cosine
ing and p ost-aliasing prop erties. Machira ju and Yagel [5 ]
term as compared to the Hann and Hamming window.
use a Hamming windowed sinc without further explanation
(
1 x
0:42 + cos ( )+
of their choice. However, we found the Hamming window
2
x
Blackman (x)= 0:08 cos (2 ) jxj <
of not being optimal. Recently, Meijering et al. [7 ] per-
0 else
formed a purely quantitative comparison of eleven windowed
(8)
sinc functions. They found the Welch, Cosine, Lanczos and
Kaiser windows yielding the b est results and noted that the
- The Lanczos window, which is the central lob e of a sinc
truncated sinc was one of the worst p erforming reconstruc-
function scaled to a certain extend.
tion lters.
x
) sin(
Goss [2] extended the idea of windowing the ideal recon-
jxj <
x
(9) Lanczos (x)=
struction lter to derivative reconstruction. The parameter
0 else
of the Kaiser windowcontrols the smo othing characteristics 1.2 1.2 rectangular window Hann window Bartlett window Hamming window 1 Welch window 1 Blackman window Parzen window Lanczos window
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
-0.2 -0.2
-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3
spatial domain
1.2 1.2 rectangular windowed sinc Hann windowed sinc Bartlett windowed sinc Hamming windowed sinc Welch windowed sinc Blackman windowed sinc 1 Parzen windowed sinc 1 Lanczos windowed sinc ideal ideal
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
pi pi
frequency domain
3.5 3.5 rectangular windowed cosc Hann windowed cosc Bartlett windowed cosc Hamming windowed cosc 3 Welch windowed cosc 3 Blackman windowed cosc Parzen windowed cosc Lanczos windowed cosc ideal Ideal 2.5 2.5
2 2
1.5 1.5
1 1
0.5 0.5
0 0
0 pi 2pi 0 pi 2pi
Figure 2: Rectangular, Bartlett, Welch, Parzen, Hann, Hamming, Blackman and Lanczos windows of width two on top, b elow
the frequency resp onses of corresp ondingly windowed sinc and cosc functions.
window. It is in its general form de ned by - The Kaiser window [3], which has an adjustable param-
eter which controls how steeply it approaches zero at
2
x