You are on page 1of 6

Mikrochim.

Acta 133, 273±278 (2000)

Fusion of 2-D SIMS Images Using the Wavelet Transform

Thomas C. Stubbings1 , Stavri G. Nikolov2, and Herbert Hutter1;


1
Institute for Analytical Chemistry, Vienna University of Technology, Getreidemarkt 9/151, Vienna 1060, Austria
2
Department of Electrical and Electronic Engineering, University of Bristol, Merchant Venturers Building, Woodland Road,
Bristol BS8 1UB, UK

Abstract. Secondary ion mass spectroscopy (SIMS) is Key words: SIMS; multispectral images; image fusion; wavelets;
wavelet transform; wavelet transform fusion.
a powerful method for element distribution examina-
tion of conducting and semi-conducting surfaces at
Secondary ion mass spectroscopy (SIMS) is a powerful
high spatial resolution and with a high sensitivity.
method for examination of conducting and semi-
Routine surface analysis produces about 8 to 15 images
conducting samples at high spatial resolution. Imaging
in a short time, each of which displays the intensity
SIMS is capable of producing two- and three-
distribution of one mass, thus generating a multispec-
dimensional spatially resolved images of element
tral SIMS image. Formation of occlusions, segrega-
distributions [1, 2]. Due to the high sensitivity of
tions, and the overall location of the elements relative
SIMS, even trace elements (including hydrogen) can be
to each other, are dif®cult to recognise when looking at
imaged [3]. Routine measurements record n images
n separate 2-D images.
simultaneously, each of which displays the intensity
Image fusion is a process whereby images obtained
distribution of one mass, thus generating multispectral,
from various sensors, or at different moments of time,
n-dimensional SIMS images. A common problem with
or under different conditions, are combined together to
the analysis of SIMS images is the channeling of the
provide a more complete picture of the object under
sample information into n images, one for each mass.
investigation. The process of combining SIMS images
This aggravates the interpretation of the element
may be viewed as an attempt to compensate for the
distribution, especially the exact location of occlusions,
inherent effect of SIMS to channel the information
segregations, etc., of certain elements within other
obtained from the sample into different images, corre-
elements, is often hard to determine.
sponding to different element phases.
Image fusion can be de®ned as the process by
The wavelet transform is a powerful method for
which several images, or some of their features, are
fusion of images. This work covers the use of wavelet
combined together to form a single image. The fusion
based fusion algorithms on multispectral SIMS images,
result has to meet certain conditions: (a) the fused
evaluating the performance of different wavelet based
image should preserve, as closely as possible, all
fusion rules on different type of image systems and
relevant information contained in the input images
comparing the results to conventional fusion techniques.
[4]; (b) the fusion process should not introduce any
An aim of this study is to increase the information, i.e.
artifacts or inconsistencies which can distract or
the number of masses, which can be merged into one
mislead a human observer or any subsequent image
image in order to enhance the perception and interpreta-
processing steps [4]; (c) in the fused image, irrelevant
tion of the SIMS surface images.
features and noise should be suppressed to a
maximum extent. Pixel level fusion algorithms vary
from very simple, e.g. image averaging, to more
 To whom correspondence should be addressed complex, e.g. principal component analysis (PCA),
274 T. C. Stubbings et al.

pyramid based image fusion [5] and wavelet trans- A0j (the output of the low-pass ®lter L at scale j)
form (WT) fusion [6, 7]. It has been shown that contains the smoothed copies of the image. The other
wavelet transform [8, 9] has several advantages over three channels, Dcj …c ˆ 1; . . . ; 3†, which are the result
similar pyramid based techniques: (a) the wavelet of applying the high-pass ®lter H in the horizontal,
transform is a more compact representation than the vertical, and horizontal-vertical directions, contain the
image pyramid; (b) the wavelet transform provides detail information about the image. The same
directional information, while the pyramid represen- convolution is performed on the low-pass channel
tation doesn't introduce any spatial orientation in the A0j at level j in order to compute the next level ( j ‡ 1)
decomposition process [10]; (c) in pyramid based of the decomposition. Thus, a multiscale wavelet
image fusion the fused images often contain blocking decomposition having J levels is computed.
effects in the regions where the input images are The fusion rule  in the case of 2-D images is
signi®cantly different. No such artifacts are observed actually a set of fusion rules cj , where j ˆ 1, . . . , J and
in similar wavelet based fusion results [10]; (d) c ˆ 0, . . . ,3, which de®ne the fusion of each set of
images generated by wavelet image fusion have better corresponding channels c for each band/scale j. There
signal to noise ratios (SNR) than pyramid based exist many fusion rules to combine the wavelet
fusion images, when the same fusion rules are used coef®cients of the n input images. Some of them,
[11]. Additionally, when subject to human observers, which have been used in this work, are described
the perception of wavelet fused images is better than below.
pyramid fused images. Since, to date, the results of averaging [12]: for each band of decomposition and
pixel level fusion are primarily intended for presenta- for each channel the wavelet coef®cients of the
tion to human observers for easier and enhanced images are averaged, i.e.
interpretation, human perception is the most important  Pn c   Pn 
kˆ1 Dj Ik
0
quality measure for judging the fusion results. c 0 kˆ1 AJ Ik
Dj I ˆ and AJ I ˆ
Besides, only a few measures for quantitative n n
evaluation of different image fusion schemes have …2†
been proposed so far [10, 11] and their application can
maximum [4, 10, 12]: for each band of decomposition
not be generalised to all fusion problems.
and for each channel the maximum of absolute values
of the respective wavelet coef®cients is taken, i.e.
Theoretical
Dcj I ˆ max…jDcj Ik j† and A0J I ˆ max…A0J Ik † …3†
k k
The general idea of all wavelet based fusion schemes
is that the wavelet transforms ! [8, 9] of the registered hi-maximum: for each band of decomposition and for
input images In(x, y) are computed and these trans- each channel the maximum of the respective wavelet
forms are combined utilising some kind of fusion rule coef®cients is taken, but for the low-pass channels
. Then the inverse wavelet transform !ÿ1 is A0J Ik the average is taken instead
computed and the fused image I is reconstructed:  Pn 0

c c 0 kˆ1 AJ Ik
Dj I ˆ max…jDj Ik j† and AJ I ˆ
I…x; y† ˆ !ÿ1 …‰!…I1 …x; y††; !…I2 …x; y††; . . . ; !…In …x; y††Š† k n
…1† …4†

The wavelet transform of an image is a multiscale Experimental


image representation where the image is convolved
Samples
with a set of translated and scaled mother functions
called wavelets. The usual way to compute the discrete Several different samples were used for evaluation of the wavelet
transform based fusion algorithms. Following is a short description
wavelet transform (DWT) of an image !(I1 (x, y)) is to of the three samples used in this paper:
convolve it with a pair of ®lters, one of which is a Sample No. 1 is a steel alloy containing about 1% Al. It was
high-pass ®lter H, while the other is a low-pass ®lter produced by hot isostatic pressing (HIP); this alloy is intended to
L. The DWT is a fast, linear, pyramidal operation, show better high-temperature oxidation resistance as well as a
more homogeneous distribution of all alloy compounds.
which decomposes the image into several bands Sample No. 2 is a Nb-Zr-W alloy coated with a Si, Cr and Fe
(scales) and into four channels. One of these channels layer by Arc-PVD and following tempering at 1600 K. This
Fusion of 2-D SIMS Images Using the Wavelet Transform 275

coating is also intended to improve the oxidation behaviour at high system has certain characteristics and showed differ-
temperatures and therefore was submitted to cyclic oxidation ent results in the fusion process:
treatment for testing purposes.
Sample No. 3 is a standard Voest production steel with average
contamination, examined for the distribution of elements and I. Fusion of type I images makes no sense with more
possible enrichments of some components at the grain boundaries. than two images, no matter what fusion method is
chosen; the large overlapping areas make distinc-
Measurement Equipment tion between more than two images very hard,
For all samples, multispectral SIMS images were recorded even two images are sometimes dif®cult to
(numerous masses for each sample were recorded simultaneously). interpret. For large area images two wavelet fusion
The intensities for each mass were added for a few seconds and methods have proven very effective: maximum
saved as an image. Images were obtained either using an O‡
primary ion beam or by using a Cs‡ primary beam to have the
2
wavelet fusion with symlet wavelets and average
advantage of visualising also the oxygen distribution (primary wavelet fusion with daubechies wavelets. The
energy: 5.5 kV, primary ion current: 2 mA). The rastered area was average wavelet fusion generally gives good results
150  150 mm. SIMS investigations were performed with a
with maximum contrast and resolution of details,
modi®ed CAMECA ims 3f ion microscope. Secondary ion images
were acquired using a Pulnix TM700 CCD camera system in a usually slightly better than the simple average
combination with a double micro-channel-plate-phosphor screen method. The maximum wavelet method has proven
assembly. The camera signal was then digitised by a Matrox Pulsar especially effective when fusing non-overlapping
image-processor and stored on the hard disc of the controlling
computer. features (see Fig. 1), because it reserves maximum
contrast and resolution for the features of the
corresponding image. The shortcoming of this
Evaluation Methods
fusion rule, on the other hand, is the loss of contrast
All sample images were tested in different combinations (fusion
in areas where both images show distinct features,
ranging from two to four element images; for better illustration of
the fusion results only fusion results with two images are shown in so for large overlapping areas (especially with high
this paper) with the following wavelet fusion rules: average, intensities in both images) this method cannot be
weighted average, maximum, and hi-maximum. All these rules recommended.
were tested with four different wavelets (daubechies 4, symlet 8,
coi¯et 3 and haar) at different decomposition levels (2, 5 and 8). II. Fusion results of type II images are slightly
The results of the tests were compared visually one to each other different to the ones of type I. First, multiple
and to the simple averaging fusion results. images can be fused, though normally two or three
images seem to be a good choice, where one
Computation image may ®ll large areas and the others show
Wavelet transform algorithms and fusion rules were implemented occlusions and/or border structures (IIa and IIb).
as Interactive Data Language (IDL) [13] scripts as part of the When the structures overlap with the main
wavelet fusion toolbox (WFT) [14] for IDL. All computations features, discernment becomes more dif®cult but
were performed on a Dual 233 MHz Pentium II computer with 196
MB RAM working under MS Windows NT 4.0 with multi-
with the correct fusion method most of the
processor kernel. information in the input images can be preserved.
In this case the use of maximum wavelet fusion
has shown best results, for this algorithm explicitly
Results accentuates the main structures of each overlaying
image (see Fig. 2). In the other case, where the
Image Types
structures are mainly separated, average wavelet
It was found that the images to be fused can be fusion has again proven most effective, sometimes
categorised coarsly into three system types that will giving similar results to simple averaging, some-
be denoted as types I, II and III. Type I systems consist times signi®cantly outperforming it (see Fig. 3).
of two or more major area phases which are partially The maximum wavelet fusion showed problems
or completely overlapping. Type II systems have one with systems containing very small occlusions,
major area phase, and several isolated minor phases where blocky effects when fusing very small
which may be occlusions [IIa] or border structures features were noticed. This handling of very small
[IIb], sometimes overlapping with the major phase. features seems to be a general problem of the
Finally, type III systems show two or more isolated maximum wavelet algorithm, which may be
minor phases which may be overlapping. Each type of overcome with more suitable wavelets.
276 T. C. Stubbings et al.

Fig. 1. Image (a) shows a recording of mass 28 (Si) and


image (b) shows mass 109 (NbO) of the Nb-sample. This
is an example of a type I image fusion system with two
large, partly overlapping areas. Image (c) shows the
result of a standard averaging fusion method, whereas
image (d) shows the result of maximum wavelet fusion
with symlet 8 wavelet and decomposition level of 2. The
details of both source images are retained much butter
with the maximum wavelet method while keeping the
overall contrast high

Fig. 2. Image (a) from the Voest steel sample shows Fe (mass
56), where Mn (mass 55, image b) is incorporated. This is an
example of a fusion system of type IIa. The maximum
wavelet fusion method highlights the Mn-occlusions in the
fused image (d) much better than the simple average method
(c), where the occlusions disappear from the base structure of
Fe. Enrichments of Mn at the grain boundaries could not be
found in this sample
Fusion of 2-D SIMS Images Using the Wavelet Transform 277

Fig. 3. Image (a) shows mass 12 (C) of the hipped Al-steel


alloy and image (b) shows AlO border structures at the
grain boundaries (mass 43). With simple average fusion
(c) of this type IIb system, the C background structure
nearly completely disappears, whereas the average wave-
let fusion (d) excellently retains the overlaying structure,
as well as the underlying background structure

III. Fusion of type III images generally did not show sometimes several wavelets have to be tested to ®nd
satisfying results in spite of the rather simple task the optimal one for a given problem.
of combining isolated minor phases. The problem All fusion methods and all systems performed best
here is the dif®culty in interpreting the fused with small decomposition levels, i.e. 2 and sometimes
images, where all features seem more or less the 3. Increasing the decomposition level usually
same and mass assignment is nearly impossible. decreases the fusion quality rapidly, due to the fact
Therefore, this case has not been further eval- that after each smoothing step (low-pass ®ltering) the
uated. In order to enable discernment of different output images (the low-frequency channels) become
masses, other fusion approaches have to be more and more alike and some small features
considered, e.g. the use of colour schemes. disappear from the images.

General Results Conclusion


Weighted wavelet averaging can be used to improve Different methods of wavelet based images fusion
perception of weak structures, especially weak border have been tested on three types of 2-D multispectral
structures or occlusions; the problem however is SIMS images. It has been shown that the wavelet
almost always a loss of overall image contrast. transform is a powerful tool for fusion of images, but
The fusion result very much depends on the correct not all methods are equally suitable for all problems.
choice of the wavelet basis function. It has been The wavelet fusion rule, as well as the choice of the
noticed that usually wavelet averaging performs well right wavelet and decomposition level, very much
with daubechies and coi¯et wavelets, and all sorts of in¯uence the result of the fusion process. In the
maximum wavelet fusion rules perform best with optimal case wavelet fusion often signi®cantly out-
symlet and coi¯et wavelets. Additionally, the nature performs classic fusion methods such as simple
of the images to be fused has to be considered and averaging, though one has to keep in mind that the
278 Fusion of 2-D SIMS Images Using the Wavelet Transform

computational cost of wavelet transform is much [4] O. Rockinger, Pixel-Level Fusion of Image Sequences Using
Wavelet Frames. In: K. V. Mardia, C. A. Gill, I. L. Dryden
higher than for simple fusion calculations. (Eds.) Proceedings in Image Fusion and Shape Variability
It was also shown that there is a limit of information Techniques. Leeds University Press, Leeds, UK, 1996, pp.
that can be merged into one image, depending on the 149±154.
[5] A. Toet, Pattern Recognition Letters 1989, 9, 245±253.
type of image system and that there exist fusion [6] I. Koren, A. Laine, F. Taylor, Proceedings 1995 IEEE Inter-
problems which cannot be solved at all with greyscale national Conference on Image Processing. IEEE, Washington
fusion approaches. Although the test images used in D.C., 1995, pp. 232±235.
this study are multispectral SIMS images, the obtained [7] L. J. Chipman, T. M. Orr, Proceedings 1995 IEEE Interna-
tional Conference on Image Processing. Washington D. C.,
results can be applied to similar registered multi- IEEE, 1995, pp. 248±251.
spectral greyscale images of the denoted types, [8] S. Mallat, IEEE Trans. on Pattern and Machine Intelligence
obtained by other analytical techniques. 1989, 11(7), 674±693.
[9] S. Mallat, S. Zhong, Wavelet Transform Maxima and Multi-
scale Edges. In: Coifman et al. (Eds.) Bartlett and Jones, 1990.
Acknowledgements. This work was ®nancially supported by the [10] H. Li, B. S. Manjunath, S. K. Mitra, Graphical Models and
Austrian Fund for Science and Research (FWF Project P13160- Image Processing 1995, 57(3), 235±245.
CHE). [11] T. A. Wilson, S. K. Rogers, L. R. Myers, Optical Engineering
1995, 34(11), 3154±3164.
References [12] J. Le Moigne, R. F. Cromp, The Use of Wavelets for
Remote Sensing Image Registration and Fusion Technical
[1] H. Hutter, C. Brunner, S. G. Nikolov, C. Mittermayr, M. Report TR-96-171, NASA Goddard Space Flight Center,
Grasserbauer, Fresenius J. Anal. Chem. 1996, 355, 585±590. 1996.
[2] H. Hutter, M. Grasserbauer, Mikrochim. Acta 1992, 107, 137 ff. [13] IDL, Research Systems Inc., Boulder, CO 80301, USA.
[3] H. Hutter, C. Brunner, P. Wilhartitz, M. Grasserbauer, [14] The Wavelet Fusion Toolbox for IDL, Dept. of Electrical
Mikrochim. Acta 1996, 122, 195±202. and Electronic Engineering, University of Bristol, UK.

You might also like