Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

JAAS

View Article Online


PAPER View Journal | View Issue

Image denoising techniques applied to glow


discharge optical emission spectroscopy elemental
Cite this: J. Anal. At. Spectrom., 2014,
29, 315 mapping†
Published on 13 November 2013. Downloaded by Lib4RI on 11/06/2015 10:45:44.

Gerardo Gamez,*a Gaurav Mohantyb and Johann Michlerb

Glow discharge optical emission spectroscopy (GDOES) is becoming a mature technique for depth profiling
analysis. The advantages it affords, including fast, quantitative, and multi-elemental analysis, as well as
allowing very high depth-resolution, have attracted the attention of the thin film community. Recently,
the use of GDOES under pulsed-mode operation and coupled to hyper-spectral imaging techniques has
been proposed to perform surface elemental mapping. Several manuscripts have reported on the
underlying mechanisms in GDOES pertaining to the spatial resolution, while other manuscripts have
reported on elemental mapping applications, for example, regarding separated proteins or thin film
combinatorial libraries. Only a couple of studies have reported image processing techniques applied to
GDOES elemental mapping and none having to do with image denoising purposes. Herein, image
denoising techniques are compared in several scenarios: (a) mapping of homogeneous samples; (b)
mapping of heterogeneous samples in two dimensions; (c) mapping of heterogeneous samples in three
dimensions. Denoising techniques compared include averaging, median filtering, principal component
analysis (PCA), and local pixel grouping-PCA. The peak signal-to-noise ratio is used to show the
efficiency of noise removal, while the full-with-half-maximum of emission from sharp features is used to
Received 24th September 2013
Accepted 13th November 2013
demonstrate the resolution degradation effects of each denoising technique. In general, it is observed
that PCA outperforms other techniques albeit with a higher cost of image processing time. Also, it
DOI: 10.1039/c3ja50312g
becomes evident that having multiple image slices (a 3rd dimension) affords more efficient noise removal
www.rsc.org/jaas while minimizing losses in spatial resolution.

et al.,10 which is several orders of magnitude faster than other


Introduction techniques. Several manuscripts have reported on the under-
GDOES is becoming a mature technique for depth proling lying mechanisms in GDOES pertaining to the spatial resolu-
analysis.1–6 The advantages it affords include fast, quantitative, tion. Webb et al. showed that direct current pulsed-GDOES
and multi-elemental analysis with minimal-to-no sample prep- affords sufficient lateral resolution to perform elemental
aration. Also, GDOES allows very high depth-resolution (several mapping.11 In addition, Gamez et al. showed that this is also the
nm), thus this technique has attracted the attention of the thin case for radio frequency pulsed-GDOES.7 In both studies, the
lm community. effect of the operating conditions pertaining to pressure and
Recently, the use of GDOES under pulsed-mode operation pulsing led the authors to conclude that the lateral resolution in
and coupled to spectral imaging techniques has been proposed GDOES is limited by the residence time of the sputtered atoms
to perform surface elemental mapping.7–12 This approach in the GD lamp as well as their mean free path. Other manu-
becomes interesting once it is realized that elemental maps of scripts have reported on elemental mapping applications.
very large sample surfaces (e.g. 100 cm2) could be obtained in a Gamez et al. reported on the application of RF pulsed-GDOES
few tens of seconds, as proposed by Voronov et al.12 and Gamez for elemental mapping of proteins blotted onto membranes.7
In addition, Gamez et al. showed the advantages, including
orders of magnitude faster analysis times, of using dc pulsed-
a
Department of Chemistry and Biochemistry, Texas Tech University, Lubbock, TX GDOES for measuring the elemental landscape of thin lm
79409, USA. E-mail: gerardo.gamez@ttu.edu; Fax: +1 806 7421289; Tel: +1 806 composition spreads used in combinatorial libraries and high
7423086
throughput screening samples.10 Only a couple of studies have
b
Swiss Federal Laboratories for Materials Science and Technology, Laboratory for
reported image processing techniques applied to GDOES
Mechanics of Materials and Nanostructures, Feuerwerkerstrasse 39, 3602 Thun,
Switzerland elemental mapping. Engelhard et al. reported on the use of an
† Electronic supplementary information (ESI) available. See DOI: algorithm to correct for the geometric distortion on the images
10.1039/c3ja50312g collected with a Czerny–Turner monochromatic imaging

This journal is © The Royal Society of Chemistry 2014 J. Anal. At. Spectrom., 2014, 29, 315–323 | 315
View Article Online

JAAS Paper

spectrometer.13 Gamez et al. used an algorithm to correct for the could be used without having a previous knowledge of sample
smile distortion on the images collected with a spectrograph at homogeneity.
the heart of a push-broom imaging system.8 To the authors’
knowledge, there are no studies investigating the effects of Samples
image denoising techniques for GDOES elemental mapping.
Several samples were used to obtain GDOES elemental mapping
Herein, image denoising techniques are compared in several
images with different characteristics. The rst sample consisted
scenarios: (a) mapping of homogeneous samples; (b) mapping
of a homogeneously deposited Ag thin lm on a Cu substrate
of heterogeneous samples in two dimensions; (c) mapping of
(99.9% Cu). The Ag lm was deposited by sputter coating (Balzers
heterogeneous samples in three dimensions. Denoising tech-
Sputter Coater SCD 050) from an Ag target (99.99% Ag) under Ar
niques compared include mean ltering, median ltering, PCA,
gas. The thickness of the lm was not determined but it was
and local pixel grouping (LPG)-PCA. The purpose is to identify
enough to prevent sputtering through to the Cu substrate during
which technique provides better noise removal while preventing
the measurements (at least 100 collected GDOES images) which
Published on 13 November 2013. Downloaded by Lib4RI on 11/06/2015 10:45:44.

image blurring or maintaining the spatial resolution. The peak


would correspond to several 100 s nm. The Ag emission at
signal-to-noise ratio (PSNR) is used to show the efficiency of
338.3 nm was measured in this sample. The second sample
noise removal, while the full-width-half-maximum (FWHM) of
consisted of the same thin lm but perforated with the tip of a
emission from sharp features is used to demonstrate the blur-
needle such that the Cu substrate was accessible. The Cu emis-
ring effects of the denoising techniques.
sion at 324.7 nm was measured in this part of the sample and was
used for determination of the FWHM of the emission prole. The
third sample consisted of a Cu lm deposited by sputter coating
Experimental and methods (Cu target 99.99% Cu) on a Ni substrate (99.9% Ni). A slant was
GDOES elemental mapping system engineered into the Cu lm thickness, i.e. the thickness was not
homogeneous. The overall lm thickness (10–100 nm) was thin
The GDOES elemental mapping system has been previously
enough to sputter through the Cu lm during the GDOES
described.8 In short, a push-broom hyper-spectral imaging
elemental mapping experiment and a small perforation was
system is used to collect temporally, spectrally, and spatially
added to incorporate a lateral heterogeneity feature. The Ni
resolved images from a glow discharge lamp. The GD lamp is an
emission at 341.5 nm was measured on this sample.
end-on Grimm-type geometry with an inner diameter of 7 mm.
The spectrograph’s (microHR manual, Jobin Yvon Horiba,
France) grating size was reduced from 32 mm  32 mm to 8 mm Image processing
 8 mm with a mask to minimize aberrations and improve Different types of image processing techniques were applied to
spectral and spatial resolution. The spectral window acquired the data. Part of the image processing was performed in MAT-
by the ICCD (PI-Max 3 1024i, Princeton Instruments, USA) was LAB (The MathWorks, Inc.) and another part, including the
approximately 68.5 nm. Part of the system optics are mounted display of 3D images, was performed with ImageJ (National
on a motorized stage such that the collected image can be Institutes of Health, USA). Single image median ltering, or
scanned through the entrance slit without having to move the median ltering in two dimensions, was performed using the
spectrograph or the GD source. An in-house developed Lab- MATLAB function “medlt2”. Median ltering in three dimen-
VIEW program (National Instruments Corporation, Austin, TX, sions was performed by arranging the images of interest into a
USA) was used for the acquisition automation, and display of stack of images and using the 3D median lter tool in ImageJ.
hyper-spectral cube images. The local pixel grouping PCA was performed following the
The measurements were performed under the following protocol from Zhang et al. and using their published MATLAB
operating conditions: direct current pulsed-power mode, 2.5ms code. The only alterations to the code consisted of bypassing the
GD pulse duration, 1 kHz GD pulse frequency, 23 Torr Argon PSNR section.
(99.9999% purity), 500 V GD pulse voltage (measured at In addition to the measured depth prole images, a set of
sample), 2  2 ICCD binning, 25  C ICCD temperature, ICCD model depth prole images was created. For this purpose, the
gain value of 80, 2 ms ICCD gate duration, start of ICCD gate average of 100 images of the Ag 338.3 nm emission of the
synchronized with start of GD pulse, 100 on-camera gate accu- homogeneous sample was used to serve as the mean of a
mulations per ICCD frame, 0.191 mm s1 motor stage speed, 8 Gaussian distribution in depth. The MATLAB function
mm motor stage scan distance. The time required to obtain a “normpdf” was used to create the distribution values. Further-
single image (42 s) corresponds to a pixel size of 10 nm in more, Gaussian noise was added to the images with the MAT-
depth for a Cu substrate. It is important to note that the LAB function “imnoise” such that the resulting noise
acquisition conditions (CCD gain, gate width, accumulations) component was comparable to the one found in measured
of the imaging system will have an impact on the signal-to-noise images.
(SNR) ratio observed. It is always best to use conditions that Full-width half maximum information was obtained by
maximize SNR within the experimental limitations. The image tting a Lorentzian curve to the laterally resolved emission
acquisition conditions implemented here were chosen to allow proles. On the other hand, the depth-resolved emission
the best spatial resolution, including the depth dimension, proles were tted to a Gaussian curve to obtain the FWHM in
under the dynamic range restrictions. Thus, these conditions depth.

316 | J. Anal. At. Spectrom., 2014, 29, 315–323 This journal is © The Royal Society of Chemistry 2014
View Article Online

Paper JAAS

Results and discussion where, in our case, the maximum possible value of the image
(MAXI) is 65 536 for a 16 bit image. The MSE is the mean square
Noise sources in GDOES have been studied before by looking at error when comparing an original image I of m  n pixels to a
the noise power spectrum (NPS) and have been shown to be noisy representation image K. Higher PSNR generally indicates
system specic. Broekaert et al. showed an NPS with white noise better quality. In our case, the “original image” or better said
and some peaks attributed to pump and line sources while the reference image, was obtained by averaging 100 GDOES
other peaks were unidentied.14 On the other hand, Winchester emission maps, at 338.3 nm, of the Ag homogeneous sample.
et al. showed only white noise and dri in the NPS of their Fig. 1 shows a montage of images corresponding to different
GDOES source.15 Here, the rotationally averaged power spectra processing techniques and parameter variations. For example,
(PS) of homogeneous sample images were determined (see ESI, Fig. 1a shows a single raw image and it is evident that the noise
Fig. 1S†) to detect the possible contribution from different types in the image is higher compared to the rest of the images, for
of noise. In short, the PS is the square of the Fourier transform example Fig. 1f which shows the average 100 reference image.
Published on 13 November 2013. Downloaded by Lib4RI on 11/06/2015 10:45:44.

of the images but in this case it is rotationally averaged to show However, it is much more difficult to discern the differences in
the more familiar magnitude vs. frequency plot. The PS of all the noise content between the other images. Here is where the
images show a 1/fn dependency at lower frequencies. This is due PSNR values become very useful and also allows for a quanti-
to the contribution of the signal which is convoluted with the tative measure. One thing to note is that the images have
noise. The PS for the single image of Ag sample levels off at horizontal lines which seem to have higher intensity than its
higher frequencies as opposed to the PS obtained from aver- neighboring lines. This is a direct result of having different
aging 100 single images. This comes from the contribution of responses from different pixels in the camera. In the push-
white noise which would give a at line in the PS if it was the broom imaging approach the image is scanned over the same
only contribution. It is instructive to show how certain things column of pixels on the camera until there are enough row
can affect the PS. For example, the PS of an image of the elements to give a 2D monochromatic image. Thus, any
emission of a Ni sample previously taken using our system with response variation of the camera pixels within such column will
a faulty gas bottle regulator shows a peak at a spatial frequency manifest as differences between rows. This can be corrected by
of 10. This comes from the intensity modulation in the corre- implementing a at eld correction.8 Otherwise, the pixel-to-
sponding image that arises from pressure changes in the GD pixel (or in our monochromatic image case row-to-row) varia-
lamp due to the faulty regulator. The images in this study do not tion can be taken into account when a quantitative calibration
have such features and the only clearly distinguishable noise for elemental composition is performed.10 In this study, the raw
has a white character. images were used such that the row-to-row variation is evident.
The purpose of comparing the image denoising techniques Fig. 2 shows the PSNR values when the images obtained by
is to identify the most efficient denoising procedure for three different processing techniques are compared to the average
different cases: (a) images from samples with homogeneous 100 reference image. It is clear that the worst, or lowest, PSNR
content, (b) images from samples with lateral heterogeneity,
and (b) images from samples with lateral and depth
heterogeneity.

(a) Images from samples with homogeneous content


The rst case (a) pertains to the denoising of images such as
those coming from standard reference materials. Such images
would be used for calibration purposes in quantitative GDOES
elemental mapping.10 The noise content would affect the stan-
dard deviation around the regression and image denoising
techniques would potentially improve quantitative results. The
PSNR parameter is used in this study to measure how effective is
the noise reduction. PSNR is typically used in image compres-
sion schemes to measure the quality of image reconstruc-
tion.16,17 It compares the maximum possible signal power
against the power of the noise and it is expressed in decibels to
accommodate large dynamic ranges:
 
MAXI 2
PSNR ¼ 10 log10
MSE

Fig. 1 GDOES emission maps at 338.3 nm of homogeneous Ag film.


1X
m1 Xn1
 2
(A) single raw image, (B) single image 2D median filter 5  5, (C) 3D
MSE ¼ Iði; jÞ  Kði; jÞ
mn i¼0 j¼0 median filter 5  5  5, (D) single image LPG-PCA, (E) image recon-
structed with PC1 & 2 after PCA analysis, and (F) average of 100 images.

This journal is © The Royal Society of Chemistry 2014 J. Anal. At. Spectrom., 2014, 29, 315–323 | 317
View Article Online

JAAS Paper

more near neighbor pixels in the protocol. It is also evident that


increasing the voxel “radius” by one, 3DMF5, results in a greater
improvement in PSNR compared to increasing the pixel square
“radius” by one in the 2DMF approach.
An interesting approach to image denoising is the applica-
tion of PCA. PCA is a multivariate linear data dimensionality
reduction technique that uses eigen vector decomposition to
capture the maximum variance in the data along a set of
orthogonal axes, termed as principal components. A detailed
description of the underlying mathematics can be found
here.18,19 PCA has been previously used to lower the complexity
of large dimensional data and reveal hidden trends, even in case
Published on 13 November 2013. Downloaded by Lib4RI on 11/06/2015 10:45:44.

of spectral datasets.20 The dataset can be denoised aer trans-


formation into the PCA domain, preserving only the most
signicant principal components21 and then retransforming the
Fig. 2 Peak signal-to-noise ratio (PSNR), for GDOES emission maps at data back into the original form. The main idea of performing
338.3 nm of homogeneous Ag film, resulting from various denoising PCA on the GDOES datasets is to capture the correlations
techniques. The average of 100 images was used as the reference related to intensity variation of individual pixels in a single
image.
image as well as along the image slices. Capturing both these
types of correlations simultaneously is possible due to covariance
statistics inherent in PCA. It is assumed that the evolution of
value is displayed by a single raw image. If we were to have only a intensity of the pixels (or image as a whole) with depth is
single raw image there are several options to improve the PSNR. systematic for all the samples used in this study. This is
One such option is to perform 2D median ltering. In this case reasonable as the evolution of pixel intensity will be gradual in
we dene an area of pixels (3  3, 5  5, etc.) and nd the median case of sample with both lateral and depth heterogeneity and
value which is used to replace the center pixel. The advantage of constant in case of sample with homogeneous content and
median ltering is that it can remove outliers efficiently so it is sample with lateral heterogeneity. Therefore, the rst couple of
oen used to remove “salt & pepper” noise or hot pixel contri- PCs will capture the bulk of the variance in the dataset. These PCs
butions. Also, it is a simple and computational-cost effective will account for all the systematic changes in the evolution of
protocol. It is evident from Fig. 2 that the PSNR values improve a intensity in the depth slices and this information can be termed
lot with 2D median ltering. Also, it can be observed that as signal. The rest of the PCs will capture very less variance in the
increasing the size of the pixel area improves the PSNR values, dataset. The variance captured by these insignicant PCs will
albeit with diminishing returns of improved PSNR. represent random variations in the data set and this can be
More ltering options are accessible when more than one termed as noise. By removing the insignicant PCs, the signal-to-
image is available. For example, one can perform averaging of noise ratio can be improved for the GDOES datasets.
the collected images when the subject can be safely assumed to In this case, 20 GDOES images of the homogeneous sample
be the same. In this case, the Ag content of the layer is homo- were collected and processed with the PCA protocol mentioned
geneous in x, y, and z dimensions so it is a reasonable in the ESI. Fig. 1e shows the reconstructed image when select-
assumption, thus the pixels in the z dimension are averaged to ing only the rst two components. In this case, the rst
denoise the image. It is clear that the PSNR improves greatly component (PC1) captures 99.71% of the variance and the
when 3 raw images are averaged. Also, having more images to second component (PC2) captures 0.017% of the variance. The
average gives a much better PSNR value which is improved by reconstructed images taking into account only PC1, PC2, PC3,
utilizing 5 or 10 images to average. This result is expected since or PC4 (see ESI, Fig. 2S†) show that PC3 and PC4 give mainly
it is well known that increasing the number of measurements noise. This is also conrmed by the loading plots (ESI, Fig. 3S†).
improves the precision of a measurement. The advantage of Thus, in this case, PC1 and PC2 are enough to capture the
averaging is that it is a simple and computational-cost effective systematic variation and the rest of the PCs can be discarded
protocol. This means that it is fast to implement. because they mainly capture random variation or noise. It can
Having more than one image also allows performing median be observed from Fig. 2 that the increase in PSNR with a PC1 & 2
ltering in three dimensions. The protocol is very similar to the reconstructed image is the highest compared to the other image
2D median ltering, except that now there is a volume of pixels, processing.
or voxel, instead of an area of pixels. In here, the volume was Recently, PCA has been applied to single images for denoising
limited to a symmetrical shape which means that it contains the purposes in a technique called local pixel grouping PCA.22 In this
same number of pixels in x, y, and z. For example, the 3DMF3, case, a pixel of interest and its neighbors are represented by a
as labeled in Fig. 2, includes 3  3  3 pixels and the median vector variable. Then, pixels with similar local structure to the
from those 27 pixels replaces the center pixel. It is clear from one in question are grouped to serve as training sets for the
Fig. 2 that the PSNR is greatly improved by use of the 3DMF3, variable. This is achieved by a method called block matching,
even more so than using a 2DMF3 which is a result from using typically used to track motion in video.23 In general, block

318 | J. Anal. At. Spectrom., 2014, 29, 315–323 This journal is © The Royal Society of Chemistry 2014
View Article Online

Paper JAAS

matching relies on computing the sum of the squared differences curve was tted, as proposed by Webb et al.,11 to the data and
(SSD) between the original block, containing the pixel of interest, the full width at half maximum (FWHM) was obtained and used
and the comparison block. The blocks with the least SSD are the to measure lateral resolution degradation (Fig. 4d).
matching training sets. In this fashion, the following PCA It is evident that the lateral resolution degrades (increased
denoising can better preserve the edge features. For further FWHM) particularly in the denoising schemes that incorporate
details on the LPG-PCA protocol please refer to Zhang et al.22 It is some sort of lateral grouping of pixels. The largest value of
evident from Fig. 2 that the LPG-PCA protocol provides very good FWHM resulted from the LPG-PCA protocol and the loss of
PSNR over the raw image and most other processing methods in lateral resolution is evident by visual comparison of the GDOES
the comparison. Also, looking back at Fig. 1g (LPG-PCA) it can be emission image (Fig. 3d) to the rest of the images (Fig. 3). In the
observed that the row-to-row variation is somewhat smeared case of the 2D median lter, the 3  3 grouping showed sub-
which suggests a loss in lateral resolution. pixel increase in FWHM. However, increasing the pixel
grouping to 5  5, 7  7, and 9  9, resulted in an increased
Published on 13 November 2013. Downloaded by Lib4RI on 11/06/2015 10:45:44.

FWHM by 1 pixel, 2.8 pixels, and 6.8 pixels, which corre-


(b) Images from samples with lateral heterogeneity sponds to a relative increase of 8%, 22%, and 53%, respectively.
The second case (b) pertains to denoising images from samples It is clear that the 7  7 and 9  9 2DMF result in an unac-
with lateral heterogeneity. GDOES elemental mapping relies on ceptable loss in lateral resolution. The loss in resolution caused
distinguishing the lateral composition heterogeneity and the by the 5  5 2DMF is signicant but its use can still be
noise components can be limiting. Denoising techniques would considered in a case-by-case basis. The 3  3 2DMF yields
potentially allow for an improvement of image quality. The higher delity images but one has to also consider that the
previous section showcased which protocols give the best PSNR improvement in PSNR was not as high as other methods.
improvements. However, this section will allow us to determine Having more than one image allows the implementation of
if the denoising protocols of interest come at the loss of lateral the other techniques cited above. The simplest averaging in the
resolution. For this purpose, an Ag thin lm sample on a Cu z-dimension produces large increases in PSNR, as shown above,
substrate was perforated to reveal the substrate and have a while maintaining the lateral resolution (Fig. 4d). It should be
laterally heterogeneous sample. Ten GDOES images of such kept in mind that the lateral heterogeneity of the sample should
sample were acquired and the denoising techniques which gave not change between the averaged images.
the best PSNR improvements were applied. Fig. 3 shows a The 3DMF protocol also yields higher delity. When the
montage of selected images from this process. To help quantify voxel is 3  3  3 the increase in FWHM is 0.1 pixel, and for
the loss of lateral resolution a prole of the Cu 324.7 nm the 5  5  5 case it is 0.3 pixels. This is a minimal loss in
emission (row of the image) going through the bright spot was resolution which comes with a great increase in PSNR, as seen
extracted from each image (Fig. 4c). In addition, a Lorentzian in the section above. Again, the great advantage of the 3D
median lter protocol is that it can be applied with low
computational cost.
The PCA was again amongst the best performers. Principal
component 1 captured 98.38% of the variance, while PC2
captured 0.1880%. The loadings plots (see ESI, Fig. 4S†) show
that most of the variance has a random character for PC3 and
PC4. This was observed for the images reconstructed from the
individual principal components up to PC4 (see ESI, Fig. 5S†).
Thus, PC1 and PC2 were both used to reconstruct the image
seen in Fig. 3f. It even showed an observable improvement in
spatial resolution, or a decrease in FWHM. This is evident from
comparison of the raw data prole to the prole from the
reconstructed image aer PCA (Fig. 4c). This is a noteworthy
aspect because spatial resolution refers to the ability to distin-
guish between structures at a given level of contrast. In this
fashion, the signal-to-noise ratio can limit the spatial resolu-
tion. Thus, depending on how poor is the SNR, denoised images
can result in an improved spatial resolution provided no blur-
ring is introduced by the denoising technique, which is
demonstrated by the PCA results.

Fig. 3 GDOES emission maps at 324.7 nm of a perforated Ag film on a (c) Images from samples with lateral and depth
Cu substrate. (A) single raw image, (B) single image 2D median filter 7  heterogeneity
7, (C) 3D median filter 5  5  5, (D) single image LPG-PCA, (E) image
reconstructed with PC1 & 2 after PCA analysis, and (F) average of 10 The third case (c) pertains to denoising samples with three-
images. dimensional heterogeneity. In this case, noise in the depth

This journal is © The Royal Society of Chemistry 2014 J. Anal. At. Spectrom., 2014, 29, 315–323 | 319
View Article Online

JAAS Paper
Published on 13 November 2013. Downloaded by Lib4RI on 11/06/2015 10:45:44.

Fig. 4 (A) Topographical profile of perforated Ag film on Cu substrate. (B) Topographical map of perforated Ag film on Cu substrate. (C)
Horizontal emission profiles at 324.7 nm through center of bright spot. (D) FWHM of Lorentzian fits through the emission profiles through center
of bright spot.

dimension, or the random variability from image to image, similar blue color at the surface because the emission is
would compromise the GDOES depth resolution. Thus, dominated by lower values. This can be compared to the blue
denoising techniques would potentially help to improve the pixels sparingly distributed outside of the GDOES emission
ability to distinguish between neighbor layers in depth. volume as a result of the random noise. Fig. 5b shows a vertical
However, as seen above, some denoising protocols come at a slice through the GDOES 3D qualitative elemental map. The
premium of resolution. The effect of denoising on depth reso- perforation in the Cu lm can be clearly seen toward the top-
lution was tested in this part of the study by performing a 3D center part of the slice. It is also very clear that the Ni emis-
qualitative GDOES elemental mapping, at 341.5 nm, of a Ni sion in the le part of the slice begins closer to the surface
substrate coated with a thin Cu lm which was in turn perfo- compared to the right part of the slice, which is a result of
rated to include a lateral heterogeneity feature. having a Cu thin lm which varies in thickness across the
Fig. 5a shows a 3D GDOES elemental map where 100 raw 2D mapped volume.
GDOES images obtained at sequential sputtering times are A pixel at (295 265), in the area where the Ni emission begins
stacked in the z dimension (the top of the stack represents the to increase closer to the surface, was chosen to obtain a depth
surface of the sample or time zero). The pixels with intensities prole for comparing the raw data to the denoised data (Fig. 6).
lower than a threshold given by the measured background are Only the rst 25 images of the 3D emission map were used for
transparent. In addition, the 341.5 nm emission captured by the this purpose to include the region where the emission changes
rest of the pixels is plotted in different colors according to the more abruptly in the depth dimension and to keep the PCA
intensity (dark blue ¼ lowest, yellow ¼ highest). Imaging of a processing time to a minimum. In this case, only the three
sample with no lateral or depth heterogeneity would give a protocols that showed the best denoising while maintaining
cylindrical colored volume in the region where the GDOES lateral resolution were used: z-moving AVG, 3DMF, and PCA. It
emission is captured. In our case, the colored volume is not is evident from Fig. 6 that all three denoising processes yield
symmetric at rst because of the lateral and depth heteroge- improved data quality with respect to noise. However, it seems
neity, but eventually it becomes cylindrical. The volume is a that the change in emission as a function of depth is not sharp

320 | J. Anal. At. Spectrom., 2014, 29, 315–323 This journal is © The Royal Society of Chemistry 2014
View Article Online

Paper JAAS

enough to assess the effects on depth resolution. Thus, a model


3D GDOES depth prole was created to represent the resulting
3D elemental map of a sample with a very thin marker layer.
This was achieved by using the background subtracted average
of the 100 images taken from the homogeneous Ag sample in
Section a (Fig. 1f) and using it as the mean of a Gaussian shape
depth prole. A Gaussian distribution has been previously
observed in GDOES depth proles.24 Furthermore, noise was
added to the model 3D elemental map, with a Gaussian distri-
bution character, to mimic the noise observed in real images.
Such sort of noise would be representative of for example, read-
out noise from the CCD camera.25 Fig. 7 shows the corre-
Published on 13 November 2013. Downloaded by Lib4RI on 11/06/2015 10:45:44.

sponding volume plot (a) and slice (b). A depth prole through
pixel (185 277), in the center of the GDOES emission map
(Fig. 8), reveals that the averaging approach comes with an
unacceptable premium of depth resolution. The 3DMF
approach gives much better results, with no observable depth
resolution degradation. The 5  5  5 voxel shows slightly
better performance than the 3  3  3 voxel evidenced by the
higher peak value. Finally, the PCA yields the best results again

Fig. 5 (A) Qualitative 3D GDOES emission map at 341.5 nm from


heterogeneous Cu film on Ni substrate. (B) Plane slice across quali-
tative 3D GDOES map in (A).

Fig. 6 Comparison of qualitative depth profile (emission at 341.5 nm Fig. 7 (A) Model qualitative 3D GDOES emission map with a Gaussian
extracted from pixel (295 265), for the first 25 sequential images) intensity distribution in depth and added Gaussian noise (mean values
before and after different denoising techniques. Raw data (solid line), taken from background subtracted Fig. 1I). (B) Plane slice across model
PCA (dotted line), 3DMF5 (dashed line), AVG10 (squares). qualitative 3D GDOES map in (A).

This journal is © The Royal Society of Chemistry 2014 J. Anal. At. Spectrom., 2014, 29, 315–323 | 321
View Article Online

JAAS Paper

minimum and there are enough images to obtain the desired


gain in PSNR. On the other hand, PCA is recommended when
computational costs are of minimal concern because it will
yield the best results. In addition, in the present study only
monochromatic maps were considered but a hyper-spectral
data cube was in fact collected for a wide wavelength range
and PCA presents the opportunity of discovering patterns
hidden by the data complexity. This is an avenue which the
authors are currently pursuing.

References
Published on 13 November 2013. Downloaded by Lib4RI on 11/06/2015 10:45:44.

1 J. Angeli, A. Bengtson, A. Bogaerts, V. Hoffmann,


V. Hodoroaba and E. Steers, J. Anal. At. Spectrom., 2003, 18,
670–679, DOI: 10.1039/b301293j.
Fig. 8 Comparison of qualitative depth profile of model data in Fig. 7 2 P. Belenguer, M. Ganciu, P. Guillot and T. Nelis, Spectrochim.
before and after different denoising techniques. Model data (solid line), Acta, Part B, 2009, 64, 623–641, DOI: 10.1016/
PCA (dotted line), 3DMF5 (dashed line), AVG10 (squares). j.sab.2009.05.031.
3 B. Fernandez, R. Pereiro and A. Sanz-Medel, Anal. Chim. Acta,
2010, 679, 7–16, DOI: 10.1016/j.aca.2010.08.031.
through not only a higher peak value but also an observable 4 R. Marcus, A. Anfone, W. Luesaiwong, T. Hill, D. Perahia and
narrower peak which results in discernible spatial resolution K. Shimizu, Anal. Bioanal. Chem., 2002, 373, 656–663, DOI:
(FWHM) improvement. 10.1007/s00216-02-1378-8.
5 M. Wilke, G. Teichert, R. Gemma, A. Pundt, R. Kirchheim,
Conclusions H. Romanus and P. Schaaf, Thin Solid Films, 2011, 520,
1660–1667, DOI: 10.1016/j.tsf.2011.07.058.
It is clear that for homogeneous samples, such as standard 6 M. Winchester and R. Payling, Spectrochim. Acta, Part B,
reference materials (SRMs), 2DMF ltering gives desired 2004, 59, 607–666, DOI: 10.1016/j.sab.2004.02.013.
improvement in PSNR, without the computational costs of LPG- 7 G. Gamez, S. Ray, F. Andrade, M. Webb and G. Hieje, Anal.
PCA, when a single image is available. However, if multiple Chem., 2007, 79, 1317–1326, DOI: 10.1021/ac061361l|
images are available, PCA gives better PSNR improvements, but 10.1021/ac061361i.
average or median ltering techniques require minimal 8 G. Gamez, D. Frey and J. Michler, J. Anal. At. Spectrom., 2012,
computational cost and can provide very good denoising 27, 50–55, DOI: 10.1039/c1ja10241a.
provided with enough images. 9 G. Gamez, M. Voronov, S. Ray, V. Hoffmann, G. Hieje and
In the case of laterally heterogeneous samples, if there is J. Michler, Spectrochim. Acta, Part B, 2012, 70, 1–9, DOI:
only a single image, one can turn to 2D pixel grouping tech- 10.1016/j.sab.2012.04.007.
niques but at a cost of lateral resolution in such a way that 10 G. Gamez, G. Mohanty and J. Michler, J. Anal. At. Spectrom.,
further gains in PSNR would result in greater loss of lateral 2013, 28, 1016–1023, DOI: 10.1039/c3ja50060h.
resolution. On the other hand, if multiple images are available, 11 M. Webb, V. Hoffmann and G. Hieje, Spectrochim. Acta,
PCA and depth averaging do not result in a noticeable loss of Part B, 2006, 61, 1279–1284, DOI: 10.1016/j.sab.2006.
lateral resolution, while 3DMF only gives sub-pixel lateral 10.018.
resolution degradation, as measured by tted Lorentzian 12 M. Voronov, V. Hoffmann, T. Wallendorf, S. Marke,
FWHM. Again, provided with enough images the depth aver- J. Monch, C. Engelhard, W. Buscher, S. Ray and G. Hieje,
aging should give sufficient gains in PSNR at a minimal J. Anal. At. Spectrom., 2012, 27, 419–425, DOI: 10.1039/
computational cost. c2ja10325g.
When the sample is heterogeneous in three dimensions one 13 C. Engelhard, S. Ray, W. Buscher, V. Hoffmann and
will have a multiple image case but with systematic changes G. Hieje, J. Anal. At. Spectrom., 2010, 25, 1874–1881, DOI:
from image to image. It is clear that when the systematic 10.1039/c0ja00068j.
changes are very sharp the depth averaging degrades the depth 14 J. Broekaert, K. Brushwyler, C. Monnig and G. Hieje,
resolution. Once again, PCA shows the best results but with a Spectrochim. Acta, Part B, 1990, 45, 769–778, DOI: 10.1016/
premium of computational cost. In this case, 3DMF can prove to 0584-8547(90)80056-o.
be very useful because it can give signicant PSNR gains with 15 M. Winchester, J. Travis and M. Salit, Spectrochim. Acta,
minimal-to-no degradation in spatial resolution while being Part B, 1993, 48, 1325–1337, DOI: 10.1016/0584-8547(93)
highly computational-cost efficient. 80121-a.
Overall, averaging in the z dimension or performing 3D 16 J. Korhonen and J. You, Quality of Multimedia Experience
median ltering (depending on the expected heterogeneity) is (QoMEX), 2012 Fourth International Workshop on, Yarra
recommended when computational costs are to be kept at a Valley, VIC.

322 | J. Anal. At. Spectrom., 2014, 29, 315–323 This journal is © The Royal Society of Chemistry 2014
View Article Online

Paper JAAS

17 Z. Wang and A. Bovik, IEEE Signal Process. Mag., 2009, 26, 98– International Journal, 2012, 3, 236–244, DOI: 10.5121/
117, DOI: 10.1109/msp.2008.930649. sipij.2012.3218.
18 S. R. Broderick, H. Aourag and K. Rajan, J. Am. Ceram. Soc., 22 L. Zhang, W. Dong, D. Zhang and G. Shi, Pattern. Recogn.,
2011, 94, 2974–2980. 2010, 43, 1531–1549, DOI: 10.1016/j.patcog.2009.09.
19 J. Shlens, Systems Neurobiology Laboratory, University of 023.
California at San Diego, 2005, http://www.snl.salk.edu/ 23 M. Jakubowski and G. Pastuszak, Opto-Electron. Rev., 2013,
shlens/pca.pdf. 21, 86–102, DOI: 10.2478/s11772-013-0071-0.
20 S. R. Broderick, C. Suh, J. Provine, C. S. Roper, 24 R. Galindo and J. Albella, Spectrochim. Acta, Part B, 2008, 63,
R. Maboudian, R. T. Howe and K. Rajan, Surf. Interface 422–430, DOI: 10.1016/j.sab.2007.12.006.
Anal., 2012, 44, 365–371. 25 D. Snyder, C. Helstrom, A. Lanterman, M. Faisal and
21 Y. Murali Mohan Babu, M. V. Subramanyam and R. White, J. Opt. Soc. Am. A, 1995, 12, 272–283, DOI:
M. N. Giri Prasad, Signal & Image Processing: An 10.1364/josaa.12.000272.
Published on 13 November 2013. Downloaded by Lib4RI on 11/06/2015 10:45:44.

This journal is © The Royal Society of Chemistry 2014 J. Anal. At. Spectrom., 2014, 29, 315–323 | 323

You might also like