standardization of line-scan nir imaging systems
TRANSCRIPT
JOURNAL OF CHEMOMETRICSJ. Chemometrics 2007; 21: 88–95Published online 2 August 2007 in Wiley InterScience
(www.interscience.wiley.com) DOI: 10.1002/cem.1038Standardization of line-scan NIR imaging systems
Zheng Liu1y, Honglu Yu2 and John F. MacGregor1*1Department of Chemical Engineering, McMaster University, Hamilton, Ontario, Canada L8S 4L72ProSensus Inc., 175 Longwood Road South, Hamilton, Ontario, Canada L8P 0A1
Received 15 September 2006; Revised 6 February 2007; Accepted 13 March 2007
*CorrespoEngineeriL8S 4L7.E-mail: my PresentHamilton
A simple and easy to use method is proposed for standardizing NIR imaging systems for differences
among detectors in the charge-coupled device (CCD) array and illumination unevenness. The
standardization equations are then used to pre-treat NIR image data to reduce the systematic errors
introduced by a line-scan NIR imaging system. The method requires only easily available homo-
geneous standards with relatively uniform spectral response. The effectiveness of the standardiz-
ation in reducing the pixel-to-pixel biases and other systematic effects is illustrated with examples,
and the improved sensitivity in results obtained from a multivariate image analysis (MIA) based on
multi-way principal component analysis (MPCA) is demonstrated. Copyright # 2007 John Wiley &
Sons, Ltd.
KEYWORDS: line-scan NIR imaging; hyperspectral images; instrument standardization
1. INTRODUCTION
NIR chemical imaging technology greatly extends the
capability of the conventional probe-based NIR spectroscopy
by adding a completely new dimension, the spatial
dimension. It enables one to obtain not only spectral but
also spatial information characterizing samples with unpre-
cedented ease, speed, and spatial and spectral resolution.
Through the fusion of multivariate calibration methods and
multivariate (or hyperspectral) image analysis, it answers the
questions about the sample such as what chemical species are
in the sample, how much of each is present, and most
importantly, where they are located. Its demonstrated utility
in solving real-world problems has encouraged its rapid
acceptance and been increasingly used in practice, for
example, in the pharmaceutical industry [1–2].
Two types of NIR imaging spectrometers are generally
used in practice according to the application situation. For a
moving sample (e.g., the sample on a moving web or a
conveyor belt), the line-scan NIR imaging spectrometer is
usually used. It builds an image cube (y� x� l dimensions, y
and x are the spatial dimensions and l is the spectral
dimension) by continuously capturing multiple lines along
the y dimension with a spatial–spectral (x� l) intensity
image for each scanned line. For a stationary sample (e.g., a
sample under a microscope), the filter-based imaging
ndence to: J. F. MacGregor, Department of Chemicalng, McMaster University, Hamilton, Ontario, Canada
[email protected]: ProSensus Inc., 175 Longwood Road South,, Ontario, Canada L8P 0A1
spectrometer is employed. It builds the image cubes by
joining the spatial–spatial (x� y) intensity images taken at
different wavelength bands (l). The wavelength bands are
selected by tuning the filter to permit the light from only a
certain wavelength band to enter the camera lens. The
charge-coupled device (CCD) area array, which is the
detector of each spatial–spectral (x� l) intensity image
recorded by the line-scan imaging spectrometer or each
spatial–spatial (x� y) intensity image recorded by the filter
based imaging spectrometer, can be considered as many
thousands of individual infrared sensors. The full spectrum
of a pixel in the image taken by the line-scan imaging
spectrometer is captured by a column of sensors on the CCD
array. On the other hand, the full spectrum of a pixel in the
image taken by the filter-based imaging spectrometer is
captured by a single sensor responding to the reflectance at
different wavelength bands.
NIR spectroscopic instruments, both the probe-based
spectrometer and the imaging spectrometer, are prone to
biases between instruments or among different pixels in the
CCD array. Each pixel in the CCD array in a NIR imaging
spectrometer can be thought of as a different instrument.
Therefore, it is important to standardize the pixels in the
CCD array to achieve consistent spectral responses across
the spatial dimension of the array. Furthermore, for multiple
instruments, differences among instruments introduce the
calibration model transfer problem [3]. Much work has been
published regarding the standardization of probe-based NIR
spectroscopic instruments and the transfer of multivariate
calibration models [3–6]. The general idea of standardization
is to model the instrumental differences. The spectral
response of a subset of samples measured on the primary
Copyright # 2007 John Wiley & Sons, Ltd.
Line-scan NIR imaging systems 89
instrument is regressed against the same subset measured on
the secondary instrument. Thus, changes in the response
variables between the two instruments can be corrected and
the original model can be used for prediction on the
secondary instrument without having to compute new
regression coefficients [4].
Standardization of NIR imaging spectrometers is subject to
all the error contributions of conventional one-dimensional
probe-based spectroscopy (noise, drift, non-linear response
of detectors, wavelength-dependent errors) as well as the
two-dimensional or spatial error components associated
with camera devices and illumination (readout errors,
inconsistent detector responses, quantization errors, and
non-uniform lighting) [7]. This requires that the standard-
ization of the imaging spectrometer must be done for each
spatial or pixel position and also for each wavelength if there
are wavelength-dependent errors.
Standardization of tunable filter-based NIR imaging
spectrometers has received attention recently. Geladi et al.
[8] published one paper addressing the standardization of a
MatrixNIRTM chemical imaging system using a liquid crystal
tunable filter (LCTF) in combination with an InGaAs diode
array detector. The results are based on recalibration against
known reference standards. Standard NIR reflectance
materials, the calibration surfaces made of Spectrolon with
different levels of reflectance (99%, 75%, 50%, and 2%
reflectance) and known ‘true’ reflectance spectra in units of
reflectance percentage, were employed in their method. Each
calibration surface was imaged several times and averaged
so that the influence from the non-uniformity on the surface
was eliminated and thus the noise in the image data is only
from the imaging instrument and illumination. Then a linear
or a quadratic regression model was fitted between the ‘true’
reflectance spectral value of the standard materials at each
wavelength band and the measurement value (in units of
signal intensity counts read out from the A/D converter of
the spectrometer) at specific spatial positions at the same
wavelength in the hyperspectral images. Thus, a linear or
quadratic calibration model cube with the same dimension as
the hyperspectral image was obtained and used to transform
the readout from the spectrometer to the reflectance image in
units of reflectance percentage. It is able to compensate for
both the sensitivity difference of the InGaAs detectors at
different wavelengths and the illumination unevenness and
detector inhomogeneities in the spatial dimensions.
Recently, Burger and Geladi [7] addressed further calibration
issues of this instrument. External standards (i.e., the
standard reflectance materials that are imaged separately
from the images to be corrected) are used to correct
pixel-to-pixel variances due to camera inconsistencies and
variation in sample illumination, and internal standards (i.e.,
a mosaic of different standard reflectance materials imaged
together with the objects of which the images need to be
corrected) are used to compensate for signal drift over time
due to changes in power or temperature effects.
A difficulty with the methods in Geladi et al. [8] and Burger
and Geladi [7] is that standard reflectance materials must be
used. The ideal standard material for their methods is a
material with spatial and spectral uniformity. However,
such standard materials are not easily found for the NIR
Copyright # 2007 John Wiley & Sons, Ltd.
region [7]. The Spectralon materials with different reflectance
values used in their papers are created by adding different
amount of carbon black to a white Teflon-based material and
appear inhomogeneous and textured at high resolution [7].
Although the material was imaged several times and the
images were averaged, it is not easy to guarantee the
uniformity of the material. Maintenance of such material is
another problem because their physical properties may
deteriorate with time due to scratches on the surface and
shape change affecting precision. Any errors in the reference
standards will ultimately compromise the standardization
result.
However, it can be argued that one does not need to
standardize the measurement of the imaging spectrometer to
the ‘true’ reflectance. Calibration does not require that the
reflectance is ‘true’ with respect to the absolute value since it
is not an absolute method. All that is needed is to correct
among pixels for differences in reflectance, and for spatial
variations due to non-uniform lighting, etc.
In this paper, we develop a simple method for standardiz-
ing the line-scan imaging spectrometer to reduce the
systematic errors along the spatial axis x and the spectral
axis l in the image data without using the expensive uniform
reflectance standard materials with known spectra. We
further demonstrate the effectiveness of this standardization
for improving the detection of subtle features in NIR images.
2. LINE-SCAN NIR IMAGING SYSTEMSAND THEIR ERROR SOURCES
The line-scan NIR imaging spectrometer used in this paper is
converted from a monochrome area NIR camera by adding
an ImSpector imaging spectrograph [9] between the front
optics lens and the back InGaAs area CCD array of the
camera. A graphic representation of the imaging system is
shown in Figure 1. For each scanned line across the sample,
the reflected light is vertically dispersed into its continuous
spectral distribution by the ImSpector spectrogragh and is
captured by the area CCD array detector as a spatial–spectral
(x� l) intensity image with the resolution 126� 110 pixels.
The 110 wavelength bands are from 933 to 1663 nm. By
moving the sample at a constant velocity in a perpendicular
direction to the scan, multiple lines are recorded by the CCD
array and a hyperspectral image (y� x� l) of the sample is
obtained.
The spectral data collected from the A/D converter of the
imaging spectrometer represents the signal intensity counts,
not actual reflectance values. The raw spectral data are
mainly influenced by the light intensity of the lamp. As the
lamp is used, the values may decrease due to decreasing light
intensity. On the other hand, the raw data do not reflect the
true intensity of the reflected light because the CCD detector
generates charges even though there is no light exposure on
the detector. These temperature-generated charges cause a
small signal, called dark current [8], typically varying from
pixel to pixel. In precise measurements, this offset must be
measured and deducted from the A/D converter counts.
In practice, the raw spectral data are transformed into
reflectance (or absorbance) units by comparing with
spectra of standard materials. The usual transformation to
J. Chemometrics 2007; 21: 88–95DOI: 10.1002/cem
Figure 1. Schematic picture of the line-scan NIR imaging spectrometer [9].
90 Z. Liu, H. Yu and J. F. MacGregor
reflectance values is obtained by correcting sample spectra
for dark current and dividing by a similarly corrected total
reflectance spectrum. This is also the inherent correction
mechanism integrated into the data acquiring software of
the NIR spectrometer, and completed at the start of an
imaging run. The procedure is described as follows: first, a
spatial–spectral (x� l) image for a scanned line of the dark
current is recorded with the lens cap in place to block light
from entering the spectrometer; second, a spatial–spectral
(x� l) image for a scanned line of a white total reflectance
standard is recorded. An unglazed white ceramic title is used
for this purpose. The sample NIR reflectance image R
captured by the spectrometer is calculated from the system
response by taking, pixel by pixel, the ratio of each sample
corrected signal to the corrected white image signal using the
following equation:
ryxl ¼syxl � dxlwxl � dxl
(1)
where ryxl is an element of the hyperspectral reflectance
image cube R in the unit of reflectance percentage, wxl is an
Figure 2. (a) The monochromatic image of
and thickness at the wavelength around 1200
the image at the locations as marked in su
Copyright # 2007 John Wiley & Sons, Ltd.
element of the raw spatial–spectral image W of a scanned line
of the total reflectance standard (i.e., the white tile) and
dxlis an element of the raw spatial–spectral image D of a
scanned line of the dark current imaged by blocking the lens.
Equation (1) is a linear standardization where the
coefficient 1= wxl � dxlð Þ is found from one standard
reference value only and therefore is often termed one-point
calibration [7]. It compensates for much of the spatial
non-uniformity across the scene line due to both lighting and
detector differences.
Figure 2(a) shows the monospectral NIR image of a red
plastic shim with uniform surface and thickness at the
wavelength band around 1200 nm. The reflectance intensity
is calculated with Equation (1). The contrast of the image is
enhanced by histogram stretching, thereby displaying the
minimum value in the momospectral image with black and
the maximum value with white.
There are evident streak lines along the direction of motion
(y) in Figure 2(a). This is caused by the differences among the
sensors in the CCD array along the spatial axis x. It can also
be seen that there is contrast difference (e.g., shadowy trends)
a red plastic shim with uniform surface
nm. (b) The spectra of the two pixels in
bpart (a).
J. Chemometrics 2007; 21: 88–95DOI: 10.1002/cem
Line-scan NIR imaging systems 91
on the right side across the x–y plane of the image. The
baseline difference between the spectra (Figure 2(b)) of the
two pixels highlighted in Figure 2(a) is caused by the
illumination difference and the CCD array sensor difference
between the two positions. Both the streak lines and the
shadow trend indicate that Equation (1) cannot totally
eliminate the spatial non-uniformity due to the detector
differences and lighting unevenness. Further standardiz-
ation is needed.
Furthermore, for a given pixel position in the x plane, if the
corresponding sensors along the l axis, giving the multi-
band spectrum of this pixel, have different sensitivity to the
light at different wavelength bands, its reflectance spectrum
measured by the spectrometer would be different from its
true reflectance spectrum. Standard illuminating sources
with peaks at precisely known wavelengths are usually used
to correct this error in the spectrum [9].
Figure 3. Average line image of the red plastic shim. This
figure is available in colour online at www.interscience.wiley.
com/journal/cem.
3. METHODOLOGY
The standardization method in this paper is a further
standardization based on the reflectance image calculated
with Equation (1). It will effectively correct the multivariate
NIR images for all of the effects discussed above. The method
is based on the use of standardization samples having
spatially uniform spectral response. Color-coded plastic
shims were used in this work. Each shim has a relatively
uniform spectral response and an even thickness at different
spatial positions. For the NIR image of a plastic shim taken by
the line-scan imaging spectrometer, each scanned line is
measured by the same sensors in the InGaAs CCD array of
the camera. Due to the uniformity of the plastic shim, it is
reasonable to assume that the variation within the scanned
lines (in the y direction) is random noise. Calculating the
average along the dimension of y in an image, we obtain the
average spatial–spectral intensity image of the scanned lines
(called average line image for the purpose of this paper, with
the spatial–spectral resolution of 126� 110 pixels.). Based on
the ‘uniformity’ assumption, the noise from the physical
variation in the plastic shim is almost eliminated by this
averaging and thereby the noise in the average line image
arises only from the difference between the sensors in the
InGaAs CCD array and the unevenness of the illumination.
Then, by calculating the average along the dimension of x in
the average line image, we obtain the average multi-band
spectrum (called average spectrum for the purpose of this
paper, with the spectral resolution 110� 1) of the 126 spectra
in the average line image. This average spectrum is the average
spectrum of all the pixels in the NIR image. The influence of
the variation between the sensors in the CCD array and the
influence of the illumination unevenness along the spatial
axis of x, which are not eliminated by Equation (1), are both
reduced by the second averaging. Thus, the average spectrum
is almost free from the influence of variations in the imaging
system and will be used as the reference spectrum for the 126
spectra in the spatial–spectral average line image. The objective
of our method is to get a correction factor for each element in
the spatial–spectral average line image, which also means
getting a correction factor for each sensor in the InGaAs CCD
Copyright # 2007 John Wiley & Sons, Ltd.
array. Each scanned line forming the hyperspectral image
will be corrected by the factors and thereby each element in
the hyperspectral image cube is standardized.
Clearly one needs standardization samples with relative
spatial uniformity. The more spatial variation one has, the
poorer will be the result. However, because of averaging over
the two spatial axes, the method will be quite insensitive to
small spatial variations. Therefore, any samples with
relatively uniform spectral response, a flat surface, and
uniform thickness can be used to perform the new
standardization method. The spatial uniformity of a sample
can be checked by looking at the consistency of several
spectra measured by a single point NIR spectrometer at
randomly selected locations on the sample.
Four different plastic shims with the colors of coral, pink,
white, and yellow were imaged. Six images were taken at six
different locations on each shim. In total 24 images were
taken. Each image had 200 scanned lines in the y direction.
Therefore, the NIR hyperspectral image had a y� x� ldimension of 200� 126� 110. The images were in the units of
% reflectance ratio to the white reference calculated with
Equation (1). The spatial–spectral average line image and the
average spectrum of each image are calculated. Figure 3
illustrates the average line image from one image of the red
shim using a false color intensity image. Figure 4 shows the
plot of two spectra on the average line image at different
locations, and Figure 5 shows the average spectrum of the
image. For visual enhancement the intensity image has been
color-coded using the color scheme shown in the color bar
toward the right of the image. It is observed that the average
spectrum in Figure 5 is less noisy than the spectra in Figure 4.
The average line images and the average spectra of the 24
images are used to estimate the correction factor for each
sensor in the InGaAs CCD array.
Figure 6(a) shows the relationship between the elements of
the 24 average line images at the spatial–spectral position
x¼ 30 and l¼ 90 and their reference, the average spectral
values at l¼ 90. Figure 6(b) shows the relationship between
the elements of the 24 average line images at the position
x¼ 100 and l¼ 70 and their reference values. Both plots
J. Chemometrics 2007; 21: 88–95DOI: 10.1002/cem
Figure 5. Average spectrum of all the pixels in the NIR image
of the red plastic shim.
Figure 4. Two spectra from the average line image of the red
plastic shim at x¼ 59 (solid line) and x¼ 60 (dotted line).Figure 7. Image presentation of the slope matrix b. This
figure is available in colour online at www.interscience.wiley.
com/journal/cem.
92 Z. Liu, H. Yu and J. F. MacGregor
indicate that there is a linear relationship between them as
denoted by the straight line in each figure. The same
relationship is also observed between the elements of the
average line images at other positions and their reference
Figure 6. The elements in the average line
(a) The elements in the 24 average line ima
l¼ 90 versus their reference, the averag
elements in the 24 average line images at t
versus their reference, the average spectr
Copyright # 2007 John Wiley & Sons, Ltd.
values. That relationship can be expressed as
sl ¼ axl þ bxllxl (2)
Where sl is the average spectral value at the wavelength
band l, lxl is the value of the average line image at the
spatial–spectral coordinate position x and l, axl is the
intercept coefficient, and bxl is the slope coefficient.
The coefficients for all the elements in the average line image
are obtained by fitting a line regression model by least
squares between the average line images and the average spectra
of the 24 images. Figures 7 and 8 illustrate the slope matrix b
and the intercept matrix a by visualizing them as false color
intensity images. It can be observed that the streak lines and
shadow trend also appear in a and b. The streak lines
indicates that the difference among the sensors along the axis
x will be corrected and the shadow trend means that the
baseline offsets in the original image caused by illumination
difference or stray light along the axis x will be compensated.
It is also seen that there are variations (although smaller
variations) among the coefficients between the factors at
different wavelength bands. That means the standardization
method also corrects the sensitivity variations between the
detectors along the spectral axis l.
images versus their reference values.
ges at the spatial location x¼ 30 and
e spectral values at l¼ 90. (b) The
he spatial location x¼ 100 and l¼ 70
al values at l¼ 70.
J. Chemometrics 2007; 21: 88–95DOI: 10.1002/cem
Figure 8. Image presentation of the intercept matrix a. This
figure is available in colour online at www.interscience.wiley.
com/journal/cem.
Line-scan NIR imaging systems 93
The slope matrix b and the intercept matrix a then can be
used to filter each scanned line in the NIR image R to
counteract systematic errors from the imaging system using
the following equation:
ryxl;corr ¼ axl þ bxlryxl (3)
This method is clearly related to the general class of
Multiplicative Scatter Correction (MSC) methods with the
exception that MSC uses the residuals following regression
on the mean, but this method uses the projection onto the
mean. It is an extension of the idea to imaging spectrometers.
4. RESULTS
The standardization method is evaluated in this section
using two examples. The first illustrates the qualitative
(visual) and quantitative (variance reduction) results
obtained by applying the standardization method to the
uniform plastic shim images. The second example illustrates
the greatly increased sensitivity, as a result of standardiz-
ation, obtained upon performing multivariate image analysis
(MIA) on a shim containing subtle surface features.
Figure 9. Correction result. (a) Corrected m
shim at wavelength band 1200nm. (b) Th
marked in Figure 2(a).
Copyright # 2007 John Wiley & Sons, Ltd.
4.1. Evaluation using the uniform shimimagesFigure 9(a) shows the corrected result, on the image in
Figure 2(a), after applying standardization to the NIR
spectrometer. It is observed that both the streaks caused
by the non-uniformity of the CCD array and the shadow
caused by the unevenness of illumination across the spatial
axis x are reduced remarkably. The small thickness variation
in the sample, which is submerged by the systematic noise in
the original images, is also shown more clearly after the
correction. Figure 9(b) shows the corrected results of the
spectra for the same two pixels shown in Figure 2. Compared
with the plots in Figure 2(b), the baseline shift is remarkably
reduced and the two spectra look more consistent with each
other. The two plots in Figure 9(b) also look less noisy than
the plots in Figure 2(b). That means that the noise along the
axis l in the image cube is also reduced by applying the
standardization (3).
Figure 10 quantifies the reduced variation in the spectra at
all the pixel locations achieved by standardization. The
standard deviation of the 126 pixels across the x axis in the
spatial–spectral intensity image for each scanned line of the
red plastic shim is shown for each wavelength band, with
and without applying the standardization. The standard
deviation across the x axis at each wavelength band is
reduced substantially after standardization, as is the
structured behavior of the standard deviation with wave-
length. The same results are also observed for the images of
the other plastic shims which are not shown here. This
illustrates that much of the variation due to pixel differences,
both in the spatial direction and in the wavelength direction
has been eliminated. This also provides a way of checking the
validity of the standardization periodically by scanning a
standard shim and computing the standard deviation plot in
Figure 10 to see that it is consistent with the lower set of
curves.
4.2. Improved detection of subtlespectral featuresThis example is provided to illustrate how the standardiz-
ation can improve the ability of MIA to uncover subtle
onochromatic image of the red plastic
e corrected spectra of the two pixels
J. Chemometrics 2007; 21: 88–95DOI: 10.1002/cem
Figure 10. Standard deviation at each wavelength band of
the 126 pixels across the x axis in the spatial–spectral inten-
sity image for each scanned line of the red plastic shim. Dark
plots: standard deviations before standardization; gray plots:
standard deviations after standardization.
94 Z. Liu, H. Yu and J. F. MacGregor
features in the NIR image. A yellow plastic shim with a
fingerprint superimposed on top of some glue residual left
from a removed piece of adhesive tape was imaged. Since the
spectral channels are highly correlated, MIA using a
multi-way principal component analysis (MPCA) decompo-
sition [10] was used to extract the variations in the
hyperspectral image. Two score images explained 99.99%
of the sum of squares of the spectral intensity variation in the
image. Figure 11(a) shows a false color composite image
obtained by combining the first two score images from the
NIR reflectance image (Equation (1)) without using the
standardization. It can be observed that the streak lines blur
Figure 11. (a) and (b) MIA result based on t
with a finger print and some glue residual a
color score image; (b) t1–t2 scattering plot,
background. (c) and (d) MIA result based on
plastic shim: (c) Combined t1þ t2 false colo
the ellipse highlights the pixels of the backg
online at www.interscience.wiley.com/journa
Copyright # 2007 John Wiley & Sons, Ltd.
the details of the fingerprint and the glue residual and make
them barely visible. This indicates that the spatial variation in
the sensors at different pixel locations is greater than the
signal arising from subtle effects of the fingerprint and the
glue residual. The corresponding t1 –t2 score plot for this
uncorrected image is shown below in Figure 11(b). It is
observed that the pixels of the background scatter over a
wide area in the t1 –t2 score space as highlighted by the ellipse
in Figure 11(b). This implies a sizable variation in the spectral
response of the background pixels of the uniform shim.
Figure 11(c) and (d) shows the result of the same
MPCA-based MIA procedure after pretreating the NIR
image by the standardization model, Equation (3). It is
shown that the fingerprint and the glue residual (vertical
band from the removed tape lying under the finger print) are
more clearly distinguished from the background and from
each other in the composite false color score image. It is also
observed that the background pixels of the uniform plastic
shim now cluster much more tightly in the scattered t1 –t2score plot as highlighted by the smaller ellipse in
Figure 11(d).This tightness of the distribution of the back-
ground pixels in the t1 –t2 score plot implies that, as expected,
the spectral response of the shim is reasonably uniform.
5. CONCLUSIONS AND DISCUSSION
A simple and easily applied standardization method is
proposed for NIR line-scan imaging spectroscopes to correct
for pixel-to-pixel differences in the CCD array and for
systematic biases due to uneven lighting, etc. No special
standards with known reflectance are needed, only homo-
geneous samples with spatially uniform reflectance spectra.
The method is an extension of the similar concept of MSC
methodology to NIR imaging.
he original NIR image of a plastic shim
t the center: (a) Combined t1þ t2 false
the ellipse highlights the pixels of the
the corrected NIR image of the same
r score image; (d) t1–t2 scattering plot,
round. This figure is available in colour
l/cem.
J. Chemometrics 2007; 21: 88–95DOI: 10.1002/cem
Line-scan NIR imaging systems 95
The performance of the method is illustrated using two
examples. From both the visual difference between the
images before and after the correction and the standard
deviation of the spectral responses across a uniform
standard, it can be seen that the instrument standardization
method proposed is an effective way to reduce the systematic
errors arising from base-line offsets and scaling differences
that cannot be eliminated using the standard single reference
reflectance calculation. Both the inconsistencies along the
spatial x axis and the spectral l axis in the imaging system are
corrected. The benefit to the result of subsequent image
analysis is demonstrated with a MIA example. It is shown
that subtle effects in the image become much more clearly
apparent after the application of standardization.
REFERENCES
1. Lewis EN, Schoppelrei J, Lee E. Near-Infrared chemicalimaging and the PAT initiative. Spectroscopy 2004; 19(4):26–36.
Copyright # 2007 John Wiley & Sons, Ltd.
2. Lewis EN, Kidder LH, Lee ENIR. Chemical imaging as aprocess analytical tool. Innov. Pharm. Technol. 2005; 17:107–111.
3. Wang Y, Veltkamp D, Kowalski B. Multivariate instru-ment standardization. Anal. Chem. 1991; 63: 2750–2756.
4. Wang Y, Kowalski B. Improvement of multivariate cali-bration through instrument standardization. Anal. Chem.1992; 64: 562–564.
5. Fearn T. Standardization and calibration transfer for nearinfrared instruments: a review. J. Near Infrared Spectrosc.2001; 9: 229–244.
6. Feudale R, Woody N, Tan H, Myles A, Brown S, Ferre J.Transfer of multivariate calibration models: a review.Chemometr. Intell. Lab. Syst. 2002; 64: 181–192.
7. Burger J, Geladi P. Hyperspectral image regression part I:calibration and correction. J. Chemometr. 2005; 19:355–363.
8. Geladi P, Burger J, Lestander T. Hypespectral Imaging:calibration problems and solutions. Chemometr. Intell.Lab. Syst. 2004; 72: 209–217.
9. ImSpector, ImSpector User Manual http://www.specim.fi/pdf/User_Manual%2BApp_2_21.pdf, 2003.
10. Geladi P, Grahm H. Multivariate Image Analysis. JohnWiley & Sons: Chichester, 1996.
J. Chemometrics 2007; 21: 88–95DOI: 10.1002/cem