Available via license: CC BY 4.0
Content may be subject to copyright.
Remote spectral imaging with simultaneous extraction of 3D topography
for historical wall paintings
Haida Liang
a,
⇑
, Andrei Lucian
a
, Rebecca Lange
a
, Chi Shing Cheung
a
, Bomin Su
b
a
School of Science and Technology, Nottingham Trent University, Nottingham NG11 8NS, UK
b
Conservation Research Institute, Dunhuang Academy, Gansu Province 736200, China
article info
Article history:
Received 22 January 2014
Received in revised form 22 May 2014
Accepted 23 May 2014
Keywords:
Cultural heritage
Archaeology
Hyper spectral
Metrology
Multispectral
Three-dimensional
abstract
PRISMS (Portable Remote Imaging System for Multispectral Scanning) is designed for in situ,
simultaneous high resolution spectral and 3D topographic imaging of wall paintings and other large sur-
faces. In particular, it can image at transverse resolutions of tens of microns remotely from distances of
tens of metres, making high resolution imaging possible from a fixed position on the ground for areas at
heights that is difficult to access. The spectral imaging system is fully automated giving 3D topographic
mapping at millimetre accuracy as a by-product of the image focusing process. PRISMS is the first imag-
ing device capable of both 3D mapping and spectral imaging simultaneously without additional distance
measuring devices. Examples from applications of PRISMS to wall paintings at a UNESCO site in the Gobi
desert are presented to demonstrate the potential of the instrument for large scale 3D spectral imaging,
revealing faded writing and material identification.
Ó2014 International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS). Published by Elsevier
B.V. All rights reserved.
1. Introduction
Spectral imaging (i.e. multispectral and hyperspectral imaging)
was first developed for remote sensing and astronomy (Goetz et al.,
1985). It is an efficient method of collecting spectral reflectance at
millions of points. By the 1990s, multispectral imaging (less than
10 spectral bands) was applied to imaging of old master paintings
in museums and galleries (Derrien et al., 1993; Martinez et al.,
1993). Initially it was used to improve colour accuracy of the
images captured and for qualitative comparison between the
bands. Later, spectral imaging was used to obtain reflectance spec-
tra for pigment identification (Baronti et al., 1998; Casini et al.,
1999; Liang et al., 2005). Other applications of spectral imaging
in art include imaging of underdrawings beneath the paint layers.
Most paints are more transparent in the near infrared and hence
images in the infrared are useful for revealing the preparatory
drawings beneath the paint (van Asperen de Boer, 1968; Liang
et al., 2013). A comparison between images in the visible spectral
range with those in the near infrared can also reveal past interven-
tions and damages to the paintings, since conservators colour
match the paint for retouching to the original without necessarily
using the same paint material. Two materials that are colour
matched do not necessarily have the same appearance in the near
infrared. Comprehensive reviews on spectral imaging applications
in art conservation and archaeology can be found in a number of
reviews (Liang, 2012; Kubik, 2007; Fischer and Kakoulli, 2006).
Spectral imaging systems in museums are usually scanning
devices used in close range (<2 m) from the paintings to obtain
high resolution spectral images. In some cases, fixed mechanical
scanners are built to move, either the imaging device or the paint-
ing, in order to scan an entire painting. The size of a painting that
can be imaged will be limited by the scanner size and by the size of
the studio for the very large movable paintings. Such imaging sys-
tems are not suitable for imaging paintings such as wall paintings
and extremely large easel paintings, since these need to be imaged
in situ. Free standing spectral imaging camera with a normal lens
will not be able to image large paintings or paintings at lofty
heights (e.g. ceiling paintings) at high resolution without scaffold-
ing. For example, a mechanical elevator weighing nearly a tonne
was built to bring a colour camera close to the target in a project
to image stained glass panels in churches (Macdonald, 2006); scaf-
folding towers were built on tracks to use a colour camera to scan
the wall paintings at Mogao caves in Dunhuang (Wallach, 2004).
Apart from the complexity of using heavy and cumbersome
mechanical structures, a major problem with these methods is that
the images cannot be correctly mosaiced together because of par-
allax since each neighbouring image is captured with the camera at
http://dx.doi.org/10.1016/j.isprsjprs.2014.05.011
0924-2716/Ó2014 International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS). Published by Elsevier B.V. All rights reserved.
⇑
Corresponding author. Tel.: +44 1158488056.
E-mail address: Haida.Liang@ntu.ac.uk (H. Liang).
ISPRS Journal of Photogrammetry and Remote Sensing 95 (2014) 13–22
Contents lists available at ScienceDirect
ISPRS Journal of Photogrammetry and Remote Sensing
journal homepage: www.elsevier.com/locate/isprsjprs
a different position. This is usually not a problem for very flat
paintings such as easel paintings, but wall paintings by nature have
significant 3D structure. In this paper, we present an imaging
system that is capable of remote high resolution imaging from a
fixed position on the ground by using a telescope on an alt-az
mount that avoids parallax.
Historical wall paintings are often found in houses, churches,
cathedrals, temples, caves and tombs. They are particularly
vulnerable and difficult to image. Wall paintings can be visually
represented by the spectral reflectance per pixel and the 3D spatial
structure. Spectral imaging provides not only material identifica-
tion but also the colour image under lighting of any emission spec-
tra. Currently, 3D imaging of wall paintings are usually achieved by
laser scanning or photogrammetry. To obtain simultaneous spec-
tral reflectance and 3D topographic measurements, the 3D data
will then have to be spatially registered with spectral imaging data
(Brusco et al., 2006; Barazzetti et al., 2010). An alternative method
is to acquire 3D texture information by using two cameras and in
different wavelengths (Mara et al., 2009). Such methods of acquir-
ing 3D multispectral imaging data needs two separate instruments
for acquisition which increases both instrumental and post-
processing complexity as well as the cost. In addition, none of
the systems has so far been able to achieve remote imaging at
transverse resolution better than 1mm at distances greater than
a few metres.
In this paper, we demonstrate a spectral imaging system that
allows automatic, in situ, remote imaging (distances up to 35 m)
of paintings at high resolution that gives not only spectral informa-
tion per pixel of the paintings, but also 3D position and distance
measurements as a by-product. Applications of the remote imaging
instrument to a UNESCO world heritage site, Mogao caves, along
the Silk Road will be discussed.
2. PRISMS – Portable Remote Imaging System for Multispectral
Scanning
PRISMS is designed for portable, flexible and versatile remote
imaging, consisting of modular components: (1) for imaging at dis-
tances >3–4 m, a telescope (focal length 1250 mm and aperture
diameter 90 mm) is used; (2) for close range imaging at distances
<3–4 m, lenses are used; (3) for imaging in the range 400–900 nm
(VIS/NIR), interference filters are used with CCD detectors (Liang
et al., 2007); and (4) for imaging in the 900–1700 nm range (SWIR),
an imaging AOTF spectrograph and an InGaAs detector are used
(Liang et al., 2010). This paper will concentrate on 3D spectral
imaging with the VIS/NIR multispectral imager, which is a simple,
low budget instrument consisting of a filter wheel with 10 filters
and a CCD camera. The first 9 filters are centred at 400–800 nm
at 50 nm spacing and 40 nm bandwidth. The last filter is at
880 nm with 70 nm bandwidth. The modest spectral resolution is
a trade-off with speed of capture. In general most natural material
such as paint have rather smooth reflectance spectra and hence lit-
tle loss in information by reducing the spectral resolution to
50 nm. Only a handful of pigments such as cobalt blue and some
red lake pigments require spectral resolutions of 15 nm. Fig. 1
shows the system: the camera, filter wheel and telescope are
placed on an alt-az mount that can be computer controlled.
PRISMS is arranged in such a way that the telescope optical axis
passes through the pivot of the two mechanical axes to avoid
parallax.
Active illumination is used in the present application of close
range remote imaging (within 35 m), unlike normal remote sens-
ing where passive illumination from the Sun is used. A tungsten
halogen light with an optical system that projects a flat illumina-
tion of 10°is used for remote illumination. For imaging at distances
>5 m a 900 W tungsten bulb with colour temperature of 3200 K is
used. The light is selected to give maximum illumination for fast
capture without causing damage to the paintings. Temperature
increase due to illumination was measured with a highly absorbing
liquid crystal thermometer (i.e. worst case scenario similar to the
effect on a black paint) at a distance of 10 m and found to be
<2 °C (the accuracy of the thermometer is 2°) above ambient tem-
perature of 22 °C for the light source using the 900 W bulb. The
light is mounted on an identical alt-az mount as the telescope
and the movements of the two mounts are synchronised. The light
and the telescope system are placed as close as possible with the
optical axis of the light at an angle of 5°to the optical axis of
the telescope such that the light will always illuminate the area
imaged at all distances >3 m.
The spectral transmission of the filters and the overall spectral
response of the system using the 900 W light are shown in Fig. 2.
3. System characteristics and remote calibration
Calibration of spectral imaging systems in a fixed laboratory
setting is well established. It generally involves dark current sub-
tract, flatfield calibration to correct for non-uniform illumination
and pixel to pixel gain variation of the CCD sensor, and spectral cal-
ibration with a spectral standard such as a Labsphere Spectralon
white target. In these cases the white standard is usually placed
at exactly the same position and orientation as the imaging target.
Remote imaging presents a number of challenges since it is not
practical to take an image of the white standard at each imaging
position. In the following sections, we will discuss the methods
used for remote calibration as well as the system characteristics.
3.1. Spatial resolution
A He–Ne laser with a 5
l
m diameter fibre output was placed at
7 m from PRISMS as a point source to measure the Point Spread
Function (PSF) of PRISMS in the configuration with the telescope
using the 650 nm filter. The intensity of the laser was adjusted to
a minimum so that the integration time was 1 ms. The FWHM
PSF of the system was found to be 1:7
00
which corresponds to
60
l
m on the target at a distance of 7 m. The PSF thus measured
is unlikely to be affected by lab ‘seeing’ (air turbulence) because of
the short integration time. However, realistic integration times are
likely to be in the range of tens of milliseconds to a few seconds
depending on the distance, the filter and the target. The ‘seeing’
was measured by taking the standard deviation of the peak posi-
tions of the laser taken successively over a period of 25 s which
provides the worst case scenario. The lab ‘seeing’ over a 25 s period
was found to be 1:6
00
FWHM measured at a distance of 4 m. Typical
integration time of PRISMS is of the order of hundreds of millisec-
onds, therefore the effect of seeing is likely to be much less than
1:6
00
. However, these measurements are taken in a temperature
regulated enclosed lab. When long distance measurements
(>10 m) were performed in the corridor outside the lab, seeing
effects were much more noticeable as will be discussed in the next
section.
3.2. Metrology
In order to perform remote calibration and measure 3D surface
topography, we need to measure the distance from the imaging
system to the target position. In this section, we show that the
spectral imaging system can determine the distance without extra
measurements such as laser scanning.
Since it is always necessary to focus the image, the target
distance can be determined by the position of focus. The unique
14 H. Liang et al. /ISPRS Journal of Photogrammetry and Remote Sensing 95 (2014) 13–22
relation between the object distance and the image distance is
given by the lens maker’s formula. In PRISMS, an image quality
metric is used to determine the sharpness of the image and hence
the position of best focus. Focusing is done in the filter that is most
efficient, that is at 650 nm (see Fig. 2b). The same focus position is
adequate for all filters when using the telescope. The following five
image quality metrics were tested:
1. Variance, defined by:
V¼ðMNÞ
1
X
x
X
y
½fðx;yÞ
l
2
;ð1Þ
where fðx;yÞis the grey level intensity of pixel ðx;yÞof a MN
image, and
l
¼1
MNX
x
X
y
fðx;yÞ:ð2Þ
2. Energy of image gradient (EoG).
EoG ¼X
x
X
y
ðf
2
x
þf
2
y
Þ;ð3Þ
where f
x
¼fðxþ1;yÞfðx;yÞand f
y
¼fðx;yþ1Þfðx;yÞ.
3. Spatial frequency (SF).
SF ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
RF
2
þCF
2
qð4Þ
where RF is the row frequency
RF ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1
MNX
M
x¼1
X
N
y¼2
½fðx;yÞfðx;y1Þ
2
rð5Þ
and CF the column frequency
CF ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1
MNX
M
x¼2
X
N
y¼1
½fðx;yÞfðx1;yÞ
2
r:ð6Þ
4. Energy of the Laplacian of the image (EoL).
EoL ¼X
x
X
y
f
2
xy
ð7Þ
where f
xy
is a mask with coefficients
141
4204
141
8
<
:9
=
;
centred on
pixel ðx;yÞ.
5. Sum-modified Laplacian (SML).
SML ¼X
xþN
i¼xN
X
yþN
j¼yN
r
2
ML
fði;jÞfor
r
2
ML
fði;jÞPT:ð8Þ
where
r
2
ML
fðx;yÞ¼j2fðx;yÞfðxstep;yÞfðxþstep;yÞj
þj2fðx;yÞfðx;ystepÞfðx;yþstepÞj:
ð9Þ
and where Tis a threshold value and Nsets the size of the window
used to calculate the metric. The step parameter allows for different
size of texture elements (Nayar and Nakagawa, 1994). In our exper-
iments, we have only used a value of 1 for the step.
Fig. 1. (a) PRISMS: lighting system on an alt-az mount (top) and imaging system (bottom) with telescope, filter wheel and camera on another alt-az mount and (b) PRISMS at
work imaging cave paintings at Mogao caves in the Gobi desert.
400 500 600 700 800 900
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
wavelength (nm)
% transmittance
(a)
400 500 600 700 800 900
0
0.2
0.4
0.6
0.8
1
1.2
1.4
wavelength (nm)
normalised count rate
(b)
Fig. 2. (a) VIS/NIR filter transmission spectra and (b) overall spectral response of the system.
H. Liang et al. / ISPRS Journal of Photogrammetry and Remote Sensing 95 (2014) 13–22 15
At high integration time, all the above metrics produce good
results, however, for imaging efficiency we need the metric that
works at lower integration times. The SML algorithm is generally
found to be the best choice in both accuracy and implementation
time (Huang and Jing, 2007). However, for low exposure times or
low reflectivity targets, the EoG was the best option and it is also
computationally the fastest.
For most surfaces we image a region of interest (ROI)
1
4
the size
of the entire image for focusing. It was found that an integration
time 10% of the maximum exposure time without saturating the
CCD detector was adequate for focusing. A higher integration time
(50% of the maximum exposure time) is needed for darker areas
with reflectance less than a few percent. Fig. 3a shows an example
plot of EoG focus measure metric versus relative distance at around
7 m. The peak position of the Gaussian fit provides the best
estimate of the focus position. The error bars were obtained by
repeating the measurements 10 times.
The focus position can be calibrated in the laboratory against
distance to target by imaging a high contrast calibration chart at
a range of distances. Fig. 3b shows the distance versus focus
position relation for PRISMS where the distance to the target was
measured with a Leica Disto D3 laser ranger (typical accuracy
1.0 mm for distances up to 100 m). The accuracy of this calibration
can be improved using a more precise distance measurement
device. Once calibrated, the relation can be used in the field to
determine distances using just the spectral imaging system.
The uncertainties in distance determination using this method
is shown in Fig. 3c for a low contrast mock wall painting target
(painted in one colour without any features other than the intrinsic
surface texture). The distance errors are smallest at short distances
but increases steadily with distance. The anomalous large scatter
between 11 and 20 m was due to bad ‘seeing’ (air turbulence and
temperature gradient) because the measurements were conducted
along a corridor that was divided into 3 sections and the 11–20 m
range intersects with other corridors leading to increased air tur-
bulence. The distance accuracy is ultimately limited by the sig-
nal-to-noise ratio (SNR) of the images and air turbulence. It is
important to note that the distance accuracy can be much better
than the depth of field, in the same way that position accuracy
can be much better than the imaging resolution if the SNR is high.
There is a trade-off between distance accuracy and speed of
imaging.
Distances to the centre of each field of view (FOV: 0:31
0:23
)
is determined as the spectral images are taken without any
additional measurements. This gives a coarse global shape mea-
surement as a by-product of spectral imaging without reducing
the speed of image scanning. For example, when imaging a mock
wall painting target at a distance of 5 m, we expect a distance
measurement, at millimetre accuracy, every 2 cm across the
wall; the relative position accuracy in the transverse direction cor-
responds to the image pixel size of 20
l
m at this distance. Such
accuracy is comparable to laser scanning (Amann et al., 2001; Fryer
Fig. 3. (a) An example of focus measure using the EoG metric versus relative distance centred around 7 m (the data points are in blue and the red curve is the Gaussian fit);
(b) calibration curve for distance to target versus focus position relation; (c) Distance error (rms) versus distance to target in the worst case scenario of imaging a featureless
mock wall painting tile. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
16 H. Liang et al. /ISPRS Journal of Photogrammetry and Remote Sensing 95 (2014) 13–22
et al., 2005). The advantage of the current method is that the
distance measurements do not require additional instruments
nor extra capture and processing time. There is also no need to
co-register the 3D spatial and spectral data as they are taken with
the same imaging instrument at the same time.
If texture mapping is required at smaller spatial scales, individ-
ual images can be saved as the system is focusing which allows
focus positions for sub-regions of each FOV to be determined
instead of the central ROI. This will slow down the data capture
as extra time is needed for processing each sub-region.
3.3. Radiometry
In conventional imaging in a lab, the radiometric calibration
involves (i) a flatfield calibration to remove the non-uniform illu-
mination and the pixel-to-pixel gain variation of the detector by
imaging a matte white card at the position of the target and (ii)
an absolute reflectance calibration by imaging a white standard
of known reflectance at the same position as the target. In remote
imaging, it is not possible to place a white card or white standard
at each target position. Instead the radiometric calibration will be
carried out in the lab by measuring a Spectralon white standard at
different distances and angles relative to the optical axis of the
imaging system.
The illumination is designed to give a uniform illumination
across the 10°beam width and the light is placed next to the camera
with its optical axis at an angle of 5°to the camera optical axis such
that over the telescope focusing range (d>3 m), the illumination
beam always overlaps the entire FOV of the camera (see Fig. 4a).
The FOV of the camera is illuminated by different parts of the light
beam at different target distances. An inverse square law, relating
the intensity of light reflected back from a white standard and its
distance from the light, is expected if the light beam is uniform in
cross-section. Fig. 4b shows that the measured back reflected inten-
sity is consistent with an inverse square law over the measured dis-
tance range between 4 m and 12 m. The data deviates from the
inverse square law at distances below 4 m. Therefore, one measure-
ment of a white standard at a given distance in the field can be used
to calibrate all measurements at different distances.
The configuration of the lighting and imaging system and the
synchronous motions of the two implies that reflectance is mea-
sured in near retro-reflection geometry with the lighting axis
always at an angle of 5
from the imaging axis.
3.4. Spectrometry
Spectral calibration can be carried out by imaging a Spectralon
white standard at a convenient distance at the start and end of the
imaging run. Spectrum of the light source was found to stay
constant within <1% over 13 days.
Fig. 5 shows the spectra of various colour patches from a mini-
Macbeth chart measured by PRISMS and an Ocean Optics HR2000
fibre optic spectrometer, with both measurements conducted at
the standard 45°/0°configuration for illumination and collection
of light. The measurements agree within the error margins.
4. Remote imaging of cave paintings at a UNESCO site
The Mogao caves near Dunhuang at the edge of the Gobi desert
is a Buddhist temple site with a history that extends over 1000
years from the 4th century to the 15th century. There are 735 caves
(492 with wall paintings) and 45,000 square metres of wall paint-
ings at the site, which is an immense resource for the study of the
history of art and architecture, religion, science and technology,
politics and cultural exchange along the Silk Road. The wall paint-
ings are vulnerable since they are exposed to the elements and
recording as much information as possible of the current state of
the paintings is both a form of digital preservation in case of future
lose of the paintings as well as a means for scientific examination
to inform conservation and art historical studies.
4.1. Large scale 3D spectral imaging
Fig. 1b shows the setup of PRISMS in one of the caves at the
Mogao site.
Fig. 6a shows the ceiling of cave 55 at the Mogao site illumi-
nated by the lighting system of PRISMS. Cave 55 is a Buddhist cave
temple constructed in the 10th–11th century. Fig. 6b shows a high
resolution image captured at the ground level of a patch of the
ceiling 11 m away that gives a detailed image of clusters of green
pigment particles. The colour image was derived from the multi-
spectral images assuming a standard D65 daylight illumination
and the CIE 1931 2°standard observer colour matching functions
(CIE Colorimetry, 2004; McLaren, 1976). The individual colour
images are automatically stitched into a seamless mosaic using a
cross-correlation routine that returns a maximum when the over-
lapped regions between the images are matched (Fig. 6c). The
imaging system is designed such that the images are captured from
a single view point. This has the advantage of avoiding parallax
effects, however, it also means that the imaging geometry is not
constant relative to the target. This is not necessarily a problem,
since the geometries of the measurements are known, the reflec-
tance measurements are accurate with respect to the known
geometries. In particular, it presents the paintings in the colour
as they would be viewed by a person standing at the position of
the camera. Fig. 7a shows the change in back-reflected light
Fig. 4. (a) A schematic diagram of the relative geometry between the light beam, the camera system and the target to determine the relation between the intensity of the light
reflected versus the distance. The point Ais at 2.3 m and the angle between the light and the camera axes is 5°. (b) Intensity of light reflected from a Spectralon white standard
as a function of distance. The straight line corresponds to an inverse square law (I/d
2
).
H. Liang et al. / ISPRS Journal of Photogrammetry and Remote Sensing 95 (2014) 13–22 17
intensity as a function of the angle between the normal of the
target surface and the optical axis of the imaging system for a
Spectralon white standard. In some caves, there are statues in
the middle which can block the view to a wall from the camera.
This will then necessitate imaging from another view point.
Alignment of scans from different view points will need to be
λ λ
λ λ
(a) (b)
(c) (d)
Fig. 5. Macbeth colour patches 8, 10, 14, 15: PRISMS measured spectra compared with spectrometer measured spectra at 0°/45°geometry (solid curve).
Fig. 6. (a) View of the east ceiling of Mogao cave 55 from the position of the imaging system at a distance of 11 m using a normal camera; (b) high resolution colour image of
the ceiling (corresponding to the area in (c) marked with a black square) derived from PRISMS images showing the green pigment particle clusters (image size corresponds to
32.4 cm
2
on the ceiling); (c) an automatically mosaiced image of part of the ceiling (area in (a) marked with a black square). (For interpretation of the references to colour
in this figure legend, the reader is referred to the web version of this article.)
18 H. Liang et al. /ISPRS Journal of Photogrammetry and Remote Sensing 95 (2014) 13–22
implemented in the future by using common areas scanned in both
view points as references.
All the image processing are performed using the VIPS library
(Cupitt and Martinez, 1996). The capturing and mosaicing process
transforms and projects the paintings onto a plane tangential to a
sphere centred at the camera with a radius equal to the distance
between the camera and the part of the painting that is used as
the reference image in the mosaicing process. The mosaiced image
can then be projected back onto the 3D surface of the painting
measured by the imaging process as detailed in Section 3.2 to
produce a virtual 3D reconstruction in colour or any of the spectral
bands, thus representing the 3D topography together with the
multispectral images and the derived colour image. Such a visual-
isation makes it a convenient tool for conservators.
4.2. Revealing detailed drawings and faded writings
The date of construction of Mogao cave 465 is still under debate
amongst art historians. The range of dates vary between the 9th
and the 13th century. Any extra information in terms of drawing
styles or hidden writings would help to inform historical studies.
Fig. 8 shows that the drawings which are not seen in the colour
image are clearly revealed in the near infrared image at 880 nm.
Fig. 9 shows the images of a region on the western parts of the
ceiling obtained with PRISMS from a distance of 17 m. No writing
can be seen in the colour image or the near infrared image
(880 nm). However, a difference image between 880 nm and
550 nm revealed the Sanskrit writing even though no writing can
be seen in either of the two images. This demonstrates the powers
of multispectral imaging in revealing faded writing. As the spectra
of the writing has a greater difference in the two spectral bands
than the background, the difference image enhances the writing
while removing the common background.
4.3. Pigment identification
A mock wall painting tile was made up in a similar manner as
those found on the cave walls. Five most common pigments mixed
in animal glue were painted out on a tile plastered with clay and
straw and prepared with a traditional white ground layer
(Fig. 10). As discussed in Section 4.1, the intensity of the reflected
light depends on the angle of incidence (see Fig. 7a) and the surface
texture of the target. It was found that this angular dependence is
much weaker for the wall painting tile compared with a matte
Spectralon white standard (Fig. 7b). Fig. 10 shows the spectral
reflectance of the 5 types of paint on the wall painting tile mea-
sured at various angles of incidence using the measurements of
the white standard at the same angle as a reference. While the
intensities change as a function of the angle, the spectral shapes
are independent of the angles and consistent with spectrometer
−50 0 50
0
0.2
0.4
0.6
0.8
1
angle (deg)
normalised cts
(
a
)
0 20406080
0.2
0.4
0.6
0.8
1
angle
normalised reflectance
white
azurite
vermilion
malachite
red ochre
(
b
)
Fig. 7. (a) Normalised intensity measured through the 880 nm filter as a function of angle between the normal of the target surface and the optical axis of the camera at a
distance of 7 m for the matte Spectralon white standard; (b) a comparison of the angular dependence of the reflected intensity of light at 880 nm for different colour patches
on a mock wall painting tile (picture shown in Fig. 10a) and the Spectralon white standard. (For interpretation of the references to colour in this figure legend, the reader is
referred to the web version of this article.)
Fig. 8. Mogao cave 465 top of the east side of the south wall (3.5 m above ground, imaged from a distance of 8 m) showing the drawing details in the 880 nm near infrared
channel (b) which were not seen in the colour image (a).
H. Liang et al. / ISPRS Journal of Photogrammetry and Remote Sensing 95 (2014) 13–22 19
Fig. 9. Revealing Sanskrit writing in Mogao cave 465: (a) Image of writing on the western ceiling in colour; (b) image at 550 nm; (c) image at 880 nm; (d) difference image
between 550 nm and 880 nm. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
Fig. 10. (a) Mock wall painting tile with colour patches of the most common paint found in Mogao caves from left to right: azurite, vermilion, malachite, red ochre and red
lead in animal glue. (b–f) PRISMS measured spectra from a distance of 8 m at 0°(black dot), 30°(red dot), 45°(green dot) and 60°(blue cross) compared with spectrometer
measured spectra at 45°/45°(blue curve) for azurite (b), vermilion (c), malachite (d), red ochre (e), red lead (f). (For interpretation of the references to colour in this figure
legend, the reader is referred to the web version of this article.)
20 H. Liang et al. / ISPRS Journal of Photogrammetry and Remote Sensing 95 (2014) 13–22
measured spectra. Therefore pigment identification will not be
affected by the angles of incidence.
Fig. 11 shows some examples of pigment identification in two
of the caves. The red paint is unambiguously identified with red
ochre. In the case of the blue pigment, areas with a similar spectral
shape at a more accessible height were examined with a portable
Niton X-ray Fluorescence (XRF) spectrometer and found to contain
copper which is consistent with the identification of azurite. XRF
gives elemental information but spectral reflectance in the visi-
ble/NIR gives the characteristics of the compound. Similarly, for
the green pigment, XRF found the presence of Cu for an area with
similar spectral reflectance which is also consistent with the
pigment malachite. However, it is known that at Mogao another
copper containing green pigment atacamite is also used. Unfortu-
nately, malachite and atacamite have very similar spectral shape
in the 400–900 nm range. Spectral imaging at longer wavelengths
or Raman spectroscopy could potentially distinguish between the
two.
5. Conclusions
The versatile remote spectral imaging system, PRISMS, is capa-
ble of high resolution imaging at a resolution of 8:2
l
rad (or 1:7
00
)
which translates to 80
l
m at a distance of 10 m. It is the first spec-
tral imaging system to simultaneously measure spectral reflec-
tance and 3D shape of the imaging target without additional
distance measurement devices. This eliminates the complexity
and cost of multiple instruments and the need for extra processing
in order to co-register 3D data and spectral images. The on-line dis-
tance measurement is a by-product of spectral imaging as passive
focusing is used to focus the images. Distance accuracy of a few
mm were achieved on wall painting targets at distances of 10 m.
(a)
400 500 600 700 800 900
0
0.2
0.4
0.6
0.8
1
λ (nm)
λ (nm)
λ (nm)
Reflectance (normalised)
(b)
(c) 400 500 600 700 800 900
0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
Reflectance (normalised)
(d)
(e) 400 500 600 700 800 900
0.2
0.4
0.6
0.8
1
Reflectance (normalised)
(
f
)
Fig. 11. Pigment identification in Mogao caves 55 and 465: (a) and (b) cave 55 red paint identified with red ochre; (c) and (d) cave 465 blue paint identified with azurite; (e)
and (f) cave 465 green paint identified with malachite. The circles are the normalised PRISMS measured reference spectra from the mock wall painting tile and the crosses are
the normalised PRISMS measured spectra from the cave wall paintings. (For interpretation of the references to colour in this figure legend, the reader is referred to the web
version of this article.)
H. Liang et al. / ISPRS Journal of Photogrammetry and Remote Sensing 95 (2014) 13–22 21
On site remote imaging with PRISMS at a UNESCO world
heritage site, Mogao caves, has demonstrated a broad range of
applications of spectral imaging in art conservation, history and
archaeology. It has found faded Sanskrit writings on the ceilings
of cave 465 that were not visible in any of the spectral bands or col-
our images, revealed invisible drawings and identified pigments.
PRISMS records the 3D shape, spectral reflectance and therefore
colour images remotely from one position on the ground making
it convenient for large scale recording of a range of quantitative
data as well as detailed examination of specific parts of a cave
painting for scientific and art historical studies. It would be inter-
esting in the future to compare the topographic measurements
obtained with PRISMS with a conventional laser scan in the same
cave.
Acknowledgements
Funding from the UK Engineering and Physical Science Research
Council (EP/E016227/1), Chinese Ministry of Science and Technol-
ogy 973 Program (2012CB720906) are gratefully acknowledged.
We would like to thank Simon Godber, David Parker and Mike
Newton of Nottingham Trent University for technical assistance,
Zhang Huabin, Mao Jiaming, Chai Bolong and other colleagues from
Dunhuang Academy for assistance with on site data collection, Cui
Qiang and Yu Zongren for the XRF data. Dunhuang Academy is in
charge of the conservation and management of the Mogao caves.
References
Amann, M., Bosch, T., Lescure, M., Myllylä, R., Rioux, M., 2001. Laser ranging: a
critical review of usual techniques for distance measurement. Opt. Eng. 40 (1),
10–19.
Barazzetti, L., Remondino, R., Scaioni, M., Brutto, M., Rizzi, A., Brumana, R., 2010.
Geometric and radiometric analysis of paintings in international archives of
photogrammetry. In: Remote Sensing and Spatial Information Science, Part 5,
Commission V Symposium, vol. XXXVIII, pp. 62–67.
Baronti, S., Casini, A., Lotti, F., Porcinai, S., 1998. Multispectral imaging system for
the mapping of pigments in works of art by use of principal-component
analysis. Appl. Opt. 37, 1299–1309.
Brusco, N., Capeleto, S., Fedel, M., Paviotti, A., Poletto, L., Cortelazzo, G., Tondello, G.,
2006. A system for 3D modeling frescoed historical buildings with multispectral
texture information. Mach. Vis. Appl. 17, 373–393.
Burmester, A., Cupitt, J., Derrien, H., Dessipris, N., Hamber, A., Martinez, K., Müller,
M., Saunders, D., 1993. The examination of paintings by digital image analysis.
In: 3rd International Conference on Non Destructive Testing, Microanalytical
Methods and Environmental Evaluation for Study and Conservation of Works of
Art, Rome 1993, pp. 210–214.
Casini, A., Lotti, F., Picollo, M., Stefani, L., Buzzegoli, E., 1999. Image spectroscopy
mapping technique for noninvasive analysis of paintings. Stud. Conserv. 44, 39–
48.
CIE Colorimetry, 2004. 3rd Ed. CIE Publication 015:2004. Central Bureau of the CIE,
Vienna.
Cupitt, J., Martinez, K., 1996. VIPS: an image processing system for large images.
Proc. SPIE 1663, 19–28.
Fischer, C., Kakoulli, J., 2006. Multispectral and hyperspectral imaging technologies
in conservation: current research and potential applications. Rev. Conserv. 3–16
(2006).
Fryer, J., Chandler, J.H., El-Hakim, S., 2005. Recording and modelling an aboriginal
cave painting: with or without laser scanning? In: El-Hakim, Sabry, Remondino,
Fabio, Gonzo, Lorenzo (Eds.), ISPRS Archives, WG V/4 3D-ARCH 2005: Virtual
Reconstruction and Visualization of Complex Architectures 22–24 August,
2005, Vol. XXXVI-5/W17. Mestre-Venice, Italy.
Goetz, A., Vane, G., Solomon, J., Rock, B., 1985. Imaging spectrometry for earth
remote sensing. Science 228, 1147–1153.
Huang, W., Jing, Z., 2007. Evaluation of focus measures in multi-focus image fusion.
Patt. Recog. Lett. 28, 493–500.
Kubik, M., 2007. Hyperspectral imaging: a new technique for the non-invasive
study of artworks. In: 2007 Physical Techniques in the Study of Art, Archaeology
and Cultural Heritage, vol. II. Amsterdam, pp. 199–255.
Liang, H., 2012. Advances in multispectral and hyperspectral imaging for
archaeology and art conservation. Appl. Phys. A 106, 309–323.
Liang, H., Keita, K., Vajzovic, T., 2007. PRISMS: a portable multispectral imaging
system for remote in situ examination of wall paintings. In: Proceedings of the
SPIE 6618, 661815.
Liang, H., Keita, K., Pannell, C., Ward, J., 2010. A SWIR hyperspectral imaging system
for art history and art conservation. In: IX CONGRESO NACIONAL DEL COLOR.
Alicante, Spain, pp. 189–192.
Liang, H., Lange, R., Peric, B., Spring, M., 2013. Optimum spectral window for
imaging of art with optical coherence tomography. Appl. Phys. B 111 (4), 589–
602.
Liang, H., Saunders, D., Cupitt, J., 2005. A new multispectral imaging system for
examining paintings. J. Imag. Sci. Technol. 49 (6), 551–562.
Macdonald, L., 2006. A robotic system for digital photography. In: Proceedings of
the SPIE 6069, 60690I.
Mara, H., Breuckmann, B., Lang-Auinger, C., 2009. Multi-spectral high-resolution
3D-acquisition for rapid archaeological documentation and analysis. In:
Proceedings of 17th European Signal Processing Conference (EUSIPCO’09).
Glasgow, UK, pp. 1205–1209.
Martinez, K., Cupitt, J., Saunders, D., 1993. High resolution colorimetric imaging of
paintings. In: Proceedings of the SPIE 1901, pp. 25–36.
McLaren, K., 1976. The development of the CIE 1976 (L
⁄
a
⁄
b
⁄
) uniform colour space
and colour-difference formula. J. Soc. Dyers Colour. 92, 338–341.
Nayar, S.K., Nakagawa, Y., 1994. Shape from focus. IEEE Trans. Pattern Anal. Mach.
Intell. 16 (8), 824–831.
van Asperen de Boer, J.R.J., 1968. Infrared reflectography: a method for the
examination of paintings. Appl. Opt. 7, 1711–1714.
Wallach, H., 2004. High resolution photography at the Dunhuang Grottoes. In:
Conservation of Ancient Sites on the Silk Road – Proceedings of the Second
International Conference on the Conservation of Grotto Sites. Dunhuang, China,
pp. 259–261.
22 H. Liang et al. / ISPRS Journal of Photogrammetry and Remote Sensing 95 (2014) 13–22
