Signal Processing - September 2016 - 136

worse as the distance between the light source and scene
narrows. Not only does uneven illumination produce poor
visualizations for relighting, but when these data are used for
photometric stereo, systematic errors introduce curl to the surface normal estimations and make quantitative surface reconstructions difficult. An algorithm to correct for captures that
violate far-light assumptions has recently become available
[25] that creates uniformly lit images (Figure 3) and, moreover, accurate photometric stereo calculations for estimation
of surface normals.

Macro XRF scanning

Some of the most exciting recent developments in cultural
heritage analysis have involved XRF. The method involves
using an X-ray source to ionize core electrons from atoms or
ions. After the generation of inner-shell electron "holes,"
higher energy electrons "fall" into those holes, leading to the
fluorescence of an X-ray. Because electron energy levels are
quantized, the fluoresced X-rays are characteristic of the elements involved. Because inner-shell electrons are involved in
these processes, the technique gives only elemental rather
than chemical information. Therefore, for better and for
worse, the spectra are simplified by their lack of chemical
Macro X-ray methods
information. X-rays of different energies are attenuated differAll macro X-ray methods in cultural heritage stem from
ent amounts when passing through a given material from
Roentgen's discovery of X-rays in the late 19th century.
emitter to detector. As a result, it is possible to make some
X-radiography has been a staple of the field for decades and
statements regarding the depth of materials
is still valuable in its original form: a
contrast image formed from the absorpSome of the most exciting relative to one another in the layers of a
painting, particularly when a model of that
tion of X-rays by high-Z contrast elerecent developments in
layered material can be computer simulatments, such as the lead associated with
cultural heritage analysis
ed [27]. For example, the difference in
the pigment lead white. X-radiographs
have involved XrF.
intensity for an element's spectral response
are routinely used by conservators and
compared to theory can indicate how close
curators to characterize the method and
to the surface of the object that element is, given information
style of painting and can be indicative of the artist's
from the spectrum about which elements might be on top of
thought process when pentimenti are observed. Within the
it. Highly portable, rugged XRF point analyzers have made it
last decade, major advances have been made in interpreting
possible to do qualitative (and under favorable conditions,
X-radiographs by computational methods. Also, there has
semiquantitative) elemental analysis nondestructively on culbeen an increasing use of macro tomographic methods, as
tural heritage objects in a matter of minutes.
well as the development of macro XRF scanning and macro
The true revolution in the field has resulted from taking
XRD scanning that have transformed our views of cultural
XRF scanning methodologies and repurposing them with
heritage objects.
transportable macro XRF scanners [5], [28]. These scanners
acquire a hyperspectral XRF data cube by scanning point by
Computational processing of X-radiographs
point in the x-y plane-each point in the x-y plane contains a
Computational processing of X-radiographs has revolutionfull XRF spectrum. As with single point analysis, depth inforized the area of thread count and thread direction analysis for
mation can often be inferred based on relative X-ray intensipaintings on fabric supports [11] and now the chain lines are
ties. As one might imagine, the amount of data involved in
impressed into the paper by the wire mesh of the molds durthese cubes has demanded computational methods that can
ing fabrication [26]. The development of the method in [11]
handle and mine this wealth of information [29]. Sometimes
hinged on realizing that a Fourier transform to the observed
scanning a painting on a canvas support from behind can
alternating light and dark X-ray contrast patterns of a canvas
provide a better data set, due to different X-ray absorption
could provide both thread count and thread direction data.
characteristics, than scanning a painting from the front. The
Prior to this insight, threads were painstakingly counted by
resulting information about elemental composition can be
hand under magnification, and those counts were limited to
used to infer pigment maps and inferred information about
only a few centimeters of a painting.
relative depth can be used in combination with those maps to
These new computational methods permit global analysis
reconstruct paintings underneath overpaint [28]. For an artist
of the entire work. The overall pattern of threads has been
such as van Gogh, whose work sold so poorly during his lifeshown to be very diagnostic for matching paintings to a single
time that he frequently reused his canvases and was supported
bolt of fabric and is now being used to date paintings. Furby his brother, this XRF scanning technique has opened vast
thermore, primary and secondary cusping in the canvas weave
new areas of research (Figure 4).
(scallop patterns caused by the stretching methods used to
prepare canvases for old master paintings) becomes obvious
after employing the computational algorithm, and not only can
Macro XRD scanning
these patterns be used to match paintings to proximal regions
As with point XRF analysis, point powder XRD has historof a bolt of cloth, their absence can be used to infer that a paintically been invaluable in the characterization of artists' piging has been trimmed. More importantly, this method provides
ments. When performed in situ, the method does not
a way to match paintings at approximately the same period to
require a sample and is considered nondestructive. Because
a single bolt of cloth [11].
the diffraction of X-rays requires a regular repeating array
136

IEEE SIgnal ProcESSIng MagazInE

September 2016

Table of Contents for the Digital Edition of Signal Processing - September 2016