3D shape measurement of objects with high

dynamic range of surface reflectivity

Gui-hua Liu,1,2,* Xian-Yong Liu,1 and Quan-Yuan Feng2

1
School of Information Engineering, Southwest University of Science and Technology, Mianyang 62010, China
2
School of Information Science & Technology, Southwest Jiao tong University, Chendu 610031, China
*Corresponding author: fengquanyuan@163.com

Received 13 April 2011; revised 26 June 2011; accepted 26 June 2011;

posted 27 June 2011 (Doc. ID 145846); published 4 August 2011

This paper presents a method that allows a conventional dual-camera structured light system to directly
acquire the three-dimensional shape of the whole surface of an object with high dynamic range of surface
reflectivity. To reduce the degradation in area-based correlation caused by specular highlights and dif-
fused darkness, we first disregard these highly specular and dark pixels. Then, to solve this problem and
further obtain unmatched area data, this binocular vision system was also used as two camera-projector
monocular systems operated from different viewing angles at the same time to fill in missing data of the
binocular reconstruction. This method involves producing measurable images by integrating such tech-
niques as multiple exposures and high dynamic range imaging to ensure the capture of high-quality
phase of each point. An image-segmentation technique was also introduced to distinguish which mono-
cular system is suitable to reconstruct a certain lost point accurately. Our experiments demonstrate that
these techniques extended the measurable areas on the high dynamic range of surface reflectivity such as
specular objects or scenes with high contrast to the whole projector-illuminated field. 2011 Optical
Society of America
OCIS codes: 150.6910, 330.1400, 100.3010, 120.6650, 080.5084.

1. Introduction A series of papers [610] explore the use of a dif-

A structured light binocular technique is be-
coming more and more popular to achieve a three-
dimensional (3D) shape of surfaces due to its fast
speed and high accuracy, as well as its nondestructive
nature. But the conventional fringe projection [13] or
Moire [4,5] techniques can only measure diffuse sur-
faces. High dynamic range (HDR) of surface reflectiv-
ity, such as with specular objects or scenes with very
bright and dark parts, may lead to dynamic ranges
that the camera(s) cannot handle; thus, the optical
signal cannot be properly retrieved. As a consequence,
parts of the surface will go undetected or are misread.
One traditional way to measure such objects involves
spraying the surface with certain materials, such as
developer; however, this is not suitable for many ob-
jects, such as clothing, sculptures, and cultural relics.
0003-6935/11/234557-09$15.00/0
2011 Optical Society of America
fusive screen of a planar monitor coupled with a
projector as a structured light source to measure
specular surface so that information for the whole
field can be obtained. The periodic fringe patterns
are displayed on the screen, reflected by the mea-
sured specular surface, and the distorted patterns
are captured with a camera. A 3D shape is obtained
by integrating the slope distribution numerically,
which can be reconstructed by analyzing these dis-
torted patterns. However, the phase values generally
depend on the surface slopes coupled with the sur-
face heights and have ambiguities in their distribu-
tions. Hongwei Guo [11] improved this method by
moving the diffusive light source vertically to two
or more known positions and measuring the phase
distribution of the deformed fringe pattern at each
position to eliminate phase ambiguities and error ac-
cumulations. But this is a time-consuming procedure
and always needs some known positions.

10 August 2011 / Vol. 50, No. 23 / APPLIED OPTICS 4557

 Scene-adapted structured light is another ap-
proach to make the colors, intensities, and shapes
of the projected patterns adapt to the scene. Caspi
et al. [12] presented the structured light approaches,
which explicitly modeled the camera-projector path.
The work [13] of the question of an optimal set of
patterns is formulated for the first time. In order to
avoid overexposure and underexposure in the image,
Koninckx [14] proposed a more complete camera-
projector model, which allows adapting patterns on-
line on a per pixel basis. This model is based on a
crude estimate of the scene geometry and a reflec-
tance characteristic, which was a bit similar to the
work proposed in [15], in which a projector was used
to change an objects appearance. But it relies on
prior knowledge of surface reflectance or on an as-
sumed reflectance model. As a result, they may cause
severe errors in the measurement of objects.
Our approach falls into another type of approach by
changing the viewpoint, exposure time of camera, or
light intensity of projection. This type of approach
seeks to avoid estimating surface roughness and
scene geometry because measuring them is often
impractical. Bhat and Nayar [16] used a trinocular
stereo system as three binocular stereo systems and
analyzed pairs of camera images to extract correct
depth estimates of scene points from different pairs
so as to yield an accurate depth map of scene. The im-
portant characteristic of these configurations is that
for each scene point in the common field of view of sen-
sors, at least one binocular pair provided the correct
depth estimate. But it involved extra efforts to deter-
mine a suitable trinocular configuration. Scharstein
[17] used a camera translating on a linear stage to
construct a binocular system, and one or two light pro-
jectors illuminated the scene from different direc-
tions. This approach enables high-confidence view
disparities at points visible in both cameras and illu-
minated by at least one source. A number of unknown
code values due to shadows in the scene could be re-
duced by using more than one illumination source.
But this paper points out that partial occlusion (visi-
ble only in one view) was unavoidable because of
highly specular surfaces or very low albedo areas.
This would ultimately lead to some holes in the recon-
struction. To compensate for the influence of specular
reflections or shadowed areas, Kowarschik et al. [18]
also used a projection system to grate structures from
up to 15 different directions, which is connected with a
simultaneous variation of the intensity of the pro-
jected grating structures. Moreover, the object can
be rotated, yielding other views of the object. A
camera acquired different patches of the object that
were transformed into a global coordinate system
to achieve complete measurement. But the param-
eters of the system are so much that the calibration
is a time-consuming procedure and is difficult to rea-
lize with high accuracy. This can result in a poor con-
ditioning of the coordinate calculation because of the
errors and their propagation. Recently, Zhang and
Yau [19] proposed a HDR 3D scanning technique
4558 APPLIED OPTICS / Vol. 50, No. 23 / 10 August 2011
based on multiple exposures and a phase-shifting
method. A sequence of fringe images with different ex-
posures was taken. The saturated pixels because of
specular reflection in the higher exposure were re-
placed by the corresponding pixels of the lower expo-
sures. Therefore, the specular or dark area can be
properly measured without affecting the rest of the
areas. However, most of the multiple-exposure fusion
methods depend on certain image properties like the
scene irradiance map [20], etc. When the irradiance
across a scene varies greatly, there will always be
overexposed and underexposed areas in the captured
image no matter what exposure time is used.
Our approach is to further optimize the third type
of method to handle high-dynamic range surfaces. A
conventional dual-camera structured light system is
used in our research. The techniques of multiple ex-
posures and HDR imaging are unified and general-
ized to lead to insights regarding solutions to the
traditional problems of dealing with HDR surfaces.
We further developed a practical technique combin-
ing both binocular and monocular reconstruction to
deal with some parts of highly specular surfaces with
large reflectivity variations that may be too shiny
or dark to measure by all the mentioned methods.
These improvements now make it suitable for a con-
ventional structured light binocular vision system to
perform high-quality whole-field surface measure-
ments of objects with HDR of surface reflectivity
without increasing the hardware cost.
The paper is organized as follows. Before giving a
more detailed explanation of our techniques in
Sections 3 and 4, Section 2 explains the effect of re-
flection mechanisms on structured light systems.
Some experimental procedures and results about ty-
pical cases of measurement of objects are given in
Section 5. The last section summarizes the main re-
sults of this work.
2. Effect of Reflection Mechanisms on Structured
Light Binocular System
Briefly, in our paper, a series of digital fringe patterns
composed of vertical straight stripes are generated
by a computer and sent to a digital video projector,
which then casts the fringe images onto the object.
These deformed fringe images are then recorded by
two cameras from different angles. The digital fringe
images acquired by the camera are then processed to
retrieve the phase using a phase-shifting algorithm.
A four-step phase-shifting algorithm with a phase
shift of =2 was utilized to determine relative phase
in this research. The projector also illuminates a
scene with a temporally varying pattern of light
stripes. A temporal phase unwrapping algorithm is
utilized to get the absolute phase map, which serves
as an important matching constraint, as well as an
epipolar geometry to get the corresponding points
between two images. The depth of corresponding
points can be determined by triangulating between
projected light patterns and serving camera view-
point [21,22].
 It is known that surface radiance may be decom- Source
Surface Image
posed into three primary reflection components: normal sensor
the diffuse lobe, the specular lobe, and the specular vector
spike. Figure 1 shows the polar plots of these three
components. The diffuse lobe represents both an Incident
ray Specular
internal scattering mechanism and multiple reflec- Image spike
tions on the surface in a random manner, and reflects sensor
hemisperically in all directions. The specular lobe
spreads for some range around the specular direc- Diffuse i i Specular
tion, which is the angle at which the incident angle lobe lobe
equals the reflected angle. The specular spike repre-
sents a mirror-like reflection, and it is nearly zero in
all directions, except for a very narrow range around Fig. 1. (Color online) Polar plots of the three reflection
the specular direction. components.
When the surface is smooth enough, the irradiance
across a scene varies greatly. The specular spike
areas are often too bright while the diffuse lobe areas highly specular and dark pixels. Those techniques
are often too dark; the captured image may be over- match the measuring range of intensity of camera
exposed or underexposed due to limitations on the with the range of reflection intensity.
measuring range of intensity of the camera. But the To start the procedure, a sequence of even white
overexposed or underexposed image cannot repre- patterns with 256-level light intensities is first pro-
sent the proper intensity of fringe images, which jected onto the inspected surface by turns. We sup-
can lead to serious phase error. These errors would pose that the gray of a given pixel in the captured
result in the loss of matching points and leave blank white image is I and set minimal and maximal high-
areas or holes in the final 3D shape. precision measurable gray scale threshold I min and
To reduce the degradation in area-based correla- I max by experiment. The thresholds can be used to
tion caused by specular highlights and diffuse dark-
reduce the effect of the other nearby pixels influ-
ness, we first acquire measurable images by multiple
enced by the overexposed and underexposed pixels.
exposures and an HDR imaging technique, disre-
The next step is to calculate the mask image se-
garding those highly specular or dark points. Then
we adopt a technique that combines the binocular quence M, which serves as the definition of measured
and monocular vision stereo to further obtain un- validity of pixels in a certain image (here, 1 repre-
matched data of the whole surface. sents validity and 0 represents invalidity). A bright,
even pattern is projected onto the inspected surface,
3. High Dynamic Range Imaging and the images are captured using multi-exposure
To meet luminous requirements of different surface times. Assume the number of multiple exposures
properties, a multi-exposure technique is adopted for is N. The entire captured images form an image se-
the cameras. The camera takes the fringe images quence I n , with n 1; 2; ; N. They are arranged
with different exposures by adjusting the aperture from brighter to darker (i.e., from higher exposure
of the camera lens or by changing the exposure time to lower exposure). The mask image sequence M
of the camera. Longer exposure may give good fringe can be achieved by the following Eq. (1) and (2).
quality of the darkest areas, but the brightest areas When n 1,
might be overexposed, while shorter exposure may
give good fringe quality of the brightest areas but
cause the darkest areas invisible.
The HDR imaging technique is also introduced to

1; I min I n x; y I max
merge each well-illuminated and unsaturated pixel Mn : 1
from different exposures together to obtain the 0; otherwise
right/left synthesized images, and disregard these

When n > 1,

1; I min I n x; y I max and M l x; y 0l 1; 2; :::; n 1
M n x; y ; 2
0; otherwise

10 August 2011 / Vol. 50, No. 23 / APPLIED OPTICS 4559

where n, l denote the sequence number of I, the
I n x; y represents the intensity of pixel x; y of im-
age I n , and M n x; y and M l x; y are the mask value
of pixels x; y in I n and I l .
Next, sinusoidal phase-shifted fringe patterns are
projected onto the surface under the same series ex-
posures as the former step. A sequence of fringe
images [Pkn x; y, with k 1; 2; 3; 4 and n 1; 2; 3;
; N] are acquired for the measurement. For each
set n, four fringe images with a phase shift of =2
are captured under the same exposure. In other
words, each set of fringe images can be used to inde-
pendently reconstruct a 3D shape for good points.
We assume that the phases of pixels whose inten-
sities lie in the range I min to I max can all be calculated
accurately. If all pixels meet the requirement, the al-
gorithm comes to an end. If not, we decrease the ex-
posure time and take images again. The overexposed
pixels in a higher exposure are replaced by the
corresponding pixels in a lower exposure, while the
rest of the pixels remain unaltered. The algorithm
stops only when the whole surface is measurable or
the exposure time reaches to the longest point of the
camera. Assume the brightness of the fringe image
sets decrease from one exposure to the next, i.e.,
Pkn x; y Pkn1 x; y:
Then the final fringe image pixel used for 3D
measurement is
Pkf x; y Pkm x; y; m minn;
with I min Pkm x; y I max k 1; 2; 3; 4 and P1m1
I min ; I max
orP2m1 I min ; I max
or P3m1 I min ; I max
or
P4m1 I min ; I max
. Here, m minn is the minimum
function of n.
The obtained high-quality pixels from all expo-
sures are then merged together to generate the com-
plete fringe image to acquire 3D coordinates. Each
pixel of the fringe images H k x; y k 1; 2; 3; 4 is
generated by selecting the brightest within the range
I min to I max unsaturated corresponding pixel from
one set of fringe images, namely,
X
N
H k x; y M n x; y Pkn x; y k 1; 2; 3; 4:
n1
For an arbitrary point on the image, its intensity
values of four fringe images use a certain exposure
so that all intensity values of this pixel are not satu-
rated while the same pixel of the previous set of
images with higher exposure is saturated for at least
one fringe image. Both the right and the left synthe-
sized images obtained in Eq. (5) can obtain the phase
pixel by pixel that can be further converted into 3D
coordinates.
4560 APPLIED OPTICS / Vol. 50, No. 23 / 10 August 2011
The multiple exposures and HDR imaging techni-
que take full advantage of the merit of retrieving the
phase values of different pixels from different shots
of fringe images to ensure the capture of high-quality
phase of each point with HDR of surface reflectivity.
4. Combination of Binocular and Monocular Vision
Technique
In reality, many observed intensities include not only
body reflectance but also specular reflectance, inter-
reflectance, shadows, etc. In specular objects, in
which the component of the specular spike domi-
nates, a reflected light beam does not expand but con-
centrates in a small region around the specular
direction. Hence, certain areas may cause very differ-
ent light captures for the two cameras. The light-
input intensity from some points is often appropriate
for only one of the two cameras, which makes it im-
possible to perform the matching of these points by
using binocular reconstruction. To solve this problem
and further obtain unmatched area data of the whole
surface, we adopted a system that combines the bi-
nocular vision and monocular vision techniques as
described below.
The setup of this system is shown in Fig. 2. Two
cameras from different viewing angles then capture
3 the deformed fringe images. Different from other tra-
ditional binocular structure, this system is also di-
vided into two monocular structures, each of which is
composed of a single camera and the same projector,
respectively. Fringe images acquired by each mono-
cular structure are used to reconstruct some specific
4 patches of the 3D geometry through the phase-
shifting algorithm, while the binocular vision struc-
ture functions normally at the same time.
A. System Calibration
System calibration requires geometric calibration of
the cameras and the projector. Binocular stereo
Camera Camera
5 Projector
Left monocular system Right monocular system
Binocular system
Fig. 2. System setup.
visual sensors are calibrated from images of a stan-
dard checkerboard shown at a few different unknown
positions [23]. The calibration of the two camera-
projector monocular systems uses the same standard
checkerboard as that of a binocular system. The key
is to treat the projector as a camera by projector
calibration, which can be referred to as the phase-
shifting method. This method [24] regards the projec-
tor as a camera, and the fringe images are used as a
tool to establish the correspondence between the
camera pixel and the projector pixel, as shown in
Fig. 3. So the projector can capture images like a
camera, thus making the calibration and reconstruc-
tion of a monocular system essentially the same as
that of a binocular system.
In order to further obtain the complete data of the
whole projector-illuminated surface, the data from
the three vision system must be well aligned and re-
gistered. So the binocular and monocular systems
are to be calibrated in the same world coordinate
system. In the process of calibration, the world coor-
dinate system may be set on the left camera coordi-
nate system for the unification of binocular and left
monocular coordinate systems. Thus it is only needed
for the right monocular coordinate system to be con-
verted to the world coordinate system, which has
been well-studied and is not discussed here.
B. Compensation of Unmatched Points in Binocular
Reconstruction
In highly specular objects, the component of the spec-
ular spike dominates, and the irradiance across this
area may vary so greatly that there will always be
overexposed and underexposed areas in the captured
image no matter what exposure time is used.
Specular reflections can cause the intensity and color
of corresponding points to change dramatically ac-
cording to different viewpoints, thus producing se-
vere matching errors for various stereo algorithms.
In order to solve the matching problem, we pro-
posed that construction of monocular vision system
can resolve the matching problem and make up
Fig. 3. (Color online) Monocular stereo vision.
the holes of specular area in binocular reconstruc-
tion. Namely, the points reconstructed by monocular
system only aim to the unmatched points. This not
only avoids overlapping of the point clouds but also
maintains the reconstruction precision of the binocu-
lar vision system since it is well-known that the pre-
cision of the binocular stereo is often better than
that of the monocular stereo.
For monocular systems, each camera captures only
a part of the 3D shape. Since all the systems are ca-
librated in a world coordinate system, the corre-
sponding 3D data from those systems are aligned.
The 3D data pieces can also be registered with the
assistance of the absolute phase [25]. Here only a sin-
gle projector is utilized and absolute phase is unique;
therefore, they should be aligned whichever way
each camera acquires it. As for unmatched points,
the absolute phases from the two monocular systems
are not the same or the epipolar geometry constraint
cannot be satisfied. So there is a problem regarding
which monocular system should be chosen to con-
struct the lost points in a binocular system accu-
rately as to a certain point.
Since the reliability of the unwrapping phase de-
pends on the gray of a pixel, we introduce an image
segmentation technique to distinguish the reliable
points that can recover the lost points accurately from
the right or left monocular system self-adaptively.
The Ostu thresholding method [26] is the simplest
approach for image segmentation through threshold-
ing, which is based on the analysis of the gray-levels
histogram. Thresholding can be explained in the fol-
lowing simple procedure. Let N be the set of natural
numbers and x; y be the spatial coordinate of a
digitized image. The thresholding operation is re-
garded as the partitioning of the pixels into two
groups, C0 f0; 1; ; tg and C1 ft 1; t 2; ;
L 1g, where L is the number of levels in the grays-
cale image, and f0; 1; ; m 1g is a set of
positive integers of gray levels. Then, from the defi-
nitions above, an image function can be defined as
the mapping:
f x; yN N ;
where f x; y is the brightness (gray level) of a pixel
with coordinates x; y. Let t be a threshold level and
B fb0; b1g be a pair of binary gray levels, i.e.,
b0; b1 . An image function f :; : at gray level t
can be derived from a binary image function
f t x; yN N B:


b0; if f x; y < t;
f t x; y : 6
b1; otherwise
From the histogram of the image, let f i be the num-
ber of ith gray level points. t is a thresholding point.
0 t is the mean for the first group C0 and 1 t is the
mean for the second group C1. Their definitions are
given as follows:
10 August 2011 / Vol. 50, No. 23 / APPLIED OPTICS 4561
 X
L1 X
t
pi f i = fi qt pi
i0 i0
Pt PL1
i0ipi 0 t it ipi t
0 1 1 ;
qt qt 1 qt 1 qt
where p1 ; ; pi represent the histogram probabilities
of the observed gray values of 1; ; I. Let 2t be the
between-group variance:
2t qt1 qt1 0 2 : 9 Fig. 4. Experimental system.
Then the optimal thresholding, t , can be deter-
mined by maximizing the between-group variance
as shown below:
2t max 2t :
Thus, the set B can be decided according to the op-
timal thresholding. All the image points in the right/
left synthesized image labeled by the value of b0 and
b1 are called labelr =labell , which are used to deter-
mine the membership of the pixel to construct the
lost points unreliably or reliably (here, b0 represents
unreliability; b1 represents reliability). If either one
label labelr =labell of unmatched pixels in the right/
left synthesized images is b1, the lost pixel Plose in
binocular system can be filled in by the reconstructed
pixel Pr =Pl in the right/left monocular system whose
label is b1 at the pixel. Otherwise, Plose cannot be re-
covered accurately even if both labelr and labell
equal to b1 because some error may be produced dur-
ing the process of unwrapping phase results in these
unmatched points. So these points are to be omitted
in this case.
8 Our approach consists of the following steps:
< Pr ; labelr b1; labell b0
Plose Pl labelr b0; labell b1 : 11
:
omitted; otherwise
In practice, there are many factors affecting the ac-
curacy of the result of monocular systems. These fac-
tors can be classified as being from uncertainty due
to sensor noise or from a nonlinear response or gam-
ma distortion in the cameraprojector pair. The two
patches from two monocular systems are often not
aligned well with binocular reconstruction. Some
techniques about elimination of nonlinear luminance
effects for digital video projection must be applied in
our system to improve the accuracy of two camera-
projector system. It is mainly the gamma distortion
that causes nonlinear luminance effects for digital
video projection. Gamma correction was used to
greatly improve the measurement accuracy of phase.
Since the gamma value is not easily measured, the
distorted phase was corrected by minimizing the en-
ergy in the harmonic components. This method
4562 APPLIED OPTICS / Vol. 50, No. 23 / 10 August 2011
7
8
involved applying a constraint derived from a gam-
ma model according to the number of projected pat-
tern. The details of this method were described in
[27]. Those techniques can make the data from mono-
10 cular and binocular vision systems merged more
smoothly.
5. Experiments
The proposed method was implemented and tested
in our previously developed 3D shape measurement
system, as shown in Fig. 4. This system includes a
Digital Light Processing projector (MD-565X) with a
resolution of 1024 768 pixels and two digital CCD
cameras (DH-HV13202UM) with an image resolu-
tion of 1280 1024 pixels. The camera lenses are
Schneider with a focal length of 23 mm. The proces-
sor used for computation has two micro core struc-
tures with main frequency of 1:6 GHZ and 2G
internal memory. To ensure good measurement re-
sults, the baseline of the two cameras is about 0:7
0:95 mm with the measurement distance being
0:9 1:2 mm, and the angle between optical axis of
the camera and the base line is around 40.
System calibration. Calibrate the binocular
and two monocular vision systems in the same world
coordinate system with the same checkerboard.
HDR imaging. The right/left synthesized
images are obtained after merging fringe images un-
der multiple exposures, while some highly specular
and dark pixels are disregarded.
Phase solution. The right/left absolute phases
of all points are acquired by unwrapping right/left
phase maps obtained from the right/left synthesized
images. The absolute phases serve as a matching
constraint, as well as an epipolar geometry to get
the corresponding points in stereo pairs.
3D coordinates acquisition of matched points in
binocular system. The corresponding points are
further converted to x, y, z coordinates of such points
in binocular system.
Lost points recovered by the reliable points
from monocular reconstruction. Image segmentation
technique is introduced to distinguish the suitable

Fig. 5. The measured vase (a), its left and right fringe images
(b, d), and its left and right synthesized fringe image (c, e).
monocular system to reconstruct the lost points
accurately as to a certain point.
We measured a china vase. The contrast between
different colors is very large, albeit its surface is al-
most diffuse. Figure 5(a) illustrated the photograph
of the measured object. Three exposures from long to
short were preformed in this test. One image pair un-
der certain exposure is seen in Figs. 5(b) and 5(d).
One pair of synthesized fringe images for a certain
phase shift after merging the results from all expo-
sure images is illustrated in Figs. 5(c) and 5(e) and
Fig. 6. Left and right unwrapped phases maps of the vase (a, b),
the 3D results reconstructed from left and right monocular system
(c, d) and binocular system (e), respectively, and the compensation
result from the combination of binocular and monocular system (f).
Fig. 7. Zoom-in view of a small region of binocular reconstruction
result (a) and compensation result from the combination of bino-
cular and monocular systems (b).
eventually used for both binocular and monocular
reconstruction.
Here, Fig. 6 illustrates the process of the reconstruc-
tion. The corresponding unwrapping phase maps ob-
tained from the right or left synthesized fringe images
are shown in Figs. 6(a) and 6(b). Figures 6(c) and 6(d)
show 3D results from the two monocular systems, re-
spectively. The 3D results from the binocular recon-
struction jointly constructed by the two unwrapping
phase maps were also shown in Fig. 6(e). It is obvious
there were some holes still existing in the binocular
reconstruction even though multiple exposures were
functional. As is shown in Fig. 6(f), the eventual 3D
result combining all the output from the three systems
extends the measurable areas on the objects to the
whole projector-illuminated field. That is, all the holes
in the binocular reconstruction are filled up with data
from monocular reconstruction. The existence of the
regular gamma distortion in monocular systems re-
sulted in the stripes shown in Figs. 6(c) and 6(d).
Although the gamma correction was performed, there
was still some gamma distortion existing in the
cameraprojector pair and canceled out in the bino-
cular system. The zoom-in views of a small region of
binocular reconstruction and its compensation result
through monocular reconstruction are illustrated in
Fig. 7. Compared with traditional binocular results
using multiple exposures, it is clear that the combined
Fig. 8. The turbine blade (a) and the reconstructed 3D results ob-
tained after single exposure from the binocular system (b), four
exposures from the binocular system (c), and four exposures from
the binocular and monocular systems (d), respectively.
10 August 2011 / Vol. 50, No. 23 / APPLIED OPTICS 4563
 Table 1. Comparison of the Reconstructed Turbine
Blade by the Tree Method
Computation Number of the
Method Time (s) Acquired Points
Single exposure 2.76 311053
Multiple exposure 4.61 563592
Multiple exposure and 5.73 694726
combined binocular and
monocular technique
result obtains more complete 3D cloud data of sur-
faces, and solves the problem of loss of 3D data.
Moreover, we measured a polished, highly specular
surface: a metal turbine blade shown in Fig. 8(a).
Both cameras captured images under four exposures.
However, it should be noted that the reconstructed
3D results after single exposure and four exposures
from the binocular system, depicted in Figs. 8(b) and
8(c), respectively, show incomplete results since only
the overlapping view field of the two cameras could
be reconstructed by the binocular system and the
matching errors resulted in the loss of matching
points and left blank areas or holes in the final 3D
shape. It is obvious that the number of the uncertain
points is decreased by increasing the number of
exposures.
In fact, some points, although not being recon-
structed, have good phases either from the left cam-
era or from the right camera. Of note, the nonpublic
area, the root as well as the top of the object that can-
not be obtained from the binocular system, have good
phases in two monocular systems. These data could
be used to compensate for the incomplete areas. As is
illustrated in Fig. 8(d), the final reconstructed result
with complete 3D cloud data was obtained from bino-
cular and monocular vision systems. Comparison of
the turbine blade reconstructed by single exposure,
four exposures, and the proposed method, including
the number of the acquired points and the computa-
tion time, is shown in Table 1. These experimental
results confirmed that the proposed method could
successfully measure full-field specular surfaces
where the projector could project fringe images onto
without increasing computation cost too much. The
rms error for all these 3D points of the turbine blade
is around 0:036 mm.
6. Conclusions
We have presented here a method that allows a con-
ventional structured light system with dual cameras
to successfully measure the 3D shape of objects with
a broad range of surface reflectivity without knowing
the property of surface reflectivity and scene geome-
try. Enjoying this advantage, however, as compared
to using a monocular and binocular reconstruction
separately, the high precision of the binocular vision
system is maintained because public field of view
is reconstructed by the binocular vision system.
Furthermore, as long as each pixel is caught by at
least one of the cameras, its correspondence in the
4564 APPLIED OPTICS / Vol. 50, No. 23 / 10 August 2011
other image (or lack of correspondence, which indi-
cates occlusion) can be unambiguously determined,
and the holes could be made up by the monocular
stereo system. In addition, we do not attempt to
avoid specular reflection like strong highlights but
rather perform accurate compensation in their pre-
sence. Thus, preprocessing of imaging, like removal
of highlights, is avoided. Our approach is not limited
by any specific reflectance model or to any correspon-
dence scheme, and this method does not require the
change of the relative position between the system
and the object, so it is therefore easy to incorporate
into existing stereo algorithms. Our experiments de-
monstrated that this proposed method could success-
fully measure full-field surfaces of typical objects
with large dynamic surface reflectivity without too
much computation cost. However, enhancing self-
adaptability of measurable images acquisition may
be further needed to make this method more
practicable.
This work is supported by the National Natural
Science Foundation of China (NSFC) under grant
number 60 990 323, and the Sichuan provinces Inter-
national Cooperation Fund under grant number
2009HH0023.
References
1. V. Srinivasan, H. C. Liu, and M. Halioua, Automated phase-
measuring profilometry of 3-D diffuse object, Appl. Opt. 23,
31053108 (1984).
2. E. Hu and Y. He, Surface profile measurement of moving
objects by using an improved phase-shifting Fourier trans-
form profilometry, Opt. Lasers Eng. 47, 5761 (2009).
3. J. Vanherzeele, S. Vanlanduit, and P. Guillaume, Processing
optical measurements using a regressive Fourier series: a re-
view, Opt. Lasers Eng. 47, 461472 (2009).
4. D. M. Meadows, W. O. Johnson, and J. B. Allen, Generation of
surface contour by Moire patterns, Appl. Opt. 9, 942947
(1970).
5. A. K. Asundi, Moire methods using computer-generated
gratings, Opt. Eng. 32, 107116 (1993).
6. P. Aswendt, R. Hoing, and S. Gartner, Industrial inspection of
specular surfaces using a new calibration procedure, Proc.
SPIE 5856, 393400 (2005).
7. O. A. Skydan, M. J. Lalor, and D. R. Burton, Three-
dimensional shape measurement of non-full-field reflective
surfaces, Appl. Opt. 44, 47454752 (2005).
8. O. A. Skydan, M. J. Lalor, and D. R. Burton, 3D shape
measurement of automotive glass by using a fringe reflection
technique, Meas. Sci. Technol. 18, 106114 (2007).
9. W. Li, T. Bothe, C. von Kopylow, and W. P. O. Juptner,
Evaluation methods for gradient measurement techniques,
Proc. SPIE 5457, 300311 (2004).
10. A. Moreno, J. Campos, and L. P. Yaroslavsky, Frequency re-
sponse of five integration methods to obtain the profile from
its slope, Opt. Eng. 44, 033604 (2005).
11. H. W. Guo, F. Peng, and T. Tao, Specular surface measure-
ment by using least squares light tracking technique, Opt.
Lasers Eng. 48, 166171 (2010).
12. D. Caspi, N. Kyriati, and J. Shamir, Range imaging with
adaptive color structured light, IEEE Trans. Pattern Anal.
Machine Intell. 20, 470480 (1998).
13. E. Horn and N. Kiryati, Towards optimal structured light
patterns, Image Vision Comput. 17, 8797 (1999).
14. T. P. Koninckx, P. Peers, P. Dutre, and L. Van Gool, Scene-
adapted structured light, in Proceedings of IEEE Conference
on Computer Vision and Pattern Recognition (IEEE, 2005), pp.
611618.
15. M. Grossberg, H. Peri, S. Nayar, and P. Belhumeur, Making
one object look like another, in Proceedings of IEEE Interna-
tional Conference on Computer Vision and Pattern Recognition
(IEEE, 2004), pp. I-452I-459.
16. D. Bhat and S. Nayar, Stereo in the presence of specular
reflection, in Proceedings of IEEE International Conference
on Computer Vision (IEEE, 1995), pp. 10861092.
17. D. Scharstein and R. Szeliski, High-accuracy stereo depth
maps using structured light, in Proceedings of IEEE Interna-
tional Conference on Computer Vision and Pattern Recognition
(IEEE, 2003), pp. 195202.
18. R. M. Kowarschik, J. Gerber, G. Notni, and W. Schreiber,
Adaptive optical three-dimensional measurement with struc-
tured light, Opt. Eng. 39, 150158 (2000).
19. S. Zhang and S-T Yau, High dynamic range scanning techni-
que, Opt. Eng. 48, 033604 (2009).
20. A. R. Varkonyi-Koczy, Improved fuzzy logic supported HDR
colored information enhancement, in Proceedings of IEEE
International Conference on Instrumentation and Measure-
ment Technology (IEEE, 2009), pp. 361366.
21. J. Batlle, E. Mouaddib, and J. Salvi, Recent progress in coded
structured light as a technique to solve the correspondence
problem: A survey, Patt. Recog. 31, 963982 (1998).
22. J. Davis, D. Nehab, R. Ramamoorthi, and S. Rusinkiewicz,
Spacetime stereo: A unifying framework for depth from tri-
angulation, IEEE Trans. Pattern Anal. Machine Intell. 27,
296302 (2005).
23. Z. Zhang, A flexible new technique for camera calibration,
IEEE Trans. Pattern Anal. Machine Intell. 22, 13301334
(2000).
24. S. Zhang and P. S. Huang, Novel method for structured light
system calibration, Opt. Eng. 45, 083601 (2006).
25. S. Zhang and S-T Yau, Absolute phase assisted three-
dimensional data registration for a dual-camera structured
light system, Appl. Opt. 47, 31343142 (2008).
26. N. Ostu, A thresholding selection method from gray-level his-
tograms, IEEE Trans. Syst. Man Cybern. A 9, 6266 (1979).
27. K. Liu, Y. Wang, D. L. Lau, Q. Hao, and L. G. Hassebrook,
Gamma model and its analysis for phase measuring profilo-
metry, J. Opt. Soc. Am. A 27, 553561 (2010).
10 August 2011 / Vol. 50, No. 23 / APPLIED OPTICS 4565