Target detection in optically scattering

media by polarization-difference imaging

J. S. Tyo, M. P. Rowe, E. N. Pugh, Jr., and N. Engheta

Polarization-difference imaging 1PDI2 was recently presented by us as a method of imaging through

scattering media 3Opt. Lett. 20, 608 1199524. Here, PDI is compared with conventional, polarization-
blind imaging systems under a variety of conditions not previously studied. Through visual and
numerical comparison of polarization-difference and polarization-sum images of metallic targets
suspended in scattering media, target features initially visible in both types of images are shown to
disappear in polarization-sum images as the scatterer concentration is increased, whereas these
features remain visible in polarization-difference images. Target features producing an observed
degree of linear polarization of less than 1% are visible in polarization-difference images. The ability of
PDI to suppress partially polarized background variations selectively is demonstrated, and discrimina-
tion of target features on the basis of polarization information is discussed. Our results show that,
when compared with conventional imaging, PDI yields a factor of 2–3 increase in the distance at which
certain target features can be detected. r 1996 Optical Society of America

1. Introduction presence of scatterers. The underwater environ-

Obtaining images of objects in optically scattering ment is a good example of a regime in which scatter-
media presents a challenging set of problems. ers are present and limit the performance of imaging
Solving these problems improves the quality of im- systems.4 Therefore it is a logical place to look for a
ages obtained when scattering particles impair sys- naturally occurring visual system that has adapta-
tem performance, as in underwater environments, in tions for improving the quality of vision in scattering
fog and rain, and in medical diagnostics. Recent media.
approaches to such problems include studies de- Inspired by the visual systems of certain fish,5 we
signed to improve images obtained by standard have suggested and developed a novel approach to
photography underwater,1 time-resolved imaging,2 underwater imaging.6 We have hypothesized that
and analysis of diffusive photon-density waves.3 certain species of fish, such as Lepomis cyanellus
Most imaging systems simply map intensity spa- 1green sunfish2, have adjacent photoreceptors opti-
tially and are not proficient at seeing through scatter- mally tuned to orthogonal linear polarizations and
ing media, since scattered light degrades image that such species may compute a polarization-
quality. Some insight into methods for improving difference 1PD2 signal by differencing the outputs of
the quality of images of objects in scattering media these receptors.5,7 A recent study by us demon-
may come from an examination of visual systems of strated that an imaging system employing the prin-
animal species that have evolved to see in the ciple of polarization differencing could uncover sur-
face features of a target suspended in a scattering
medium, even though these features were not visible
The authors are with the University of Pennsylvania, Philadel- in conventional images.6
phia, Pa. 19104; J. S. Tyo is with the Moore School of Electrical This paper extends our earlier work in several
Engineering, M. P. Rowe is with the Institute of Neurological ways. In Sections 2–4 we review the concept of
Science, E. N. Pugh, Jr., is with the Institute of Neurological polarization-difference imaging 1PDI2 and describe
Science and the Department of Psychology, and N. Engheta is the experimental setup and scattering medium in
with the Moore School of Electrical Engineering and the Institute
detail. In Section 5 we extend the generality of the
of Neurological Science.
Received 16 August 1995; revised manuscript received 21 previous work by presenting images obtained at
November 1995. various scatterer concentrations. In Section 6 we
0003-6935@96@111855-16$06.00@0 explore the common-mode rejection feature of PDI.
r 1996 Optical Society of America In Section 7 we address the discriminability of

10 April 1996 @ Vol. 35, No. 11 @ APPLIED OPTICS 1855

targets on the basis of polarization information. them or subtracting them, and transforming the
In Section 8 we examine the intrinsic detectability of resulting sum or difference image can be imple-
these targets with signal-detection theory. Finally, mented rapidly for both PS and PD images. We
in Section 9 we discuss future possibilities for PDI. chose the affine transformation because it is logical,
linear, and easy to implement. By applying the
2. Concept of Polarization-Difference Imaging specific affine transformations defined in Eqs. 132, we
Conventional imaging systems create images by can use the entire available dynamic range of the
measuring and recording the signal intensity as a display without losing any information contained in
function of position within a frame. While conven- the raw images.
tional imaging systems do not measure the polariza-
tion components, the overall intensity can in prin- 3. Experimental Setup
ciple be decomposed into the intensities of the The experimental setup we use is detailed in Fig. 1.
portions of light linearly polarized along two orthogo- Two white-light, tungsten-filament slide projectors
nal directions. Theoretically this decomposition can 1Prado Universal, 250-W2 are used as light sources.
be performed by capturing images through an ideal The light is diffused by a 0.64-cm-thick Plexiglass
linear polarization analyzer. The intensity distribu-
tions obtained at orthogonal linear polarizations are
represented by I\1x, y2 and I'1x, y2, where 1x, y2 identi-
fies the position on the image and the symbols 6 and
' represent the two orthogonal directions. The
choice of a reference for the polarization axes 6 and '
is discussed in Subsection 6.B. Polarization-sum
1PS2 and polarization-difference 1PD2 intensity distri-
butions are formed as6

psI1x, y2 5 I\1x, y2 1 I'1x, y2, 112

pdI1x, y2 5 I\1x, y2 2 I'1x, y2. 122

Ideally, an image based on Eq. 112 is equivalent to one
obtained by a conventional, polarization-blind imag-
ing system that measures intensity only. Further-
more, such an image is not dependent of the choice of
axes. In contrast, PD images, as defined by Eq. 122,
are sensitive to the choice of axes. The dependence
of PD images on the choice of axes can be used to
minimize the effects of a partially polarized back-
ground in a PD image, as we show below. Fig. 1. Experimental setup 1not drawn to scale2: A. Setup
After PS and PD images of a scene are formed, it is used for all backillumination experiments. Light is emitted by
desirable to display them in a way that optimizes the two slide projectors shown at the right side of the figure. The
light is diffused by a translucent white Plexiglass sheet before
use of the available dynamic range of the output de-
entering a glass tank of linear dimension 30 cm 3 30 cm 3 15 cm
vice. This makes it easier to locate small intensity 113.5 L in volume2. The tank is filled with a milk@water mixture
variations within a scene. We carry out this process that acts as a scattering medium. Inside the tank is a target 1the
by applying an affine transformation to the intensity targets are described in the caption of Fig. 22. The target is
distributions psI1x, y2 and pdI1x, y2. Specifically, we suspended from above by a clear Plexiglass sheet. There is a rod
determine the transformed quantities6 fitted to the back face of the target, i.e., the face away from the
camera. This rod is then attached to the clear Plexiglass sheet
psI1x, y2trans 5 gps3psI1x, y2 2 psI1x, y2min4, for support. The clear Plexiglass is transparent at optical wave-
lengths and has a linear extent larger than the area of view of the
pdI1x, y2trans 5 gpd3pdI1x, y2 2 pdI1x, y2min4. 132 camera. Since it is oriented perpendicular to the line of sight, it
does not introduce a partial polarization and it is completely
The quantities on the left-hand side of Eqs. 132 are the undetectable in all of the images. The images are captured by a
PS and PD images transformed for optimal display, CCD camera. In front of the camera are a twisted-nematic
psI1x, y2min and pdI1x, y2min represent the smallest value liquid-crystal cell 1TNLC2, a linear polarization analyzer 1A2, and a
of intensity in the original PS and PD images, and gps narrow-band filter 1F2. For a description of the polarizing proper-
and gpd are stretching factors, defined as ties of the optical path, see the description in Rowe et al.6 B.
The modified setup used for the side-illumination experiments
gps 5 C@3psI1x, y2max 2 psI1x, y2min4, described in Subsection 6.B. One projector is removed, and the
other is moved so as to provide light in a direction perpendicular
gpd 5 C@3pdI1x, y2max 2 pdI1x, y2min4, 142 to the line of sight of the camera. This geometry gives rise to a
partial polarization in the vertical direction as seen by the camera
where C is the maximum available display level 1e.g., and is effectively similar to horizontal viewing underwater at
255 in an 8-bit display monitor2. The simple, four- noon when the sun is directly overhead. All other features of the
step process of capturing two raw images, adding imaging system remain unchanged from Fig. 1A.

1856 APPLIED OPTICS @ Vol. 35, No. 11 @ 10 April 1996

sheet before it enters a glass tank of linear dimen- 151, Imaging Technologies, Inc.2 mounted on a PC
sion 30 cm 3 30 cm 3 15 cm 113.5 L2 filled with platform. The image-processing system is equipped
diluted milk, which acts as a scattering agent. with a 16-bit frame buffer 1accumulator2, which is
Suspended approximately in the center of the tank is used for summing and averaging of images. Linear-
one of two 3.8-cm-diameter aluminum disks that are polarization directions are selected by a twisted-
used as targets. The targets are depicted and de- nematic liquid-crystal cell 1TNLC2 1Liquid Crystal
scribed in Fig. 2. The images are captured with a Institute, Kent State University2 in conjunction with
CCD camera 1Hamamatsu XC-772 whose output is a linear-polarization analyzer, which can be rotated
digitized by an 8-bit image-processing system 1Series mechanically. The details of the operation and the
polarization properties of the TNLC@analyzer pair
have been described elsewhere.6,8 A narrow-band
filter 110-nm FWHM centered at 610 nm2 is placed in
the optical path to eliminate wavelengths outside
the operating waveband of the TNLC. The direc-
tions referred to by the symbols 6 and ' introduced
in Eqs. 112 and 122 are taken as parallel and perpen-
dicular to the analyzer axis, respectively.
4. Characteristics of the Medium
It is known that, roughly speaking, there are four
different paths that light can follow after undergoing
a scattering event at the target plane,9 as shown in
Fig. 3: 112 image-forming light comes from the tar-
get plane and travels directly to the camera without
being scattered off of the path; 122 veiling light9 is
A
scattered into the camera by particles in front of the
target plane; 132 light is lost by being scattered away
from the detector after leaving the target; and 142
light is absorbed by the medium.10 The relation-
ships between the intensity of light following each of
these different paths dictates how well a given
imaging system can perform in the presence of
scatterers. These relationships can be character-
ized in terms of measurable optical properties of the
medium. To characterize our medium, we briefly
examine two properties, the beam-attenuation coeffi-
cient c1l, Cs2, typically a function of wavelength 1l2

B
Fig. 2. Targets: A. Aluminum disk with scratched patches.
The disk is 3.8 cm in diameter, and the patches are 1 cm 3 1
cm. Except for the two patches, the disk face is sandblasted,
rendering it nearly Lambertian. The patches are raised a few
thousandths of an inch above the face of the disk and are abraded
lightly in orthogonal directions with emory paper. The target is
shown with the same orientation used in the images presented in
Fig. 5; i.e., the left patch is scratched vertically as we look at it, the
right patch is scratched horizontally. Light reaches the front of
the disk through multiple scattering events. Light from the
projectors is scattered by milk particles toward the face of the
disk, then is scattered at the face of the disk toward the camera.
The scratched patches have facets that act like mirrors at Fig. 3. Major paths of light in the scattering medium: 112 Im-
different orientation angles. These facets selectively reflect light age-forming light is scattered at the target plane and travels
incident upon the target from specific directions toward the directly to the camera head without being scattered off the
camera. The polarization is caused at least in part by the initial path. 122 Veiling light is ultimately scattered by the particles in
scattering event involving the milk particle. B. The Aluminum the medium between the target plane and the image plane, then
disk without scratched patches. This target was used as the reaches the camera head. 132 Image-forming light lost because of
nonsignal 1blank2 target for the signal-detection-theory experi- scattering. This light can be scattered back into the line of sight,
ments described in Section 8. It is identical to the target of Fig. but when this happens, it becomes veiling light. 142 Image-
2A except for the lack of scratched patches. The entire face is forming and veiling light can both be partially absorbed through-
sandblasted, so light reflecting from its surface is nearly unpolar- out the medium. This figure is an adaptation of a figure pre-
ized everywhere under the illumination conditions of Fig. 1A. sented by Lythgoe 1Ref. 9, p. 586, Fig. 132.

10 April 1996 @ Vol. 35, No. 11 @ APPLIED OPTICS 1857

and scatterer concentration 1Cs2, and the diffuse- image-forming light 1path 3, Fig. 32 is almost com-
attenuation coefficient K1l, r, j, Cs2, typically a func- pletely caused by scattering, and the lack of strong
tion of wavelength 1l2, position within the medium 1r2, absorption 1path 42 leads to a significant amount of
direction of view 1j2, and scatterer concentration 1Cs2. veiling light 1path 22 at the camera. The intensity of
Since we use only narrow-band light, hereafter we the veiling light makes detection of the image-
suppress the wavelength dependence of these param- forming light 1path 12 difficult, and hence the quality
eters in the notation. In addition, we are concerned of images obtained in the milk@water mixtures is
with the diffuse-attenuation coefficient in the direc- deteriorated.
tion of propagation only, so dependence of K on the
direction j is suppressed as well.
5. Comparison of Polarization-Sum and
The beam-attenuation coefficient c is a measure of
Polarization-Difference Images Obtained at Varying
the rate of loss of intensity of a narrow, collimated
Scatterer Concentrations
beam in the direction of propagation. It is defined
by the relation In our earlier study, PDI was described and con-
trasted with a conventional 1PS2 imaging system
1 dI with the target positioned 1.5 attenuation lengths
c52 , 152 from the front of the tank.6,12 In the present study
I dz we compare the quality of PD and PS images ac-
where z is the direction of propagation of the beam, quired with various amounts of milk added resulting
dz is the extent of an infinitesimal volume element in in effective distances of from 0 to 5 attenuation
the z direction, I is the intensity of the collimated lengths.
The images presented in Fig. 5 are PS and PD
beam, and dI is the intensity lost over the length
images of the target 1Fig. 2A2 suspended in diluted
dz.11
milk. These images were obtained at effective tar-
The reciprocal of the beam-attenuation coefficient,
get distances of 0.13 1Ref. 132, 2.8, and 4.8 attenua-
c21, provides a natural length scale in a scattering
tion lengths. For all six images in Fig. 5 the target
medium. This measure is the attenuation-length
was in the same physical location in the tank, but as
constant of the medium4 and is the distance at which milk was added, the attenuation-length constant of
the intensity of the beam is attenuated by a factor of the medium decreased, and hence the total effective
1@e. Throughout this paper we report the distance length of the scattering medium between the target
between the target and the front face of the tank in and the front of the tank increased. At 0.13 attenu-
attenuation lengths in order to indicate the strength ation length both the left and right patches are
of the scattering medium, which we vary systemati- clearly visible in both the PS and PD images 1panels
cally. A and B2. At 2.8 attenuation lengths the right patch
While the beam-attenuation coefficient character- is no longer visible in the PS image 1panel C2 and the
izes the rate of loss of image-forming light, the left patch is barely discernible, whereas both patches
diffuse-attenuation coefficient K characterizes the are still clear in the PD image 1panel D2. At 4.8
total change of irradiance along the line of sight as a attenuation lengths, both patches are essentially
function of position within the medium. This coeffi- invisible in the PS image 1panel E2 but are still visible
cient is defined by the relation in the PD image 1panel F2.
In the PD images presented in Fig. 5 the relative
1 d E 1r2 amount of polarized light forming the images of the
K1r2 5 2 , 162
E 1r2 dz target patches is quite small. The degree of linear
polarization, a function of the field quantities, is a
where E 1r2 represents the scalar irradiance at posi- measure of the fraction of light intensity that is
tion r, and z is the position along the direction of linearly polarized. We measured the average ob-
view.11 In our experiment it is assumed that the served degree of linear polarization 1ODLP2 on the
scattering medium is homogeneous and that any images of the target patches. For a particular
effects of the tank walls are negligible in the region region on an image this quantity is equal to14
of interest, i.e., the center of the tank. Under these
assumptions, K is independent of position in the 7PD I1x, y28region
7ODLP8region 5 , 172
region of interest. 7PS I1x, y28region
Panels A and B of Fig. 4 show the experimental
setup used to estimate the values of the beam and where the brackets, 7I8region, indicate an average of
the diffuse-attenuation coefficients. For the milk@ the variable I over the region. The 7ODLP8 is the
water mixtures we used, the values of c and K are fraction of light intensity in a given region that is
shown in panels C and D of Fig. 4. We find that c is linearly polarized in the direction of the PDI axes 16
approximately 3 orders of magnitude larger than K and '2. With our choice of axes, 7ODLP8 should be
for each milk concentration used in this study. This very close to the actual degree of linear polarization,
large difference in magnitude shows that the milk@ since the scratches are parallel and perpendicular to
water mixture is a highly scattering medium but one the PDI axes. For the images obtained at an effec-
that absorbs relatively little light. Thus, loss of tive distance of 0.13 attenuation length 1panels 5A

1858 APPLIED OPTICS @ Vol. 35, No. 11 @ 10 April 1996

Fig. 4. Measurement of approximate values of c and K: A. Schematic of the setup used to measure the beam-attenuation coefficient.
A He–Ne laser 1632.8 nm2 emits a pencil beam, which is attenuated by the scattering medium. The photodiode is used to measure
intensity, and Eq. 152 is used to solve for c. Attenuation length is c21. B. Schematic of setup used for measuring the diffuse-
attenuation coefficient in the direction of propagation. The sources are placed on one side of the tank, and the CCD camera is placed on
the other side. The CCD camera is calibrated with the photodiode. Intensity measurements are made with various amounts of milk
added to the 13.5-L tank full of water, and Eq. 162 is used to estimate K. C. Beam-attenuation coefficient. As the amount of milk added
to the tank increases, the reliability of the measurement decreases because we approach the lower limit of the photodiode sensitivity.
The points plotted are the values of c calculated from the experimental data; the line is fitted through the points corresponding to milk
additions of less than 8 mL. The fitted line has the form c 5 3.6 3 1milk in milliliters2 1 1.8 1m212. To determine the effective length 1in
attenuation lengths2 of the scattering medium, multiply the beam-attenuation coefficient c by the path length within the medium between
the target and the front face of the tank, i.e., 0.075 m 1effective length 5 c 3 7.5 3 10222. In this study, for values of added milk greater
than 8 mL, we used the extrapolated values of c along the fitted line to determine the effective length. D. Diffuse-attenuation
coefficient. The fitted line has the form K 5 33.8 3 1milk in milliliters2 1 0.64 3 1023 1m212. Notice that c is approximately 3 orders of
magnitude greater than K. This indicates that a large amount of veiling light is being scattered into the camera, thus degrading the
images.

and 5B2 the left patch has 7ODLP8L 5 0.0302 and the caused by increased or decreased scattering by the
right patch has 7ODLP8R 5 20.0256. Since the target patches. Contrast is often referred to as the
water with no milk added is not highly scattering, we magnitude of an intensity variation relative to the
expect that the average degree of linear polarization average intensity in an image.16 Contrast is a
at the target surface is of the same order. For the quantity that is not defined locally—it depends on
images obtained at 2.8 attenuation lengths 3Figs. 5C the intensity distribution across the entire scene.
and 5D4, 7ODLP8L 5 0.0101 and 7ODLP8R 5 20.0136. For the PS images of Fig. 5, such a measure would
For the images obtained at 4.8 attenuation lengths not adequately describe the signals arising from the
3Figs. 5E and 5F4, 7ODLP8L 5 0.0096 and 7ODLP8R 5 scratched patches, since a comparison between the
intensity that is due to the patches and the average
20.0068. These values of 7ODLP8 show that our
intensity across the screen has no clear meaning.
PDI system can render target features visible when
Instead, we want to quantify the incremental change
less than 1% of the captured light intensity is in intensity that is due to scattering at the patch
linearly polarized. The images of many object sur- relative to what the intensity would have been, had
faces in natural environments predictably may have the patch not been present.
ODLP’s of considerably higher magnitude.15 The curves in Fig. 6 are numerical plots of the
The ODLP quantifies the amount of polarized light average intensity taken across the bands through
available for PDI. For comparing PDI with conven- the center of the images in Fig. 5 1see the caption of
tional imaging, we need to compute a measure of the Fig. 6 for details2. By examining panels A, C, and E
local change in intensity in the PS images that is of Fig. 6, we see that the patches give rise to small

10 April 1996 @ Vol. 35, No. 11 @ APPLIED OPTICS 1859

Fig. 5. PS and PD images after transformation for optimal Fig. 6. Intensity plots across the center of the images presented
display 3Eqs. 1324 at different effective lengths. The images in the in Fig. 5. We obtained the plots by averaging ten vertically
left-hand column are the PS images, the images in the right-hand adjacent pixels within the band shown in Fig. 5 for each horizontal
column are PD images. Panels A and B were obtained at a position in the display. The plots corresponding to the PS images
distance of 0.13 attenuation length 1no milk added to the tank2. 1A, C, and E2 show clearly the intensity gradients that are
Panels C and D were obtained at 2.8 attenuation lengths 110 mL of corrupting the images, i.e., the drop from the background to the
milk2. Panels E and F were obtained at 4.8 attenuation lengths disk face and the gradient across the disk. The plots of the PD
117.5 mL of milk2. The PD images in the right column were images show that these variations were removed. The large
transformed with the maximum and the minimum values on the noiselike variations in the PD line plots are largely due to
screen in original PD images, while the PS images were trans- quantization. Despite this, the areas corresponding to the
formed with the maximum and the minimum intensities on the scratched patches can clearly be made out in the PD intensity
disk face alone. We obtain the untransformed images as follows: plots.
the vertically and the horizontally polarized images are obtained
by adding 128 consecutive images into the 16-bit accumulator and
dividing the resulting sum by 8. This results in a linear amplifi-
cation factor of 16. This procedure is used because the narrow- the blank target shown in Fig. 2B, we compute the
band filter eliminates the majority of the light, and since the gain local contrast, defined as
mechanism on the camera is highly nonlinear, we could not com-
pensate for the loss of intensity by increasing the gain. There- 7PS I1x, y2target8 2 7PS I1x, y2blank8
fore we set the gain to its minimum setting at which linearity is Cl1x, y2 5 2 , 182
verified. The linearly polarized images are then filtered with a 7PS I1x, y2target8 1 7PS I1x, y2blank8
spatial low-pass filter with an effective linear width of ,5
pixels. This filter is used to eliminate an artifact of the image- where 7PS I1x, y2target8 is the ensemble average of the
processing system. We then obtain PS images by adding the two value of PS I at the position 1x, y2 over several images
raw images together and dividing by 2 and obtain PD images by of the target. For the PS images obtained at 0.13
subtracting the horizontally 1'2 polarized image from the verti- attenuation length the average value of the local
cally 162 polarized image and adding an offset of 128 to the
contrast in the area of the image plotted in Fig. 6 is
difference. The offset was used to compensate for the negative
values introduced by the subtraction operation because such 7Cl8L 5 0.0751 for the left patch and 7Cl8R 5 0.0239 for
values cannot be displayed on our system. The bands across the the right patch. For the images obtained at 2.8
centers of the images represent the area within which the attenuation lengths, 7Cl8L 5 0.0113 and 7Cl8R 5 0.0036.
intensity plots displayed in Fig. 6 are taken. For the images obtained at 4.8 attenuation lengths,
irregularities in the data prevented calculation of
perturbations from a large intensity gradient that is the average perturbation contrast, but from visual
due to the shadow of the target. By comparing inspection, 7Cl8L , 0.005 and 7Cl8R , 0.002. These
numerical plots of the images of the scratched target values show that even when 7Cl8 and 7ODLP8 are
shown in Fig. 2A with numerical plots from images of comparable, the PD images of the target patches

1860 APPLIED OPTICS @ Vol. 35, No. 11 @ 10 April 1996

appear qualitatively better than the PS images wanted variations to be eliminated, differential-
obtained in scattering media of varying strengths. mode amplification makes the small variations cre-
In the remainder of this report we examine three ated by the target patches more visible. In any
aspects of PDI that contribute to the better visibility imaging system, the dynamic range is fixed, and this
of the patches in the PD images than in the PS finite range, in our case 256 discrete levels, limits
images of Fig. 5, enabling us to see target features our ability to apply a large stretching factor gps as is
that produce ODLP’s of less than 1%: 112 the com- shown in Eq. 142. In order to increase the stretching
mon-mode rejection@differential mode amplification factor, one must first decrease the range of intensi-
that is inherent to PDI but absent in conventional ties contained in an image. This is done by elimina-
imaging systems, 122 the polarization signatures of tion of unwanted common-mode variations. In the
targets that can be revealed by PDI but are not images used to generate panels A, C, and E of Fig. 5,
available to conventional imaging, and 132 the quality gps varied between 8 and 13; in panels B, D, and F, gpd
of the signals in the PD and the PS images arising varied between 23 and 65. In the images presented
from the target patches. in Fig. 5, the PS images are transformed with the
maximum and the minimum intensities on the disk
6. Common-Mode Rejection face alone to obtain the stretching factors 3Eq. 1424.
This means that any information contained in the
A. Unpolarized Background and Veiling Light area of the image outside the disk was eliminated a
Common-mode rejection@differential-mode amplifica- priori in transforming the PS images. Had the
tion is a signal-processing strategy used in electron- stretching factors been calculated with the maxi-
ics to extract small signals that would otherwise be mum and the minimum intensities over the entire
masked by other signals having larger amplitude.17 image, gps would have been even smaller, and the
In general, the smaller signal that is being sought scratched patches would be even less visible to the
out has a different frequency content 1either tempo- human eye in the transformed PS images 1data not
ral as in electronics applications or spatial as in our shown2. In contrast, the PD images were trans-
experiments2 than does the larger signal. In Fig. 5 formed with the maximum and the minimum over
there are two relatively large-amplitude intensity the entire frame. Despite the fact that we are
variations reducing the visibility of the small signals ignoring a significant area of the PS images in
created by the light scattered from the scratched transforming them, a larger stretching factor gpd can
target patches: 112 the abrupt change in intensity be applied to the PD images than to the PS images.
from the background to the disk face that is due to Introducing a large stretching factor is not useful
the shadow of the disk and 122 the large-amplitude if the intensity of the corrupting variations18 is high
intensity gradient across the shadow face of the disk. when compared with the signal intensity. Our expe-
The first intensity variation renders visible the disk rience has shown that the PS images tend to be at
itself, but it constrains our ability to optimize the use least as corrupted, if not more so, than corresponding
of the dynamic range of the display by forcing us PD images. As long as the relative intensities of the
either to use a small stretching factor 3gPS, Eqs. 1424 or signal and corrupting variations are comparable in
to saturate portions of the images. The second the PS and PD images, the larger stretching factors
variation tends to overwhelm the local contrast 3Eq. used in the PD images make the targets more visible
1824 produced by the target patches. In addition to in the resulting transformed images.
these two gradients, as the effective length of the
scattering medium increases, more veiling light is
scattered into the camera, causing the images to B. Partially Polarized Backgrounds and Veiling Light
become further degraded. Since the background The illumination conditions of the experiment of
light, the light from the sandblasted portions of the Figs. 5 and 6 were designed so that the veiling light
disk face, and the veiling light are unpolarized under and background gradients were largely unpolarized.
this illumination condition, the variations just de- In general, this condition is not necessary for PDI to
scribed are nearly identical in both linearly polar- be useful. As long as the polarization states of the
ized 16 and '2 images. Subtracting the images background and the veiling light are nearly uniform,
removes these gradients, leaving the signals that are a judicious selection of the orthogonal 6 and ' axes
different in the two linearly polarized images. enables the effects of corrupting variations to be
Panels A, C, and E of Fig. 6 clearly show the removed by common-mode rejection. A working
variations described earlier. In panel 6A, the pertur- model of a system designed to select a proper set of
bations introduced by the target patches are large axes automatically was developed by one of the
enough to be seen on the numerical plot, but in authors.8 This system is called the automatic rota-
panels 6C and 6E, the gradient across the disk face tional polarizer 1ARP2 and has been used to eliminate
makes it virtually impossible to detect the target the effects of partially polarized background varia-
patches. Panels B, D, and F of Fig. 6 do not show tions in certain circumstances. The ARP can be
the masking gradients. These plots show a rela- used to select an appropriate set of axes under two
tively flat background with the signals introduced by different conditions: 112 if the background polariza-
the target patches superimposed. tion state is nearly uniform or 122 if the direction of
While common-mode rejection enables the un- partial polarization remains constant, even if the

10 April 1996 @ Vol. 35, No. 11 @ APPLIED OPTICS 1861

degree varies. If either of these conditions is met,
and the area of the target is small compared with the
total image area,19 the ARP rotates the analyzer
until the averaged projections of the background
polarization state onto the two orthogonal directions
are nearly equal.20 This ensures that gradients in
the background illumination are suppressed in the
final PD image.
To demonstrate that PDI is useful when the back-
ground intensity variations are partially polarized,
we modified the setup of Fig. 1A so that a single
projector illuminates the target from the left of the
tank as seen by the camera. The modified setup is
shown in Fig. 1B. Since the direction of illumina-
tion is perpendicular to the direction of view, the
background and the veiling light seen by the camera
is partially vertically polarized.21 With this modi-
fied setup the direction of partial polarization across
the frame is constant 1vertical2 even though the
degree of polarization decreases from left to right
across the scene. The target disk is rotated by 45°
about the line of sight for reasons that become
apparent shortly, and the images shown in this
section are obtained at an effective distance of 4.1
attenuation lengths. Fig. 7. Side-illuminated images transformed for optimal display.
With the setup modified as just described, the The target is rotated by 45° about the line of sight so that the
transformed PS image shown in Fig. 7A is obtained signals created by the patches are not subtracted out in the PD
with the analyzer oriented vertically. The plots in images after the analyzer is rotated. A. PS image. Notice
Fig. 7B show that the intensity decays across the that the upper-left patch is visible in this PS image, but the
image from left to right. Panels 7A and 7B reveal lower-right patch is practically invisible. B. Intensity plots
that the upper-left patch is barely visible and the across the bands shown in panel A. There is a pronounced decay
of intensity across the screen that masks the presence of the
lower-right patch is practically invisible. The decay-
signal caused by the patches. C. PD image created with the
ing-intensity gradient is an important factor contrib- analyzer oriented vertically. D. Intensity plots across the bands
uting to our inability to see the target patches. In shown in panel C. Since the light is predominantly vertically
addition, there is partially polarized veiling light polarized, the decaying intensity gradient is not common to both
that further degrades the image. raw images. Therefore it is not subtracted out in the final
By forming a transformed PD image with the image. E. PD image created with the analyzer oriented at
analyzer oriented vertically, we obtain the image 45°. F. Intensity plots across the bands shown in panel E.
shown in panel 7C. This PD image is not an The decaying intensity gradient is suppressed, and both target
patches are easily visible.
improvement over the PS image. In fact, the patches
are even less visible in panel 7C than they are in
panel 7A. There are two reasons for this: 112 since comparing panel 7E to panels 7A and 7C, we see that
the target is rotated by 45°, the orientation of the PDI successfully suppresses the effects of the par-
scratched patches falls exactly between the horizon- tially polarized background. The patches are clearly
tal and the vertical directions, making detection more visible in panel 7E than they are in either
difficult even in the absence of scatterers, and 122 panel 7A or 7C, and the intensity gradient that is
since the light being scattered by the milk is par- seen in both panels 7B and 7D is eliminated in panel
tially vertically polarized 162, the decaying-intensity 7F. For the images obtained with the analyzer
gradient and the veiling light seen in the PS image oriented vertically 1panels A and C2, the 7ODLP8 over
are mostly contained in the vertically polarized 162 the entire screen is 0.0920, and the partial polariza-
image. The gradient across the horizontally polar- tion across the image masks the signal produced by
ized 1'2 image is weaker, so subtracting I' from I\ the patches. In contrast, for the PD image shown in
does not eliminate the intensity gradient, i.e., the panel E with the analyzer rotated, the 7ODLP8 over
gradient is not common to both linearly polarized the screen is reduced to 0.0023, and the signals
images. produced by the patches with 7ODLP8L 5 0.0155 and
In order to realize the common-mode rejection 7ODLP8R 5 20.0172 are visible.
potential of PDI, we use the ARP to select a new set
of polarization axes, and we obtain the image shown 7. Sign of the Polarization-Difference Signal
in panel 7E. The final orientation of the analyzer An important feature of PDI is the ability to differen-
used in obtaining panel 7E was 45° from the vertical; tiate between image regions having different polar-
thus 6 refers to 45° and ' refers to 245°. By ization signatures. In both panels A and B of Fig. 5

1862 APPLIED OPTICS @ Vol. 35, No. 11 @ 10 April 1996

the two target patches are easily discernible, but patches is imaged, with its performance on nonsig-
there is no clear difference between the left and the nal trials, in which a target without scratched patches
right patches in the PS image. In the PD image it is is imaged. To make this comparison, we used the
apparent that the light scattered by these two areas two targets depicted in Fig. 2. The target with
have different polarization states than that of the scratched patches shown in Fig. 2A was present on
light scattered by the unscratched portions of the the signal trials, and the blank target shown in Fig.
disk, and the two patches can be distinguished from 2B was present on the nonsignal trials 1hereafter
each other on the basis of this different polarization. called blank trials2.
The difference between these two images arises The fundamental principle behind SDT is that a
because conventional imaging systems detect inten- decision is based on a random variable calculated on
sity only and do not represent the polarization cues both signal and blank trials. The probability distri-
that differentiate the two patches. PDI, on the bution of the random variable on blank trials is of
other hand, is a two-state system that preserves noise only, and on signal trials the distribution is of
information about both states, therefore making it the signal plus noise. These probability distribu-
possible to distinguish between targets if the polar- tions limit the performance of the observer. The
ization states of the light scattered from them are greater the amount of overlap between the noise and
different. This is demonstrated by the light patch the signal-plus-noise distributions, the worse the
on the left side of the images in Fig. 5, which is observer performs.22 The sensitivity index d8a, shown
partially vertically polarized 1parallel to the ana- in Eq. 192 below, is a measure of the separation of the
lyzer2, in contrast to the dark patch on the right, signal-plus-noise distribution and the noise-alone
which is partially horizontally polarized 1perpendicu- distribution in decision space. It is defined as
lar to the analyzer2.
µs 2 µn
d8a 5 Œ2 , 192
8. Intrinsic Detectability of Signals Arising from the
Target Patches
Πs2s 1 s2n

A. Signal-Detection Theory where µs and s2s are the mean and the variance of the
We maintain that PDI is qualitatively better than signal-plus-noise distribution and µn and s2n are the
conventional imaging systems for rendering visible mean and the variance of the noise distribution.22
certain target features in scattering media. The As d8a increases, the proportion of correct decisions is
common-mode rejection capability inherent to PDI expected to increase for any given decision process
plays a significant role in this improvement of visibil- since the amount of overlap between the distribu-
ity. Although common-mode rejection is effective at tions is decreasing.
removing global variations and creating an overall Most SDT experiments measure the performance
increase in visibility, application of a variety of of an observer under different conditions. From
image-processing algorithms, such as the Laplacian these measurements, d8a is determined, and the
operator, might confer the same advantages on con- underlying distributions are inferred.22 In our inves-
ventional imaging. A fundamental question re- tigation, however, we directly measure variables
mains: does PDI have any unique capabilities for that contribute to the decision-making process, esti-
imaging? A partial answer was given in Section 7, mate their probability distributions, and use these
in which we pointed out that polarization signatures data to calculate d8a for PS and PD images like those
of targets would always be invisible to conventional shown in Fig. 5.
imaging. In this section we provide a more compre- Figure 8 illustrates an example of a SDT experi-
hensive answer by comparing quantitatively the ment that could be carried out to test the decision
information available locally in PS and PD images process that we study. In such an experiment a
for making decisions about the presence of target human observer would be presented with two con-
features. This approach neutralizes the global ad- secutive images and would be asked to identify the
vantage of common-mode rejection and focuses in- one having the scratched patches. For the images
stead on the quality of the signals available for presented in Fig. 8 the decision would likely be more
unequivocally identifying the presence of the target reliably made based on the PD images 1panels C and
patches in PS and PD images. The statistical@ D2 rather than the PS images 1panels A and B2.
analytical method we employ to make the compari- Although we do not perform this specific experiment,
son is known as signal-detection theory22 1SDT2. the analysis that follows is based on the principles of
SDT characterizes the performance of an observer SDT and describes the factors that allow a human
1man or machine2 carrying out the decision task of observer to locate the scratched patches in PD
detecting a target in the presence of noise.22 Here, images more reliably than in PS images like those
the task is to determine whether there are scratches shown in Fig. 8.
at a given location on the face of the target presented
in an image in a fixed interval of time, called a trial. B. Experiment
Characterization of a system’s ability to make such a At each of several milk concentrations corresponding
decision requires comparison of its performance on to effective distances of 0 to 5 attenuation lengths,
signal trials, in which a target having scratched we collect 30 images of the signal target shown in

10 April 1996 @ Vol. 35, No. 11 @ APPLIED OPTICS 1863

 The spatial units must be small when compared with
the size of the target patches so that the effects of the
patch edges are minimized in the final analysis.
In order to have as large a unit as possible while still
meeting criteria 132 and 142 above, we used a spatial
unit size of 5 3 5 pixels. Since the target patches
are roughly 78 3 90 pixels, there are approximately
280 spatial units within each target patch.
To determine d8a for a given spatial unit, we use
Matlab 1Mathworks, Inc., version 4.2c2 to carry out
the calculations. For each of the 30 PS and PD
images we import the numerical intensity arrays
corresponding to the two linearly polarized 16 and '2
images and convolve them with a two-dimensional,
spatial-low-pass filter to eliminate an artifact of the
imaging system. The filtering process was intro-
duced in order to improve the visual quality of the PS
and the PD images, and it was retained in this
experiment for consistency.24 After filtering, each
array is normalized to its minimum value in order to
Fig. 8. Example of a SDT experiment designed to test our
decision process. In such an experiment an observer would be
compensate for variations in intensity from long-
presented with two stimuli in succession and would be asked to term source fluctuations. After applying Eqs. 112
determine which was the signal stimulus and which was the and 122 to create PS and PD images, respectively, we
nonsignal stimulus. This paradigm is referred to as two- determine the average intensity of the 25 pixels in
alternative forced choice because the observer must choose one of the spatial unit for each of the 30 PS and PD images.
the two stimuli. Panels A and B are the PS images, and panels C We then calculate the mean and the variance across
and D are the PD images. These images were created at a the 30 images, plot histograms of the average inten-
distance of 4.8 attenuation lengths. It is clear that panel C is the sity, and obtain d8a using Eq. 192. Results for a single
signal stimulus for the PD images, but it is not at all clear whether
spatial unit are illustrated in Fig. 9.
panel A or panel B is the stimulus for the PS images. 1It is in fact
panel A2. Even though the quality of the image in panel C does
The particular spatial unit considered in Fig. 9 lies
not seem as good as the images in Figs. 5B and 5D, it is still near the middle of the right 1horizontally scratched2
obvious that the target shown in panel C of this figure has target patch. The data were obtained at an effec-
scratched patches, whereas the target shown in panel D does not. tive distance of 2.9 attenuation lengths. For the PS
images, the histogram data from the signal and the
blank trials are bunched together so that no clear
Fig. 2A over a period of 2 h using the setup described separation exists between the points corresponding
in Section 3 and depicted in Fig. 1A.23 These im- to the blank trials and the points corresponding to
ages comprise the signal trials. Next, we replace the signal trials. This indicates that the average
the signal target with the blank target shown in Fig. normalized pixel intensities on signal and blank
2B and collect 30 more images over an additional 2 h. trials in the PS case are almost statistically indistin-
With the exception of the targets, all other conditions guishable for this spatial unit. Furthermore, the
under which the two sets of images are collected Gaussian curves in Fig. 9 clearly illustrate the
remain unchanged. The order of collection of the statistic d8a. The Gaussians fitted to the noise-alone
two sets of 30 images is reversed from day to day. distribution and the Gaussian fitted to the signal-
The statistics we measure for this decision task plus-noise distribution for the PS case are nearly
are the spatially averaged intensities in particular identical 1d8a 5 0.112.
spatial units on the target. Several factors dictate In contrast, the histogram data for the PD case are
the size of the unit to consider: 112 In order to clustered, which indicates that the average normal-
achieve statistical reliability, we would like to have ized pixel intensity in this spatial unit is likely to be
as large a unit as possible. 122 The spatial units greater on the blank trials than on the signal trials.
must be large enough to ensure that the sources of This separation is also demonstrated by the fitted
noise are approximately uncorrelated between units, Gaussian curves, which are clearly distinguishable
allowing us to treat the average intensity in each in decision space 1d8a 5 21.982. If this increased
spatial unit as an independent random variable. statistical separation of the average normalized pixel
132 To evaluate the intrinsic detectability without the intensity in the PD images over the PS images,
effects of common-mode variations, we must pick a characterized by the magnitude of d8a, is shown to be
spatial unit that is small compared with the extent of consistent across the target patches, we can under-
the intensity gradients described in Section 6. By stand why an observer can more reliably distinguish
taking small enough units, we can assume that the the scratched patches in PD images than in PS
effects of the intensity gradients are negligible. 142 images such as the ones shown in Fig. 8.

1864 APPLIED OPTICS @ Vol. 35, No. 11 @ 10 April 1996

Fig. 9. Determination of d8a for one spatial unit of an image. The average intensity within the 5 3 5 pixel unit on each of the 30 PS and
PD images is calculated for both the signal and the nonsignal cases. These data are collected and plotted as a histogram that is
normalized to yield an integrated value of unity. The mean and the standard deviation are calculated for each of the histograms, and,
from these values, d8a is determined for the PS and PD images. The PS images have d8a 5 0.11, and the PD images have d8a 5 21.98.
The minus sign in the value of d8a for the PD case is there because the right patch is darker 1more negative2 in the signal trials than in the
blank trials. The curves plotted through the histograms are normal distributions with the appropriate means and standard deviations.
These data were obtained from the 5 3 5 pixel region surrounding the point 1314, 2432 3the image is 512 3 480 pixels, and the point 11, 12 is
in the upper-left corner of the image4 in the PS and PD images obtained at 2.8 attenuation lengths, which are shown in panels C and D of
Fig. 5. The position pointed out at the top of this figure is the approximate location of the spatial unit considered.

To show that the statistical separation of the the sensitivity index as a function of position. They
average value of PDI 3Eq. 1224 is consistent across the are not intensity plots. The large-amplitude inten-
entire PD image, we compare the overall perfor- sity gradients shown in Figs. 5 and 6 are functions of
mance of the two imaging systems by spatially the illumination conditions and are not affected by
mapping the sensitivity index d8a within a rectangu- the presence or the absence of the scratched patches,
lar region surrounding the center of the target. The which only give rise to small perturbations from
region is subdivided into 5 3 5 pixel spatial units, these gradients. Since the illumination conditions
and the process described above is used to calculate are the same for all images obtained at each effective
d8a in each of these units. In Fig. 10 we plot distance, the effects of common-mode variations that
three-dimensional surface maps of the value of d8a as would be seen in an intensity map are not present in
a function of unit position for both PS and PD images the sensitivity-index maps of Fig. 10.
at varying effective distances. The images shown While small spatial units are needed to obviate the
in Fig. 5 are taken from the set of images used to effects of the illumination-induced intensity gradi-
generate the values of d8a mapped in Fig. 10. As was ents, an intelligent observer would base a decision
shown in Fig. 9, the higher the magnitude of d8a, the about the presence or the absence of the target
smaller the amount of overlap between the noise and patches on the combined signals from neighboring
the signal-plus-noise probability distributions in the regions of the target surface. If the decision were
decision space. Examination of Fig. 10 verifies that being made automatically, it would be based on a
the spatial sensitivity-index maps highlight the ar- statistic combining a finite sample of independent
eas that correspond to the scratched patches in the observations. In either case a comparison must be
images of Fig. 5. We emphasize that the maps in made between the values of d8a for the regions
Fig. 10 are of d8a 3Eq. 1924; i.e., they show the value of corresponding to the patches and the value of d8a for

10 April 1996 @ Vol. 35, No. 11 @ APPLIED OPTICS 1865

Fig. 10. Maps of d8a in the shaded region shown at the top of the figure. The rectangular region shown shaded at the top of the figure is
245 3 135 pixels. It is completely on the face of the disk and includes the center of the disk, both scratched patches, and a significant
area of the disk face outside the scratched patches. It is subdivided into 5 3 5 pixel spatial units; the process depicted in Fig. 9 is carried
out for each spatial unit, and the resulting values of d8a are plotted as a function of position. The maps correspond to the images in Fig. 5.
Panels A and B are the sensitivity-index maps of the PS and PD images obtained at 0.13 attenuation length. Panels C and D show the
maps of d8a for the PS and PD images obtained at 2.8 attenuation lengths. Panels E and F are the maps of the PS and PD images
obtained at 4.8 attenuation lengths. Maps were formulated with Matlab 1see text2. Panels A and B are presented with the same d8a
scale, as are panels C and D. Panels E and F are presented on different scales, but note that the range of the sensitivity index presented
in each of these panels has the same magnitude 13.52. We are currently unable to explain the slope in panel E, and we believe that it may
be an artifact. However, it is interesting to note that the map in panel F, obtained from the same data as the map in panel E, shows no
sign of any artifact.

1866 APPLIED OPTICS @ Vol. 35, No. 11 @ 10 April 1996

the background region, i.e., the region known not to of the spatial sensitivity index of the variable
contain the scratched patches. For each effective d8a, D 81d8a2, is calculated as
distance the averaged value of d8a over the patches is
7d8a8p 2 7d8a8f
0Π0
computed and compared with the average value of d8a
in an area that immediately surrounds the patches. D 81d8a2 5 Œ2 , 1102
These background regions have approximately the s2p 1 s2f
same total area as the patches and are diagrammed
where the brackets, 7 8, indicate a spatial average,
in the insets of Fig. 11. From these data a measure
the p and the f subscripts represent the patch and
the surrounding background frame, respectively, and
s2p and s2f are the variances of d8a over the patch and
the surrounding background frame. The results of
these calculations are presented in Fig. 11. In Fig.
11 the value of D 81d8a2 is shown at each effective
distance for both patches in the PS and the PD
images. As the effective length of the scattering
medium increases, the value of D 81d8a2 decreases in
all four plots. Exponential curves are fitted to the
calculated values of D 81d8a2 1see the caption of Fig. 11
for details2. For both target patches these fitted
curves intercept the vertical axis at approximately
the same value of D 81d8a2 in the PS and the PD cases,
but the value of D 81d8a2 decays faster in the PS
images than it does in the PD images as the effective
length of the scattering medium is increased.

C. Analysis of Signal-Detection-Theory Data

By examining the maps presented in Fig. 10 and the
curves in Figs. 9 and 11, one can see how the
performance of PDI compares with conventional
imaging systems. The most obvious feature of the
PD maps in Fig. 10, like the corresponding images in
Fig. 5, is that they show positive values for the left
target patch and negative values for the right target
patch. This again demonstrates the ability of PDI
to represent different polarization states. This fea-
ture alone indicates that PDI preserves information
Fig. 11. Spatial sensitivity index of the variable d8a, D 81d8a2,
that conventional imaging does not.
computed at each effective distance for both types of images. A second important feature of PDI that stands out
Panel A presents the data from the left target patch; panel B in Fig. 10 is that the distribution of d8a in the PD
presents the data from the right target patch. At all times, PDI images has a consistent shape over attenuation-
performs as well or better than the conventional 1PS2 system. In length variation. At low effective target distances,
addition, the PS images degrade faster as a function of scatterer both patches are clearly detectable in both the PD
concentration. The exponential curves fitted to the data indicate and the PS images, and the spatial variance of the
that if any particular value of D 81d8a2 is chosen as the threshold of sensitivity index on the patch and on the background
detectability, then the PD system will reach this level at distances is small compared with the difference in mean.
at least double that of conventional 1PS2 systems. The exponen-
However, as the milk concentration is increased, the
tials have the form D 81d8a2 5 A exp12x @x2, where x is the effective
length of the scattering medium and x is the decay constant. For
detectability of the patches degrades more rapidly in
the left target patch, APD < APS 5 5.5, and xPD 5 6.3 while xPS 5 the PS images than in the PD images. The right
3.2. For the right target patch, APD < APS 5 5.1, and xPD 5 7.1 target patch is practically undetectable in panel 10C,
while xPS 5 2.3. The points shown as diamonds correspond to while it is clearly detectable in panel 10D. In
measurements at 1.40 attenuation lengths. Examination of the addition, the variance of the background seems to be
PS sensitivity-index map for this single effective distance indi- smaller in the PD sensitivity-index maps than it is in
cates that errors may have occurred during the collection of the PS sensitivity-index maps. An increase in the
images. Furthermore, these points lie well below the curve fitted spatial variation of the sensitivity index d8a on the
to the data 1excluding the points corresponding to 1.40 attenua- background and on the patch means an increase in
tion lengths2 in the PS case, especially for the right patch. For
the rate of misses and the rate of false alarms that
this reason, the points shown as diamonds corresponding to the
single case of 1.40 attenuation lengths were left out in the
would be reported by an observer.22 Examination of
exponential fitting; however, they are presented here for complete- the maps for the PD images at large effective target
ness. Note that the points from the PD images obtained at 1.40 distances shows a remarkable resistance to this
attenuation lengths 1open diamonds2 lie much closer to the fitted increase in spatial variance. As the number of
curve. attenuation lengths between the target and the front

10 April 1996 @ Vol. 35, No. 11 @ APPLIED OPTICS 1867

of the tank is increased, an overall degradation of the B. Mechanisms Underlying
PS maps is apparent before any such effect is seen in Polarization-Difference-Imaging Robustness
the corresponding PD maps. Figure 11 shows clearly One mechanism that contributes to the robust qual-
that at low effective target distances the statistical ity of PD images is the ability of PDI to suppress the
distributions of d8a on the patches and the surround- effects of background light with slowly varying polar-
ing frames are at least as separated in the PD ization state. Previously, we have referred to this
images as they are in the PS images. In addition, as common-mode rejection, a term borrowed from
the separation of these distributions decays more signal processing. In Fig. 5 we demonstrated the
slowly in the PD case than in the PS case as effective elimination of unpolarized light. In Fig. 7 we
distance is increased. Figure 11 shows that if any showed that the common-mode rejection feature can
given value of D 81d8a2 is chosen as a threshold of also be used to suppress the effects of partially
detectability, D 81d8a2 reaches this value in the PD polarized background light. This feature has en-
images at effective lengths 2–3 times greater than in abled us to see small signals 17ODLP8 , 1%; Figs. 5D
the PS images. This detectability measure is ob- and 5F2 in PD images whereas signals of comparable
tained in a way that suppresses the effect of the magnitude 1local contrast of ,1%; Figs. 5C and 5E2
common-mode variations discussed above and is are invisible in PS images.
based on the analysis of raw PS and PD images, i.e., We hypothesize that a second mechanism contrib-
before any transformation for display. uting to the increased distance at which the target
patches can be detected with PDI may be polariza-
9. Discussion
tion-dependent properties of the point-spread func-
In this investigation we have qualitatively and quan- tion 3or, alternatively, the modulation transfer func-
titatively compared the performance of PD and tion 1MTF24 of the medium. Kuga and Ishimaru
conventional 1PS2 imaging in optically scattering have studied the MTF in optically scattering media
media as we approach the threshold of visibility. similar to ours.25,26 However, the polarization-
The images presented in this paper have shown that dependent properties of the MTF are probably impor-
the targets can be seen in PD images even when tant. Since researchers have shown that the MTF
their 7ODLP8 3Eq. 1724 is less than 1%. In contrast, for radar images of the sea surface are polarization
the targets are virtually invisible in the PS images at dependent27,28 and that the optical transfer function
the same effective distances even when the local of a lensing system with polarization masks is
contrast 3Eq. 1824 is comparable to the 7ODLP8. Table dependent on the polarization state of the incident
1 summarizes the results presented in this study for light,29 there is reason to believe that the point-
the backilluminated conditions. spread function of the milk@water mixture used in
this study may be polarization dependent as well.
A. Robustness of Polarization-Difference Images
We are currently investigating the possibility that
An important feature of practical visual systems is the point-spread function for PDI may be narrower
the ability to produce consistent images of the same than the point-spread function for conventional imag-
scene under varying external conditions. For the ing systems, implying that higher-resolution images
scattering medium used in this study our investiga- can be obtained.25,30
tion shows that PDI is capable of producing qualita-
tively good images of the scratched target patches
under backillumination and side-illumination condi- C. Polarization-Difference Imaging in Conjunction with
tions and at varying effective distances. Thus, al- Conventional Imaging
though the PD images in panels B, D, and F of Fig. 5 Throughout this paper we have emphasized the
and panel F of Fig. 7 are different, they are qualita- important and useful features of PD images. How-
tively similar in form. In contrast, the correspond- ever, PS images also contain important information
ing PS images do not show the same consistency. not in the PD images. In Fig. 5 the disk face is more
This robustness of PDI is further emphasized by a detectable with conventional imaging techniques
comparison of the numerical plots in Figs. 6 and 7 than with PDI. This is because the disk was de-
and the maps of the sensitivity index shown in signed so that the light reflected from its un-
Fig. 10. scratched portions would be unpolarized, causing it

Table 1. Summary of the Numerical Results for the Backillumination Condition Organized by Attenuation Lengtha

Left Patch Right Patch

Attenuation
Length 7ODLP8 Cl D P8 S D P8 D 7ODLP8 Cl D P8 S D P8 D

0.13 0.030 0.075 4.93 5.63 20.026 0.024 4.12 4.75

2.8 0.010 0.011 3.34 4.27 20.014 0.004 2.10 3.74
4.8 0.009 ,0.005 1.04 2.63 20.007 ,0.002 0.39 2.36
aThe quantity 7ODLP8 is the observed degree of linear polarization 3Eq. 1724, C is the local contrast 3Eq. 1824, and D 8 is the spatial
l
sensitivity index 3Eq. 11024. The patches are labeled left and right based on their positions in Fig. 5.

1868 APPLIED OPTICS @ Vol. 35, No. 11 @ 10 April 1996

to blend into the unpolarized background light in the with other image-processing techniques. The algo-
PD images. We would not recommend replacing rithm presented in this paper can be implemented
conventional imaging systems with PDI because to without any advanced methods, and in the present
do so would eliminate such information. Instead, study, sophisticated algorithms have intentionally
we are showing that PDI, like color vision, is capable not been developed in order to preserve this simplic-
of providing information about a scene that is not ity. Standard processes and more complicated algo-
available with conventional total-intensity methods. rithms that are typically used to enhance conven-
Previously, studies were undertaken that incorpo- tional images can also be used with PDI. Moreover,
rated polarization information into a visual dis- PDI utilizes common-mode rejection, which is one of
play.31–36 Walraven31 and Garlick et al.32 developed the most important features for producing rapid,
systems that mapped functions of the Stokes param- recognizable images near the threshold of detectabil-
eters into a monochrome display. Walraven31 and ity and providing useful information that can be
Solomon33 developed transformations designed to used in conjunction with intensity information ob-
map Stokes parameters into red-green-blue 1RGB2 tained by conventional imagery. Hopefully, future
display. Garlick et al.,32 Halaijan and Hallock,34 development in hardware design and display proce-
Wolff and Boult,35 and Egan et al.36 map pdI@psI or dures will make PDI a feasible option for use in
some form of percent polarization and angle into the real-time target-recognition applications in scatter-
display. In contrast, in PDI we display polarization- ing media.
difference intensity alone, which allows common-
mode rejection@differential-mode amplification to be We thank Fred Letterio and Michael Suplick of the
used optimally for mapping PD signals. Unlike Institute of Neuroscience Machine Shop for techni-
percent polarization maps, PD images are orthogo- cal advice on the design of the apparatus as well as
nal to PS images over the ensemble of all possible the construction of the elements of the system.
images in the same sense that the color pathways in This work was supported by the U.S. Department of
the human visual system are orthogonal when aver- Navy, Office of Naval Research, under grant N00014-
aged over all possible images.37 Furthermore, the 93-1-0935 and in part by the University of Pennsylva-
experiments presented in this paper examine the nia Research foundation. J. S. Tyo is supported by
performance of PDI near the limit of target detection a National Science Foundation Graduate Research
in optically scattering media. When the results Fellowship and a U.S. Air Force Institute of Technol-
that we have presented for PDI are considered along ogy Sponsored Fellowship.
with other techniques suggested by previous investi-
gators, there is every indication that incorporation of
References and Notes
PD information into the visual display of images
could significantly improve the detectability of tar- 1. S. Q. Duntley, ‘‘Underwater visibility and photography,’’ in
Optical Aspects of Oceanography, N. G. Jerlov, ed. 1Academic,
gets near and below current thresholds.
London, 19742, pp. 138–149.
D. Potential Applications 2. L. Wang, P. P. Ho, C. Liu, G. Zhang, and R. R. Alfano, ‘‘Ballistic
2-D imaging through scattering walls using an ultrafast Kerr
Although we have tested PDI in a manner simulat- gate,’’ Science 253, 769–771 119912.
ing an underwater environment, we believe that PDI 3. A. Yodh and B. Chance, ‘‘Spectroscopy and imaging with
should be useful in any application in which scatter- diffusing light,’’ Phys. Today 48132, 34–40 119952.
ers serve to blur the image and the available targets 4. S. Q. Duntley, ‘‘Light in the sea,’’ J. Opt. Soc. Am. 53, 214–233
reflect light with a different polarization signature 119632.
than the background. Under water and in fog, 5. D. A. Cameron and E. N. Pugh, ‘‘Double cones as a basis for a
suspended particles act as scatterers. In both of new type of polarization vision in vertebrates,’’ Nature 1Lon-
these situations, dielectric and conducting targets don2 353, 161–164 119912.
6. M. P. Rowe, E. N. Pugh, Jr., J. S. Tyo, and N. Engheta,
may be present. Dielectric targets give rise to
‘‘Polarization-difference imaging: a biologically inspired tech-
specular and diffuse reflections that have different nique for observation through scattering media,’’ Opt. Lett.
polarization signatures.35 Conducting and dielec- 20, 608–610 119952.
tric targets may also polarize light through a mecha- 7. M. P. Rowe, N. Engheta, S. S. Easter, Jr., and E. N. Pugh, Jr.,
nism similar to the one discussed in the caption of ‘‘Graded-index model of a fish double cone exhibits differen-
Fig. 2. Additionally, PDI could be useful for applica- tial polarization sensitivity,’’ J. Opt. Soc. Am. A 11, 55–70
tions in which the state of polarization of the inci- 119942.
dent light is under the control of the experimenter, 8. J. S. Tyo, ‘‘Automatic rotational polarizer for the polarization
such as in microscopy, medical imaging, and detec- differencing camera,’’ undergraduate senior design project
tion of airport runway approach lighting that has 1Moore School of Electrical Engineering, University of Penn-
sylvania, Philadelphia, Pa., 19942.
been specifically polarized. By use of PDI in conjunc-
9. J. N. Lythgoe, ‘‘The adaptation of visual pigments to the
tion with existing techniques such as time-resolved photic environment,’’ in Handbook of Sensory Physiology,
imaging, additional benefits may be realized. H. J. A. Dartnall, ed. 1Springer-Verlag, New York, 19712, Vol.
VII@1, pp. 566–603.
E. Additional Benefits of Polarization-Difference Imaging
10. We do not explicitly discuss the effects of forward-scattered
Other potentially useful features of PDI are that it light and multiple scattering. For simplicity we assume that
can be very fast and that it can be used in parallel light that is truly forward scattered acts as image-forming

10 April 1996 @ Vol. 35, No. 11 @ APPLIED OPTICS 1869

 light. Furthermore, light that is scattered out of the imaging that the TNLC has had time to relax to its unexcited state
path and then back into it acts as veiling light. after collection of a pair of linearly polarized images, one
11. H. G. Gordon, ‘‘Modeling and simulating radiative transfer in needs to wait approximately 3 min after turning off the
the ocean,’’ in Ocean Optics, R. W. Spinrad, K. L. Carder, and excitation before capturing a subsequent pair of images.
J. J. Perry, ed. 1Oxford U. Press, New York, 19942, pp. 1–46. Typically, collection of the 30 images took just less than 2 h.
12. As a reference, for Atlantic Ocean water in the vicinity of 24. The filter is designed by hand to be used with the Imaging
Gibraltar at a depth of 25 m for light of wavelength 465 nm, Technologies, Inc., software. It is 7 3 7 pixels, is peaked in
the attenuation length is 20 m.4 The wavelength of light the middle, has a summed value of unity, and falls off to a
used in our study is 610 nm. A wavelength of 610 nm was value of 10.125 max2 in 3 pixels.
chosen only because the TNLC was designed to operate at 610 25. Y. Kuga and A. Ishimaru, ‘‘Modulation transfer function and
nm. There is no relative advantage to using light of wave- image transmission through randomly distributed spherical
length 610 nm, and the results presented here should be particles,’’ Appl. Opt. 25, 2330–2335 119852.
generalizable to any wavelength of electromagnetic radiation. 26. Y. Kuga and A. Ishimaru, ‘‘Modulation transfer function of
We performed studies using white light in the present setup layered inhomogeneous random media using the small-angle
with nominally similar results 1data not shown2. approximation,’’ Appl. Opt. 25, 4382–4385 119862.
13. When there is no milk in the tank, the water and the particles 27. C. Brüning, R. Schmidt, and W. Alpers, ‘‘Estimation of the
suspended in the water act as scatterers. This is why the ocean wave–radar modulation transfer function from syn-
beam-attenuation coefficient is not equal to zero when there is thetic aperture radar imagery,’’ J. Geophys. Res. C 99, 9803–
no milk. 9815 119942.
14. Using the notation of Stokes parameters, 7ODLP8 can be 28. A. Schmidt, V. Wismann, R. Romeiser, and W. Alpers, ‘‘Simul-
written as S1@S0, where S0 and S1 are the first and the second taneous measurements of the ocean wave–radar modulation
Stokes parameters. It is inherent in Eqs. 112 and 122 that transfer function at L, C, and X bands from the research
PS I 5 S0 and PD I 5 S1. For discussion of Stokes parameters, platform Nordsee,’’ J. Geophys. Res. C 100, 8815–8827 119952.
see, for example, C. H. Papas, Theory of Electromagnetic Wave 29. K. Bhattacharya, A. Ghosh, and A. K. Chakraborty, ‘‘Vector
Propagation 1Dover, New York, 19882, pp. 118–134. wave imagery with a lens masked by polarizers,’’ J. Mod. Opt.
15. G. P. Können, Polarized Light in Nature 1Cambridge U. Press, 40, 379–390 119932.
London, 19852, pp. 74–99. 30. A. Ishimaru, Wave Propagation and Scattering in Random
16. R. Shapley and C. Enroth-Cugell, ‘‘Visual adaptation and Media 1Academic, New York, 19782, pp. 301–306, 448–454.
retinal gain controls,’’ Prog. Retinal Res. 3, 263–346 119842. 31. R. Walraven, ‘‘Polarization imagery,’’ in Optical Polarimetry:
17. A. S. Sedra and K. C. Smith, Microelectronic Circuits 1Saun- Instrumentation and Applications, R. M. Azzam and D. L.
ders College, Ft. Worth, Tex., 19912, pp. 48–114. Coffeen, eds., Proc. SPIE 112, 164–167 119772.
18. Rather than using the term noise, we use the term corrupting 32. G. F. J. Garlick, C. A. Steigmann, and W. E. Lamb, ‘‘Differen-
variations. This refers to all variations other than the tial optical polarisation detectors,’’ U.S. patent 3,992,571 116
signals produced by the scratched patches. In the PD case November 19762.
this is mostly noise that is due to the system and the medium. 33. J. E. Solomon, ‘‘Polarization imaging,’’ Appl. Opt. 20, 1537–
In the PS case we are referring to noise as well as the 1544 119812.
unwanted intensity variations described earlier. 34. J. Halaijan and H. Hallock, ‘‘Principles and techniques of
19. In our study the target is thought of as the scratched target polarimetric mapping,’’ in Proceedings of 8th Symposium on
patches, and the background includes the unscratched por- Remote Sensing and Environment 1Environmental Research
tions of the disk as well as the area outside the disk. Institute of Michigan, Ann Arbor, Mich., 19722, Vol. 1, pp.
20. The ARP was developed as an undergraduate senior design 523–540.
project by J. S. Tyo.8 We plan to test its performance under a 35. L. B. Wolff and T. E. Boult, ‘‘Constraining object features
variety of conditions, and we hope to present the details of its using a polarization reflectance model,’’ IEEE Trans. Patt.
operation in a subsequent paper. Anal. Mach. Intell. 13, 635–657 119912.
21. T. H. Waterman, ‘‘Polarization sensitivity,’’ in Handbook of 36. W. G. Egan, W. R. Johnson, and V. S. Whitehead, ‘‘Terrestrial
Sensory Physiology, H. Autrum, ed. 1Springer-Verlag, New polarization imagery obtained from the Space Shuttle:
York, 19812, Vol. VII@6b, pp. 283–311. characterization and interpretation,’’ Appl. Opt. 30, 435–442
22. N. A. Macmillan and C. D. Creelman, Detection Theory: 119912.
A User’s Guide 1Cambridge U. Press, London, 19912, pp. 7–28. 37. G. Buchsbaum and A. Gottschalk, ‘‘Trichromacy, opponent
23. In the present arrangement of the PDI system, the TNLC colours coding and optimum colour information transmission
limits the speed with which the system performs. To ensure in the retina,’’ Proc. R. Soc. London Ser. B 220, 89–113 119832.

1870 APPLIED OPTICS @ Vol. 35, No. 11 @ 10 April 1996