Color-Accurate Underwater Imaging Using Perceptual Adaptive

Color-Accurate Underwater Imaging Using
Perceptual Adaptive Illumination
Iuliu Vasilescu, Carrick Detweiler, and Daniela Rus
Distributed Robotics Laboratory, CSAIL
Massachusetts Institute of Technology
Cambridge, MA 02139
Email: iuliuv@tiaresearch.com, {carrick, rus}@csail.mit.edu
Abstract— Capturing color in water is challenging due to the
heavy non-uniform attenuation of light in water across the visible
spectrum, which results in dramatic hue shifts toward blue.
Yet observing color in water is important for monitoring and
surveillance as well as marine biology studies related to species
identification, individual and group behavior, and ecosystem
health and activity monitoring. Underwater robots are equipped
with motor control for large scale transects but they lack sensors
that enable capturing color-accurate underwater images. We
present a method for color-accurate imaging in water called
perceptual adaptive illumination. This method dynamically mixes
the illumination of an object in a distance-dependent way
using a controllable multi-color light source. The color mix
compensates correctly for color loss and results in an image
whose color composition is equivalent to rendering the object in
air. Experiments were conducted with a color palette in the pool
and at three different coral reefs sites, and with an underwater
robot collecting image data with the new sensor.
I. I NTRODUCTION
Underwater monitoring of natural as well as humanengineered environments can be greatly enhanced by automated capture of underwater color images using underwater
robots [8, 13]. Our goal is to develop color-accurate automated
imaging of underwater environments using robots capable of
capturing, mosaicing, mapping, and analyzing color images of
the space. In our previous work [12] we have described a robot
capable of such a task. Figure 7 shows this underwater robot at
the end of an underwater color imaging mission. In this paper
we discuss an algorithm and instrument that enables robots to
image objects such as coral heads and wrecks by automatically
collecting a succession of underwater color-accurate images.
Color encodes important information and plays a key role in
underwater object identification, monitoring and surveillance,
and marine biology—for example a healthy sea anemone is
pigmented by zooxenthella, while an unhealthy sea anemone
is bleached.
Current cameras are not able to capture color in water
at distances greater than 1m, which makes color-dependent
underwater studies challenging [6]. Part of the challenge
for underwater imaging is that color is unavailable as an
accurate data parameter for distances greater than 1m due
to the uneven attenuation of the color spectrum as light
travels through water. Longer wavelengths (e.g., red light)
attenuate more rapidly than shorter wavelengths (e.g., blue
light), resulting in dramatic hue shifts toward blue. Absorption,
Fig. 1. Comparison of underwater imaging using white flash and adaptive
illumination in a coral reef setting and a wreck setting. The color row in
the middle shows color patches extracted from these images for visualization
without the scene context. The left image shows a clown-fish and sea anemone
in Moorea with white flash (Top) and adaptive illumination (Bottom). The
pictures were taken at 20m depth by Melissa Holbrook Schmitt. The scene
was at a distance of 3m. The right image shows a coral formation on a
ship wreck in the Grand Cayman with existing white flash (Top) and adaptive
illumination (Bottom). The pictures were taken at 15m depth by Cathy Church.
The scene was at a distance of 3m.
the primary cause for color loss in water, is exponential with
respect to the propagation distance. The exponential varies
according to the wavelength of the light [9]. Other factors
such as scattering contribute to image loss but are substantially
independent of the wavelength of the visible light and do
not affect color. Current approaches to underwater imaging
rely on flooding objects with white light from a very close
distance (e.g. less than 0.5m), using color filters, and doing
manual post-processing. These techniques are cumbersome, do
not render accurate colors, and work only for fixed setups, as
color loss is distance-dependent. For example, Figure 1(Top)
shows underwater images taken with current techniques at 3m
distance.
In this paper we describe a new method for color-accurate
imaging in water called perceptual adaptive illumination. Our
method uses knowledge about the physical processes that
lead to the color shift. It computes light loss given distance
to the imaged object and compensates for light loss using
an illuminant whose radiation spectrum is controlled to be
approximately the inverse of the water transfer function. The
human color perception models are used for computing the
optimal radiation spectrum. The full dynamic range of the
camera is used and post-processing is not necessary. The end
result is the color-accurate presentation of the object’s image to
the camera’s CCD. Color-accurate imaging renders the object
as if the image was taken in air, without the color loss side
effects of water.
Figure 1 shows the comparison between images taken in
water, using white flash and using perceptual adaptive illumination. The top pictures were taken with a white Xenon flash.
The bottom figures were taken using the perceptual adaptive
illumination method described in this paper. The color row in
between the pictures shows color patches extracted manually
from the images, presented without the scene context.
Underwater imaging with perceptual adaptive illumination
requires a controllable illumination source capable of estimating distances. We have developed an instrument, algorithm, and software for capturing color-accurate images underwater by adaptive illumination which includes a spectrumcontrollable light source. The light composition is calculated
so that this composition is transformed into white light by the
water between the camera and the subject. The energy content
of the light source is calculated using the optical properties of
water and the distance to the subject. The light is generated
by a source composed of several filtered Xenon light bulbs.
Varying the relative power of the Xenon flashes effectively
results in a light source with variable spectrum. Intuitively, the
device senses the distance of the object and mixes the light of
a multi-color flash so as to compensate for each component
of the color spectrum according to known physics of how that
particular light frequency dissipates in water. Distance sensing
is accomplished with an acoustic sonar, by using the distance
information from the camera’s auto-focus system or can be
manually entered by the operator.
We present data from a suite of pool experiments and
ocean experiments using our system. We use a color palette
and the L*a*b* color metric for evaluating color accuracy.
Our experiments demonstrate color-accurate imaging in one
image plane at distances ranging from 1m to 5m and at
depths ranging from 5m to 30m. Distance to the imaged
object was measured in three different ways: using a measuring
tape, using an external distance sensor, and using auto-focus
information from the camera. We have attached the imaging
apparatus to our underwater robot (see Figure 7(Right)) and
used it to collect underwater images. We do not report in detail
on the underwater robot missions because the focus of the
paper is the underwater imaging system.
A. Related Work
Color plays a very important role in underwater monitoring
and surveillance, as well as in marine biology studies related
to species and behavior identification [2]. Many recognized
species of coral reef fishes exhibit two or more color variants
[7]. A recent study [1] found that fish use color to communicate.
Current approaches to color imaging underwater rely on
flooding objects with white light from close distances (e.g. less
than 0.5m [3]), possibly followed by post-processing [15, 11].
The post-processing step approximates the color of the image
by manually setting a white point and correcting the image
uniformly so that the selected point appears white. Since the
color attenuation is not uniform this technique will not render
accurate colors. These methods address the blue color shift by
making assumptions about the properties of the picture (e.g.
tagging known white patches in the picture). They do not use
any knowledge of the physical process that led to color shift.
When objects of known color are present in the picture,
various statistical or learning approaches have been applied
[11]. Another option is to add a color filter to the camera
to block much of the blue light, correcting for the loss of
red [3]. Since absorption is distance-dependent, a different
filter is needed for each distance. All these current techniques
are cumbersome, do not render accurate colors, and produce
color approximations for fixed distances and configurations
only. What is needed is a method for adaptively illuminating
an underwater scene with the correct light mix for each
distance and depth. This will achieve color-accurate images
without post processing. Accurate coloring in water refers to
the equivalence between the color spectrum that reaches the
imaging device in water to the color spectrum for the same
object in air.
A recent paper by Yamashita et al. [15] considers the
problem from a fundamental perspective and models the water
effects on the color using the wavelength-dependent absorption coefficients and distance between camera and subject.
The computed inverse function is applied to the captured
image. The function is only a very coarse approximation
that takes into account only the effect on three particular
light wavelengths instead of the continuous spectrum. This
simplification limits the performance and is not scalable to
significant distances.
B. Outline
This paper is organized as follows. Section II introduces the
physics of color perception in water. Section III discusses color
perception by the human visual system and its implications
to computation. Section IV presents the adaptive illumination
algorithm. Section V describes the adaptive illumination instrument. Finally, Section VI presents our experimental data
and evaluation.
II. C OLOR P ERCEPTION IN WATER
In this section we discuss the physics of color perception
in water and the intuition behind how knowledge about colorabsorption in water can be used to do color-accurate imaging
in water.
There are two important phenomena that affect imaging in
water: scattering and absorption. Scattering is the physical process whereby light is forced to deviate from straight trajectory
the 530nm green light; thus, the light source should output 5
times more power at 650nm than at 530nm. In general the
spectral output power of the light source should be:
Spectral radiance W*m−2*nm−1
2.5
2
← air
1.5
← 1m
Iec (λ, dsc ) =
1
← 2m
0.5
← 5m
0
400
450
500 550 600 650
wavelength (nm)
700
750
Fig. 2. The simulated spectrum of sunlight after it travels through 1m, 2m,
and 5m of water. The simulation uses published sun spectrum and published
light attenuation coefficients of water [9].
6
10
air
1m
2m
3m
5
Spectral radiance W*m−2*nm−1
10
4
10
3
10
2
10
1
10
0
10
−1
10
400
450
500
550
600
wavelength (nm)
650
700
= ID65 (λ)e2aw (λ)dsc
Such a light source compensates exactly for the light loss
at the specified camera to subject distance dsc . The subject
appears as though illuminated by mid-day Sun light (D65
illuminant in air).
The usefulness of the brute-force method that relies on
Iec (λ, dsc ) is very limited in practice. Fabricating a light
source with the required spectral power distribution is challenging as the amount of optical power needed rises very
sharply with distance (e.g. at 3m the power required to take the
picture in water is over 105 times the power required to take
the picture in air). In addition, existing light sources generally
have fixed output spectra, which implies the use of filters to
generate the required spectral distribution. This increases the
power requirements by another order of magnitude.
Alternatively, we can exploit the scope of human color
vision to save power.
750
Fig. 3. The light required to compensate for color absorption in water. This
is the inverse of water light transfer function plotted in Figure 2. The data is
presented on a logarithmic scale due to its high dynamic range.
as it travels through water. The effect is caused predominantly
by solid particles suspended in water, but also by the water
molecules and substances dissolved in water. Scattering is
wavelength independent and does not affect the color balance
of the underwater image. This paper is not concerned with
scattering.
Absorption is caused by water molecules and dissolved
substances which convert light energy into heat. The water absorption coefficient, aw , is wavelength dependent. Within the
visible spectrum, longer wavelengths are attenuated stronger
than shorter wavelengths (see Figure 2). Absorption is responsible for the color shift in underwater imaging. The
absorption law describing the light energy transmitted by water
is exponential.
I(λ, d) = I(λ, 0)e−aw (λ)d
where:
λ
I(λ, 0)
aw (λ)
I(λ, d)
ID65 (λ)
−2a
e w (λ)dsc
light wavelength
spectral power distribution at source
water absorption coefficient
spectral power distribution at
distance d from the source, in water
The distance light travels through water (and is attenuated)
is double the distance between subject and camera dsc .
One solution is to compensate exactly for light absorption
(see Figure 3). For example, given the camera to subject
distance, the 650nm red light is attenuated 5 times more than
III. C OLOR P ERCEPTION BY THE H UMAN V ISUAL S YSTEM
A light source that can compensate exactly for light loss
in water yields color-accurate pictures (at least for relatively
planar scenes, where the distance between the subject and the
camera or light source is relatively constant). However, this
method is both wasteful in terms of power and resources, and
not necessary as the human vision system is not a “perfect”
light sampler. This section discusses the physics of the color
vision system of humans. We use the properties of this system
to create a low complexity and power efficient light source for
color-accurate underwater imaging.
Creating the exact inverse of water attenuation will concentrate the majority of the energy in the deep red part of the
spectrum where the water attenuation is the highest. However,
it turns out that the human eye sensitivity is very low in
this part of the spectrum. Very little extra accuracy would be
achieved by exactly compensating for the deep red attenuation.
Exact compensation is not necessary since the eyes (and
cameras) sample the spectrum with only three discrete color
sensors, collapsing the infinite-dimensional color space into
a three dimensional space. To preserve the human-perceived
color accuracy, it is only necessary to preserve the color coordinates for naturally occurring colors in the three dimensional
color space of humans. This allows for some tolerance in the
light source spectrum.
Under typical conditions the objects around us are illuminated by a white light source I0 (λ), which contains energy
at all visible spectrum wavelengths. The color of an object is
given by its wavelength-dependent reflection coefficient R(λ)
which represent the fraction of incident power that is reflected.
Our eyes receive the reflected light as shown in Figure 4.
The spectral distribution of the reflected light is given by the
equation: E(λ) = R(λ)I0 (λ).
0.5
S (λ)
ρ
1
0.4
Sγ(λ)
0.35
Sβ(λ)
0.8
0.3
normalized sensitivity
Reflected Light E(λ) W*m−2*nm−1
0.45
0.25
0.2
0.15
0.1
0.05
0.6
0.4
0.2
0
400
450
500
550
600
wavelength (nm)
650
700
750
0
400
450
Fig. 4. Reflected spectral power of red color sample when lighted by a
standard light (CIE illuminant D65).
500
550
600
wavelength (nm)
650
700
750
Fig. 6. Normalized sensitivities of the three types of cone cells present in
human retina (ρ, γ and β) [10].
1.8
1.6
normalized response
1.4
←X
1.2
1
Y→
0.8
0.6
←Z
0.4
0.2
0
400
450
500
550
600
wavelength (nm)
650
700
750
Fig. 5. The human visual system color response as modeled by the CIE
1931 2◦ Standard Observer [4]. The 3 curves model the response of the three
types of cone cell in human retina.
The human vision system perceives colors by sampling the
light spectrum with three different types of cone cells (ρ, γ and
β). Each type of cell is sensitive to a different region of the
visible spectrum — long, medium and short wavelengths. The
three normalized sensitivity curves Sρ (λ), Sγ (λ) and Sβ (λ)
are plotted in Figure 6. The responses of these three types of
cells Tρ , Tγ and Tβ are used by the central nervous system
to associate colors to objects. For example, in the case of the
Munsell 5Y 8/10 color sample, illuminated with white light the
responses of the three types of cone cells can be calculated
as:
Tρ5Y 8/10
Z
750nm
=
Sρ (λ)E(λ)dλ =
380nm
Z
750nm
Sρ (λ)R5Y 8/10 (λ)I0 (λ)dλ
380nm
Tγ5Y 8/10 =
Z
The brain receives a tuple of three numbers (Tρ , Tγ , Tβ ) and
associates colors to objects based on this tuple. This results in
an apparent three dimensional color space for humans. Any
two objects that generate the same respective response in the
three types of cone cells will appear to human observers as
having the same color. This property can be exploited to reduce
the complexity and power requirements of the underwater light
source.
Using this observation, we develop a light source that preserves the coordinates of most naturally occurring colors in the
human three dimensional color space, when used underwater.
With this approach, the coordinates are also preserved in the
color space of current cameras, since they closely model the
human vision system color space.
Artificial light sources do not have the same spectral
composition as the sun light. Therefore, they do not render
the color of objects in the same way sunlight does. The
Color Rendering Index (CRI) was developed by Commission
Internationale de l’Éclairage (CIE) as a metric for comparing
ability of light sources to render color accurately. The metric
is based on comparing the color coordinates of 15 sample
colors illuminated with the reference light and the tested light.
The color samples have been chosen as representative for
naturally occurring colors. The Euclidean distance between
the color coordinates of two colors represented in CIE 1964
U ∗ V ∗ W ∗ color space [14] is used as a quantitative value for
color difference. U ∗ V ∗ W ∗ is a uniform color space, meaning
the Euclidean distance is a consistent measure of perceptual
color difference, across the entire color space.
Our algorithm selects the power levels of a multi-color
flash to maximize the CRI for objects illuminated in water.
Formally, CRI is defined as:
750nm
Sγ (λ)E(λ)dλ =
CRI = 100 − 4.6∆ĒU ∗V ∗W ∗
380nm
Z
750nm
15
Sγ (λ)R5Y 8/10 (λ)I0 (λ)dλ
380nm
5Y 8/10
Tβ
Z
Z
750nm
=
Sβ (λ)E(λ)dλ =
380nm
750nm
Sβ (λ)R5Y 8/10 (λ)I0 (λ)dλ
380nm
∆ĒU ∗V ∗W ∗ =
∆Ei =
1 X
∆Ei
15 i=1
q
∆Ui2 + ∆Vi2 + ∆Wi2
CRI has values between 0% and 100%. A CRI of 100%
corresponds to a perfect light source which renders colors
identically to the reference light source.
IV. P ERCEPTUAL A DAPTIVE I LLUMINATION
Perceptual adaptive illumination is a power-efficient method
for color-accurate imaging underwater. Given the distance to
the object, the wavelength-dependent attenuation is computed
using the known optical properties of water [9]. The light
source spectrum is adjusted to compensate for this attenuation.
The required light spectra is generated by a variable spectrum
light source composed of several basic light sources, with fixed
but distinct spectral power distributions. The variable spectrum
is achieved by varying the relative power of the composing
light sources. The spectral power distribution is optimized such
that when filtered by the water between the camera and the
subject, it will render the subject’s human-perceived colors in
the same way that natural light would render it in air. The
optimization is performed by maximizing the apparent CRI of
the light source when filtered by the water.
Given the camera-subject distance dsc , adaptive illumination
computes the optimal spectrum that can be generated by a
source composed of n light sources with fixed output spectra
Ii (λ) and independently adjustable output power pi . The total
output spectrum is a linear combination of the composing light
sources:
I(λ) =
n
X
Algorithm 1 Light optimization
Require: Ii (λ, ti ) Output characteristics of the component
light sources
Require: dmax Maximum distance for which the light will
be used
1: for dsc = 0m to dmax in 0.1m steps do
2:
compute t1 (dsc ) for optimal illumination at distance dsc
in air
3:
compute
t2 (dsc)..tn (dsc )
=
P6
arg maxt2 ..tn ∈[0..1] CRI e−2aw (λ)dsc i=1 Ii (λ, ti )
4: end for
pi Ii (λ)
i=1
The spectra of the component light sources can be viewed
as the basis functions that are used to generate the required
spectral power distribution. The scaling factors, pi , correspond
to the output power setting of the component light sources.
The generated light is filtered by the water between the light
source and the subject, and also by water as it travels back
from the subject to the camera. Thus, the total travel distance
is 2dsc , where dsc is the distance between subject and camera.
Water attenuation makes the illumination equivalent to a light
source in air:
ˆ
I(λ)
= e−2aw (λ)dsc I(λ) = e−2aw (λ)dsc
n
X
pi Ii (λ)
i=1
From the camera (and observer) perspective, the subject
ˆ
looks as though it is illuminated by I(λ)
in air. The problem
ˆ
can be posed as adjusting I(λ)
so that the colors are rendered
accurately in the observer’s color space.
The power settings are found by solving the following optimization problem for the subject to camera distance parameter
dsc :
ˆ
p1 ..pn = arg max CRI I(λ)
=
p1 ..pn ∈[0..1]
arg max CRI e
p1 ..pn ∈[0..1]
−2aw (λ)dsc
n
X
!
pi Ii (λ)
i=1
This optimization problem can be solved using any existing
numerical optimization software. Algorithm 1 summarizes the
computation for the optimal light.
Fig. 7. (Left) The adaptive illumination device consists of six colored flashes
with adjustable power. The device is attached to a standard digital camera in
an underwater housing. (Right) Imaging apparatus attached to an underwater
robot at the end of a mission.
ˆ
The range of obtainable Color Rendering Index for I(λ)
depends on the choice of the source light components
I1 (λ)..In (λ). This is restricted by the available light source
technologies and filters, and is discussed in Section V. The
aim is to obtain a CRI of at least 90%, which is equivalent to
the best artificial light sources in air.
V. I MAGING A PPARATUS
We developed and built a light source prototype capable
of adaptive illumination called AquaLight. The light source
can be used as an external flash for a digital SLR camera
(as seen in Figure 7(Left)), as a free-standing illumination
unit, or attached to an underwater robot or sensor node.
Figure 7(Right) shows the apparatus attached to an underwater
robot.
The light source is composed of 6 independently controlled
Xenon strobes. Each strobe is capable of energy discharge
up to 50J. The strobes have different but fixed spectra. One
is unfiltered and capable of full light spectrum illumination.
The other five strobes are filtered with increasingly longer cutoff long pass optical filters. The choice of filters was based
on availability and simulation. We optimized for maximum
corrected light output for objects at practical underwater photographic range (e.g distances dsc between 0m and 5m). The
output spectrum of each flash was measured with a calibrated
radiometric spectrometer. Adaptive illumination is achieved
by controlling the strobes’ output power, or, equivalently, the
energy discharged by each individual flash. Energy discharge
is controlled by timing the flash so as to achieve the desired
CRI.
More specifically, a Xenon discharge lamp is used as the
light source. We chose for a Xenon lamp for its uniform output
power across the visible spectrum, which is very similar to sun
light. The Xenon lamp is capable of the highest instantaneous
output power for its size (thus its widespread use in general
photography). One of the flashes is unfiltered and provides
the power required for the short wavelengths of the spectrum.
Each of the other 5 component light sources has a Xenon
discharge lamp and a long-pass filter. Filters are needed to
generate different spectra for the 6 light sources. Long-pass
filters are used in order to improve the overall efficiency of
the device. Since the attenuation in water increases with the
wavelength, more power is needed at longer wavelengths than
the shorter wavelengths for any distance. The filters were
chosen using simulation and maximization of the CRI and
illumination power. The cut-off wavelengths for the 5 filters
are 475nm, 575nm, 600nm, 600nm, and 630nm.
For the duration of their discharge, the Xenon lamps are
powered by a capacitor bank. Each lamp has an associated
1080µF capacitor, for a maximum discharge energy of 33J.
The capacitors are charged to 350V by an inverter. The entire
unit is powered by a 11.7V, 2.1Ah Lithium-Polymer battery
which provides enough power for approximatively 300 full
power flashes.
An NXP LPC2148 CPU together with Xilinx XC2C256
CPLD control the timing of the 6 flashes. The timing is based
on the distance to the subject and a precomputed lookup
table of distance vs power requirements. In our system, the
distance to subject can be input to the unit through a user
interface (e.g. distance can be set manually using magnet to
set a counter implemented with magnetic switches), camera
auto-focus information, or it can be automatically determined
using a distance sensor such as an ultrasonic pinger. The
data we report in this paper was collected using manually set
distance and auto-focus information using an Olympus camera.
We found that the ultrasonic pinger (Teledyne/Benthos PSA916) performed poorly for our application (swimming pool
conditions as well typical reef environment).
The flash electronics are housed in a water and pressure
resistant enclosure made out of clear acrylic. Two underwater
connectors are placed on the back of the enclosure. One is used
for the electrical connection between the camera and the flash
(for synchronization). The other is used for interface with the
ultrasonic pinger, programming and battery re-charging. The
flash is attached to the camera through a rigid fixture, which
ensures they point in the same direction (see Figure 7).
VI. E VALUATION
Adaptive illumination has been evaluated in a series of the
pool and ocean experiments. The goal was to measure the color
accuracy and to compare the results with existing methods.
ambient
white
processed
adaptive
1
2
3
4
D/2 distance (m)
5
Fig. 8. Results of imaging the color palette in water. The columns correspond
to the distance between the light source and the object: 1m, 2m, 3m, 4m, and
5m (note this corresponds to dsc in the main document. The rows show
the image obtained using ambient light (first row), white strobe (second
row), post-processed white strobe image (third row), and adaptive illumination
(fourth row).
We ran the adaptive illumination optimization algorithm
for distances between 0.1m and 5 m (in 0.1m increments)
and determined the optimal power ratios between the flashes.
Table I shows the relative power of six flashes for illuminating
an object at distance dsc away in water and the resulting CRI.
We note that the predicted CRI is above 95% in all cases. The
best available artificial light sources are halogen incandescent
bulbs which have a CRI of 98% to 100%, while fluorescent
illumination has a CRI of 75%. In Table I F1 is the white
flash which was kept at a constant power. Very little power is
required from F1 relative to the other flashes. F2 is the yellow
flash which contributes to the overall high CRI, but for longer
distances a light red flash would have been more effective.
The choice of filters was done empirically.
The flash power settings for each distance are stored in the
microcontroller’s memory of the adaptive illumination device.
When taking a picture, given the distance to the desired object,
the corresponding power settings are selected from this table.
The experimental methodology consisted of imaging the test
object at distance dsc varying between 1m and 5m (in 1m
increments) using (1) ambient light, (2) white light strobe,
(3) white light strobe and post-processing, and (4) adaptive
illumination. The test object was also imaged in air for groundtruth. The test object was a custom made waterproof color
palette with 15 colors distributed evenly in the color space (see
Figure 8). Distance was determined with a measuring tape and
was set manually using the system’s magnetic switches. We
also used the camera’s auto-focus information to automatically
TABLE I
T HE RELATIVE POWER
OF THE
6
FLASHES AT DIFFERENT DISTANCES AND THE EXPECTED
CRI. N OT THAT THE OBJECT IS AT DISTANCE
D
FROM THE
2
LIGHT SOURCE .
determine the distance to the imaged object. The standard
white flash pictures were post processed with the current state
of the art method: equalization using a manually marked white
sample patch as reference. This method adjusts the R, G, B
channels proportionally to yield the white color on the selected
sample.
Color accuracy was measured by computing the color
distance between manually selected patches of the color palette
and the corresponding patches of the image captured in air.
For each pair of color patches we converted the colors to the
L*a*b* color space [5] and computed the resulting Euclidean
distance. Since the L*a*b* color space was specifically designed to preserve the perceptual color distance, the Euclidean
distance is an accurate representation of the perceptual color
difference. The smaller the distance the better the color accuracy. We used L*a*b* color space due to the difficulty of
measuring CRI experimentally (i.e. procuring the precise color
samples used in CRI’s definition).
Figure 10 shows the average L*a*b* error for all 15 colors
in the palette for each distance. The numerical values for these
averages is shown in Table II. The ambient light performs
the worst as expected. White flash performs well at 1m (and
below), but the performance decreases significantly for greater
distances. The post processed white flash under-performs
adaptive illumination. Its performance decreases steadily with
a lower slope than the white flash. Adaptive illumination has
no significant decrease in quality up to the measured distance.
Figures 9 show experimental data for four colors from the
palette: yellow, red, green and blue. We note that the images
taken with the adaptive illumination method have low constant
L*a*b* distances from ground-truth for all distances (accurate
colors), while the images taken with the other methods have
L*a*b* distances that increase with distance (that is, the
rendering is increasingly inaccurate in color).
As expected, not all the colors are affected in the same way.
(see Figure 9). Red is most affected by water and it is hardly
recoverable by post-processing. Blue is the least affected color.
All methods render blue well. Green and the brown - the most
commonly occurring natural colors are significantly distorted.
This figure also shows distance-dependent L*a*b* values and
the perceived colors the L*a*b* values map to, as computed
by ambient light (top row), white flash (second row), postprocessed white flash (third row), and adaptive illumination
(fourth row).
An additional suite of pool experiments was done using
the camera auto-focus information for automatically setting
F4
0.0285
0.0710
0.1518
1.0000
F5
0.0272
0.0517
0.1425
0.9802
F6
0.0188
0.0302
0.1090
0.6584
CRI
99.26
98.56
97.97
96.16
100
100
NL
80
60
WF
40
PPWF
20
0
1
L*a*b* color distance
F3
0.0405
0.0789
0.1798
1.0000
AF
2
3
4
5
D/2 distance (m)
100
60
WF
40
PPWF
20
AF
2
3
4
5
dsc distance (m)
1
2 3 4 5
dsc distance (m)
1
2 3 4 5
dsc distance (m)
100
NL
80
60
WF
40
PPWF
20
0
1
NL
80
0
1
1 2 3 4 5
D/2 distance (m)
L*a*b* color distance
F2
0.0227
0.0317
0.0421
0.0475
L*a*b* color distance
F1
0.0652
0.0652
0.0652
0.0652
L*a*b* color distance
D
2
1m
2m
3m
4m
AF
2
3
4
5
dsc distance (m)
1
2 3 4 5
dsc distance (m)
NL
80
60
WF
40
PPWF
20
0
1
AF
2
3
4
5
dsc distance (m)
Fig. 9. The L*a*b* color error (left graphs) and visual appearance (right
graphs) for four colors extracted from the color palette (yellow, red, green
and blue), as captured by the four imaging methods: ambient light (blue
curve), white strobe (green curve), white strobe followed by post-processing
(red curve), and adaptive illumination (cyan curve). In the left graphs, the
smaller the error, the more accurate the color. The ambient light method
performs poorly, the white flash and postprocessed white flash methods
degrade with distance, and the adaptive illumination curve remains constant
for all distances. The right graphs present the colors extracted from the color
palette, as captured by the camera. The columns of the matrix correspond
to distance. The rows correspond to the method used (NL for ambient light,
WF is white flash, PPWF is postprocessed white flash and AF is adaptive
illumination).
the distance to the imaged object. The camera used in these
experiments was an Olympus E520 SLR camera. Data from
these tests is very similar to the data from the tests where the
distance was measured externally.
Underwater imaging experiments have also been conducted
in the field at four sites: Fiji, Tahiti, Hawaii, and Grand
Cayman. In each experiment images were taken at measured
distances using the white flash and the adaptive flash. The
images were compared visually. Figure 1 shows typical images
using white strobe and adaptive illumination.
We installed the adaptive illumination device on our underwater robot and commanded the robot to swim around a coral
head taking a dense sequence of color-accurate photographs.
These photographs can be mosaiced to reconstruct the coral
head. The robot experiment demonstrated that color-accurate
underwater images can be gathered automatically.
VII. C ONCLUSIONS
Adaptive illumination computes the color dissipation at a
given distance and compensates for color loss by introducing
the correct color mix for that distance into the scene. Perceptual adaptive illumination was tested in a variety of environments and was shown to significantly outperform existing
methods for underwater photography with respect to image
Dist/2
ambient
white
white post-processed
adaptive
1m
59.4586
14.9747
9.2605
8.3305
2m
52.2793
26.4679
14.3705
11.0851
3m
53.8530
37.4749
18.9246
11.3560
4m
54.2241
43.9231
27.9815
8.7389
5m
58.6122
51.0177
34.8542
9.9966
TABLE II
T HE L* A * B * VALUE AVERAGED
Natural Light
White Flash
Post Processed White Flash
Adaptive Flash
L*a*b* average color distance
80
70
60
50
40
30
20
10
0
1
2
3
D/2 distance (m)
4
15
COLORS .
R EFERENCES
100
90
OVER
5
Fig. 10. L*a*b* average values for 15 colors at 1m, 2m, 3m, 4m, and
5m, as computed by 4 different imaging methods: ambient light (blue curve),
white strobe (green curve), white strobe with post-processing (red curve), and
adaptive illumination (cyan curve).
color accuracy using the L*a*b* metric space for measuring
colors. Color encodes important information about underwater
environments and habitats. This work enables the capture of
color-accurate underwater images and thus opens the door to
automating underwater monitoring and surveillance operations
based on color.
Our work demonstrates that color-adaptive illumination is
effective for color-accurate imaging in water for objects at
a specified or computed distance. Our experimental data
provides support for this claim for imaging objects that
are up to 5m away from the camera. Current methods are
cumbersome and do not produce accurate colors at distances
greater than 2m. Our imaging instrument was prototyped in our
lab using inexpensive components to demonstrate the concept.
The instrument was not optimized for distance. Imaging at
distances greater than 5m is achievable if the instrument uses
more powerful light sources.
Perceptual adaptive illumination renders color accurately for
all the objects at the given distance d in the image. Objects
further away or closer to the camera will not be rendered
correctly because adaptive illumination is distance-dependent.
We are currently extending perceptual adaptive illumination
from color-accuracy in one image plane to color-accuracy in
multiple image planes. We are also working on extensions
from still images to video.
Our current work includes field experiments to mosaic
underwater scenes using robots equipped with the perceptual
adaptive illumination device described in this paper.
[1] D.
Bellwood.
The
’lost’
language
of
fishes.
http://www.coralcoe.org.au/news stories/fishcolour.html.
[2] A.C. Crook. Colour patterns in a coral reef fish: Is background
complexity important? J. of eperimental marine biology and ecology,
217(2), 1997.
[3] Martin Edge. The underwater photographer, third edition: Digital and
traditional techniques (paperback), 2006.
[4] Commission internationale de l’éclairage. Cie 1931 2◦ standard observer, 1931.
[5] Commission internationale de l’éclairage. Supplement no. 2 to cie
publication no. 15, colorimetry, 1976.
[6] Jules Jaffe, Kad Moore, John McLean, and Michael Strand. Underwater
optical imaging: Status and prospects. Oceanography, 14, 2007.
[7] V. Messmer, G.P. Jones, L. van Herwerden, and P.L. Munday. Genetic
and ecological characterisation of colour dimorphism in a coral reef fish.
Evironmental Biology of Fishes, 74, 2005.
[8] Hanumant Singh, Roy Armstrong, Fernando Gilbes, Ryan Eustice, Chris
Roman, Oscar Pizarro, and Juan Torres. Imaging coral i: Imaging coral
habitats with the seabed auv. Subsurface Sensing Technologies and
Applications, 5, 2004.
[9] R. C. Smith and K. S. Baker. Optical properties of the clearest natural
waters (200–800 nm). Appl. Opt., 20:177–184, 1981.
[10] A. Stockman, D.I.A. MacLeod, and N.E. Johnson. Spectral sensitivities
of human cones. Journal of the Optical Society of America, 10:2491–
2521, 1993.
[11] L.A. Torres-Mndez and G. Dudek. A statistical learning-based method
for color correction of underwater images. Research on Computer
Science, 17, 2005.
[12] Iuliu Vasilescu, Carrick Detweiler, Marek Doniec, Daniel Gurdan,
Steffan Sosnowski, Jan Stumpf, , and Daniela Rus. Amour v: a hovering
energy efficient underwater robot capable of dynamic payloads. The
International Journal of Robotics Research, 29(5):547–570, 2010.
[13] Stefan B. Williams, Oscar Pizarro, Martin How, Duncan Mercer, George
Powell, Justin Marshall, and Roger Hanlon. Surveying noctural cuttlefish
camouflage behaviour using an auv. In ICRA, pages 214–219, 2009.
[14] G. Wyszecki. Proposal for a new color-difference formula. JOSA,
53:1318–1319, 1963.
[15] A. Yamashita, M. Fujii, and T. Kaneko. Color registration of underwater
images for underwater sensing with consideration of light attenuation.
Robotics and Automation, 2007 IEEE International Conference on,
pages 4570–4575, April 2007.