The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XL-5/W5, 2015

Underwater 3D Recording and Modeling, 1617 April 2015, Piano di Sorrento, Italy

A NEW COLOR CORRECTION METHOD FOR UNDERWATER IMAGING

G. Biancoa, M. Muzzupappaa , F. Brunoa , R. Garciab , L. Neumann b,c
a

DIMEG, University of Calabria, 87036 Rende (CS), Italy - (gianfranco.bianco, muzzupappa, f.bruno)@unical.it
b
Computer Vision and Robotics Group, University of Girona, 17001, Spain - rafael.garcia@udg.edu
c
ICREA, Barcelona, Spain - lneumann@silver.udg.edu
Commission V

KEY WORDS: Underwater Imaging, Color Correction, Ruderman space, Computational Color Constancy

ABSTRACT:
Recovering correct or at least realistic colors of underwater scenes is a very challenging issue for imaging techniques, since illumination
conditions in a refractive and turbid medium as the sea are seriously altered. The need to correct colors of underwater images or
videos is an important task required in all image-based applications like 3D imaging, navigation, documentation, etc. Many imaging
enhancement methods have been proposed in literature for these purposes. The advantage of these methods is that they do not require the
knowledge of the medium physical parameters while some image adjustments can be performed manually (as histogram stretching) or
automatically by algorithms based on some criteria as suggested from computational color constancy methods. One of the most popular
criterion is based on gray-world hypothesis, which assumes that the average of the captured image should be gray. An interesting
application of this assumption is performed in the Ruderman opponent color space l, used in a previous work for hue correction
of images captured under colored light sources, which allows to separate the luminance component of the scene from its chromatic
components. In this work, we present the first proposal for color correction of underwater images by using l color space. In
particular, the chromatic components are changed moving their distributions around the white point (white balancing) and histogram
cutoff and stretching of the luminance component is performed to improve image contrast. The experimental results demonstrate the
effectiveness of this method under gray-world assumption and supposing uniform illumination of the scene. Moreover, due to its low
computational cost it is suitable for real-time implementation.

1.

INTRODUCTION

Recovering correct or at least realistic colors of underwater scenes

is a very challenging issue for the imaging techniques since illumination conditions in a refractive and turbid medium as the
sea are seriously altered. Progressive attenuation with the depth
of the visible spectrum at the different wavelengths, scattering
and absorption of the incoming radiation from the ambient can
cause relevant modifications of image colors (Schettini and
Corchs, 2010). The need to correct colors of underwater images
or videos is an important task required in all image-based applications like 3D imaging, navigation, image segmentation, documentation, etc. For dissemination purposes, it would be desirable to show 3D models of archaeological sites as represented
on the land, or to improve the vision during underwater exploration and navigation, for example (Mahiddine et al., 2012; Gallo
et al., 2012). The use of imaging devices with special optical filters or polarizers in front of the lens (Schechner and Averbuch,
2007) can attenuate part of the incoming radiation and often alter
the real colors of the scene to be acquired. Moreover, the use
of artificial light sources can introduce more non-uniformities
in the lighting of the scene (Garcia et al., 2002). Many imaging techniques have been proposed in the literature for this task
(Schettini and Corchs, 2010), based on a physical model of the
image formation process (image restoration techniques) or qualitative, subjective criteria to produce a more realistic image (image enhancement methods). In the first case the knowledge of
the medium physical parameters is required while in the other
case some image adjustments can be performed manually (as histogram stretching) or based on some criteria as suggested from
computational color constancy methods (Gijsenij et al., 2011).
One of the most popular criteria is based on the gray-world hy Corresponding

author

pothesis, which assumes that the average of the captured image

should be gray (Buchsbaum, 1980). This assumption allows to
implement automatic methods which achieve good results. In
(Reinhard et al., 2001) an interesting application of this assumption has been presented, by performing hue correction of images
captured under colored light sources in the Ruderman opponent
color space l (Ruderman et al., 1998). This color space, proposed to study the human eye perception of natural images, is
composed of a luminance component l (achromatic) and two opponent color components and , which encode yellow-blue and
red-green chromaticities, as used for color discrimination in the
human color visual mechanism. Until now, l space has been
only used for color transfer applications (Reinhard et al., 2001),
therefore, its ability to correct images taken under colored lighting is not yet deeply investigated, especially in underwater imaging.
In this work we present the first proposal for color correction
of underwater images by using the l color space. In particular, the chromatic components are changed moving their distributions around the white point (white balancing) and histogram
cutoff and stretching of the luminance component is performed
to improve image contrast. Our main goal is to demonstrate the
effectiveness of the color correction in this space for underwater
scenes. Thus the improvement of image contrast and noise reduction are not investigated since we retain that such techniques
are widely studied and proposed in literature, however they can
be applied independently. The experimental results, presented below in this paper, demonstrate the effectiveness of this method under the hypothesis of gray world and uniform illumination of the
scene. These assumptions are acceptable in many real cases for
close-range acquisition in downward direction, like seafloor mapping, or capturing objects in the foreground (underwater photography) or less-complicated scenes where there are not strong light

This contribution has been peer-reviewed.

doi:10.5194/isprsarchives-XL-5-W5-25-2015

25

The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XL-5/W5, 2015
Underwater 3D Recording and Modeling, 1617 April 2015, Piano di Sorrento, Italy

variations. As example of a possible application, we have processed an image set of an underwater finding with our method in
order to obtain the 3D reconstruction by using a structure-frommotion technique. The proposed method requires just a single
image and does not need any particular filters or prior knowledge
of the scene. Moreover, due to its low computational cost it is
suitable for real-time implementation.
The remaining of paper is organized as follow: in section 2 the
main image enhancement methods for underwater imaging are
summarized. To follow, the color alteration problem in underwater environments is described in section 3 and a few details
as background are presented in section 4. The color correction
method is proposed in section 5 and the experimental results are
reported in section 6. Finally, conclusions and future works are
discussed in section 7.
2.

RELATED WORK

In this section the most important works about color correction

of underwater images are presented. We focus on color correction strategies and methods used in the image enhancement techniques, which work on a single image without any knowledge
about physical parameters or medium properties (image restoration methods). Essentially the different methods are composed by
two steps: contrast enhancement and color correction, performed
separately on RGB space or another one. From the analysis of
the State of the Art, it is apparent that do not exist calibrated
datasets of underwater images in order to compare and evaluate quantitatively the different methods from the point of view
of computational color constancy (Gijsenij et al., 2011). In fact,
it is difficult to obtain the ground truth, that is, a reference image
taken with standard illumination to compare the results. Hence,
the performance is evaluated qualitatively by visual inspection.
For a more comprehensive and exhaustive treatise of the methods
applied to improve quality of underwater images we suggest to
refer to (Schettini and Corchs, 2010).
A histogram equalization method in RGB space is presented in
(Shamsuddin et al., 2012). Color correction is carried out through
both auto level adjustment and manual adjustment techniques.
Hence, the authors found that the manual enhanced technique
(done iteratively) is more precise as compared to the auto enhanced technique due to the levels of significance of 1% and 5%.
The images reported in the paper were taken in the shallow water
(at depth from 1 m to 5 m) in a coastal area of Malaysia.
Iqbal et al. (2007) proposed an approach based on slide stretching. Firstly, contrast stretching of RGB algorithm to equalize the
color contrast in the images is used. Secondly, the saturation and
intensity stretching of HSV (Hue, Saturation, Value) is applied in
order to increase the true colors and solve the lighting problem.
The advantage of applying two stretching models is that it helps
to equalize the color contrast in the images and also addresses the
problem of lighting. The results are illustrated with images of
fishes and marine flora captured in the foreground.
The HSV color space has been also used more recently in a twostage method (Ghani et al., 2014), for enhancing of coral images. First, the contrast correction technique is applied to the image with a modified Von Kries hypothesis and stretching it into
two different intensity images at the average value with respect
to Rayleigh distribution. At the second stage, color correction is
performed in the HSV color model by stretching Saturation and
Value (brightness) components. The stretching processes are applied to the histograms of both Saturation and Value at 1% to
99%, respectively.
Bazeille et al. (2006) proposed an automatic algorithm to pre-

process underwater images. It involves several successive independent processing steps which correct non-uniform illumination, suppress noise, enhance contrast and adjust colors. In particular, color correction is achieved by equalizing the mean of each
color-channel. A linear translation of the histogram (for each
RGB channel) is performed by adding to each pixel the difference
between the desired mean value and the mean of the channel. Images taken under different conditions are showed and elaborated.
A different type of equalization is proposed in (Chambah et al.,
2003), by using Automatic Color Equalization (ACE), for automatic live fish recognition. This method merges two computational color constancy principles: gray world and retinex white
patch. Lightness constancy makes stably perceive the scene regardless changes in mean luminance intensity and color constancy
makes stably perceive the scene regardless changes in color of the
illuminant. ACE is able to adapt to widely varying lighting conditions.
An approach based on the fusion principle has been presented
by Ancuti et al. (2012) for enhancement of images and videos
acquired in different lighting conditions. Two inputs are represented by color corrected and contrast enhanced versions of the
original underwater image and also four weight maps that aim to
increase the visibility. Color correction is performed by a white
balancing process able to remove the color casts but also to recover the white and gray shades of the image, based on the grayworld assumption. In particular, the authors propose to increase
the average value estimated with a percentage [0, 0.5].
A very different approach has been used by Petit et al. (2009),
which propose a method based on light attenuation inversion after processing a color space contraction using quaternions. Applied to the white, the attenuation gives a hue vector characterizing the water color. Using this reference axis, geometrical transformations into the color space are computed with quaternions.
Pixels of water areas of processed images are moved to gray or
colors with a low saturation whereas the objects remain fully colored. Thus, the contrast of the observed scene is significantly
improved and the difference between the background and the rest
of the image is increased. Moreover, Principal Component Analysis (PCA) calculation allows to extract a dominant color of the
image, hence, in most cases, the water color provides also good
results in term of color enhancement by using the method described above. The results are showed on images with a large
presence of a water area.
In (Gouinaud et al., 2011) the authors introduce a Color Logarithmic Image Processing (CoLIP) framework. This novel framework expands the logarithmic image processing theory to color
images in the context of the human color visual perception. The
color correction performs a CoLIP subtraction of the color in the
image that should be white to the eye under daylight illumination,
that is, a pixel of the image that is recognized as white under standard illumination. From the LMS space expressed following the
LIP approach, a three-component function is defined (color tone
function). Its first component is the achromatic tone function,
and the last components are the two chromatic antagonist tone
functions. This method has been successfully evaluated on both
underwater images (blue illuminant) and indoor reflected light
images (yellow illuminant).
Color recovery is also analysed by Torres-Mendez and Dudek
(2005) but from a different perspective: it is formulated as an energy minimization problem using learned constraints. An image
is modeled as a sample function of Markov Random Field and the
color correction is considered as a task of assigning a color value
to each pixel of the input image that best describes its surrounding structure using the training image patches. Training images
are small patches of regions of interest that capture the maximum
of the intensity variations from the image to be restored. The im-

This contribution has been peer-reviewed.

doi:10.5194/isprsarchives-XL-5-W5-25-2015

26

The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XL-5/W5, 2015
Underwater 3D Recording and Modeling, 1617 April 2015, Piano di Sorrento, Italy

ages proposed as result regard fishes in foreground or scene in

shallow water. To evaluate the performance of the algorithm the
authors use a dataset (available on a website) of underwater images taken with a professional camera and in most of the cases
they were also enhanced by using a commercial software. They
consider these dataset images as ground truth and apply an attenuation filter to obtain a color depleted image. As metric the mean
absolute residual error in CIELab space is computed between the
ground truth and the color corrected images.
Also dehazing techniques have been applied in underwater imaging. Chiang and Chen (2012) propose a novel dehazing algorithm for wavelength compensation and image dehazing in order
to remove distortions caused by light scattering and color change,
respectively. Energy compensation for each color channel is carried out after water depth estimation to adjust the bluish tone to
a natural color. The algorithm performance is evaluated by utilizing ground-truth color patches and video available on internet. In particular, a board with six color patches is captured in
the close range before diving (ground truth) and at a depth of 5
and 15 m, and as measurement of the error the authors report the
SNR (Signal-to-Noise) ratio. In (Carlevaris-Bianco et al., 2010)
a dehazing algorithm for removing the effects of light scattering
in underwater images is presented. The underwater model takes
into account the absorption as function of both scene depth and
wavelength, which attenuating the red color channel much more
severely than blue or green. Color correction is provided for the
reduction in the red channel caused by absorption. Moreover the
estimation of the airlight (blue in water) includes the effects of
absorption on the airlight. Therefore, when the additive airlight
has removed, the color balance of the imagery is improved, without having to additionally account for absorption effects. In order
to test the color accuracy, multiple images of a color wheel target
are collected at varying distances between 3 and 6 m from the
camera in a fresh water test tank. Then, an image of the color
wheel taken in air is used as the ground truth.
3.

UNDERWATER COLOR ALTERATIONS

The sunlight that penetrates from the air-water interface into the
water is altered from several causes. Firstly, the visible spectrum
is modified with the depth because of the radiation absorption at
the different wavelengths, in particular, the radiations with lower
frequency are more absorbed. Thus, the red components disappear already within about 5 m (shallow waters), the orange at 7.5
m, the yellow in the range of 10-14 m, and the green at about 21
m. A schematic representation of this color loss is illustrated in
Figure 1. It is evident that more depth increases, more the lower
temperature colors disappear, producing greenish-blue scenes. It
needs also to take into account the scattering (backward and forward) and absorption effects due to the particles dissolved and
suspended in the medium, like salt, clay and micro-organisms.
The greater the dissolved and suspended matter, the greener (or
browner) the water becomes. These particles not only alter the
colors but also attenuate and diffuse more the light with consequent decreasing of contrast and increasing of blur and noise in
the captured images. Let us remember that when we capture horizontally an underwater image, we have to consider as the light
path the horizontal distance from the target (working distance)
and the depth which we are, hence, the radiation suffers twice
the absorption. As depicted in Figure 1 light intensity that penetrates into the water already decays of 50% at 10 m. Moreover,
under water the color light depends on other factors as the coastal
proximity, the season, the water surface conditions, the time of
the day, the cloudiness of the sky, the presence of marine vegetation or the soil type. For the reasons listed above, it is clear that

underwater environments offer a non-constant and multi-variate

typology of light conditions. Since many of the aforementioned
factors are constantly changing, we cannot really know all the
color variations in water.

Figure 1: Schematic representation of underwater lighting

conditions that can cause color alterations.

4.

BACKGROUND

Before to describe the proposed color correction method, we provide some details about color constancy, white balancing, image formation model and gray world assumption. The l color
space is also introduced.
4.1

Formulation of the color correction problem

Recognizing correct casts of objects in a scene illuminated by a

colored light source is a capability of the human visual system,
called color constancy (Ebner, 2007). In fact, the human visual
system can perceive the true colors of objects even if not illuminated by a neutral light source, i.e. white. For example, a white
object illuminated with a high temperature light source becomes
bluish, but the human observer will still recognize it as white. On
the other hand, for digital systems as imaging devices, it is necessary to adjust color casts after acquisition, in order to recover
the correct colors, because sensors measure the incoming light
from the objects, and they are not able to distinguish between
the color of the illuminant and the reflectance of the objects. In
digital photography, the process (automatic or manual) to remove
unrealistic color casts is called white balancing, which allows to
obtain a balanced color distribution around a white point taken as
reference. Digital cameras are able to balance color images under a wide variety of illumination conditions (i.e. neon light, sun
light, cloudy sky) by internal processes but these settings often
fall in underwater environments. For a more accurate correction
a calibration chart or a color checker can be placed and captured
in the scene. Then in post-processing the images are corrected.
The use of these calibrated charts underwater is not suitable and
practical for applications like mapping and navigation, in which
a large amount of data is collected.
In this work we tackle the color correction problem of underwater
images from the point of view of computational color constancy,
which aims to perform a color correction that closely mimics the
performance of the visual system. We suppose that the scene
is illuminated by a colored light source uniformly. This means
that the light intensity is approximately constant over whole image. Moreover we not consider degradation effects which cause
noise and blur in the image due to scattering and we suppose

This contribution has been peer-reviewed.

doi:10.5194/isprsarchives-XL-5-W5-25-2015

27

The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XL-5/W5, 2015
Underwater 3D Recording and Modeling, 1617 April 2015, Piano di Sorrento, Italy

that the object surfaces are Lambertian, that means the surface
reflects the light equally in all directions. These assumptions
can be considered too restrictive, but allow to use a simply image formation model, and however, they are acceptable in many
real cases for close-range acquisition in downward direction, like
seafloor mapping, or capturing objects in the foreground or lesscomplicated scenes where there are not strong light variations.
Thus, we can formulate the problem in a simplified manner following the computational color constancy approach. In the image formation model (Gonzales and Woods, 1992), as represented
in Figure 2, we can consider an underwater image fi (m, n) expressed by:

reason we have chosen to adopt a perceptual color space as we

described below. The l color space has been proposed by Ruderman et al. (1998) studying the human eye perception of natural
images. It is composed of a luminance component l (achromatic)
and two opponent color components and , as in the human
color visual mechanism. Until now this color space has been
only used for color transfer applications. In particular, Reynard
et al. (2001) have showed that it can be used for hue correction,
presenting the results obtained to correct an image with uniform
illumination, under red, tungsten, and blue illuminants. Therefore color correction by using this color space is not yet deeply
investigated, especially in underwater imaging.

fi (m, n) = gi (m, n) ri (m, n)

Here, we describe l space following the transformations matrices suggested in (Reynard et al., 2001), that we have reported in
Appendix for the readers convenience. The several steps to convert RGB coordinates in l space are reported in the flow-chart
of Figure 3. First of all, RGB image has to be corrected by the
non linearity (gamma correction), to work with linear RGB coordinates. This step is very important and necessary to pass to the
next: conversion in the XYZ tristimulus values. This conversion
is achieved by multiplying the linear RGB coordinates fi (m, n)
with Txyz,ij matrix, where the i and j are the matrix indeces in
according to the Einstein notation:

where

(1)

gi (m, n) = illuminant
ri (m, n) = reflectance
(m, n) = image coordinates

with i {1, 2, 3} we indicate the three color components of image, corresponding in RGB space to {R, G, B} channels. Correcting the color casts of an image means computing the illuminant gi and removing its chromatic components from fi . In order
to compute the color of illuminant further assumptions are needed
and for this purpose several computational color constancy methods have been proposed in literature (Gijsenij et al., 2011). In
the next section we describe the formulation of the gray-world
assumption used in our method.

xj (m, n) = Txyz,ij fi (m, n)

(3)

where xj is the vector of coordinates in XYZ space and (m, n)

are the pixel coordinates. From this device-independent XYZ
space, we convert the image to LMS space (lj (m, n) vector) using the Tlms,ij matrix:
lj (m, n) = Tlms,ij xi (m, n)

(4)

Now, we transform the data in logarithmic space:

llog,j (m, n) = log li (m, n)

(5)

and finally we can proceed to decorrelate these axes as suggested

by Ruderman et al. (1998), by using Principal Components
Analysis (PCA), that means multiply llog,j (m, n) vector by the
decorrelation matrix Tpca,ij :

Figure 2: Underwater reflectance model.

ll,j (m, n) = Tpca,ij llog,i (m, n)

4.2

The gray-world assumption

The gray-world assumption is a well-known approach, proposed

by Buchsbaum (Buchsbaum, 1980). In its simplest version it assumes that the information in the average of each channel of the
image is the representative gray level. In other words, the average
of the captured image should be gray (achromatic). Following the
image formation model described above in Eq. 1 and taking into
account that the illuminant gi can be approximate (Ebner, 2007)
by the average of image fi , the gray world assumption can be
formulated as:
fi (m, n) =

fi (m, n)
fi (m, n)

gi (m, n)
f i (m, n)

(2)

i {1, 2, 3} is the number of channels, corresponding in RGB

space to {R, G, B} and f i = mean{fi (m, n)}. The main drawback of gray-world methods is that they will fail to produce correct colors if sufficiently large numbers of hues are not present in
the scene.
4.3

Thus, we have obtained the vector coordinates ll,j (m, n) of

image in l space. The three resulting principal axes are orthogonal, and it is found that the first axis l represents an achromatic channel, while the second () and the third () channels
are chromatic yellow-blue and red-green opponent channels:
luminance : Achromatic (r + g + b)
: Yellow-blue (r + g b)
: Red-green (r g)
Hence, in this space the color components are decoupled from the
luminance, that is, the brightness level of the scene. The l is
a perceptual color space, that is, based on the response functions
of the cones in the retina of the eye, because it is derived by the
perceptual color space LMS (Long, Medium, Short) mostly used
in psychometric research.
5.

l color space

The proposed method aims to perform a color correction that

closely mimics the performance of the visual system. For this

(6)

COLOR CORRECTION METHOD

In this section we describe the color correction method that we

propose for underwater imaging. The method is based on the

This contribution has been peer-reviewed.

doi:10.5194/isprsarchives-XL-5-W5-25-2015

28

The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XL-5/W5, 2015
Underwater 3D Recording and Modeling, 1617 April 2015, Piano di Sorrento, Italy

Figure 3: Flow-chart: conversion of a non linear RGB image in

l space.

assumptions of gray world and uniform illumination of the scene,

as explained in the previous section. Thus, we formulate the Eq. 2
in l space. Let us express the RGB color coordinates fi (m, n)
in the LMS space (li (m, n)) as:
li (m, n) = Dij fj (m, n);

(7)

where Dij is the transformation matrix obtained as product of the

two transformation matrices Dij = Txyz,ij Tlms,ij . Expressing
the terms of Eq. 2 in LMS space in according to Eq. 7 and
computing the logarithm of both sides of the Eq. 2, we have:
log li (m, n) = log

li (m, n)
;
li

(8)

(a)

(b)

(c)

(d)

Figure 4: Two cases of underwater images analysed in this work

(a), (b), and the relative results obtained with our algorithm (c),
(d).
6.

EXPERIMENTAL RESULTS

In order to demonstrate the effectiveness of the proposed method,

we have tested it on two underwater images captured in different
locations (Figures 4a,b). The first case that we analysed is an
image corresponding to a part of a brick wall, taken in the archaeological site of Baia (Naples-Italy) at the depth of about 5 m,
in which a greenish component is present. The second case is an
image representing an amphora (Kas, Turkey) in which a bluish
component is dominant. This image is taken from a website of
divers community (Worldadventuredivers, 2015).

where llms,i
(m, n) are the corrected coordinates in LMS space

of fi (m, n), llms,i (m, n) are the coordinates of fi (m, n) and li

its average in according to gray-world assumption. Applying the
logarithm quotient rule and multiplying by the Tpca,ij matrix, we
obtain:

Tpca,ij log li (m, n) = Tpca,ij log li (m, n)Tpca,ij log li ; (9)

Each term in this equation is similar to Eq. 6, that states the
coordinates conversion in the l space. Thus, we have obtained
the desired expression of Eq. 2 in l coordinates:

ll,i
(m, n) = ll,i (m, n) ll,i ;

(a)

(10)

where ll,i
(m, n) are the corrected coordinates, ll,i (m, n)
the component of RGB image and ll,i its average. In order
to correct color casts of an image the and components (i
{2, 3}) must be changed, as suggested in (Reynard et al., 2001),
while the luminance component remains unchanged, at least regarding the translation of the mean, but it will be elaborated in a
different way, to improve the contrast image, as described in the
next section. After color correction processing we must transfer
the result back to RGB to display it by applying the inverse operations. It has to be noted that the white point ({1, 1, 1} in RGB
space) has {0, 0, 0} coordinates in l space. This means that
the method performs a shift of and components respect to
the coordinates {0, 0}, putting the distributions of these coordinates around the white point. Correcting image casts respect to
the white point in this space seems to be more realistic respect to
adjust the distributions of color channels around a gray point as
in RGB space. We can conclude that the gray-world assumption
is transformed in a white-world assumption in the l space.

(b)

Figure 5: Distribution of chromatic components of image in

Figure 4a in plane before and after color correction.
The two RGB images are converted in the l space, following
the steps discussed in section 4.3. Then we apply the Eq. 10 to

This contribution has been peer-reviewed.

doi:10.5194/isprsarchives-XL-5-W5-25-2015

29

The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XL-5/W5, 2015
Underwater 3D Recording and Modeling, 1617 April 2015, Piano di Sorrento, Italy

Figure 6: Luminance histograms of images in Figure 4.

Figure 7: Image of the luminance channel before (at left) and

after (at right) contrast enhancement by histogram stretching.
and components. According to our experience, it is preferred
for underwater images to use channel-wise median values instead
of the mean, since the median is not affected by extreme values
in luminance or chrominance. In this way the median would ignore, for instance, overlighted close objects or the blue parts of
the image corresponding to farther distances, which, on the other
hand, would affect the mean. As example of the result of the
white balancing process, we plot in Figure 5 in the plane
the distribution of the chromatic components for the image of
Figure 4a. It is apparent that after color correction, the distribution of and is centered on the white point coordinates {0,
0}. Moreover, the channel is more affected by the illumination color, while channel remains approximately unchanged. In
fact, we have found that for underwater images the median value
of is always negative. In order to verify the hypothesis of uniform illumination, we have computed the luminance histograms,
of both images as depicted in Figure 6. The values of luminance
distribution are sparse from darker values to the lighter values:
the image histogram of Figure 4a has a bell-shape distribution
concentrated on lighter values, while the one of Figure 4b has a
distribution more equalized. In case of the luminance values are
concentrated around two or more picks, especially in the brighter
area, the uniform-illumination hypothesis has to be rejected.

Figure 8: Comparison among different color correction methods.

to process underwater images with the proposed method. It takes

0.795 sec on a Mac Laptop (Processor 2.7 GHz Core i7, 16 GB
1600 MHz DDR3 Memory, NVIDIA GeForce GT 650M 1024
MB) to elaborate a 591 x 892 image. We point out that the implementation of this algorithm can be improved by using parallel
computing or C implementation, thus we think that the method is
suitable for real-time applications.
6.1

In Figure 7a the image of the luminance channel is illustrated. It

has been obtained by setting to zero the chromatic components
and reporting back the coordinates in RGB space. The result of
this process is a gray-level image, as we expected because the
luminance channel is achromatic. In order to improve the image
contrast, in this work we propose to perform a cutoff of 1% of
both histogram limits and stretching them. This is a simple way
to improve the contrast image, but also other contrast enhancement techniques can be applied like CLAHE (Contrast Limited
Adaptive Histogram Equalization) or histogram equalization, for
example. The result of the luminance histogram stretching is illustrated in Figure 7. Finally, we combine the results of luminance enhancement and hue correction, then we apply the inverse
transformations of the flow-chart in Figure 3 to report the results
in RGB space, as shown in Figures 4c,d. Regarding computaR code
tional considerations, we have implemented a MATLAB

Comparisons with other color spaces

From the study of the State of the Art we found that underwater image datasets provided with calibrated ground truth are not
available in literature as for the in-air applications of computational color constancy algorithms (Gijsenij et al., 2011). Hence
we limit the performance evaluation of our method to a qualitative evaluation by visual inspection. We have compared the white
balancing results of our method with the ones obtained with other
similar color spaces, in order to investigate if the l space allows to give similar or better results. For this reason we have performed this analysis on CIELab, YCrCb, CIECAM97, CIELuv
spaces, which are composed from a luminance component and
two chromatic components like l space. Similarly to the white
balancing done in l space, we have shifted the median values of chromatic components around its own white point. Moreover we have elaborated the underwater images of Figure 4 with a

This contribution has been peer-reviewed.

doi:10.5194/isprsarchives-XL-5-W5-25-2015

30

The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XL-5/W5, 2015
Underwater 3D Recording and Modeling, 1617 April 2015, Piano di Sorrento, Italy

gray-world algorithm in RGB space (Colorconstancy, 2015) and

with Photoshop software by using Auto Color tool. The outputs of each aforementioned methods are illustrated in Figure 8.
From the comparison it is apparent that the l space gives the
best results of color correction removing completely in the whole
image the greenish and bluish casts. In the case of gray-world result, the correction seams globally acceptable and the colors are
most saturated, but by a more accurate inspection we can see that
in the darker and lighter areas (pebbles on the seafloor), remain
colored casts. Color correction in l space outputs images with
realistic color casts both in gray shades and bright areas.
6.2

3D reconstruction application

As example of a possible application, we have processed an image set of an underwater finding with our method in order to
obtain the 3D reconstruction by using a structure-from-motion
software. The dataset has been taken in the archaeological site
of Kaulon in Monasterace (Calabria - Italy), capturing a remainder of an ancient pillar. The dataset is composed by 45 images
with resolution of 980 x 1479 pixels. We have pre-processed all
the images with our method and then input in a structure-frommotion algorithm. The results of color correction and the 3D
model obtained by the corrected dataset is shown in Figure 9. As
an alternative we can process directly the texture applied on the
3D model and obtained by UV mapping.

methods proposed in literature. Thus, a qualitative comparison of

white balancing process, in other perceptual color spaces, shows
that the hue correction in l space outputs the best results, also
respect to a RGB gray-world algorithm and a commercial software.
As example of a possible application, we have proposed to preprocess underwater image dataset in order to improve the results
of structure-from-motion software, obtaining 3D models with color
corrected texture. These promising results convince us to deepen
the investigations of this method for underwater imaging. In particular, further works will regard the comparisons with the other
methods presented in literature, studying an experimental protocol to acquire an appropriate dataset in order to evaluate quantitatively the error in the color correction, implement a new version of our method to take into account the non-uniform lighting
conditions and apply better contrast enhancement techniques to
reduce the degradation effects, like noise and blur. We will study
instead average color an appropriately weighted median color,
too.
ACKNOWLEDGEMENTS
This work has been supported by the project ARUE (MIDAVI
55), funded with support from the European Commission, European Social Fund and Regione Calabria. This publication reflects the views only of the author, and the European Commission
and the Regione Calabria cannot be held responsible for any use
which may be made of the information contained therein. Moreover, this work has been partially supported by the Project CoMAS (ref. PON01 02140), financed by the MIUR under the
PON R&C 2007/2013 (D.D. Prot.n.01/Ric.18.1.2010).

(a)

REFERENCES
Ancuti, C., Ancuti, C. O., Haber, T., Bekaert, P., 2012. Enhancing
underwater images and videos by fusion. In Computer Vision and
Pattern Recognition (CVPR), IEEE Conference on, pp. 81-88.

(b)

Bazeille, S., Quidu, I., Jaulin, L., Malkasse, J.P., 2006. Automatic
Underwater Image Pre-Processing. CMM06 - CARACTERISATION DU MILIEU MARIN, 16-19 Oct.

Figure 9: 3D model (b) obtained from a color corrected dataset

of underwater images (a).

Buchsbaum, G., 1980. A spatial processor model for object colour

perception, Journal of The Franklin Institute-engineering and Applied Mathematics, 310:126.

7.

CONCLUSIONS AND FUTURE WORKS

We presented a new method for color correction of underwater

images. The method is based on gray-world and uniform illumination assumptions applied in the l space. Color correction is performed by white balancing of chromatic components
( and ), while the luminance component l is processed to improve the image contrast. We have tackled the color correction
problem with a computational color constancy approach, developing the mathematical formulation of the hue correction suggested by Reynhard et al. (2001). From the experimental results,
we demonstrated that this method is suitable to remove undesired
color casts in underwater imaging: greenish-blue components in
the images are completely removed also in the gray shades. From
the analysis of the State of the Art, we have found out that until
now it does not exist a calibrated underwater dataset provided of
ground truth, in order to perform quantitative evaluations of the

Carlevaris-Bianco, N., Mohan, A., Eustice, R. M., 2010. Initial

results in underwater single image dehazing. In OCEANS 2010,
pp. 1-8.
Chambah, M., Semani, D., Renouf, A., Courtellemont, P., Rizzi,
A., 2003. Underwater color constancy: enhancement of automatic live fish recognition. In Electronic Imaging, International
Society for Optics and Photonics, pp. 157-168.
Chiang, J. Y., Chen, Y. C., 2012. Underwater image enhancement
by wavelength compensation and dehazing. Image Processing,
IEEE Transactions on, 21(4), 1756-1769.
Colorconstancy, 2015. http://colorconstancy.com/, Last accessed
on 2015, March 14.
Ebner., M., 2007. Color Constancy, Wiley 1st edition.
Gallo, A., Angilica, A., Bianco, G., De Filippo, F., Muzzupappa,
M., Davidde, B., Bruno, F., 2012. 3D Reconstruction and Virtual

This contribution has been peer-reviewed.

doi:10.5194/isprsarchives-XL-5-W5-25-2015

31

The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XL-5/W5, 2015
Underwater 3D Recording and Modeling, 1617 April 2015, Piano di Sorrento, Italy

Exploration of Submerged Structures: a Case Study in the Underwater Archaeological Site of Baia (Italy). In VAST: International
Symposium on Virtual Reality, Archaeology and Intelligent Cultural Heritage, Eurographics Association, pp. 121-128.
Garcia, R., Nicosevici, T., Cufi, X., 2002. On the way to solve
lighting problems in underwater imaging, In OCEANS 02 MTS /
IEEE, pp.1018-1024 vol.2, 29-31 Oct.
Ghani, A. S. A., Isa, N. A. M., 2014. Underwater image quality
enhancement through composition of dual-intensity images and
Rayleigh-stretching. SpringerPlus, 3(1), 757.

Conversion from linear RGB to XYZ space:

xj (m, n) = Txyz,ij fi (m, n)

"

0.5141 0.3239 0.1604

0.2651 0.6702 0.0641
0.0241 0.1228 0.8444

Txyz,ij =

(11)

#
(12)

Conversion from XYZ space to LMS:

lj (m, n) = Tlms,ij xi (m, n)

Gijsenij, A., Gevers, T., van de Weijer, J., 2011. Computational

Color Constancy: Survey and Experiments, Image Processing,
IEEE Transactions on, vol.20, no.9, pp.2475-2489.
Gonzalez, R.C., Woods, R.E., 1992. Digital Image Processing.
Addison-Wesley, Reading, MA.
Gouinaud, H., Gavet, Y., Debayle, J., Pinoli, J. C., 2011. Color
correction in the framework of color logarithmic image processing. In Image and Signal Processing and Analysis (ISPA), 7th
International Symposium on, pp. 129-133.
Iqbal, K., Abdul Salam, R., Osman, M., Talib, A. Z., 2007. Underwater Image Enhancement Using An Integrated Colour Model.
IAENG International Journal of Computer Science, 32(2), 239244.
Mahiddine, A., Seinturier, J., Boi, D.P.J., Drap, P., Merad, D.,
Luc Long, 2012. Underwater image preprocessing for automated
photogrammetry in high turbidity water: An application on the
Arles-Rhone XIII roman wreck in the Rhodano river, France,
In Virtual Systems and Multimedia (VSMM), 2012 18th International Conference on, pp.189,194, 2-5 Sept.

"
Tlms,ij =

0.3897 0.6890 0.0787

0.2298 1.1834 0.0464
0.0000 0.0000 1.0000

(13)

#
(14)

Conversion from log(LMS) space to l space:

ll,j (m, n) = Tpca,ij llog,i (m, n)

Tpca,ij

= 0
0

0
1
6

"

0
0

1
2

1
1
1

1
1 2
1 0

(15)

#
(16)

Revised March 2015

Petit, F., Capelle-Laize, A., Carr, P., 2009. Underwater image enhancement by attenuation inversionwith quaternions. In Acoustics, Speech and Signal Processing, IEEE International Conference on, pp. 1177-1180.
Reinhard, E., Ashikhmin, M., Gooch, B.,Shirley, P., 2001. Color
Transfer between Images, IEEE Computer Graphics and Applications, vol. 21, no. 5, pp. 34-41.
Ruderman, D., Cronin, T., Chiao, C., 1998. Statistics of cone
responses to natural images: implications for visual coding, J.
Opt. Soc. Am. A 15, 2036-2045.
Schechner, Y., Averbuch, Y., 2007. Regularized image recovery
in scattering media. IEEE Trans Pattern Anal Mach Intell.
Schettini, R., Corchs, S., 2010. Underwater image processing:
state of the art of restoration and image enhancement methods.
EURASIP Journal on Advances in Signal Processing 2010, 14.
Shamsuddin, N. B., Fatimah, W., Ahmad, W., Baharudin, B. B.,
Kushairi, M., Rajuddin, M., Mohd, F. B., 2012. Significance
Level of Image Enhancement Techniques for Underwater Images.
In IVIC 2012.
Torres-Mndez, L. A., Dudek, G., 2005. Color correction of underwater images for aquatic robot inspection. In Energy Minimization Methods in Computer Vision and Pattern Recognition
pp. 60-73, Springer Berlin Heidelberg.
Worldadventuredivers, 2015. http://worldadventuredivers.com/,
Last accessed on 2015, March 14.
APPENDIX
We report the conversion matrices used by Reinard et al. (2001).

This contribution has been peer-reviewed.

doi:10.5194/isprsarchives-XL-5-W5-25-2015

32