Computers and Electronics in Agriculture 141 (2017) 181–195

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture

journal homepage: www.elsevier.com/locate/compag

Original papers

Automatic system for improving underwater image contrast and color

through recursive adaptive histogram modification
Ahmad Shahrizan Abdul Ghani a,⇑, Nor Ashidi Mat Isa b
a
Innovative Manufacturing, Mechatronics, and Sports Lab (IMAMS), Faculty of Manufacturing Engineering, Universiti Malaysia Pahang, 26600 Pekan, Pahang, Malaysia
b
School of Electrical & Electronics Engineering, Engineering Campus, Universiti Sains Malaysia, 14300 Nibong Tebal, Penang, Malaysia

a r t i c l e i n f o a b s t r a c t

Article history: Contrast and color are important attributes to extract and acquire much information from underwater
Received 11 October 2016 images. However, normal underwater images contain bright foreground and dark background areas.
Received in revised form 29 June 2017 Previous enhancement methods enhance the foreground areas but retain darkness and blue-green illumi-
Accepted 27 July 2017
nation of background areas. This study proposes a new method of enhancing underwater image, which is
Available online 4 August 2017
called recursive adaptive histogram modification (RAHIM), to modify image histograms column wisely in
accordance with Rayleigh distribution. Modifying image saturation and brightness in the hue–satura
Keywords:
tion–value color model increases the natural impression of image color through the human visual system.
Underwater image
Contrast enhancement
Qualitative and quantitative evaluations prove the effectiveness of the proposed method. Comparison
Color improvement with state-of-the-art methods shows that the proposed method produces the highest average entropy,
Recursive overlapped area measure of enhancement (EME), and EME by entropy with the values of 7.618, 28.193, and 6.829,
Dual-intensity image respectively.
Ó 2017 Elsevier B.V. All rights reserved.

1. Introduction ronment. Underwater activities, such as discovery and recognition,

result in new challenges and address new problems related to light
Underwater image processing is undeniably applied in various absorption and diffusion (Schechner and Karpel, 2005; Church
fields, such as monitoring of marine habitats. Researchers involve et al., 2003; Shamsudin et al., 2012). Nevertheless, the proposed
in maritime research and ecological evaluation of offshore system methods and techniques to improve the quality of underwater
may benefit from underwater image processing when recording images are still limited.
sea flora and fauna (Cronin et al., 2003; Horváth and Varjú, The water surface structure causes the light to bend and create
1995). This application enables the possibility of determining the crinkle patterns when passing through the water medium. The
size and type of living creatures and organisms to a great extent. reflected light on the water surface will be horizontally or verti-
Underwater imaging also simplifies plant and piping inspection cally polarized as it propagates into the water medium (Anthoni,
in the engineering field. A camera-based imaging method for 2012). When rays of light pass through the water vertically, objects
underwater applications is an appropriate solution and provides in the water appear less shiny and the captured image turns bluish
a potential application of underwater image processing. At present, (Shamsudin et al., 2012). This phenomenon is caused by the atten-
many consumer devices, such as camera and video camera, are uation of light energy as it travels through water. Fig. 1 shows the
integrated with such application to widen its utilization underwa- water surface effects.
ter. In addition, underwater vision system is used in underwater Scattering and absorption effects produced from light attenua-
imaging (Webster, April 1998), archaeology (Coleman et al., tion results in low color and contrast of underwater images,
2000), ocean basement mapping (Moore and Jaffe, 2002), sub- thereby reducing information from these images. The use of addi-
merged robots (Xu and Negahdaripour, 1999), nuclear reactors, tional artificial lighting has been proposed to improve the lighting
and mine detection (Schechner and Karpel, 2005). source, but this method produces another lighting phenomenon.
Underwater imaging is challenging and attracts attention of An artificially lighted image presents a non-uniform illumination
researchers because of the physical properties of underwater envi- because a bright spot is created at the image center and light inten-
sity is reduced as it moves away from the image center.
⇑ Corresponding author. Underwater image processing has faced various problems in
E-mail addresses: [email protected] (A.S. Abdul Ghani), [email protected] dealing with the aforementioned phenomenon, including low
(N.A. Mat Isa).

http://dx.doi.org/10.1016/j.compag.2017.07.021
0168-1699/Ó 2017 Elsevier B.V. All rights reserved.
182 A.S. Abdul Ghani, N.A. Mat Isa / Computers and Electronics in Agriculture 141 (2017) 181–195

lighting of underwater image; (ii) Fourier method for rotation,

scale, and translation recovery to provide robustness toward dis-
similarity of image regions; (iii) local correlation for image regis-
tration; and (iv) multi-resolution pyramidal blending of the
image to form a composite seamless mosaic. CLAHS is the exten-
sion of contrast limited adaptive HE (CLAHE) and is designed for
medical imaging. CLAHS divides the image into equal regions,
and each region is applied with histogram specification method.
Meanwhile, Iqbal et al. (2010), Iqbal et al. (2007) proposed the
integration of color model (ICM) and unsupervised color correction
method (UCM). In Iqbal et al. (2010), the image histograms in the
red–green–blue (RGB) color model are stretched to the entire
intensity level before they are converted into the hue–satura
tion–intensity (HSI) color model. In the HSI color model, the satu-
ration and intensity components are stretched before they are con-
verted back into the RGB color model to produce an enhanced
output image. In Iqbal et al. (2007), two color channels, namely,
red and green, are modified in accordance with the von Kries
Fig. 1. The effect of water surface (Shamsudin and N., 2012). hypothesis to minimize color cast. Contrast correction is applied
in the RGB color model. In addition, the histograms are stretched
on the basis of the minimum and maximum intensity values of
contrast, limited visibility, non-uniform lighting, noise, bright arti-
each channel at 0.2% and 99.8% of the original histogram. Contrast
facts, floating particles, blurring, and elimination of true color
stretching toward a high intensity value is applied to the lowest
(Shamsudin et al., 2012; Schettini and Corchs, 2010). Therefore,
intensity channel, which is red. Contrast stretching at both his-
digital image processing is necessary to restore the contrast, color,
togram sides is applied to the green channel, and contrast stretch-
and lost information from images as well as reduce these prob-
ing toward the lower side is implemented to the blue channel that
lems. Consequently, various underwater imaging techniques have
represents the major effect of color cast (Iqbal et al., 2007). Then,
been proposed to deal with the said issues.
the image is converted into the HSI color model, and saturation
In this paper, recursive adaptive histogram modification
and intensity components are stretched at the lower and upper
(RAHIM) is proposed as a new technique of histogram-based
sides of the histograms. Abdul Ghani and Mat Isa (2015) extended
underwater image processing. The RAHIM method uses overlap-
the ICM method by introducing RD (ICM-RD) in histogram stretch-
ping histogram processing with respect to Rayleigh distribution
ing and modification of saturation and brightness of the HSV color
(RD) and modifies image histogram column wisely. This method
model. The extended ICM-RD method improves the overall image
integrates the modification of saturation and brightness compo-
contrast and reduces the noise in the resultant images.
nents in the hue–saturation–value (HSV) color model to improve
Petit et al. (2009) proposed underwater enhancement method
image color. Quantitative and qualitative results show significant
using quaternions to improve image contrast and color rendition.
and adequate improvement of resultant images mostly in terms
The proposed method works on the basis of light attenuation
of contrast and color. The RAHIM method can increase the contrast
inversion while processing a color space contraction through
of the image especially at the background areas. The rest of the
quaternions. The attenuation process produces a hue vector when
paper is organized as follows. Section 2 presents the current
applied to white color and consequently characterizes the color of
research in similar areas of histogram-based underwater image
the water medium. Using this reference axis, geometric transfor-
processing. Section 3 describes the motivation of the proposed
mations of the color space are computed by use of quaternions.
method. Section 4 explains the detail of the RAHIM method. Sec-
The results show that the water areas of the output image are
tion 5 demonstrates the qualitative and quantitative comparison
transformed into gray with low saturation values whereas the
results of the proposed and state-of-the-art methods. The last sec-
objects are fully colored. The object color is restored and its visibil-
tion elaborates the conclusions of the study.
ity is increased. The reason is due to the improvement of separa-
tion between the objects and their background. In addition, the
2. Previous work blue color of the water is significantly reduced. Saturation correc-
tion is not well performed whereas the physical-based correction
Image enhancement uses subjective and qualitative criteria to shows a good performance.
produce output image that is easily captured by the human visual Pixel distribution shifting color correction (PDSCC) is proposed
system (Schettini and Corchs, 2010). This technique does not by Naim and Isa (2012) to correct the white reference point of
require any physical model for image formation and involves mod- image while ensuring it is achromatic. PDSCC is designed to
ification of image intensity value that contributes toward changes improve the quality of underwater images. The technique
in image color and contrast. enhances the 3D rotational matrix by modifying the 2D two-
Histogram equalization (HE) is a popular image enhancement color channel plane. The method rotates the pixels three times
method. HE works by stretching and equalizing the histogram (yaw, pitch, and roll) using three different angles. In PDSCC, two
through the intensity range by means of probability distribution methods of color correction are designed to shift the distribution
function (PDF) and cumulative distribution function (CDF). HE is of the image pixels; thus, the insignificant illumination is reduced
used by previous researchers as a basic method in the enhance- from the output image.
ment process. Dwivedi et al. (2015) proposed an enhancement method based
In 2002, Eustice et al. (2002) extended the available method in on distance factor estimation. Underwater image is nearly identical
the image processing toolbox from MATLAB for optical image pro- to haze image and can be modeled as integration of attenuation
cessing and mosaicking algorithm. The main methods that have and scattering components. However, the proposed technique pro-
been introduced are as follows: (i) contrast limited adaptive his- duces various parameters that should be determined prior to the
togram specification (CLAHS) to handle inherent non-uniform implementation. Nevertheless, the resultant images still suffer
 A.S. Abdul Ghani, N.A. Mat Isa / Computers and Electronics in Agriculture 141 (2017) 181–195 183

from over-enhancement and blue-green illumination. Ancuti et al. because histogram modification considers not only the current tile
(2012) proposed a fusion method to enhance underwater image but also the neighboring areas. This process becomes relevant if
based on a single image or video of dynamic scenes. The method the large-sized object occupied more than one tiles. Therefore,
derives the input and weight measures from the degraded image. the adaptive modification of tiles occupied by an object becomes
Two inputs that represent the color- and contrast-corrected ver- accurate considering the neighboring tiles. Unlike the conventional
sions of the original underwater image are defined together with CLAHE processing 8
8 tiles (i.e., default value of tiles for adaptive
four weight maps that aim to improve visibility of distant objects. HE) independently, neighbor tiles are not considered into the tile
Output images and videos possess low level of noise while the process. Given that an object occupies more than one tiles, the pro-
overall contrast is increased. cess considers these tiles independently. Therefore, adaptive mod-
Meanwhile, Hitam et al. (2013) proposed the mixture CLAHE ification will become less accurate. The image modification in the
(CLAHE-Mix). CLAHE-Mix can reduce noise introduced by CLAHE HSV color model improves the overall image color.
in underwater images. First, CLAHE is applied to the image in the Previous methods, such as ICM (Iqbal et al., 2007), UCM (Iqbal
RGB color model. Then, CLAHE is applied to the image after conver- et al., 2010), and ICM-RD (Abdul Ghani and Mat Isa, 2015), apply
sion into the HSV color model. In the HSV model, the saturation global enhancement in consideration of the modification process
and brightness components of the image are applied with CLAHE. of the entire image. Meanwhile, CLAHE (Eustice et al., 2002) and
Then, these channels are created and converted back into the DIRS-CLAHS (Abdul Ghani and Mat Isa, 2015) consider the
RGB color model. During the process, pixel distribution is set in enhancement process by dividing the image into identical sizes
accordance with the RD for CLAHE in both color models. These pro- of n
n windows (tiles). The number of tiles (n) is dependent of
cesses produce two output images: CLAHE–RGB and CLAHE–HSV the user, but the default number of n is 8. However, as the number
images. Both images are combined to produce an output image of tiles increases, the processing time also increases. On the basis of
with low noise by use of the Euclidean norm. CLAHE and DIRS-CLAHS, image tiles are individually processed
An enhancement method using inherent optical properties is pro- without overlapping each other.
posed by Li et al. (2015) to enhance shallow ocean images and videos The RAHIM method applies the enhancement process by divid-
using fast dark channel prior to descattering. The proposed method ing the image into column-wise areas or denoted as n
1 windows
begins by estimating the depth map through dark channels before (tiles) with identical column size. The division of column-wise
considering the positions of the lighting lamp, camera, and imaging areas is based on the nature of the captured images of which most
plane. Furthermore, the proposed method successfully eliminates have bright areas at the foreground located near the capturing
the scattering effect by employing weight-guided median filtering. device while having dark areas at the background areas. The tiles
Finally, color correction is applied through spectral properties. The are processed when half of the previous tile is overlapped with half
comparison results show that the proposed method reduces the of the next tile. The overlapped tile is processed two times. The dif-
blue-green illumination but insignificantly increases object visibility. ferences of the RAHIM method with the aforementioned methods
The background objects are still barely observed. Identical results are are summarized in Table 1.
obtained using adaptive dehazing framework as proposed by Qing Fig. 2 shows the initial comparison results of state-of-the-art
et al. (2015). The proposed method inadequately increases the visi- methods with the RAHIM method mentioned in Table 1. Detailed
bility of the background areas. methodology and comparison of the RAHIM method are described
CLAHS was originally developed from CLAHE and is used for in Sections 4 and 5, respectively. As shown in the figure, the RAHIM
enhancing medical image. Specific pixel distribution through CLAHS method provides better results in terms of visibility, contrast, and
was introduced in 2002 by Eustice et al. (2002) and is used by Hitam under- and over-enhancement.
et al. (2013). The initial points of the darkest and brightest pixels are
important in a digital image. Shamsudin et al. (2012) and Rizzi et al. 4. Methodology
(2004) revealed that the black pixel will produce an initial intensity
value of 5% set to the brightness value whereas the white point is set As shown in Fig. 3, the RAHIM method begins with decompos-
to 95% of the initial brightness value. Shamsudin et al. (2012) ing the image color into respective channels, namely, red, green,
reported that the significant effect of applying color correction tech- and blue. For each color channel, the image is divided into regions.
niques is at 5% intensity level. Abdul Abdul Ghani and Mat Isa (2015) Then, the histogram of each region is generated with the corre-
proposed a dual-image Rayleigh-stretched CLAHS (DIRS-CLAHS) of sponding look-up table (LUT). The histogram of each region is
two main stages, namely, contrast and color corrections. DIRS- applied with global stretching process. In addition, the clip-limit
CLAHS combines global and local corrections and produces good out- process is applied to the histogram to limit the number of pixels
put images in terms of contrast and color. of certain intensity levels and thus restrain the image from having
concentrated dark or bright areas. Then, gray-level mapping with
respect to RD is employed to the histogram through transforma-
3. Motivation tion function. In each respective region, recursive overlapped pro-
cesses (including histogram stretching, clip-limit process, and
Underwater images are captured with bright areas at the fore- intensity level mapping with respect to RD) are applied for the
ground and dark areas at the background. These bright and dark entire image. The image regions are then combined through bilin-
areas result from the light propagation in water medium. As the ear interpolation to reduce the artificial effect between the sepa-
background area is far from the capturing device, the light from rated regions. Then, the image is converted into the HSV color
the objects is ‘‘filtered” by the water medium before it reaches model before stretching the saturation and brightness components
the capturing device. Thus, the background area is dark in the to improve the image color. Image channels are converted back
image. On the contrary, the objects near the capturing device into the RGB color model after the stretching process.
reflect much light to the capturing device and less reflected from
its path. The enhancement of the image will provide the actual sce- 4.1. Decomposition of image channels and splitting the image into
nario of the overall underwater image by considering the dark area smaller regions
at the background and bright area at the foreground. Thus, image
processing will become adaptive to the actual underwater sce- The image in the RGB color model is first decomposed into
nario. In addition, overlapping process is a good adaptive process its respective color channels. For each color channel, the image
184 A.S. Abdul Ghani, N.A. Mat Isa / Computers and Electronics in Agriculture 141 (2017) 181–195

Table 1
Comparison of the RAHIM and current state-of-the-art methods.

Method Methodology of the available state-of-the-art methods Methodology of the RAHIM method
(CLAHE) Eustice and - n
n windows (tiles) are processed for enhancement. Although the num- - n
1 windows (tiles) are processed for enhancement.
R. (2002) ber of row and column can be modified, the minimum value of column or - The tile is overlapped with the adjacent tiles.
row is 2. - Half of the neighbor tile will be processed two times when
- The tile is processed individually where no overlapping processed areas the previous tile is overlapped with the consecutive tiles.
exist. - Image modification based on the HSV color model for color
- All tiles are processed once individually and recursive process is not correction is applied.
applied. - Prior to the recursive overlapping process, histogram gen-
- Color model conversion for color correction is not applied. eration and implementation of clip limit are considered for
(UCM) Iqbal et al. - UCM applies histogram stretching as global process, and tiles are not overlapping areas.
(2010) divided.
- The stretching process is applied to the entire image intensity levels.
- The image is converted into the HSI color model before saturation and
intensity components are stretched individually.
(ICM) Iqbal et al. - Red and green channels of the image are modified on the basis of the Von
(2007) Kries hypothesis to minimize color cast; in this procedure, the percentages
of all color channels are equalized.
- Histograms are stretched toward high or low intensity levels based on its
color channels between 0.2% and 99.8%.
- Image is converted into the HSI color channel; in this procedure, saturation
and intensity components are stretched.
(ICM-RD) Abdul - The inferior color channel (red) is stretched toward the high intensity
Ghani and Mat value whereas the dominant color channel (blue) is stretched toward
Isa (2015) the low intensity value.
- Both directions of stretching are applied to intermediate color channel
(green).
- The image is then converted into the HSV color model before saturation
and brightness components are stretched.
(DIRS-CLAHS) Abdul - DIRS-CLAHS techniques involve global contrast correction, local contrast
Ghani and Mat correction, and color correction.
Isa (2015) - In global contrast correction, the image channels are divided into two
regions at its midpoint before they are individually stretched to produce
dual-intensity images.
- The images are then combined on the basis of the average intensity value
before they are applied with CLAHS.
- Color correction is applied and converted into the HSV color model to
improve the saturation and brightness of the image.

Original image CLAHE UCM ICM ICM-RD DIRS-CLAHS RAHIM

Underwater image A

Underwater image B

Underwater image C
Fig. 2. Initial comparison of state-of-the-art methods with the RAHIM method shown in Table 1.

is divided into eight small regions (i.e., columns). The number package. Fig. 4 illustrates the division of image channel into
of column is based on the default number of CLAHE applied small columns.
by Eustice et al. (2002), which determines the best default value As shown in Fig. 4, the image is divided into eight columns. A
of tiles for local histogram specification. This value is also short explanation on how the overlapped process is applied to
applied in the function ‘‘adapthisteq” in MATLABTM software these regions is given for better understanding of the process.
 A.S. Abdul Ghani, N.A. Mat Isa / Computers and Electronics in Agriculture 141 (2017) 181–195 185

Fig. 3. Flowchart of the implemented RAHIM method.

half of the first and second columns are processed together. Nota-
bly, half of the first column is processed two times. Similar process
is also applied to the third process, in which half of the second col-
umn is combined with the half of the third column and applied
with the local histogram specification. The identical overlapping
process is also applied to the next overlapped column.

4.2. Histogram stretching and clip-limit process

Under- and over-contrast occur in an underwater image

whereas the amount of pixels is cumulatively concentrated at
low and high intensity levels. Hence, stretching and clip-limit pro-
cesses are applied to the image histogram of respective regions to
prevent under- and over-contrast effects. For this purpose, the
histogram of a region from the previous step is generated and
the LUT is built. The clip-limit visual process is shown in Fig. 5,
in which the spikes in the histogram higher than the clip limit
will be cut off. The excessive numbers of pixels are equally dis-
tributed to all intensity levels, thereby increasing the number of
pixels at all intensity levels. In this case, a normalized value of
the clip limit is set at 0.01 in accordance with the finding by
Fig. 4. Image tile division and recursive overlapped process.
Eustice et al. (2002), who also implemented the MATLAB software
package.
In accordance with Iqbal et al. (2007), the linear stretching is
given by Eq. (1).
For the first process of local histogram specification, only the first

column (purple rectangular) is involved without combining with omax  omin
Pout ¼ ðPin  imin Þ þ omin ð1Þ
other columns. However, for the second process (blue rectangular), imax  imin

Fig. 5. Applying clip limit to the image histogram.

186 A.S. Abdul Ghani, N.A. Mat Isa / Computers and Electronics in Agriculture 141 (2017) 181–195

On the basis of bilinear interpolation, image tiles are composed

(Hurtik and Madrid, 2015) to produce a resultant image with less
artificial-induced borders. Then, HSV-based color correction is
applied to the image.

4.4. Conversion into HSV color model and dividing and stretching S
Number of pixels

and V components at mid-points

Contrast correction degrades color performance such that the

object color becomes light and dark. In addition, the overall bright-
ness and saturation are reduced. The experiment shows that
under- and over-brightness areas are produced from contrast cor-
rection. As a result, the saturation and brightness of image need to
be improved. Hence, the image is first converted into the HSV color
model before the saturation and brightness components of the
image are modified. Hitam et al. (2013) mentioned that HSV can
reduce the noise of the resultant image. Nevertheless, saturation
and brightness components increase the visibility and clearness
Intensity level (0:255)
of the image while having a natural look.
Fig. 6. RD with most of the pixels concentrated at the middle intensity values. In the proposed method, saturation and brightness are
stretched in accordance with RD. In the RAHIM method, a differ-
ent global stretching approach is implemented. Referring to Abdul
4.3. Gray-level mapping through transformation function and bilinear Ghani and Mat Isa (2014) who implemented dual-intensity
interpolation (image tiles composition) images in the RGB color model, the method can improve the
image contrast significantly while increasing the image detail.
RD is the most appropriate distribution for underwater imaging However, in the RAHIM method, such technique is applied in
(Eustice et al., 2002). It refers to the bell-shaped histogram distri- the HSV color model for the respective saturation and brightness
bution in which most of the pixels are concentrated at the middle histograms. Thus, this process will produce two images with dif-
of the intensity level. The pixel number at the minimum and max- ferent intensities in the HSV color model. In addition to the rele-
imum sides of the distribution is the lowest to minimize the pixel vance of applying dual-intensity histogram modification, the
amount from having too low or too high intensity values. There- applied technique is based on under- and over-saturation in the
fore, RD reduces the pixel number of under- and over-contrasted resultant image from the previous step. By applying global
areas that may be produced in the resultant image. Fig. 6 illustrates stretching to the image, under-saturated regions will be improved
the RD in which most of the pixels are concentrated around the by producing bright areas (increased intensity values). However,
middle intensity values. for the (originally) good saturation regions, excessive brightness
The clip-limit process is applied to the image histogram to will result in over-saturation. An identical scenario will happen
reduce excessive pixels for the dominant intensity level. Image his- when applying global stretching to over-saturated regions. Never-
togram reflecting the PDF and CDF is then mapped with the RD. theless, under- and over-saturation are normally detected in an
The PDF and CDF of the RD is given by Eqs. (2) and (3), respectively underwater image. Limited stretching range for saturation and
Bibalan and Amindavar, 2015. brightness is available for global stretching technique in some
  cases, as shown in Fig. 7.
x x2
Therefore, dual-intensity histogram modification is imple-
PDF Rayleigh ¼ 2 e 2a2
; ð2Þ
a mented in the RAHIM method to improve the stretching process
  capability. In addition to the dual-intensity histogram modifica-
x2 tion, the stretching process is applied in accordance with RD. As
CDF Rayleigh ¼ 1  e 2a2
; ð3Þ
where x refers to the input data while a indicates the RD parameter.
In this study, a value is set at 0.4 as suggested by Eustice et al.
(2002).
To obtain the PDF of Rayleigh-stretched distribution, PDF Rstretch ,
Eq. (1) is integrated with Eq. (2) to obtain Eq. (4). The stretching
process of the output histogram occupied is from 0 to 255; thus,
the value of omin can be substituted with 0 to derive Eq. (5).

 !
P out P2out
PDF Rstretch ¼ exp ð4Þ
a2 2a2

 !
2
ððPin  imin Þomax Þ ððPin  imin Þomax Þ
PDF Rstretch ¼ exp ð5Þ
a2 ðimax  imin Þ 2a2 ðimax  imin Þ

Therefore, the CDF of Rayleigh-stretched distribution, CDF Rstretch ; is

given by
!
2
ððPin  imin Þomax Þ
CDF Rstretch ¼ 1  exp ð6Þ
2a2 ðimax  imin Þ
Fig. 7. Limited stretching range for brightness and saturation components.
 A.S. Abdul Ghani, N.A. Mat Isa / Computers and Electronics in Agriculture 141 (2017) 181–195 187

Fig. 8. Dividing and stretching of saturation and brightness components.

reported and proven by Eustice et al. (2002) and implemented by Apart from the Rayleigh-distributed dual-intensity histogram
Hitam et al. (2013), RD is the best distribution for underwater specification, a stretching limit of 1% is also applied from the max-
image. Hitam et al. (2013) also implemented RD in the HSV color imum and minimum intensity levels to minimize under- and over-
model. Fig. 8 illustrates the dividing and stretching of saturation saturation or brightness of the image. This claim is proven by
and brightness components of the HSV color model. Abdul Ghani and Mat Isa (2014) because the process can reduce

1% from the
upper limit

1% from the
Center = white lower limit
S

Original dynamic
range of S New dynamic
range of S

1%
New dynamic
range of V

1%

(a) (b)
Fig. 9. HSV color model. (a) Hue, saturation, and value elements of the HSV color model. (b) Modified range of saturation and brightness elements within 1% from minimum
and maximum values applied in the proposed method.
188 A.S. Abdul Ghani, N.A. Mat Isa / Computers and Electronics in Agriculture 141 (2017) 181–195

Underwater image 1 Underwater image 2

Fig. 10. Sample of the input and output of underwater images processed using the RAHIM method.

a) Original image b) ICM c) UCM

d) HE e) PDSCC f) CLAHS

g) CLAHE-Mix h) ICM-RD i) RAHIM

Fig. 11. Brown stone image. (a) Original image and resultant images of (b) ICM, (c) UCM, (d) HE, (e) PDSCC, (f) CLAHS, (g) CLAHE-Mix, (h) ICM-RD, and (i) RAHIM. (For
interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
 A.S. Abdul Ghani, N.A. Mat Isa / Computers and Electronics in Agriculture 141 (2017) 181–195 189

h   i
negative saturation and brightness. On the basis of the above- omax omin
ðP in  imin Þ imax imin
þ omin
mentioned method, midpoint, imid, is calculated using the following Rayl:  stretched ¼

a2 
equation: 2 omin
3
ðPin  imin Þ oimax i
þ omin
imax  imin
exp 4 max min
5 ð8Þ
imid ¼ þ imin ð7Þ 2a2
2
where imin and imax are the minimum and maximum intensity val- Fig. 8 shows the sample division and stretching of S and V com-
ues, respectively. ponents. Fig.9(a) and (b) show the components of the HSV color
Dividing the input histograms of saturation and brightness at model and the new range of saturation and brightness components
the midpoint produces upper and lower regions for each compo- within the 1% limit that is applied in the HSV color model.
nent, respectively. The division of these components is based on Then, the image in the HSV color model is converted back to the
the following procedure: RGB color model. The following section elaborates the comparison
of the RAHIM method with other state-of-the-art methods.
i. Upper regions: input value is taken from imid to imax and
stretched to a new output intensity ranging from 1% to 99%.
ii. Lower region: input value is taken from imin to imid and 5. Result and discussion
stretched to a new output intensity ranging from 1% to 99%.
Seven underwater-focused histogram-based methods, namely,
Rayleigh-stretched histogram for respective regions is applied ICM (Iqbal et al., 2007), UCM (Iqbal et al., 2010), CLAHS (Eustice
in accordance with Eq. (8). et al., 2002), CLAHE-Mix (Hitam et al., 2013), PDSCC (Naim and

a) Original image b) ICM c) UCM

d) HE e) PDSCC f) CLAHS

g) CLAHE-Mix h) ICM-RD i) RAHIM

Fig. 12. Red coral image. (a) Original image and resultant images of (b) ICM, (c) UCM, (d) HE, (e) PDSCC, (f) CLAHS, (g) CLAHE-Mix, (h) ICM-RD, and (i) RAHIM. (For
interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
190 A.S. Abdul Ghani, N.A. Mat Isa / Computers and Electronics in Agriculture 141 (2017) 181–195

Isa, 2012), ICM-RD (Abdul Ghani and Mat Isa, 2015), and HE, are contrast (Liang et al., 2012), and natural image quality evaluator
compared with the RAHIM method to validate its capability. (NIQE) (Mittal et al., 2012). Entropy H(X) is defined by
UCM and PDSCC outperform the famous gray-world and white
X
k
patch methods in enhancing image color. HE is selected as it is a HðXÞ ¼  pðxÞlog2 pðxÞ ð9Þ
popular image processing method than can solve most image prob- x¼1
lems. In the comparison, the parameters for CLAHS and CLAHE-Mix
EME and EMEE are two quality metrics for blind-image evalua-
are set following the default values set by the authors.
tion and measure the overall quality of underwater image. EME
The RAHIM method is compared qualitatively and quantita-
evaluates image by dividing the image into k1k2 blocks. For a given
tively with other methods. In terms of image processing, visual
class {U} of orthogonal transforms of an image x(n,m) which is split
observation (qualitative evaluation) is taken as a first priority for
into k1k2 blocks of size I1
I2, the EME value of the image is given
overall image quality evaluation. On the contrary, quantitative
by Eq. (10). Imin;k;l ð/Þ and Imax;k;l ð/Þ are the minimum and maximum
measurement is used to support qualitative perceptions. Contrast,
intensity levels of the image inside the block, respectively. c indi-
visibility, color cast, and under- and over-enhancement are criteria
cates a small constant with a value of 0.0001 to avoid dividing
in deciding the quality of the final output image. Output images are
by 0. EMEE is an extension of EME and measures image enhance-
compared in terms of entropy value (Beghdadi and Negrate, 1989)
ment on the basis of entropy value. In accordance with Agaian
that describes image details and information content, measure of
et al. (2001), EMEE is calculated using Eq. (11). a is a constant with
enhancement (EME) (Agaian et al., 2000; Agaian et al., 2007),
a value of 0.8 as reported by previous authors.
EME by entropy (EMEE) (Agaian et al., 2000; Agaian et al., 2001),

a) Original image b) ICM c) UCM

d) HE e) PDSCC f) CLAHS

g) CLAHE-Mix h) ICM-RD i) RAHIM

Fig. 13. Green coral image. (a) Original image and resultant images of (b) ICM, (c) UCM, (d) HE, (e) PDSCC, (f) CLAHS, (g) CLAHE-Mix, (h) ICM-RD, and (i) RAHIM. (For
interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
 A.S. Abdul Ghani, N.A. Mat Isa / Computers and Electronics in Agriculture 141 (2017) 181–195 191

1 X 2 X
k k 1
Imax;k;l ðUÞ Meanwhile, NIQE is a no-reference image quality assessment
EME ¼ 20log ð10Þ that measures the deviation from statistical regularities observed
k1 k2 l¼1 k¼1 Imin;k;l ðUÞ þ c
in natural image without training on human-rated distorted
images (Mittal et al., 2012). NIQE is based on identical size compu-

a
1 X
k2 Xk1 tation and feature patches of the image and fitting with multivari-
I ð/Þ Imax;k;l ð/Þ
EMEE ¼ a max;k;l log ð11Þ ate Gaussian (MVG) model. The quality of image, D, is given by
k1 k2 l¼1 k¼1 Imin;k;l ð/Þ Imin;k;l ð/Þ
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u 0X X11
Two other quality metrics are used in evaluating the output u þ
u
images, namely, contrast, C contrast , and NIQE. Contrast is used to u TB 1 2 C
measure the contrast improvement based on the following equa-
D¼u ðð
t 1 v  v Þ
2 @ A ðv 1  v 2 ÞÞ ð13Þ
2
tion (Liang et al., 2012):
P P

1 X
2
 where v 1 , v 2 , and 2 2 are the mean vectors and covariance of
1 XP X Q

P X Q
 the natural MVG model and the MVG model of the distorted image,
C contrast ¼ g ðx; yÞ  
2
gðx; yÞ ð12Þ
PQ x¼1 y¼1 PQ x¼1 y¼1  respectively. A low NIQE value is preferable as it results in
improved quality of the image.
where P and Q are the width and height of the image, respectively; The samples of the output image produced by the RAHIM
gðx; yÞ indicates the gray level at location ðx; yÞ: Large value of method are shown in Fig. 10. As shown in the figure, output images
C contrast indicates large dynamic range of gray level and thus produce better results in terms of visibility and contrast. The back-
improved contrast and much detailed information of image. ground objects can be clearly observed and distinguished with the

a) Original image b) ICM c) UCM

d) HE e) PDSCC f) CLAHS

g) CLAHE-Mix h) ICM-RD i) RAHIM

Fig. 14. Blue fish image. (a) Original image and resultant images of (b) ICM, (c) UCM, (d) HE, (e) PDSCC, (f) CLAHS, (g) CLAHE-Mix, (h) ICM-RD, and (i) RAHIM. (For
interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
192 A.S. Abdul Ghani, N.A. Mat Isa / Computers and Electronics in Agriculture 141 (2017) 181–195

Table 2
Quantitative measurement of output images by methods compared in Fig. 11–14.

Image Method Quantitative analysis

Entropy EME EMEE Contrast NIQE
Brown stone Original 7.379 17.653 1.941 41.19 3.80
ICM 7.798 24.528 9.065 40.27 4.00
UCM 7.806 21.512 6.385 40.29 3.92
HE 5.986 26.154 8.534 42.05 3.81
PDSCC 7.248 15.593 1.504 41.78 3.79
CLAHS 7.575 24.046 4.080 41.37 3.94
CLAHE-Mix 7.098 15.640 1.619 41.01 3.69
ICM-RD 7.650 25.633 8.589 40.22 3.95
RAHIM 7.736 30.725 11.250 40.35 4.13
Red coral Original 7.388 22.575 2.634 40.95 2.95
ICM 7.836 31.893 11.154 40.55 3.16
UCM 7.815 28.970 5.534 40.54 3.16
HE 5.988 28.153 7.201 42.05 3.08
PDSCC 7.313 20.410 2.077 41.33 2.91
CLAHS 7.693 31.490 6.554 41.31 3.14
CLAHE-Mix 7.158 19.943 2.347 41.08 3.35
ICM-RD 7.710 31.967 8.656 40.36 3.01
RAHIM 7.790 36.127 13.159 40.48 3.34
Green coral Original 7.403 23.468 3.569 40.59 3.38
ICM 7.722 30.392 11.595 38.87 3.43
UCM 7.653 27.606 6.027 39.26 3.36
HE 5.968 22.481 6.840 42.04 3.35
PDSCC 7.200 15.469 1.318 41.77 3.26
CLAHS 7.466 24.780 3.944 40.59 3.32
CLAHE-Mix 7.141 16.934 1.899 40.44 3.51
ICM-RD 7.632 32.295 8.161 38.60 3.41
RAHIM 7.637 38.640 14.728 38.78 3.33
Blue fish Original 7.612 12.627 1.137 41.70 2.56
ICM 7.826 19.681 2.931 41.22 2.61
UCM 7.820 14.135 1.726 41.45 2.58
HE 5.985 20.572 4.081 42.04 2.56
PDSCC 7.235 9.355 0.624 43.03 2.51
CLAHS 7.423 16.544 1.661 41.07 2.50
CLAHE-Mix 7.074 9.794 0.681 40.46 2.42
ICM-RD 7.631 20.760 5.515 40.36 2.53
RAHIM 7.707 24.230 7.330 39.73 2.53

Note: Values in bold represent the best result obtained in the comparison.

foreground objects. In addition, blue-green color cast is signifi- effect detected in the output image of UCM. The RAHIM method
cantly removed while contrast and color are highly improved. shows the best output image by minimizing the blue-green illumi-
Fig. 10 shows the example of the output images produced by the nation effects at the background areas. In addition, no saturated
RAHIM method. Notably, a clear difference between the original areas are observed as the image contrast and color look more nat-
and resultant images is observed. ural compared with those of other output images. Over- and
Figs. 11–14 show the resultant images of the compared meth- under-saturated areas are also significantly reduced.
ods. The highlighted rectangular areas in each of the figures show HE (Fig. 11(d)) produces low-quality image as the resultant
the clear differences among the compared methods. In addition to image is over-bright with too bright areas. The background areas
qualitative evaluation (visual observation), Table 1 shows the are also filled with blue-green color cast. These effects reduce the
quantitative evaluation of the resultant images in Figs. 11–14 in image details. Rectangular areas in the image highlight these draw-
terms of entropy, EME, EMEE, contrast, and NIQE. backs. This argument is supported by the quantitative evaluation
The brown stone image in Fig. 11 shows that the resultant image as HE produces the lowest entropy with a value of 5.986 (i.e., lower
produced by ICM (Fig. 11(b)) improves the overall image contrast. than that of the original image). HE produces the second highest
Blue-green illumination is significantly improved. However, visual EME and third highest EMEE values. However, these values are
observation shows that the background areas of the image are insignificant compared with the visual evaluation of the output
inadequately improved. Some blue-green illumination effects of image. PDSCC (Fig. 11(e)) improves limited contrast as the ‘‘veil”
the background areas are retained in the output image, and the of the blue-green illumination remains in the output image. Object
areas appear dark. The brown stone exhibits an over-saturated visibility in the image is low compared with that for other meth-
effect as it looks yellowish. Nevertheless, ICM produces the second ods. Quantitative evaluation indicates that image detail (i.e.,
highest entropy value of 7.798 after the UCM method, as shown in entropy value) is reduced as it produces lower entropy value than
Table 2. These effects are highlighted in the rectangular areas in the that of the original image. PDSCC produces the lowest EME and
image. The overall output image produced by UCM (Fig. 11 (c)) is EMEE values compared with other state-of-the-art methods.
brownish, and over-saturation is also detected at the brown stone. CLAHS (Fig.11(f)) improves image contrast as object visibility in
However, the blue-green illumination at the background areas is the image is improved. However, CLAHS reduces the image color,
significantly reduced compared with that in ICM. Quantitative thereby resulting in low image saturation and brightness. Visual
evaluation indicates that UCM produces the highest entropy value observation indicates that the objects in the image are dimmed.
of 7.806. This finding may be attributed to the over-saturation CLAHE-Mix (Fig.11(g)) produces greenish image and reduces the
 A.S. Abdul Ghani, N.A. Mat Isa / Computers and Electronics in Agriculture 141 (2017) 181–195 193

Table 3
Dataset for average quantitative measurement of 300 underwater images in comparison with state-of-the-art methods.

Method Quantitative analysis

Entropy EME EMEE Contrast NIQE
Original 7.157 15.674 2.259 32.97 3.831
ICM 7.608 19.599 4.455 36.74 3.699
UCM 7.600 17.366 3.538 36.52 3.692
HE 5.893 21.765 5.287 42.07 3.780
PDSCC 7.053 9.782 0.947 33.16 3.799
CLAHS 7.282 16.025 1.799 38.66 3.378
CLAHE-Mix 7.124 14.559 1.522 35.94 3.970
ICM-RD 7.514 27.751 7.853 39.27 3.910
RAHIM 7.618 28.193 6.829 39.39 3.930

Note: Values in bold represent the best result obtained in the comparison.

overall image color. The output image exhibits limited contrast from the following Google drive address:
enhancement. The visual observation is supported by the quantita- https://drive.google.com/drive/folders/0By7h39AYN85uTnFOeFh-
tive measurement as all quantitative values (i.e., entropy, EME, and PamJQQlk?usp = sharing
EMEE) of CLAHE-Mix are lower than those of the original image. The aforementioned effects are also observed in the green coral
ICM-RD improves the output image but retains the color cast and blue fish images in Figs. 13 and 14. These effects are clearly
effect as blue-green illumination is still observed in the back- highlighted by the rectangular areas in the respective images.
ground areas. The RAHIM method (Fig. 11(h)) produces the best These effects are also supported by quantitative evaluation. The
resultant image as the image exhibits significantly improved con- RAHIM is the best method and shows the best visually observed
trast and color. Blue-green illumination especially at the back- image compared with other methods. In addition, the proposed
ground areas is adequately minimized. In addition, no under- and method produces the highest EME and EMEE values for each sam-
over-enhanced areas are observed in the resultant image. This find- ple image. In addition to the individual quantitative values, Table 3
ing is supported by the quantitative evaluation as the RAHIM presents the dataset for the average values of 300 underwater
method produces the highest EME and EMEE values (i.e., 30.725 images in terms of entropy, EME, EMEE, contrast, and NIQE. As
and 11.250, respectively) and produces the third highest entropy shown in Table 3, the RAHIM method produces the highest average
value after ICM and UCM. In terms of contrast and NIQE, the quan- entropy, EME, and EMEE with the values of 7.618, 28.193, and
titative evaluation does not produce a good result for the RAHIM 6.829, respectively, among all the compared methods. The contrast
method. This finding may be due to the inappropriate evaluation and NIQE values of 39.39 and 3.930, respectively, exhibit negative
of these parameters given that most of the tested images that effect on the output image. Nevertheless, the contrast value of the
use these evaluation parameters are normal image captured above proposed method is ranked second after HE. However, these values
the water surface. However, visual observation of the resultant are not matched with the visual observation. The visual observa-
image produced by the proposed method emphasizes that RAHIM tion and significant quantitative evaluation prove that the RAHIM
method produces the best output image. Tables 2 and 3 show that method successfully improves underwater image contrast, color,
HE produces the best value for contrast. However, under- and over- and visibility.
saturated effects are clearly observed in the output images of HE. Additional sample images are attached in Appendix A. The
For the red coral image in Fig. 12(b), ICM improves the overall quantitative evaluation of these sample image can be found in
image contrast. However, blue-green illumination retains at the Appendix B.
background areas. Some areas are also over-saturated. UCM
improves better image contrast at the foreground but produces 6. Conclusion
over-saturated areas that look reddish. On the other hand, HE
over-enhances the image as many bright areas are produced. Color The RAHIM method is proposed to solve underwater image
cast is also retained in the background areas. On the contrary, problems, especially image contrast and color. Underwater image
PDSCC improves limited image contrast as blue-green illumination presents two main contrast areas, namely, foreground and back-
retains in the image. As shown in Fig. 12(f), CLAHS improves image ground areas. The proposed method considers overlapping and
contrast as the resultant image exhibits improved visibility. How- column-wise modification of image histogram. Image color is sig-
ever, CLAHS reduces image color as the objects look dimmed. In nificantly improved through the stretching process of the dual-
addition, some areas are still covered with blue-green illumination. intensity histogram of saturation and brightness components in
CLAHE-Mix diminishes the output image as the image color is dis- the HSV color model. The effectiveness and capability of the pro-
torted and the image contrast and visibility are insignificantly posed method are verified by the qualitative and quantitative eval-
improved. Moreover, blue-green illumination retains in the output uation results.
image. RAHIM is the best method and improves the image contrast
and color. The visibility of the resultant image is significantly Acknowledgement
improved at the foreground and background areas. Quantitative
evaluation indicates that the proposed method produces the high- We thank all the reviewers for the suggestions and comments
est EME and EMEE values at 36.127 and 13.159, respectively. to improve the paper. The study is supported by Universiti Malay-
Table 3 provides a dataset for the average value of 300 samples sia Pahang (UMP) Research Grant (RDU170392) entitled ‘‘Dual
of underwater images for the compared methods. The values in Image Fusion Technique for Enhancement of Underwater Image
bold represent the best results obtained in the comparison. The Contrast” and Fundamental Research Grant Schema from the Min-
image database can be accessed from the following open link: istry of Higher Education of Malaysia entitled ‘‘Elucidation of Light
https://drive.google.com/open?id = 0By7h39AYN85uTHBneTk- Attenuation in Water Medium for Enhancement of Underwater
wUHRWVzg. The resultant images of the samples can be accessed Image Contrast and Color” (FRGS/1/2015/TK04/TATI/02/1).
194 A.S. Abdul Ghani, N.A. Mat Isa / Computers and Electronics in Agriculture 141 (2017) 181–195

Appendix A.

Additional resultant images and comparison with state-of-the-art methods.

Original ICM UCM HE PDSCC CLAHS CLAHE-Mix ICM-RD RAHIM

Image 1

Image 2

Image 3

Image 4

Image 5

Image 6

Appendix B.

The following dataset shows the quantitative evaluation of the images in Appendix A.

Image Original ICM UCM HE PDSCC CLAHS CLAHE-Mix ICM-RD RAHIM

Image 1 Entropy 6.36 7.73 7.62 5.64 6.42 6.88 6.96 7.37 7.52
EME 16.255 21.079 17.302 23.333 4.240 12.512 13.806 20.456 26.311
EMEE 1.59 1.05 2.93 6.79 0.24 0.96 1.26 5.481 4.97
Contrast 40.46 40.11 40.80 42.03 42.60 40.40 42.28 39.52 39.22
NIQE 5.72 4.26 4.08 4.16 5.44 4.61 5.03 4.26 4.32

Image 2 Entropy 6.05 7.35 7.52 5.64 6.35 6.88 6.83 7.403 7.51
EME 20.960 18.519 22.858 26.740 8.104 13.462 15.744 21.642 25.766
EMEE 2.63 3.13 4.72 11.39 1.13 1.03 1.57 4.197 4.68
Contrast 40.28 39.30 40.56 42.05 42.81 39.92 42.99 39.19 38.37
NIQE 8.20 6.45 5.52 5.50 6.48 6.60 6.78 6.22 5.92

Image 3 Entropy 5.78 7.29 6.98 5.45 6.04 6.34 6.30 7.111 7.41
EME 10.915 9.428 5.192 28.892 19.941 15.442 18.600 34.411 28.168
EMEE 1.89 1.35 0.45 4.22 2.45 1.22 1.93 9.918 4.12
Contrast 39.99 42.03 43.16 42.10 42.85 39.50 40.79 40.55 38.66
NIQE 6.71 5.44 6.76 5.62 6.63 5.55 6.26 5.17 5.13

Image 4 Entropy 7.38 7.70 7.69 5.96 7.04 7.47 7.20 7.473 7.62
EME 22.160 27.256 28.480 26.970 12.936 23.061 15.520 33.185 31.805
EMEE 4.08 7.07 7.86 9.12 1.01 3.59 1.87 11.330 9.82
Contrast 39.80 38.96 39.20 42.05 41.27 40.68 40.52 37.63 39.24
NIQE 4.17 4.02 3.92 4.06 3.92 3.91 3.88 3.83 3.92
 A.S. Abdul Ghani, N.A. Mat Isa / Computers and Electronics in Agriculture 141 (2017) 181–195 195

Appendix B. (continued)

Image Original ICM UCM HE PDSCC CLAHS CLAHE-Mix ICM-RD RAHIM

Image 5 Entropy 6.26 7.42 7.25 5.48 5.64 6.31 7.10 7.018 7.26
EME 4.494 10.287 9.049 12.160 2.308 6.258 6.581 18.791 17.723
EMEE 0.32 1.13 1.35 0.79 0.13 0.40 0.42 6.572 3.42
Contrast 40.62 44.45 44.53 42.04 42.39 40.79 44.39 41.05 41.59
NIQE 6.95 4.28 4.44 4.51 5.84 5.54 5.80 4.110 4.96

Image 6 Entropy 6.93 7.64 7.68 5.95 6.75 7.16 7.26 7.553 7.65
EME 13.136 17.819 13.324 23.916 3.386 11.391 14.244 18.056 28.062
EMEE 0.89 1.45 0.91 2.49 0.16 0.72 1.01 2.830 8.14
Contrast 40.55 41.23 42.03 42.04 42.53 40.55 41.93 41.02 39.34
NIQE 3.33 2.64 2.70 2.76 3.13 2.98 3.14 2.811 2.73
Note: The values in bold represent the best result obtained in the comparison.

References Dwivedi, P.K., Paul, B., Ghoshal, D., 2015. Underwater image enhancement using
distance factor estimation. In: IEEE International Conference on Electrical,
Electronics, Signals, Communication and Optimization (EESCO).
Cronin, T.W., Shashar, N., Caldwell, R.L., Marshall, J., Cheroske, A.G., Chiou, T.H.,
Visakhapatnam, 24–25 Jan. 2015. pp. 1–5. http://dx.doi.org/10.1109/EESCO.
2003. Polarization vision and its role in biological signalling. J. Integr. Comp.
2015.7254047.
Biol. 43 (4), 549–558.
Ancuti, C., Ancuti, C.O., Haber, T., 2012. Enhancing underwater images and videos by
Horváth, G., Varjú, D., 1995. Underwater refraction-polarization patterns of skylight
fusion. Computer Vision and Pattern Recognition (CVPR). In: 2012 IEEE
perceived by aquatic animals through Snell’s window of the flat water surface. J.
Conference on. IEEE, pp. 81–88.
Vis. Res. 35 (12), 1651–1666.
Hitam, M.S., Yussof, W.N.J.W., Awalludin, E.A., Bachok, Z., 2013. Mixture contrast
Webster, M., April 1998. The Art and Technique of Underwater Photography.
limited adaptive histogram equalization for underwater image enhancement.
Fountain Press Ltd., Surray.
In: IEEE International Conference on Computer Applications Technology
Coleman, D.F., Newman, J.B., Ballard, R.D., 2000. Design and implementation of
(ICCAT). Sousse, pp. 1–5. 20–22 January 2013.
advanced underwater imaging systems for deep sea marine archaeological
Li, Y., Lu, H., Li. J., Li, X., Serikawa, S., 2015. Underwater image enhancement using
surveys. Proceeding IEEE International Conference and Exhibition OCEANS of
inherent optical properties. In: Proceeding of the 2015 IEEE, International
Marine Technology Society (MTS), vol. 1. Providence, RI, pp. 661–665.
Conference on Information and Automation. Lijiang, China. August 2015. pp.
Moore, K.D., Jaffe, J.S., 2002. Time-evolution of high-resolution topographic
419–422. http://dx.doi.org/10.1109/ICInfA.2015.7279324.
measurements of the sea floor using a 3-D laser line scan mapping system.
Qing, C., Huang, W., Zhu, S., Xu, X., 2015. Underwater image enhancement with an
IEEE J. Ocean. Eng. 27 (3), 525–542.
adaptive dehazing framework. In: IEEE International Conference on Digital
Xu, X., Negahdaripour, S., 1999. Automatic optical station keeping and navigation of
Signal Processing (DSP). Singapore. 21–24 July 2015. pp. 338–342. http://dx.doi.
an ROV: Sea trial experiment. Proceeding of Marine Technology Society (MTS)/
org/10.1109/ICDSP.2015.7251888.
IEEE OCEANS, vol. 1, pp. 71–76. Seattle, WA.
Rizzi, A., Marini, D., Gatta, C., 2004. From retinex to automatic color equalization:
Schechner, Y.Y., Karpel, N., 2005. Recovery of underwater visibility and structure by
issues in developing a new algorithm for unsupervised color equalization. J.
polarization analysis. IEEE J. Ocean. Eng. 30 (3), 570–587.
Electron. Imag. 13 (1), 75–84.
Church, S.C., White, E.M., Partridge, U.C., 2003. Ultraviolet Dermal Reflection and
Abdul Ghani, A.S., Mat Isa, N.A., 2015. Enhancement of low quality underwater
Mate Choice in the Guppy. Elsevier Science Ltd, pp. 693–700.
image through integrated global and local contrast correction. Appl. Soft
Shamsudin, N., Ahmad, W.F.W., Baharudin, B., Rajuddin, M.K.M., Mohd, F., 2012.
Comput. 37, 332–344.
Significance level of image enhancement techniques for underwater images. In:
Bibalan, M., Amindavar, H., 2015. A new alpha and gamma based mixture
International Conference on Computer & Information Sciences (ICCIS), 12–14
approximation for heavy-tailed rayleigh distribution. In: IEEE International
June 2012, Kuala Lumpur, vol. 1, pp. 490–494.
Conference on Acoustics, Speech, and Signal Processing (ICASSP). South
Anthoni, J.F., 2012. Water and Light in Underwater Photography. http://www.
Brisbane, QLD. 19–24 April 2015, pp. 3711–3715. http://dx.doi.org/10.1109/
seafriends.org.nz/phgraph/water.htm, excessed on September 2012.
ICASSP.2015.7178664.
Schettini, R., Corchs, S., 2010. Underwater image processing: state of the art of
Hurtik, P., Madrid, N., 2015. Bilinear Interpolation over fuzzified images:
restoration and image enhancement methods. EURASIP J. Adv. Signal Process.
enlargement. In: IEEE International Conference on Fuzzy Systems (FUZZ-IEEE).
2010, 746052. February 2010.
Istanbul, 2–5 August 2015. pp. 1–8. http://dx.doi.org/10.1109/FUZZ-IEEE.2015.
Eustice, R., Pizarro, O., Singh, H., Howland, J., 2002. UWIT: Underwater image
7338082.
toolbox for optical image processing and mosaicking in MATLAB. In:
Abdul Ghani, A.S., Mat Isa, N.A., 2014. Underwater image quality enhancement
Proceedings of the 2002 International Symposium on Underwater Technology,
through Rayleigh-stretching and averaging image planes. Int. J. Nav. Archit.
Tokyo, Japan, April 2002, pp. 141–145.
Ocean Eng. 2014 (6), 840–866. http://dx.doi.org/10.2478/IJNAOE-2013-0217.
Iqbal, K., Odetayo, M., James, A., Salam, R. A., Talib, A. Z. H., 2010. Enhancing the low
Beghdadi, A., Negrate, A.L., 1989. Contrast enhancement technique based on local
quality images using unsupervised color correction method. In: International
detection of edges. J. Comp. Visual, Graph. Image Process. 46 (2), 162–274. May
Conference on System Man and Cybernetics (SMC), Istanbul, 10–13 October
1989.
2010, pp. 1703–1709.
Agaian, S.S., Lentz, K.P., Grigoryan, A.M. 2000. A new measure of image
Iqbal, K., Salam, R.A., Osman, A., Talib, A.Z., 2007. Underwater image enhancement
enhancement. In: Proceedings of the International Conference of Signal
using integrated color model. IAENG Int. J. Comp. Sci. 34 (2). IJCS 34-2-12,
Processing and Communication (IASTED), pp. 19–22.
November 2007.
Agaian, S.S., Silver, B., Panetta, K.A., 2007. Transform coefficient histogram-based
Abdul Ghani, A.S., Mat Isa, N.A., 2015. Underwater image quality enhancement
image enhancement algorithms using contrast entropy. IEEE Trans. Image
through integrated color model with Rayleigh distribution. Appl. Soft Comput.
Process. 16 (3), 741–758.
27, 219–230.
Agaian, S.S., Panetta, K., Grigoryan, A.M., 2001. Transform-based image
Petit, F., Capelle-Laize, A.S, Carré P., 2009. Underwater image enhancement by
enhancement with performance measure. IEEE Trans. Image Process. 10 (3),
attenuation inversion with quaternions. In: IEEE International Conference on
367–381.
Acoustics, Speech and Signal Processing (ICASSP’09), Taipei, Taiwan, 19–24
Liang, K., Ma, Y., Xie, Y., Zhou, B., Wang, R., 2012. A new adaptive contrast
April 2009, pp. 1177–1180.
enhancement algorithm for infrared images based on double plateaus
Naim, M.J.N.M., Isa, N.A.M., 2012. Pixel distribution shifting color correction for
histogram equalization. Infrared Phys. Technol. 55 (2012), 309–315.
digital color images. Elsevier: J. Appl. Soft Comput. 12 (9), 2948–2962.
Mittal, A., Soundararajan, R., Bovik, A.C., 2012. NIQE Software Release. URL: http://
September 2012.
live.ece.utexas.edu/research/quality/niqe.zip.