Professional Documents
Culture Documents
To cite this article: G. Jiang, C.Y. Wong, S.C.F. Lin, M.A. Rahman, T.R. Ren, Ngaiming Kwok, Haiyan Shi, Ying-Hao Yu &
Tonghai Wu (2015): Image contrast enhancement with brightness preservation using an optimal gamma correction and
weighted sum approach, Journal of Modern Optics, DOI: 10.1080/09500340.2014.991358
Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”) contained
in the publications on our platform. However, Taylor & Francis, our agents, and our licensors make no
representations or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of the
Content. Any opinions and views expressed in this publication are the opinions and views of the authors, and
are not the views of or endorsed by Taylor & Francis. The accuracy of the Content should not be relied upon and
should be independently verified with primary sources of information. Taylor and Francis shall not be liable for
any losses, actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities whatsoever
or howsoever caused arising directly or indirectly in connection with, in relation to or arising out of the use of
the Content.
This article may be used for research, teaching, and private study purposes. Any substantial or systematic
reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any
form to anyone is expressly forbidden. Terms & Conditions of access and use can be found at http://
www.tandfonline.com/page/terms-and-conditions
Journal of Modern Optics, 2015
http://dx.doi.org/10.1080/09500340.2014.991358
Image contrast enhancement with brightness preservation using an optimal gamma correction
and weighted sum approach
G. Jianga , C.Y. Wonga , S.C.F. Lina , M.A. Rahmana , T.R. Rena , Ngaiming Kwokad∗ , Haiyan Shib ,
Ying-Hao Yuc and Tonghai Wud
a School of Mechanical and Manufacturing Engineering, The University of New South Wales, Sydney, Australia; b School of Computer
Science and Technology, Shaoxing University, Shaoxing, China; c Department of Electrical Engineering, National Chung Cheng
University, Min-Hsiung Township, Taiwan; d Key Laboratory of Modern Design and Rotor Bearing System of Ministry, Xi’an Jiaotong
University, Xi’an, China
(Received 8 October 2014; accepted 17 November 2014)
The enhancement of image contrast and preservation of image brightness are two important but conflicting objectives in
Downloaded by [Selcuk Universitesi] at 04:46 16 January 2015
image restoration. Previous attempts based on linear histogram equalization had achieved contrast enhancement, but exact
preservation of brightness was not accomplished. A new perspective is taken here to provide balanced performance of
contrast enhancement and brightness preservation simultaneously by casting the quest of such solution to an optimization
problem. Specifically, the non-linear gamma correction method is adopted to enhance the contrast, while a weighted sum
approach is employed for brightness preservation. In addition, the efficient golden search algorithm is exploited to
determine the required optimal parameters to produce the enhanced images. Experiments are conducted on natural colour
images captured under various indoor, outdoor and illumination conditions. Results have shown that the proposed method
outperforms currently available methods in contrast to enhancement and brightness preservation.
Keywords: image contrast enhancement; gamma correction; weighted sum averaging; brightness preservation
modifications to cater for minimum brightness error. A pop- imposes a non-linear manipulation of the image intensity
ular method separated the image into two sub-images and while outperforms histogram equalization with regard to its
then performed HE on each sub-image independently. The ability to steer the mean brightness of the processed image
separation threshold was chosen as the mean brightness [16] to a desired value. In the following, some backgrounds and
or the median brightness [17]. However, due to the design related work are briefly reviewed.
limitations, exact brightness preservation was not accom-
plished. Refinements on the image separation concept were
later made where the histogram was modified before con- 2.1. Contrast enhancement
ducting the HE process [18,19]. In these works, the peaks in
the histogram were clipped according to either the mean or Contrast enhancement using the histogram equalization
the median of individual sub-images. Although additional strategy is a class of global enhancement methods [1]. Firstly,
control over brightness manipulation was presented, the the statistics of the input image brightness Iin are collected
algorithms were not designed for optimal settings. in the form of a histogram. It is then normalized to give
Another class of approaches integrates the above two the probability density function, and a cumulative density
strategies aiming at providing a better performance with re- function is further generated. Finally, an output image of
spect to the goals in contrast to enhancement and brightness enhanced contrast is obtained from
preservation. For instance, in the work reported in [20], the Ien (i) ← (L − 1) × cd f (i), (1)
Downloaded by [Selcuk Universitesi] at 04:46 16 January 2015
Clearly, the output image brightness depends on the content range. On the other hand, non-linear manipulations on the
of the input image but the division into sub-images is not image brightness could well be employed to steer the mean
generally possible to provide an exact match of brightness brightness to a desired value while enhancing the contrast
between the input and output images. [1].
Consider the special case that p(lo) = p(hi) = 1/2, Considering that the input image intensities are normal-
the dividing threshold becomes the median of the image ized in [0 1] and risen to power γ ∈ [0 ∞]. Then we have,
brightness [17]. The output mean brightness is as example, cases,
μ(Ien ) = (2μ(Iin ) + L − 1)/4, ∞
(7) Ien,0 = Iin
0
= 1, Ien,1 = Iin
1
= Iin , Ien,∞ = Iin ≈ 0.
which only holds when μ(Iin ) = (L − 1)/2. (10)
The other class of methods modify the histogram by If the parameter γ can be properly determined, then each
clipping the peaks to some chosen threshold in order to enhanced pixel brightness can be manipulated to desired
steer the output image brightness to match that of the input values and contribute to maintaining to overall brightness.
[18,19]. The essence of these methods rests on the modi- For instance, when γ = 0, then all pixel brightness are
fication of the target histogram to be used in equalization. changed to unity, hence, the mean output image brightness
When the target histogram is no more a uniform distribution, is also unity. When γ = 1, there would be no bright-
the mean brightness can be tailored. The reported works, ness change. If γ = ∞, then except when Iin = 1, all
however, still adopt the sub-image division approach. brightness are steered to zero and mean image brightness
Let the image be divided into two sub-images and their also tends to zero. Further developments on this idea are
corresponding histograms are constructed as formulated in the proposed method presented below.
3.1. System description is carried out using the efficient golden section search for its
Asystem block diagram for contrast enhancement and bright- implementation simplicity [22]. When an optimal parameter
ness preservation is depicted in Figure 1. Given a colour set is obtained, the interim image is reconverted to RGB
image in the red–green–blue (RGB) space, from HSI as the output for display or for further processing.
over all pixels in the image. The green and blue channels where Icp is the compressed image and Iex is the expanded
are also stretched in the same way. Figure 2(a) shows the image. The weighting factor α is adopted to adjust the
histogram of the input image brightness, and the stretched resultant image mean brightness. The compressed interim
histogram is shown in Figure 2(b). It is noticeable that image is obtained from
brightness has covered the permissible magnitude range. γ
Icp = Iin , γ > 1, (14)
This process enriches the image colour content and serves
as pre-processing for contrast enhancement. while the expanded interim image is given by
After magnitude stretching, all colour signals are con-
Iex = 1 − (1 − Iin )γ . (15)
verted to the Hue–saturation–intensity (HSI) space. The
I-signal representing the intensity or brightness is processed By limiting the value of α in the range [0 1], it can be
in the system. The extracted I-signal Iin is fed simulta- assured that the magnitudes of the enhanced pixels are con-
neously into a compressor and an expander, both adopt- fined to the permissible bounds. In the extreme cases, when
ing the gamma correction approach with the factor γ , α = 0, the compressed interim image has no contribution
giving interim outputs Icp and Iex , respectively. An ad- to the result. On the other hand, when α = 1, the output
ditional interim-enhanced image Iˆen is derived from the does not contain any contribution from the expansion. The
compressed and expanded images using a weighting factor exponential power factor determines the degree of com-
α. This newly obtained image is tested for an objective pression or expansion on the pixel intensities. When the
function composed of maximum entropy and a penalty term pixel intensities are normalized to unity as the maximum
depending on the brightness error between the input and the value, the gamma-corrected pixel magnitude is reduced for
interim-enhanced image. The governing factors, γ and α are all γ > 1, thus, providing compression towards the dark
optimized for a highest objective function. The optimization regions. On the contrary, if γ < 1, pixel magnitudes are
Figure 1. System block diagram for contrast enhancement with brightness preservation.
Journal of Modern Optics 5
Input Stretched
(a) 4500 (b) 4500
4000 4000
3500 3500
No. of pixels
No. of pixels
3000 3000
2500 2500
2000 2000
1500 1500
1000 1000
500 500
0 0
0 50 100 150 200 250 0 50 100 150 200 250
Intensity Intensity
Compressed Expanded
(c) 4500 (d) 4500
4000 4000
3500 3500
No. of pixels
No. of pixels
3000 3000
2500 2500
2000 2000
Downloaded by [Selcuk Universitesi] at 04:46 16 January 2015
1500 1500
1000 1000
500 500
0 0
0 50 100 150 200 250 0 50 100 150 200 250
Intensity Intensity
Weighted
(e) 4500
4000
3500
3000
No. of pixels
2500
2000
1500
1000
500
0
0 50 100 150 200 250
Intensity
Figure 2. Plots of histograms in different processing stages; (a) input; (b) stretched; (c) compressed; (d) expanded; and (e) weighted. (The
colour version of this figure is included in the online version of the journal.)
increased. However, in order to use only one parameter for In the compress–expand process the same amount of non-
both compression and expansion, the latter is obtained from linear adjustment is applied to obtain two interim enhance-
inverting the pixel magnitudes before rising to power γ and ments, referring to Equations (14) and (15). This strategy
reinverted to give the desired result. ensures that equal modification is applied to both of the
The brightness distributions of the manipulated image interim images in order to cover a widened region of the
signals are shown in Figure 2(c)–(e). In Figure 2(c), it can be permitted intensity range. In particular, notice that the ex-
seen that the peaks of the histogram had shifted to the lower panded interim image is first inverted, by which it is treated
end of the brightness spectrum as a result of magnitude as the compression such that the same gamma factor can be
compression. On the other hand, a shift of the peaks to the applied.
higher end can be seen in Figure 2(d) denoting an image of This secondary interim image Iˆen , from Equation (13),
higher overall brightness. The histogram of weighted sum of is checked for maximum contrast, measured by the image
image brightness is illustrated in Figure 2(e). It is observed entropy, and penalized by the brightness error between the
that the weighted combination of dark and bright images is input and itself. In order to obtain the best performance with
able to produce an image of higher contrast whose mean regard to the dual objectives of contrast enhancement and
brightness can be steered to a given value by adjusting the brightness preservation, the determination of the parameters
gamma correction and weighted sum values. is undertaken as an optimization problem described below.
6 G. Jiang et al.
13:
μ(Iin ) g(Iin , μ(Iin ), γ1 , α)
14: end if
where μ( Iˆen ) is the mean brightness of the interim-enhanced 15: update r = γ2 − γ1
images, μ(Iin ) is the mean brightness of the input image. If 16: end while
there is no brightness error, the image entropy H is restored 17: return objective function J = (J1 + J2 )/2, interim
to its original value. On the contrary, when brightness error enhanced image Iˆen = ( Iˆen,1 + Ien,2 )/2
exists, the entropy is reduced as a penalty. The proposed
contrast enhancement and brightness preservation process
are summarized in the following algorithms. Algorithm 3 is the core process which performs the image
Algorithm 1 is the outer loop responsible for determining enhancement process and produces the objective function. It
the optimal weighting parameter α. It callsAlgorithm 2 in an accepts parameters γ and α from Algorithm 2 and produces
attempt to obtain the optimal gamma correction parameter the interim-enhanced image. The objective function is also
γ . Before entering the inner loop, the mean brightness value calculated in this sub-routine.
of the input image is first calculated as a reference to obtain
the brightness error.
Algorithm 3 Image enhancement
1: Input: input image Iin , mean brightness μ(Iin ), gamma
Algorithm 1 Optimization of weighting parameter α correction parameter γ , weighting parameter α
1: Input: input image Iin 2: Output: objective function J , interim enhanced image Iˆen
γ
2: Output: enhanced image Ien 3: calculate compressed image Icp = Iin
3: calculate input image mean brightness μ(Iin ) 4: calculate expanded image Iex = 1 − (1 − Iin )γ
4: set weighting parameter α1 , α2 , range√r = α2 − α1 5: calculate weighted image Iˆen = α Icp + (1 − α)Iex
5: set golden section parameter ρ = ( (5) − 1)/2 = 0.618, 6: calculate absolute brightness error δ = μ( Iˆen ) − μ(I
tolerance τ = r/100 in )
6: calculate objective functions and interim images using 7: calculate objective function J = H × 1 − μ(Iδ )
in
Algorithm 2; 8: return objective function J , interim enhanced image Iˆen
7: [J1 , Iˆen,1 ] = f(Iin , μ(Iin ), α1 ), [J2 , Iˆen,2 ] =
f(Iin , μ(Iin ), α2 )
8: while r > τ do
9: if J1 > J2 then The golden section search method is basically an iterative
10: set α2 ← α1 + r × ρ process. An initial search region is defined according to
11: calculate objective function [J2 , Iˆen,2 ] = the problem domain, while the search space can be ob-
f(Iin , μ(Iin ), α2 ) tained by carrying out pilot trials. Furthermore, a termina-
12: else
tion condition is defined according to the required solution
13: set α1 ← α1 + r × (1 − ρ)
14: calculate objective function [J1 , Iˆen,1 ] =
precision. In this work, pilot tests had revealed that the
f(Iin , μ(Iin ), α1 ) gamma factor would fall within [1 4], hence, this is the initial
15: end if search region. The termination condition is set to r/100
16: update r = α2 − α1 corresponding to the required solution precision. Higher
17: end while accuracy can be obtained by compromising efficiency with
18: return enhanced image Ien = ( Iˆen,1 + Iˆen,2 )/2
extended iterations.
Journal of Modern Optics 7
Figure 3. Example test images: (a) outdoor – insufficient illumination, (b) outdoor – poor environment condition, (c) outdoor – non-ideal
condition and (d) indoor – colour cast. (The colour version of this figure is included in the online version of the journal.)
shown in images 4(e)–(h), respectively, have little success and the edge around objects are also sharpened with the
enhancing the contrast of the input image. This is primarily boundary of each object clearly shown.
as a consequence of the clipping operation these methods In Figure 5, another hazy input image is presented. Im-
employed to preserve the mean brightness level of the input ages from 5(b)–(d) using UNFHE, BBHE and DSIHE suffer
image. from the same brightness distortion as found in the previous
Image 4(i) using SMHEG method shows an acceptable test image. Image 5(e) and (f) using BHEPL and BHEPLD
balance between brightness preservation and contrast en- has only marginal contrast improvement over the input im-
hancement, while in certain areas such as around the tower, age while preserving brightness level well. Images in 5(h)–
degradation by a fog-like effect is observed. Image 4(j), (j) with AIEBHE, MMSICHE, SMHEQ and CEGAMMA,
where the proposed CEGAMMA is used, gives a clean respectively, show comparative contrast-enhanced result
and neutral image without much distortion of the input with Image (j) having least brightness deviation from the
brightness level. Furthermore, image contrast is enhanced original input image.
with details of image observable in all the areas of the scene,
Journal of Modern Optics 9
(a)
Figure 4. Test results – image 1: (a) input; (b) UNFHE; (c) BBHE; (d) DSIHE; (e) BHEPL; (f) BHEPLD; (g) AIEBHE; (h) MMSICHE;
(i) SMHEQ; (j) CEGAMMA. (The colour version of this figure is included in the online version of the journal.)
To sum up, UNFHE, BBHE and DSHIHE are most prone 4.4. Quantitative evaluation
to brightness distortion due to the way they map input bright- On top of objective viewing, metrics including brightness
ness level linearly according to a pre-set distribution. error, entropy, contrast and gradient are also examined for
BHEPL, BHEPLD, AIEBHE and MMSICHE are better at quantitative comparison of performance among the
preserving image brightness level by posing clipping limits described methods. With 200 colour images evaluated in
around appropriate brightness levels. However, the contrast this experiment, box plots of above-mentioned metrics are
enhancement performance of these methods is not ensured selected for clear illustration of the performance difference.
in all images. Both SMHEQ and CEGAMMAshow good re- The box plot of brightness error in Figure 6(a) indicates that
sults with balancing performance of contrast enhancement the proposed CEGMMA method is preferable over other
and brightness preservation in the example test images by approaches. The CEGAMMA performs better in preventing
utilizing different enhancement philosophies. the generation of viewing artefacts and scene distortions in
the resultant images.
10 G. Jiang et al.
(a)
Downloaded by [Selcuk Universitesi] at 04:46 16 January 2015
Figure 5. Test results – image 2: (a) input ; (b) UNFHE; (c) BBHE; (d) DSIHE; (e) BHEPL; (f) BHEPLD; (g) AIEBHE; (h) MMSICHE;
(i) SMHEQ; (j) CEGAMMA. (The colour version of this figure is included in the online version of the journal.)
It also can be clearly seen in the entropy box plot, In the aspect of gradient performance as shown in
Figure 6(b) that CEGAMMA method leads other methods Figure 6(d), UNFHE, BBHE and DSIHE give best results
in maximizing entropy, which is desirable to convey more with regards to the mean gradient values. However, this
scene information. These results demonstrate an obvious result is predictable as these three methods tend to cre-
advantage of the proposed method to formulate parameters ate large amount of discontinuities in the histogram with
quest as an optimization problem. their linear mapping of brightness value. In spite of having
In the contrast box plot, Figure 6(c), the performance sharper edges around objects in the scene, these methods
of CEGAMMA is slightly worse than UNFHE, DSHIHE also create a lot of viewing discomfort and artefacts. The
and SMHEQ, while better than methods such as BHEPL, proposed CEGAMMA method performs similar to the rest
BHEPLD, AIEBHE and MMSICHE that are based on of discussed methods in this comparison with a mean gra-
setting limits around wanted brightness level to ensure dient value of 0.032.
brightness preservation. The proposed method’s ability on From the above analysis, the proposed CEGAMMA
enhancing contrast is deemed well with its above-average method demonstrates the best performance in brightness
performance. retention and entropy maximization, which is highly
Journal of Modern Optics 11
μ= 0.000 0.022 −0.010 −0.004 −0.029 −0.027 −0.045 −0.010 0.028 −0.000 μ= 7.341 7.183 7.199 7.201 7.120 7.120 7.039 7.127 7.175 7.481
(a) (b)
0.2
7.8
0.15
7.6
0.1 7.4
0.05 7.2
Brightness
Entropy
0
6.8
−0.05
6.6
−0.1
6.4
−0.15 6.2
−0.2 6
5.8
−0.25
A
E
E
t
EQ
E
LD
E
A
E
EQ
L
LD
pu
pu
BH
FH
CH
BH
EP
FH
CH
EP
M
M
In
In
BB
SI
H
EP
BB
SI
H
EP
M
M
BH
BH
N
N
IE
SI
IE
SI
SM
SM
D
D
A
BH
A
BH
U
U
M
M
A
A
G
G
M
M
CE
CE
(c) μ= 0.520 0.573 0.518 0.531 0.461 0.466 0.437 0.493 0.578 0.530
(d) μ= 0.030 0.042 0.041 0.041 0.031 0.031 0.031 0.033 0.036 0.032
0.11
1.2
Downloaded by [Selcuk Universitesi] at 04:46 16 January 2015
0.1
0.09
1
0.08
0.07
0.8 Gradient
Contrast
0.06
0.6 0.05
0.04
0.4 0.03
0.02
0.2 0.01
t
t
E
A
E
E
EQ
EQ
L
LD
LD
pu
pu
H
BH
BH
FH
H
CH
FH
CH
EP
EP
M
M
In
In
BB
SI
BB
SI
H
H
EP
EP
M
M
BH
BH
N
N
IE
IE
SI
SI
SM
SM
D
D
A
A
BH
BH
U
U
M
M
A
A
G
G
M
M
CE
CE
Figure 6. Statistics of test results in box plots: (a) brightness error; (b) entropy; (c) contrast; (d) gradient. (The colour version of this figure
is included in the online version of the journal.)
beneficial in contrast enhancement techniques to avoid scene values, is 0.374. The shift of the mode from the mean value
artefacts and to convey maximum scene information. Fur- may be attributed to the fact that most images require com-
thermore, comparable performance against other referred pression or modification to lower intensities because of the
methods in contrast and gradient is also observed in the box characteristics of the image content. By inspecting the test
plots. Overall, CEGAMMA achieves superior image en- images shown in Figures 4 and 5, it can be seen that images
hancement and brightness preservation performance against captured under foggy or hazy environment tend to cast
compared methods. the scene into white or towards the high-intensity region.
Thus, the proposed algorithm modifies the images toward
the darker intensities as a result of the enhancement process.
4.5. Parameter settings A plot of the distribution of the gamma correction factor
Based on the 200 test images, statistics of the obtained is given in Figure 7. The mean value is 1.365 and the mode
weighting factor α and the gamma correction factor γ are is 1.075 with a second peak located at 1.400. This factor
shown as distributions in Figure 7(a) and (b). For the weight- is responsible for the degree of enhancement produced as
ing factor, Figure 7(a), the average value is 0.491. It can the intensity ranges were compressed or expanded to the
be conjectured that approximately equal efforts had been two ends of the intensity region. We can see that a certain
exerted in increasing image intensities to the higher amount of power law correction is needed to produce an
magnitude region as well as in decreasing the intensities enhanced image. However, this parameter depends largely
to the lower magnitude region. The net effect resulted in on the image content. If the input image has sufficient con-
expanding the intensity range, thus, covering the allowed trast, then a small gamma is needed. On the contrary, when
magnitude range uniformly and increasing the informa- the input image contrast is low, then a larger correction is
tion content. The mode or the largest number of parameter required.
12 G. Jiang et al.
μ=0.491, η=0.374 μ=1.365, η=1.075
(a) 70 (b) 35
60 30
50 25
No. of images
No. of images
40 20
30 15
20 10
10 5
0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1 1.5 2 2.5 3 3.5 4
Factor α Factor γ
Figure 7. Distribution of algorithm parameters: (a) weighted-sum parameter α; (b) gamma correction parameter γ . (The colour version
of this figure is included in the online version of the journal.)
Downloaded by [Selcuk Universitesi] at 04:46 16 January 2015
5. Conclusion [6] Wang, X.; Zhang, D. Expert Syst. Appl. 2013, 40, 5854–
5866.
Anew procedure for image contrast enhancement and bright-
[7] Wu, T.; Wu, H.; Du, Y.; Kwok, N.; Peng, Z. Wear 2014, 316,
ness preservation had been presented. The proposed method 19–29.
was designed aiming at overcoming the limitations of his- [8] Su, F.; Fang, G.; Kwok, N. Shadow Removal Using
togram equalization-based approaches in providing close Background Reconstruction. In 5th International Congress
matches of input image brightness. Unlike the currently on Image and Signal Processing (CISP), 2012, Chongqing,
pp 154–158.
available techniques, this work adopted the gamma cor-
[9] Yao, Y.; Abidi, B.; Doggaz, N.; Abidi, M. Evaluation of
rection strategy together with weighted sum to derive Sharpness Measures and Search Algorithms for the Auto
high-contrast images with a significant improvement in Focusing of High-Magnification Images. In Defense and
brightness preservation. In particular, image intensity mag- Security Symposium, Orlando, FL, 2006, pp 62460G-1–
nitudes were compressed, expanded and then aggregated. 62460G-12.
[10] Tanaka, G.; Suetake, N.; Uchino, E. Opt. Rev. 2010, 17, 130–
High-performance results were accomplished by casting
138.
the problem into an optimization context where algorithmic [11] Choi, H.H.; Yun, B.J. Opt. Rev. 2011, 18, 389–393.
parameters were iteratively searched and optimally deter- [12] Kwok, N.; Shi, H.; Ha, Q.; Fang, G.; Chen, S.; Jai, X. Eng.
mined. The effectiveness of the proposed method was veri- Appl. Artif. Intell. 2013, 26, 2356–2371.
fied using a collection of a large number of images captured [13] Yoon, B.W.; Song, W.J.J. Electron. Imag. 2007, 16, 033005–
033005.
in natural scenes under various illumination and environ-
[14] Kong, N.S.P.; Ibrahim, H. Comput. Electr. Eng. 2011, 37,
ments. Results had shown that the developed algorithm 631–643.
outperformed others in terms of brightness preservation, [15] Kwok, N.; Shi, H.; Fang, G.; Ha, Q.P. Intensity-Based
information content increment as well as the reduction in Masking Gain Adaptive Unsharp for Image Contrast
viewing artefacts produced. Enhancement. In 5th International Congress on Image and
Signal Processing (CISP), 2012, Chongqing, pp 529–533.
[16] Kim, Y.T. IEEE Trans. Consum. Electron. 1997, 43(1), 1–8.
References [17] Wang, Y.; Chen, Q.; Zhang, B. IEEE Trans. Consum.
[1] Gonzalez, R.C.; Woods, R.E. Digital Image Processing, 3rd Electron. 1999, 45, 68–75.
ed.; Prentice-Hall Inc, Upper Saddle River, NJ, 2006. [18] Ooi, C.H.; Kong, N.S.P. IEEE Trans. Consum. Electron.
[2] Sun, W.; Han, L.; Guo, B.; Jia, W.; Sun, M. J. Mod. Optic 2009, 55, 2072–2080.
2014, 61, 1–12. [19] Ooi, C.H.; Isa, N.A.M. IEEE Trans. Consum. Electron. 2010,
[3] Zhang, G.; Jia, X.; Kwok, N. Super Pixel Based Remote 56, 2543–2551.
Sensing Image Classification with Histogram Descriptors [20] Singh, K.; Kapoor, R. Optik Int. J. Light Electron Opt. 2014,
on Spectral and Spatial Data. In International Geoscience 125, 4646–4651.
and Remote Sensing Symposium (IGARSS), Munich; 2012, [21] Tang, J.R.; Isa, N.A.M. Comput. Electr. Eng. 2014, 21,
pp 4335–4338. 86–103.
[4] Guo, W.; Xia, X.; Xiaofei. Expert Syst. Appl. 2014, 41, 6446– [22] Chong, E.K.P.; Zak, S.H. An Introduction to Optimization,
6458. John Wiley: Hoboken, NJ, 2013.
[5] Sundaram, M.; Ramar, K.; Arumugam, N.; Prabin, G. Appl. [23] Kwok, N.M.; Jia, X.; Wang, D.; Chen, S.Y.; Fang, G.; Ha,
Soft. Comput. 2011, 11, 5809–5816. Q.P. Comput. Electr. Eng. 2011, 37, 681–694.