See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/312963233

Image Enhancement Based on Equal Area Dualistic Sub-image and Non-

parametric Modified Histogram Equalization Method

Conference Paper · December 2016

DOI: 10.1109/ISCID.2016.1110

CITATIONS READS

14 1,020

3 authors, including:

Zhijun Yao Zhongyuan Lai

Yangzhou University JiangHan University
22 PUBLICATIONS 151 CITATIONS 26 PUBLICATIONS 134 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Zhongyuan Lai on 14 October 2017.

The user has requested enhancement of the downloaded file.

 2016 9th International Symposium on Computational Intelligence and Design

Image Enhancement Based on Equal Area Dualistic

Sub-Image and Non-parametric Modified Histogram
Equalization Method

Zhijun Yao Zhongyuan Lai* Chun Wang

The 723 Institute of China Shipbuilding Institute for Interdisciplinary Research Institute of Computer Application,
Industry Corporation Jianghan University China Academy of Engineering
Yangzhou, China Wuhan, China Physics
[email protected] [email protected] Mianyang, China
[email protected]

Abstract—This paper presents a new contrast enhancement The first type is the methods without image division.
method using Histogram Equalization named Dualistc Sub- They equalize the original image directly. To further
Image and Non-parametric Modified Histogram Equalization enhancethe image contrast, they often modified the original
(DSINMHE). Our proposed method consists of three steps: (i) histogram before equalization. The representative methods
The original image is segmented into two sub-images by the
median value of the image. (ii) Then we use the histogram
include Histogram Modification Framework [2] and Non-
modification technique to maximize entropy and control over parametric Modified Histogram Equalization (NMHE) [3].
enhancement problem. (iii) Two sub images are enhanced by the These methods are of low complexity and easy
traditional histogram equalization method independently and implementation. However, they have the following three
then are merged into one output enhanced image. The problems. Firstly, the mean value is easy to be moved to the
experimental results show that BPNMBHE is superior to other middle of the range of pixel values, which leads to the
contrast enhancement methods according to image subjective brightness variation. Secondly, these methods cannot be
evaluation and various image objective evaluation measures, i.e. adaptive to the contrast variations over different regions, and
Entropy, AMBE,PSNR, and SSIM. may result in over enhancement. Thirdly, these methods cannot
Keywords: Histogram modification; Entropy; Histogram
well describe the local details and may introduce noise when
equalization; Image objective evaluation measure;Brightness the object is relatively small, making the original image
preserving i n f o r m a t i o n d e s t r o y e d .
The second type is the methods with image division.
I. INTRODUCTION These methods firstly divide image into many sub-images and
then equalize them individually. The representative methods
Image contrast enhancement is a important problem in include BBHE [4] and DSIHE [5]. It has been proved that these
low level image processing. In many cases, the quality of the two methods can well preserve the brightness. To further
acquired image is not good enough due to the restrictions of preserve the details, RMSHE [6] and RSIHE [7] were proposed.
background and light conditions as well as those of the camera However, it is difficult to determine the number of iterations.
devices. With the help of image enhancement technology, one When it is too large, the enhanced image is almost the same as
can improve the visual effect of the image or enhance some the original one, which sacrifices the overall effect of contrast
specific information in image, so as to benefit the object enhancement [8].
recognition and image understanding. For the purpose of better enhancing the contrast while
Histogram Equalization (HE) is one of spatial domain maintaining the brightness and protecting local details, we
contrast enhancement method and extensively utilized because propose a image enhancement method, namely Dualistc Sub-
of its simplicity, ease of implementation and effectiveness [1]. Image and Non-parametric Modified Histogram Equalization
The principle is that the information content and image contrast (DSINMHE). Firstly, the original image is segmented into two
will reach their maxima when the pixel values are uniformly sub-images by means of its median. Then, we individually
distributed. In light of this principle, the HE methods stretch modify the histogram of each sub image using the NMHE
the pixel values with dense distribution and merge those with method. After that, we equalize the two sub-images according
spare distribution. In this way, the distribution of pixel values to the two modified histograms, and finally merge them into a
will be closer to uniform distribution, and the image contrast complete image. Our method takes the advantages of the
can be enhanced. histogram division and non-parameter modified histogram
In general, there are two types of HE methods. equalization, and can perform well in both contrast
enhancement and brightness keeping. Moreover, the local
* Corresponding author

2473-3547/16 $31.00 © 2016 IEEE 447

DOI 10.1109/ISCID.2016.109
details are well preserved after enhancement, which makes the histogram is gained by using only those pixels that have a two-
output images look natural. lagged horizontal diversity larger than a threshold value given
in (8).
II. THE PROPOSED METHOD
hi' ( k ) = p ª¬ k C º¼ (8)
This section will describle our proposed DSINMHE
method in detail. DSINMHE includes three steps: Median where p ª¬ k C º¼ is the probability of occurrence of the given
Calculation and Image Division, Histogram Modification and event C, and C represents a horizontal contrast strength. Like
Sub-Histogram Equalization. The following subsections will [3], C is set to six empirically in our experiments. According
describe these three steps. to Id, we can segment the modified histogram into two sub
A. Median Calculation and Image Division histograms. And then two sub modified histogram are
normalized to keep the value in between 0 and 1, and denoted
Assume that the gray level range of the original image I is
as h i' ,low ( k ) and h i' ,up ( k ) .
[0, L-1], where L is the number of gray scale. Therefore, the
PDF of the original image, hi(k), can be written as: The next task is to compute the weight factor w in (7),
hi ( k ) = H ( k ) / N (1) which is used to balance hi and u. Because there are two sub
where H(k) is the total number of pixels whose gray value images Ilow and Iup, we have to compute two weighting factors
L −1 wlow and wup respectively. Firstly, according to the mean Id, we
¦
equals to k, and N = H ( k ) . obtain two uniform PDFs ulow and uup as given in (9) and (10).
ulow = ones ( I d + 1,1) ( I d + 1)
k =1
(9)
Like [3], we use the median of the image to segment the
original image. The median is denoted as an gray scale Id uup = ones ( L − I d − 1,1) ( L − I d − 1) (10)
where the cumulative density function equals to 0.5 [3]. For Two sub-histograms are obtained from hi by the following
computing Id, consider a variable x(k) as calculated in (2): two equations respectively
x ( k ) = x ( k − 1) + hi ( k ) for k = 0,1,..., L − 1 and x ( 0 ) = h i ( 0 ) hi ,low ( k ) = λlowhi ( k ) , k = 0,1,..., I d (11)
(2) hi ,up ( k ) = λup hi ( k ) , k = I d + 1,..., L − 1 (12)
Id can be computed as (3) by above two equations. Id L −1
I d = arg min x ( k ) − 0.5
k
(3) where λ = 1
low ¦ h ( k ) and λ
k =0
i up =1 ¦ h (k ) . i
k = I d +1
Then, according to Id, the image I is divided into two sub Then two sub-histograms are given as (13) and (14)
images Ilow and Iup, which are given in (4) to (6). c
h low = min ( h i ,low ( k ) , u low ( k ) ) , k = 0,1,..., I d (13)
I = I low ∪ I up (4)
h = min ( h i ,up ( k ) , uup ( k ) ) , k = I d + 1,..., L − 1
c
up
(14)
where
Two weighting factors wlow and wup are now calculated by
{ }
I low = I ( i, j ) I ( i, j ) ≤ I m , ∀I ( i, j ) ∈ I (5)
the following two equations respectively
and wlow = sum ( ulow − hlow
c
) (15)
{
I up = I ( i, j ) I ( i, j ) > I m , ∀I ( i, j ) ∈ I} (6)
wup = sum ( u up − hup
c
) (16)
Please note that Ilow is made up of {I0, I1, ..., Id }, and Iup is The respective PDF of two sub-images Ilow and Iup are
made up of { Id+1, Id+2, ..., IL-1}. computed by
B. Histogram Modification m
h low ( k ) = (1 − wlow ) hi' ,low ( k ) + wlowulow ( k ) , k = 0,1,..., I d (17)
The traditional HE often yields over enhanced image m
h up ( k ) = (1 − wup ) hi' ,up ( k ) + wup uup ( k ) , k = I d + 1,..., L − 1 (18)
Because HE technique tries to create a uniform distribution
histogram. For overcome this disadvantage, we can modify the C. Sub-Histogram Equalization
input original histogram. Then, we use the modified histogram The CDFs of two sub-images Ilow and Iup are defined as
to create the mapping function of HE. Here, the modification k
process can be viewed as a double standard problem Clow ( k ) = ¦ hlow
m
( q ) ,k = 0,1,..., I d (19)
optimization process. The objective is to look for a modified q =0
k
histogram that approaches the uniform distribution histogram
(uu) as required, and makes the remainder hi - u little [7]. A
Cup ( k ) = ¦
q = I d +1
m
h up ( q ) ,k = I d + 1,..., L − 1 (20)

solution to this optimization can be written as According to (19) and (20), the transformation functions of
h = (1 − w ) ⋅ h i + w ⋅ u (7) two sub-images for histogram equalization are defined as

where h is the modified histogram, w is a weighting factor flow ( k ) = I d × Clow ( k ) , k = 0,1,..., I d (21)
whose range is 0 to 1. fup ( k ) = I d + 1 + ( L − I d − 2 ) × Cup ( k ) , k = I d + 1,..., L − 1 (22)
For computing hi, Arici et al. [7] proposed a simple and
intuitive method that only adopts those pixels which have Then, these two transformation functions are used to
dissimilarity with their neighbors. Specifically, the modified equalize the divided sub-images independently. In the end,

448
the enhanced image of DSINMHE, E, is described by (23) to TABLE II. PSNR MEASURE OBTAINED FROM FOUR SAMPLE IMAGES
(25)  Tank Field Landscape Girl Average
E = Elow ∪ Eup (23)
HE 13.778 14.913 14.303 13.003 13.999
Where
BBHE 13.893 15.328 14.432 13.304 14.239
{
Elow = flow ( I low ) = flow ( I ( i, j ) ) ∀I ( i, j ) ∈ I low } (24)
DSIHE 13.758 15.410 14.308 12.936 14.103
and RMSHE 17.336 19.915 17.396 28.024 20.668
{
Eup = fup ( I up ) = fup ( I ( i, j ) ) ∀I ( i, j ) ∈ I up } (25) RSIHE 17.432 19.358 18.111 19.859 18.690
NMHE 17.155 18.447 17.334 14.490 16.856
III. EXPERIMENTAL RESULTS
DSINMHE 24.655 23.326 22.553 31.667 25.550
In this section, we quantitatively and qualitatively
substantiate the advantages of our proposed DSINMHE TABLE III. ENTROPY MEASURE OBTAINED FROM FOUR SAMPLE IMAGES
method compared to existing other HE methods. The above
substantiation is mainly specified in the aspects of the  Tank Field Landscape Girl Average
brightness preservation, noise level, contrast enhancement Original 5.496 6.563 6.299 5.594 5.988
level, detail preservation and visual effects. To this end, we HE 4.966 5.956 5.463 4.676 5.265
carefully select four images, namely Tank, Field, Landscape,
and Girl, for our tests. These four images contain a variety of BBHE 5.430 6.465 6.227 5.291 5.853
backgrounds, lighting conditions, and object sizes. For DSIHE 5.395 6.470 6.215 5.293 5.843
comparison, BBHE, DSIHE, RMSHE, RSIHE, and NMHE RMSHE 5.471 6.497 6.126 5.354 5.862
are used. For RMSHE and RSIHE, the number of the RSIHE 5.453 6.516 6.159 5.223 5.838
iterations are set to 2, i.e. the original histogram is divided into
NMHE 5.258 6.345 5.924 5.444 5.743
four sub-histograms.
DSINMHE 5.460 6.475 6.263 5.576 5.944
A. Quantitative Comparison
In this subsection, we confirm the effectiveness of our TABLE IV. SSIM MEASURE OBTAINED FROM FOUR SAMPLE IMAGES
proposed DSINMHE method from a qualitative point of view.
In order to comprehensively reflect the advantages of our  Tank Field Landscape Girl Average
method, we select the following four performance assessment HE 0.425 0.682 0.467 0.302 0.469
indices, namely AMBE, PSNR, Entropy, and SSIM. BBHE 0.533 0.678 0.520 0.414 0.536
Tables I-IV list the above four indices for the Field,
DSIHE 0.454 0.678 0.508 0.301 0.485
Landscape, Girl and Tank using various HE methods. From
Table I, we can see that, except the Landscape, the proposed RMSHE 0.626 0.853 0.578 0.960 0.754
method produces the lowest AMBE values on the remaining RSIHE 0.609 0.824 0.580 0.725 0.685
three test images, which shows the superiority of our method NMHE 0.723 0.837 0.904 0.933 0.849
in terms of brightness preservation. From Table II, it can be DSINMHE 0.884 0.880 0.894 0.970 0.907
observed that DSINMHE produces the lowest PSNR values on
all test images, which indicates that our method introduces the SSIM value of our method on the Landscape is slightly lower
least noise. From Table III, we can see that, except that the than the best one, those on the remaining three images are the
entropy values of our method on the Tank and Field are highest, which shows the superiority of our method in terms of
slightly lower than the best ones, those on the other two test structure preservation. Furthermore, our proposed method has
images are the lowest, which indicates that our method is the best average AMBE, PSNR, Entropy, and SSIM values
competitive to existing methods in terms of information over the four test images, which shows the advantages of our
preservation. From Table IV, we can see that except that the method in the overall performance.
B. Qualitative Comparison
TABLE I. AMBE MEASURE OBTAINED FROM FOUR SAMPLE IMAGES
In this subsection, we confirm the effectiveness of our
 Tank Field Landscape Girl Average method from a quantitative point of view. Figures 1-4 show
HE 4.812 21.040 3.471 12.193 10.379 the enhancement results for the Field, Landscape, Girl and
Tank with different HE methods. From Figure 1, we can see
BBHE 21.252 14.705 11.281 22.090 17.332
that the contrast of the original image is relatively low, but the
DSIHE 5.343 13.638 6.879 14.321 10.045 HE, BBHE and DSIHE methods significantly over enhance
RMSHE 13.520 8.282 9.503 0.747 8.013 the original one. RMSHE and RSIHE, by contrast, have better
RSIHE 6.728 7.506 2.253 4.828 5.329 results, but still over enhance the backgrounds. Compared
NMHE 30.506 17.441 28.166 45.846 30.490
with the original image, the result produced by NMHE is too
dark. Our method can produce the result with moderate
DSINMHE 5.295 1.531 5.988 0.004 3.205 brightness, natural background, and good detail preservation.

449
 (a) (b) (c) (d)

(a) (b) (c) (d)

(e) (f) (g) (h)

Figure 3. Enhancement results of Landscape image: (a) Original, (b) HE, (c)
BBHE, (d) DSIHE, (e) RMSHE, (f) RSIHE, (g) NMHE, (h) DSINMHE

(e) (f) (g) (h)

Figure 1. Enhancement results of Tank image: (a) Original, (b) HE, (c) BBHE,
(d) DSIHE, (e) RMSHE, (f) RSIHE, (g) NMHE, (h) DSINMHE

(a) (b) (c) (d)

(a) (b) (c) (d)

(e) (f) (g) (h)

Figure 4. Enhancement results of Girl image: (a) Original, (b) HE, (c) BBHE,
(d) DSIHE, (e) RMSHE, (f) RSIHE, (g) NMHE, (h) DSINMHE

(e) (f) (g) (h) ACKNOWLEDGEMENT

Figure 2. Enhancement results of Field image: (a) Original, (b) HE, (c) BBHE,
(d) DSIHE, (e) RMSHE, (f) RSIHE, (g) NMHE, (h) DSINMHE The authors thank Hongmei Chen, Bo Wang, Wu Xia, Ying
Yang, and Zhixin Pan from Jianghan University for their help
The cases in Figure 2 are similar to those in Figure 1. From in experiments. This work is supported in part by the National
Figure 3, we can see that the HE, BBHE and DSIHE methods Natural Science Foundation of China (grant No. 61501208)
significantly over enhance the snow mountains. The results and the Scientific Research Foundation of Jianghan University
produced by MSHE and RSIHE are slightly better than the to Dr. Zhongyuan Lai.
formers, whereas the results produced by NMHE is too bright.
Our method, by contrast, has the best comprehensive effects. REFERENCES
From Figure 4, we can see that the HE, BBHE and DSIHE
[1] R. C. Gonzalez, and R. E.Woods, Digital Image Processing, 2nd ed.,
methods make much background noise. This noise is reduced Prentice Hall, 2002.
by RSIHE, but the girl’s clothes region is over enhanced. [2] T. Arici, S. Dikbas, Y. Altunbasak, “A histogram modification
NMHE changes the original light condition, and the overall framework and its application for image contrast enhancement,” IEEE
result is too bright. To view the image as a whole, the effects Trans. Image Processing, Vol.18, No.9, pp. 1921-1935, 2009.
of DSIHE and the proposed method are similar. However, [3] S. Poddar, S. Tewary, D. Sharma, et al., “Non-parametric modified
histogram equalisation for contrast enhancement,” IET Image
when we carefully observe Figures 4 (e) and (h), we can find Processing, Vol.7, No.7, pp. 641-652, 2013.
that our light on the girl’s is softer and thus the face is more [4] Y. T. Kim, “Contrast enhancement using brightness preserving bi-
nature. These fully demonstrate the advantages of our method histogram equalization,” IEEE Trans. on Consumer Electronics, Vol.43,
in brightness preservation, noise level, contrast enhancement, No.1, pp.1-8, 1997.
detail preserving and visual effects. [5] Y. Wan, Q. Chen, B.M. Zhang, “Image enhancement based on equal
area dualistic sub-image histogram equalization method,” IEEE Trans.
IV. CONCLUSIONS Consumer Electronics, Vol.45, pp.68-75, 1999.
[6] S. D. Chen, A. R. Ramli, “Contrast enhancement using recursive mean-
This paper presents a novel image enhancement method separate histogram equalization for scalable brightness preservation,”
called DSINMHE. In the proposed method, the histogram IEEE Trans. on Consumer Electronics, Vol.49, No.4, pp. 1301-1309,
modification approach is used to control over enhancement and 2003.
maximize entropy in the histogram equalization process. [7] K. S. Sim, C. P. Tso, Y. Y. Tan, “Recursive sub-image histogram
Experimental results obviously show that DSINMHE equalization applied to gray scale images,” Pattern Recognition Letters,
outperforms other contrast enhancement methods based on four Vol.28, No.10, pp. 1209-1221, 2007.
image objective evaluation measures. Subjective evaluation [8] H. Ibrahim, N. S. P. Kong, “Brightness preserving dynamic histogram
equalization for image contrast enhancement,” IEEE Trans. on
results also demonstrate the superiority of DSINMHE over Consumer Electronics, Vol.53, No.4, pp.1752-1758, 2007.
other methods according to natural appearance.

450

View publication stats