You are on page 1of 10

IET Image Processing

Research Article

New adaptive histogram equalisation ISSN 1751-9659


Received on 22nd January 2019
Revised 1st August 2019
heuristic approach for contrast enhancement Accepted on 16th September 2019
E-First on 5th April 2020
doi: 10.1049/iet-ipr.2019.0106
www.ietdl.org

Shubhi Kansal1 , Rajiv Kumar Tripathi1


1Electronicsand Communication Department, National Institute of Technology, Narela, New Delhi, India
E-mail: kansalshubhi@yahoo.com

Abstract: Contrast enhancement of an image can be performed by using a simple histogram equalisation (HE) technique.
However, there are some drawbacks of HE like immense brightness change, artificial effects, over-enhancement, which make it
unsuitable to be used in many applications. To resolve these issues a new adaptive heuristic HE approach is proposed in this
study. First, probability distribution function (PDF) of the image is calculated. Second, an adaptive parameter is calculated based
on the mean and maximum values of that PDF. Thereafter, PDF and cumulative distribution function (CDF) are modified by
applying a threshold limit to that adaptive parameter. Finally, another adaptive parameter is finding out by using modified CDF
and a new CDF is obtained by using this second adaptive parameter. Traditional HE is then applied with the new CDF to getting
the enhanced image. The visual and quantitative results of the proposed method outperform all other state-of-the-art papers and
works well both for low and bright contrast images simultaneously. After rigorous experiment, it is concluded that the authors’
method enhances the image contrast very well with no over-enhancement or artificial effects in the images and also preserves
the original characteristics of the input images.

1 Introduction function (PDF) of the image was changed according to those


values. This method has limited the brightness of the image.
Image enhancement is an essential and frequent process in image In 2010, quadrants dynamic HE [8] was proposed for contrast
processing [1]. Before an image can be utilised for any further enhancement. This method divided the histograms into four sub-
application, image enhancement is required for a clear vision. It is histograms based on the median and then every four histograms
extensively utilised for medical image processing, remote sensing were equalised independently. This method lost the natural
imaging, character recognition, biometric recognition surveillance characteristics of the images.
system and many more. Histogram equalisation (HE) is a Adaptive gamma correction with weighting distribution
straightforward and frequently used technique for image (AGCWD) [9] was proposed in 2013. This method applied gamma
enhancement purposes due to its effectiveness and immediate correction and the probability distribution of luminance pixels for
results. It generally distributes the image grey levels evenly contrast enhancement. This method worked good for plain images,
throughout the range, but it also changes the overall brightness of but it failed on the images, which contained very less bright pixels.
the image due to which artefacts are introduced in the image. As a A quantile-based HE [10] method was proposed for brightness
result instead of enhanced look, the image looks unnatural and preservation and contrast enhancement in 2014. This method
noisy. To overcome these issues numerous global HE (GHE) segmented the histogram on the basis of the quantile values.
methods have been proposed after modifying the basic HE method. Though this method preserved the image brightness, it failed to
Brightness preserving bi HE (BBHE) was given in 1997 by bring out the details of the image.
Kim [2] for preserving the image mean brightness. BBHE divided The exposure-based sub-image HE (ESIHE) [11] method was
the original image histogram into two portions based on its mean also proposed in 2014. This method enhanced the contrast of the
and then each individual histogram was equalised separately. image very well and also avoided the over enhancement problem
Dualistic sub-image HE (DSIHE) [3] was proposed in 1999 of HE by utilising the clipping process. However, it failed to
similar to BBHE, but here the histograms were divided on the basis preserve the image brightness.
of the median instead of mean. Both of these techniques succeeded A piecewise HE method was proposed for dark image
in preserving the image mean brightness, but failed to conserve the enhancement [12]. This method focused on maintaining the image
natural features. intensity as well as contrast boosting. However, since this method
Recursive mean separate HE [4] and recursive sub-image HE was specially designed for dark images, it did not work on bright
[5] were proposed in 2003 and 2007, respectively. Both methods and normal light images.
preserved image brightness much better than BBHE and DSIHE. Entropy maximisation histogram modification [13] scheme was
Basically image histograms were separated recursively on the basis proposed in 2015 for contrast enhancement. It comprised pixel
of mean and median, respectively. The number of separations of populations mergence (PPM) step and the grey-levels distribution
image histogram preserved the image mean brightness up to step for merging of non-zero bins. This method resulted in
significant extent, but it increased the complexity of the system. saturation artefacts in the images.
In 2009, the histogram modification framework (HMF) [6] was In 2016 entropy-based dynamic sub-HE [14] was proposed for
proposed for contrast enhancement. In this method, the level of contrast enhancement. In this method, the histograms were divided
contrast enhancement was adjusted based on some optimisation recursively based on the entropy, and then equalisation was
parameters which reduced the cost function. This method enhanced performed independently. Since this method also involved
the image, but it introduced artefacts in some images with slope recursive division, it increased the complexity of the system.
histogram spikes. In 2017, another method was proposed based on gamma
Again, in 2010, a histogram modification scheme was proposed correction and addition histogram modification (GCAHM) [15]. In
[7]. Two boundary values were found and probability density this sum of the input histogram and its standard deviation were
computed. Then, gamma correction was applied to the result sum

IET Image Process., 2020, Vol. 14 Iss. 6, pp. 1110-1119 1110


© The Institution of Engineering and Technology 2019

Authorized licensed use limited to: Auckland University of Technology. Downloaded on June 08,2020 at 03:17:28 UTC from IEEE Xplore. Restrictions apply.
to generate a modified histogram. And finally, the traditional HE H(Ik)
p(Ik) = (1)
was applied for enhancement. This method enhanced the contrast M×N
of some images very well, but showed limited enhancement for
some images. where M × N denotes the size of the image. Now, we find the mean
Recently in 2018, a plateau limit-based tri-HE (PLTHE) [16] and maximum values of the above PDF as
was proposed for image enhancement. In this method, the
histogram of the image was clipped based on the plateau limit, and pmean = mean[p(Ik)] (2)
then it was divided into three parts using histogram subdivision
limit parameters that were calculated on the basis of the standard pmax = max[p(Ik)] (3)
deviation of the image.
Some heuristic methods have also been proposed for
enhancement. 3.1 Calculation of adaptive parameter (R) and PDF
A heuristic-based face recognition technique was proposed [17] modification
in 2011. In this method, multiple faces were studied based on Now, we define an adaptive parameter (R) as
different criteria over a period of time, and then heuristic
supplemented PCA algorithm was applied to find the optimal pmax
solution. R= (4)
pmean
In 2014, a heuristic algorithm for extracting useful information
from an image was proposed [18]. This method applied a heuristic
approach for edge detection from the images. It used very basic It is adaptive in the sense as pmean and pmax values will be different
concepts of image processing to reach its goal. for different images. Now, we apply a threshold limit to this
In this paper, a new and simple adaptive HE method is proposed adaptive parameter and modify the PDF accordingly.
for image contrast enhancement. It is a heuristic approach. If R > = 10 than
Basically, the heuristic technique is not based on any specified
model or algorithm; instead it is a practical approach adopted for pmean, if (p(Ik) ≥ pmean)
p(Ik) =
finding the solution to the problem. Here, trial and error heuristic p(Ik), otherwise
method is used to find the best results after testing on number of
images. First of all, PDF of the image is calculated. Then, an This threshold limit is chosen heuristically after testing on various
adaptive parameter is calculated based on the mean and maximum images. If the value of R is <10, then PDF of the image will remain
value of that PDF. Thereafter, PDF and CDF are modified by as it is without any change.
applying a threshold limit to that adaptive parameter. Finally,
another adaptive parameter is finding out by using modified CDF 3.2 Calculation of CDF and its modification
and a new CDF is obtained by using this second adaptive
parameter. Traditional HE is then applied with the new CDF to Now, the CDF of the image is calculated as
getting the enhanced image.
The proposed algorithm is implemented and compared to other L−1
state-of-the-art papers visually and using image quality metrics CDF(Ik) = ∑ p(Ik) (5)
such as entropy, quality index (QI), quality-aware relative contrast k=0
measure (QRCM), structural similarity index measure (SSIM) and
colour quality enhancement (CQE). The experimental results show Again, the above condition is followed and CDF has modified
that the proposed method provides excellent visual quality of the accordingly:
images and also balanced parameter results. If R > = 10
This paper is organised as follows: Section 2 gives the
motivation behind the proposed work. Section 3 describes the CDF(Ik) = CDF(Ik) − min[CDF(Ik)] (6)
proposed technique. Section 4 gives experimental results and
discussion, and finally, Section 5 concludes the paper. where min[CDF(Ik)] is the minimum value of CDF(Ik). Also, we
normalise this CDF
2 Motivation
CDF(Ik)
The purpose of proposing this particular technique is to enhance CDF(Ik) = (7)
max[CDF(Ik)]
the contrast of both low and bright images. As discussed in the
above section all state-of-the-art methods have some or the other Hence, it is clear that if R < 10, then CDF will be standard CDF of
drawbacks linked with them such as the image without any changes in it.
• Mean-shift in the brightness of the image.
3.3 Calculation of second adaptive parameter (α) and new
• Over-enhancement of the images.
CDF
• Loss of natural characteristics of the input image.
• Increased complexity of the system due to a large number of Now, we find another adaptive parameter α using above CDF as
calculations.
• Saturation artefacts in the images. sum[CDF(Ik)]
α= (8)
• Limited enhancement of the images. 255

The proposed technique can overcome these problems and results α is adaptive as sum[CDF(Ik)] value will be different for different
in a visually appealing image with high contrast and natural look. images. This α value is also chosen heuristically after testing on a
number of images.
3 Proposed adaptive HE method Now, a new CDF is computed using the value of α
This section describes the proposed adaptive HE method. Fig. 1
CDFn(Ik) = CDF(Ik)α (9)
shows the flowchart of the proposed work. Let an image input Iin
having intensity from I0, I1, …, IL − 1, where L is 256. Its histogram
Finally, HE transfer function is applied as
is calculated and represented as H(Ik). The PDF of kth intensity
level p(Ik) is given as Iout = I0 + (IL − 1 − I0)*CDFn(Ik) (10)

IET Image Process., 2020, Vol. 14 Iss. 6, pp. 1110-1119 1111


© The Institution of Engineering and Technology 2019

Authorized licensed use limited to: Auckland University of Technology. Downloaded on June 08,2020 at 03:17:28 UTC from IEEE Xplore. Restrictions apply.
where Iout is the final enhanced image. It is clearly seen from above 4.1 Visual assessment
that if R < 10 than standard HE procedure will be followed and
Fig. 3 shows the bird image. The original figure is the dark image.
then after (8) and (9) will be executed for enhancement. The above
HE method enhanced the image and lost the natural characteristics.
procedure is a complete heuristic approach that is adopted after
HMF scheme enhanced the image up to some extent, but, overall
testing on various images from different databases.
the image does not provide very pleasant look. AGCWD shows
Since this method is specifically applied on colour images,
very good contrast enhancement effect for this image with a clear
therefore, RGB image is first converted into an HSV colour space,
background and foreground. ESIHE, GCAHM and PLTHE
and then the V component (matrix) is converted into an intensity
produced similar images and showed very less enhancement effect.
image whose values range from 0 to 255. Then the proposed
The best visualisation is provided by the proposed method.
method is applied to this converted intensity image. In the end, the
Fig. 4 shows the gate image. HE has again over enhanced the
image is back converted into RGB image to get the final enhanced
image. HMF enhanced the image well. The image generated by the
image.
AGCWD method is enhanced, but it gives dull look. ESIHE
Algorithm 1 (see Fig. 2) shows the pseudocode for the proposed
generated some noise in the background portion. GCAHM
method.
produced similar image to AGCWD. PLTHE showed very less
enhancement effect. The proposed method enhanced the image
4 Experimental results and discussion very well with sharp and clear output.
The proposed method has been implemented and compared with Fig. 5 shows the result for a plane image. As it can be observed
other state-of-the-art papers like HE, HMF, AGCWD, ESIHE, that the image generated by HE is completely distorted. HMF has
GCAHM and PLTHE. Images are taken from the different datasets, enhanced the image, but it introduced distortion in the ground area.
i.e. the Berkeley image database [19], Kodak lossless true colour AGCWD also introduced too much distortion in the ground area.
image suite [20] and computer vision group (CVG-UGR) [21] ESIHE does not show much enhancement effect for this image.
databases.

Fig. 1  Flowchart of the proposed method

1112 IET Image Process., 2020, Vol. 14 Iss. 6, pp. 1110-1119


© The Institution of Engineering and Technology 2019

Authorized licensed use limited to: Auckland University of Technology. Downloaded on June 08,2020 at 03:17:28 UTC from IEEE Xplore. Restrictions apply.
Fig. 2  Algorithm 1: The proposed algorithm

Fig. 3  ‘Bird’ image and the contrast enhancement results of different algorithms

GCAHM and PLTHE enhanced the image well. However, again enhancement. The proposed method generated enhanced image
the best result and image are generated by the proposed method. with very sharp features.
Fig. 6 shows the girl's image. HE produced a noise and Fig. 9 shows the airplane image. The image generated by the
saturation artefact in the image. HMF has over brightened the HE method is completely distorted. HMF also introduced noise in
image. AGCWD produced distortion in the image. ESIHE, the image. AGCWD generated well enhanced image with clear sky
GCAHM and PLTHE produced good enhancement results for this as well as an airplane. ESIHE, GCAHM and PLTHE also created
image. The proposed method enhanced the image very well with artefacts in the image. Image processed by proposed method is
pleasant look. very smooth and clear with adequate amount of contrast
Fig. 7 shows the result of the house image. HE again introduced enhancement.
noise in the image. It can be observed that for this image HMF, Fig. 10 shows the result for the seen image. HE has over
AGCWD, ESIHE, GCAHM and PLTHE, all methods generated enhanced the image. HMF enhanced the image well. AGCWD
similar kinds of images, but if we clearly observe house bricks are showed an enhancement in some areas while other areas remained
clearer in ESIHE method. The proposed method generated image dark. ESIHE enhanced the image, but less as compared to HMF.
with enhanced background and foreground. The image generated by GCAHM is very dull. PLTHE showed
Fig. 8 shows the window image. HE has produced a saturation similar results to ESIHE. The proposed method has enhanced the
artefact in the image. HMF, AGCWD and ESIHE produced almost overall image very well with a smooth and clear look.
similar outputs for this image with adequate amount of Fig. 11 shows the result of the ruins image. HE distorted the
enhancement. GCAHM and PLTHE do not provide much image. HMF also introduced some noise in the background portion.
AGCWD showed very less enhancement effect. ESIHE showed

IET Image Process., 2020, Vol. 14 Iss. 6, pp. 1110-1119 1113


© The Institution of Engineering and Technology 2019

Authorized licensed use limited to: Auckland University of Technology. Downloaded on June 08,2020 at 03:17:28 UTC from IEEE Xplore. Restrictions apply.
Fig. 4  ‘Gate’ image and the contrast enhancement results of different algorithms

Fig. 5  ‘Plane’ image and the contrast enhancement results of different algorithms

Fig. 6  ‘Girl’ image and the contrast enhancement results of different algorithms

very good results for this image with clear sky and wall bricks. proved. Three metrics, such as entropy, QI and QRCM
GCAHM and PLTHE also produced enhanced images but less as measurements are used here to compare the performance of the
compared to ESIHE. The proposed method showed reasonably proposed method with others.
well-enhanced output with clear details.
4.2.1 Entropy: Entropy [14] is the measurement of information
4.2 Quantitative assessment content present in the image. A high value of entropy means richer
details provided by the image. Entropy is calculated with the help
Quantitative assessment is a necessary part of the image of PDF as
enhancement without which the validation of the method cannot be

1114 IET Image Process., 2020, Vol. 14 Iss. 6, pp. 1110-1119


© The Institution of Engineering and Technology 2019

Authorized licensed use limited to: Auckland University of Technology. Downloaded on June 08,2020 at 03:17:28 UTC from IEEE Xplore. Restrictions apply.
L−1 seen images. For house, room and ruins images, ESIHE gives the
Entropy = − ∑ p(Ik)log p(Ik)
2 (11) highest entropy values. It can also be seen that the average entropy
k=0 value is the highest for the proposed method.
where p(Ik) is the probability value of the kth intensity level.
4.2.2 Quality index (QI): The second parameter tested here is QI
Table 1 shows the results for various parameter measures. The [22] value. Image QI is evaluated for determining the amount of
highest values are bold and second-highest values are underlined. distortion present in an image. Basically, this parameter is
As we can observe from the table that proposed method provides calculated in reference to the original image. A higher value of QI
highest entropy for bird, gate, plane, girl, window, airplane and

Fig. 7  ‘House’ image and the contrast enhancement results of different algorithms

Fig. 8  ‘Window’ image and the contrast enhancement results of different algorithms

Fig. 9  ‘Airplane’ image and the contrast enhancement results of different algorithms

IET Image Process., 2020, Vol. 14 Iss. 6, pp. 1110-1119 1115


© The Institution of Engineering and Technology 2019

Authorized licensed use limited to: Auckland University of Technology. Downloaded on June 08,2020 at 03:17:28 UTC from IEEE Xplore. Restrictions apply.
Fig. 10  ‘Seen’ image and the contrast enhancement results of different algorithms

Fig. 11  ‘Ruins’ image and the contrast enhancement results of different algorithms

means less distortion in the image. Let p =  pi, where (i = 1, 2, …, 1
N) and q =  qi, where (i = 1, 2, …, N) be the original and enhanced
Q=1−
G ∑ ∑ GMS x, y − μ w2 x, y (14)
∀x ∀y
images, respectively. The QI is defined as
1
4σ pq p̄q̄ w2 x, y = (15)
1 + Go x, y
QI = 2 (12)
(σ p + σq2 )((p̄)2 + (q̄)2)
2Gin x, y Gout x, y + T
N N
GMS x, y = (16)
where p̄ = 1/N ∑i = 1 pi and q̄ = 1/N ∑i = 1 qi. Gin x, y 2 + Gout x, y 2 + T

N N where subscripts in and out represent the original and enhanced


1 1
N − 1 i∑ N − 1 i∑
σ 2p = (pi − p̄)2 and σq2 = (qi − q̄)2 images, T = 255/ 2.
=1 =1

N
RCM = ∑ ∑ Gout in x, y w
, 1 x, y (17)
σ pq = 1/ N − 1 ∑i = 1 (pi − p̄)(qi − q̄). As we can observe from ∀x ∀y
Table 1 that expects room, airplane and ruins images proposed
method provides the highest value of QI for other images. Also, the Gin x, y
w1 x, y = (18)
average value produced by the proposed method is the highest. ∑ ∑ Gin m, n
∀m ∀n

4.2.3 Quality-aware relative contrast measure: QRCM [23] Gout x, y − Gin x, y


utilises the image pixel gradients to measure the image quality. It is Gout, in x, y = (19)
further incorporated with two more weighting factors to generate Gout x, y + Gin x, y + ε
the final score
In a 3 × 3 patch, the derivatives are Δy and Δx in the vertical and
RCM × Q, RCM ≥ 0, horizontal directions, the gradient is given by G = Δx2 + Δy2, and
QRCM = (13) QRCM ∈ [−1,1]. When QRCM is close to unity, it represents better
(1 + RCM) × Q − 1, RCM < 0,
image quality. As can be observed from Table 1 that for gate, room,
where airplane and ruins proposed method does not provide the highest
value for QRCM, but for the rest of the images it gives best values

1116 IET Image Process., 2020, Vol. 14 Iss. 6, pp. 1110-1119


© The Institution of Engineering and Technology 2019

Authorized licensed use limited to: Auckland University of Technology. Downloaded on June 08,2020 at 03:17:28 UTC from IEEE Xplore. Restrictions apply.
Table 1 Various parameter measures
Metric IMG Input HE HMF AGCWD ESIHE GCAHM PLTHE Proposed
entropy bird 7.06 5.96 7.42 7.78 7.08 7.05 7.01 7.80
gate 7.14 5.98 7.53 7.78 7.12 7.60 7.11 7.83
plane 6.64 5.86 6.89 6.60 6.62 7.07 6.58 7.49
girl 5.97 5.48 6.39 5.40 6.35 6.14 6.16 6.46
house 7.06 5.85 7.04 7.17 7.41 7.18 7.05 7.33
room 5.96 5.69 5.94 7.07 7.61 7.26 5.98 6.97
window 7.30 5.97 7.24 7.64 7.61 7.29 7.29 7.77
airplane 6.23 5.87 6.20 6.65 6.70 6.59 6.22 6.90
seen 5.65 5.71 5.98 5.67 5.80 5.60 5.80 7.08
ruins 6.40 5.86 6.85 6.81 7.27 6.95 6.93 7.05
average 6.54 5.82 6.74 6.85 6.95 6.87 6.61 7.26
Q.I. bird — 0.54 0.65 0.75 0.68 0.65 0.67 0.89
gate — 0.61 0.70 0.64 0.60 0.69 0.59 0.85
plane — 0.40 0.50 0.47 0.59 0.75 0.67 0.79
girl — 0.20 0.46 0.28 0.52 0.65 0.60 0.75
house — 0.46 0.70 0.75 0.72 0.79 0.77 0.86
room — 0.13 0.39 0.21 0.50 0.92 0.50 0.90
window — 0.69 0.75 0.86 0.80 0.75 0.78 0.95
airplane — 0.35 0.59 0.85 0.56 0.62 0.64 0.71
seen — 0.13 0.32 0.28 0.51 0.67 0.52 0.70
ruins — 0.36 0.50 0.64 0.88 0.78 0.84 0.78
average — 0.38 0.55 0.57 0.63 0.72 0.65 0.81
QRCM bird — 0.05 0.05 0.17 0.07 0.12 0.15 0.17
gate — 0.12 0.13 0.29 0.14 0.23 0.15 0.25
plane — 0.10 0.11 0.14 0.11 0.13 0.21 0.27
girl — 0.05 0.03 0.06 0.08 0.02 0.08 0.11
house — 0.03 0.03 0.08 0.13 0.02 0.10 0.06
room — 0.15 0.19 0.18 0.22 0.27 0.14 0.26
window — 0.14 0.13 0.11 0.14 0.08 0.08 0.16
airplane — 0.11 0.16 0.21 0.11 0.18 0.18 0.19
seen — 0.39 0.21 0.37 0.42 0.26 0.40 0.43
ruins — 0.13 0.25 0.17 0.32 0.18 0.21 0.27
average — 0.12 0.12 0.17 0.17 0.15 0.17 0.21

compared to other methods. However, again the average value is images it is less compared to one of the other methods. Again, the
the highest among all. average value is the highest for both PLTHE and proposed method.

4.2.4 Structural similarity index measure: The SSIM is a well- 4.2.5 Colour quality enhancement: We have also calculated a
known quality metric used to measure the similarity between two parameter to check the colour quality [25] of images used in this
images [24]. The SSIM is designed by modelling any image paper. Basically, CQE is a no-reference quality assessment metric
distortion as a combination of three factors that are loss of and is calculated by combining colorfulness, sharpness and contrast
correlation, luminance distortion and contrast distortion. The SSIM metrics linearly. It is given as
is defined as
σγ2 σλ2
SSIM f , g = l f , g c f , g s f , g CQE = 0.02 × log 0.2 × log (20)
μγ μλ 0.2
where
where γ  = R−G and λ = 0.5(R + G)−B, colour components of a
2μ f μg + C1 colour image.
l( f , g) = σγ and σλ are standard deviation of γ and λ, respectively.
μ2f + μg2 + C1
Similarly, μγ and μλ are their means. From Table 2, it can be
2σ f σg + C2 observed that except girl, house and ruins images, the proposed
c( f , g) = method provides the highest CQE value for other images and also,
σ 2f + σg2 + C2 the average value is the highest among all.
Fig. 12 shows the average result of 30 images taken from the
σ f g + C3 CVG-UGR database. The blue line shows the result for entropy. As
s( f , g) =
σ f σg + C3 it can be clearly observed that average entropy is highest for the
proposed method.
μ f and μg are the mean luminance of two images. σ f and σg are the The violet line shows the result for CQE. Again the proposed
standard deviations. The positive constants C1, C2 and C3 are used method provides the highest average value for CQE.
to avoid a null denominator. Table 2 shows the value of SSIM. As The green line gives the result for SSIM. The proposed method
can be observed that for bird, gate and plane images the value of gives the highest average SSIM while PLTHE and GCAHM are on
SSIM obtained by proposed method is the highest and for rest the second number.
The red line provides the result for QRCM. As can be observed
that there is very minute difference in the result of QRCM for

IET Image Process., 2020, Vol. 14 Iss. 6, pp. 1110-1119 1117


© The Institution of Engineering and Technology 2019

Authorized licensed use limited to: Auckland University of Technology. Downloaded on June 08,2020 at 03:17:28 UTC from IEEE Xplore. Restrictions apply.
Table 2 Various parameter measures
Metric IMG HE HMF AGCWD ESIHE GCAHM PLTHE Proposed
SSIM bird 0.21 0.65 0.68 0.72 0.75 0.93 0.94
gate 0.53 0.80 0.87 0.56 0.85 0.93 0.93
plane 0.36 0.68 0.52 0.70 0.95 0.84 0.96
girl 0.29 0.46 0.30 0.93 0.90 0.94 0.92
house 0.63 0.78 0.82 0.95 0.90 0.95 0.90
room 0.25 0.43 0.50 0.59 0.69 0.90 0.86
window 0.39 0.91 0.85 0.88 0.89 0.98 0.93
airplane 0.39 0.55 0.95 0.82 0.92 0.92 0.88
seen 0.14 0.36 0.60 0.27 0.75 0.62 0.69
ruins 0.41 0.72 0.56 0.83 0.79 0.93 0.81
average 0.36 0.63 0.66 0.72 0.83 0.89 0.89
CQE bird 0.18 0.20 0.45 0.22 0.42 0.52 0.62
gate 0.12 0.28 0.15 0.29 0.30 0.17 0.30
plane 0.18 0.35 0.22 0.41 0.50 0.57 0.65
girl 0.00 0.09 0.02 0.14 0.12 0.17 0.14
house 0.35 0.45 0.40 0.52 0.40 0.51 0.43
room 0.30 0.35 0.59 0.49 0.62 0.51 0.77
window 0.45 0.60 0.61 0.59 0.64 0.67 0.73
airplane 0.00 0.01 0.04 0.03 0.02 0.02 0.05
seen 0.40 0.58 0.49 0.69 0.65 0.69 0.75
ruins 0.11 0.18 0.22 0.27 0.25 0.23 0.26
average 0.20 0.30 0.31 0.36 0.39 0.40 0.47

Fig. 12  Average result of 30 images for different methods

Table 3 Average computation time (in sec)


Image size HE HMF AGCWD ESIHE GCAHM PLTHE Proposed
256 × 256 × 3 0.04 1.45 2.47 0.22 1.09 0.90 0.18
321 × 481 × 3 0.05 2.98 5.81 0.23 2.42 2.03 0.19
365 × 490 × 3 0.07 3.21 6.33 0.25 3.11 2.42 0.23
512 × 768 × 3 0.09 9.87 14.12 0.34 7.66 5.44 0.32
600 × 893 × 3 0.10 10.98 15.37 0.39 9.52 6.59 0.37

different methods, but still the proposed method provides the processors, 4 GB RAM and Matlab 2015a. For comparison, we
highest average value. Similar are the results for QI, which is given have taken the images of different sizes and all are colour images.
by the dark blue line. As can be observed the HE method requires minimum time to
process all the images as compared to other methods; it is because
4.2.6 Computation complexity: Table 3 compares the HE method does not apply any modification scheme and is a
computation time complexity of the different methods. The simple equalisation process. After HE, proposed method requires
experiments were conducted on a computer with Intel i5, 2.5 GHz minimum time for processing the images. ESIHE method is also

1118 IET Image Process., 2020, Vol. 14 Iss. 6, pp. 1110-1119


© The Institution of Engineering and Technology 2019

Authorized licensed use limited to: Auckland University of Technology. Downloaded on June 08,2020 at 03:17:28 UTC from IEEE Xplore. Restrictions apply.
very close to proposed method. Rest all methods computation time [8] Ooi, C.H., Isa, N.A.M.: ‘Quadrants dynamic histogram equalization for
contrast enhancement’, IEEE Trans. Consum. Electron., 2010, 56, (4), pp.
is very high as compared to proposed method. Hence, the proposed 2552–2559
method can be very well utilised in real-time applications due to its [9] Huang, S.C., Cheng, F.C., Chiu, Y.S.: ‘Efficient contrast enhancement using
fast speed and good results, which are not obtained by the standard adaptive gamma correction with weighting distribution’, IEEE Trans. Image
HE method. Process., 2013, 22, (3), pp. 1032–1041
[10] Tiwari, M., Gupta, B., Shrivastava, M.: ‘High-speed quantile-based histogram
equalisation for brightness preservation and contrast enhancement’, IET
5 Conclusion Image Process., 2014, 9, (1), pp. 80–89
[11] Singh, K., Kapoor, R.: ‘Image enhancement using exposure based sub image
In this paper, a simple heuristic adaptive HE algorithm is proposed histogram equalization’, Pattern Recognit. Lett., 2014, 36, pp. 10–14
for image contrast enhancement. Basically, PDF of the image is [12] Ling, Z., Liang, Y., Wang, Y., et al.: ‘Adaptive extended piecewise histogram
equalisation for dark image enhancement’, IET Image Process., 2015, 9, (11),
calculated. Then, an adaptive parameter is calculated based on the pp. 1012–1019
mean and maximum values of that PDF. Thereafter, PDF and CDF [13] Wei, Z., Lidong, H., Jun, W., et al.: ‘Entropy maximisation histogram
are modified by applying a threshold limit to that adaptive modification scheme for image enhancement’, IET Image Process., 2014, 9,
parameter. Finally, another adaptive parameter is finding out by (3), pp. 226–235
[14] Parihar, A.S., Verma, O.P.: ‘Contrast enhancement using entropy-based
using modified CDF and a new CDF is obtained by using this dynamic sub-histogram equalisation’, IET Image Process., 2016, 10, (11), pp.
second adaptive parameter. Traditional HE is then applied with the 799–808
new CDF to getting the enhanced image. The proposed method, [15] Wang, X., Chen, L.: ‘An effective histogram modification scheme for image
when compared to other state-of-the-art papers visually resulted in contrast enhancement’, Signal Process.:Image Commun., 2017, 58, pp. 187–
198
superior quality of images. The proposed method enhanced both [16] Paul, A., Bhattacharya, P., Maity, S.P., et al.: ‘Plateau limit-based tri-
the low as well as bright contrast images very well. Quantitative histogram equalisation for image enhancement’, IET Image Process., 2018,
assessment with metrics like entropy, QI, QRCM, SSIM and CQE 12, (9), pp. 1617–1625
also proved the supremacy of the proposed method. The [17] Kumar, G.S., Reddy, P.V.G.D.P., Gupta, S., et al.: ‘Heuristic based face
recognition using image processing techniques’, Int. J. Comput. App., 2011,
computation complexity of the proposed algorithm is also much 34, (6)
less as compared to other methods, which make it suitable to be [18] Florez, L.R., Ruiz-Zea, C.A., Bolaños, F., et al.: ‘A heuristic algorithm for
used in real-time applications. extracting useful information from an image’. III Int. Congress of Engineering
Mechatronics and Automation (CIIMA), Cartagena, Colombia, 2014
[19] Arbelaez, P., Maire, M., Fowlkes, C.: ‘Contour detection and hierarchical
6 References image segmentation’, IEEE Trans. Pattern Anal. Mach. Intell., 2011, 33, (5),
pp. 898–916
[1] Gonzalez, R.C., Woods, R.E.: ‘Digital image processing’ (Prentice-Hall, NJ, [20] Kodak Lossless True Color Image Suite. Available at http://r0k.us/graphics/
USA, 2007) kodak/. Accessed 2 September 2018
[2] Kim, Y.T.: ‘Contrast enhancement using brightness preserving bi-histogram [21] USC-SIPI Image Database. Available at http://sipi.usc.edu/database/.
equalization’, IEEE Trans. Consum. Electron., 1997, 43, (3), pp. 1–8 Accessed 2 September 2018
[3] Wang, Y., Chen, Q., Zhang, B.: ‘Image enhancement based on equal area [22] Wang, Z., Bovik, A.C.: ‘A universal image quality index’, IEEE Signal
dualistic sub-image histogram equalization method’, IEEE Trans. Consum. Process. Lett., 2002, 9, (3), pp. 81–84
Electron., 1999, 45, (1), pp. 68–75 [23] Celik, T.: ‘Spatial mutual information and pagerank-based contrast
[4] Chen, S.D., Ramli, A.R.: ‘Contrast enhancement using recursive mean- enhancement and quality-aware relative contrast measure’, IEEE Trans.
separate histogram equalization for scalable brightness preservation’, IEEE Image Process., 2016, 25, (10), pp. 4719–4728
Trans. Consum. Electron., 2003, 49, (4), pp. 1301–1309 [24] Bhandari, A.K., Maurya, S., Meena, A.K.: ‘MFO-based thresholded and
[5] Sim, K.S., Tso, C.P., Tan, Y.Y.: ‘Recursive sub-image histogram equalization weighted histogram scheme for brightness preserving image enhancement’,
applied to gray scale images’, Pattern Recognit. Lett., 2007, 28, (10), pp. IET Image Process., 2019, 13, (6), pp. 896–909
1209–1221 [25] Panetta, K., Gao, C., Agaian, S.: ‘No reference color image contrast and
[6] Arici, T., Dikbas, S., Altunbasak, Y.: ‘A histogram modification framework quality measures’, IEEE Trans. Consum. Electron., 2013, 59, (3), pp. 643–651
and its application for image contrast enhancement’, IEEE Trans. Image
Process., 2009, 18, (9), pp. 1921–1935
[7] Chang, Y.C., Chang, C.M.: ‘A simple histogram modification scheme for
contrast enhancement’, IEEE Trans. Consum. Electron., 2010, 56, (2), pp.
737–742

IET Image Process., 2020, Vol. 14 Iss. 6, pp. 1110-1119 1119


© The Institution of Engineering and Technology 2019

Authorized licensed use limited to: Auckland University of Technology. Downloaded on June 08,2020 at 03:17:28 UTC from IEEE Xplore. Restrictions apply.

You might also like