-
Keywords — Cuckoo Search, Contrast Enhancement, Gray
Image, Metaheuristic, Particle Swarm Optimization,
I. INTRODUCTION
Nowadays, digital images have become a typical way of
acquiring, storing and communicating information among
people, corporations, businesses and security outfits [
1
].
Thus, it has become pertinent to ensure the integrity of digital
images, particularly those used for sensitive purposes in
pattern recognition, forensics, and a host of other applications.
In this regard, an important area of focus is gray image
enhancement [
2
]. Several works have tried to improve the
contrast of Gray Images (GI), however, these techniques
have been either fully manual, that is, humans are required
to identify areas for improvement, or partially automated,
where humans need to assess the enhancement performance
to make conclusions. For most automated techniques, it has
been observed that they often depend only on the global
information of the image, without consideration for local details
[
3
]. Furthermore, because of their dependence on the specific
image being processed, most automated techniques lack the
capacity for generalization. Despite these limitations, full
automation is evidently required for most new applications
in areas such as pattern recognition, forensics and robotics,
and thus the need for better techniques.
In this paper, we report on the investigation of two
metaheuristic algorithms for autonomous GI enhancement.
To achieve this, we adopted the transformation and evaluation
functions in [
3
] and applied them for GI enhancement.
First, we investigated the Particle Swarm Optimization (PSO)
technique based on certain parameter settings. Secondly, we
explored the Cuckoo Search (CS) algorithm for the same
task. These algorithms were chosen owing to their respective
high performance output, as noted in the literature [
4
]. Each
algorithm was modified and details of the modifications are
presented in appropriate sections. The results of the different
algorithms were analyzed using a set of newly proposed
metrics and findings are presented herein to justify the
effectiveness of the metaheuristic algorithms. An algorithm
for computing these metrics is also presented and readers are
provided with output images to enable them cross evaluate
between the proposed metrics and the reader’s perception of
the enhanced images.
The rest of the paper is organized as follows: Section II
provides a brief review of the relevant literature. In Section
III, we present details of the methodology used, while results
and analysis are provided in Section IV. Conclusion is drawn
in Section V.
II. REVIEW OF RELEVANT LITERATURE
There are several reported works on Gray Image (GI)
enhancement. These methods can be broadly divided into
point operations, spatial operations, transform operations, and
pseudocolouring methods [
3
]. Techniques under point
operation (also termed indirect method) include contrast stretching,
window slicing and histogram modeling [
3
]. These are the
simplest and most popular methods for GI enhancement,
thus, they are widely deployed in the literature. However,
they have more global effect than local effect; thus, they
suffer from over stretching of the image contrast. Indirect the image, I, can be resized to a smaller dimension to
methods typically adjust the image histogram to improve improve processing speed, however, in this work, we used
the entropy. On the other hand, spatial operations (or di- the original dimension given as (R C). The image, I, is
rect methods) establish criterions of contrast measurement converted to gray scale, G, with same dimension, (R C).
and enhance the image by improving the measure [
5
]. In By using a local window size, LW = 3, we computed the
[
5
], Fuzzy logic was used as an adaptive direct enhance- local mean, using the LW LW window size, the global
ment method based on fuzzy entropy principle and fuzzy mean of the entire image, G, and the local standard deviation,
set theory. Authors claimed that the proposed technique . These parameters served as the basic requirements for
performed better than the Adaptive Contrast Enhancement running the GI process.
technique; however, evaluation was done subjectively. In
2004, Munteanu and Rosa [
3
] proposed a transformation B. The Image Transformation and Evaluation Process
function for contrast enhancement and used Evolutionary To begin, it is worth noting that an image transformation
Algorithm (EA) as a global search strategy for the best function typically changes the intensity value of a gray image
enhancement. Authors became one of the first to use heuristic pixel, Gi;j for i = 1; 2; :::; R and j = 1; 2; :::; C, to a different
algorithm for GI enhancement, and they used both subjec- value, Fi;j for i = 1; 2; :::; R and j = 1; 2; :::; C. Thus, the
tive and objective methods for evaluation, and showed the transformation function we used in this work is given as [
3
]
superiority of their method over Linear Stretching (LS) and
Histogram Equalization (HE). In 2005, Russo [
6
] proposed Fi;j = T (Gi;j ) 8i 2 R; j 2 C (1)
tahneoebdjgecetigvreadeiveanltusa.tiTohnotuegchhnsihqouwenbatoseoduotpnetrhfoerhmisbtoogthramlinseoafr = i;jG+b Gi;j c i;j + i;j a
and nonlinear unsharp masking technique, Rosso’s technique where a; b; c, and are the parameters of the enhancement
produces overshoots along the object contours. Kwok et kernel to be optimized, having the following typical values:
faol.r, c[o7n]trianst2e0n0h6anpcroempoesnetdusainnginPteanrtsiictlye-pSrwesaerrmvinOgpttiemcihznaitqioune 00:5 c< 1 [3<].T1h:e5;tra0nsformaed or e2n;hanGced<imbag<eis 0o:b5t;aiannedd
(PSO). The PSO technique was used to obtain proper gamma- by computing Fi;j for all i; j using (1). Subsequently, the
factor values for the enhancement process. The use of mean- number of edges, Ne in Fi;j is computed using a Sobel
intensity as the objective measure of evaluation is insufficient detector. The Sobel detector produces an edge image, Ei;j
to make conclusions, as such, broad measures are required. In containing ones at pixels describing the image’s edge pixels
[
8
], authors used PSO to maximize the information content of in Fi;j , and zeroes at other non-edge pixels. Because Ei;j
an enhanced image using Munteanu’s functions in [
3
]. Gorai contains only binary representation of pixels corresponding
et al., [
8
] showed that PSO performed better than GA, LS to edges in Fi;j , the total number of edges in Ei;j can be
and HE. Particularly, Ghosh et al., in [
4
] explored the use computed as:
of Cuckoo Search (CS) algorithm for image enhancement. It R C
was concluded that CS provides better performance compared Ne = X X Ei;j (2)
to PSO, Genetic Algorithm (GA), Linear Contrast Stretching i=1 j=1
(LCS) and Histogram Equalization. Similarly, other works in
[
9
][
10
][
11
][
12
][
13
] have made efforts to enhance GIs, and The intensity, i;j of each pixel is obtained as
the trending conclusion is that metaheuristic algorithms tend
to provide better image enhancement based on the use of i;j = Ei;j Fi;j ; 8i 2 R; j 2 C (3)
transformation functions. However, these methods often lack where ( ) denotes element wise multiplication. Thus, the
standard objective measures for measuring their effectiveness. total intensity of Fi;j is given as
Thus, in addition to investigating both PSO and CS in our
work, we proposed new objective measures for quantifying R C
an algorithm’s performance. = X X i;j (4)
III. METHODOLOGY
In this section, we provide details on the two metaheuristic
based algorithms used here for GI enhancement, namely
Particle Swarm Optimization (PSO) and Cuckoo Search (CS)
algorithms. In addition, we provide an algorithm for
computing a new set of metrics for evaluating any GI enhancement
algorithm. Details of these algorithms are provided in their
respective subsections.
A. Input Parameters
Let the image to be enhanced be denoted as I, with
dimensions (R C). Now, based on the user’s requirement,
where H(Fi;j ) is the entropy of the enhanced image,
Fi;j . Thus, having established the requisite functions for
transformation in (1) and for the evaluation in (5), the two
metaheuristic algorithms, PSO and CS were then used to
obtain the optimum values of a; b; c and , in accordance
with their constraints, using (5) as the fitness function.
C. Use of Particle Swarm Optimization (PSO) Algorithm
We present here the use of Particle Swarm Optimization
(PSO) algorithm [
14
] as modified for the enhancement
process. First, we state the necessary functions used in PSO as
vtn+1
n
xt+1
=
=
wnvtn + c1nr1(gbest
xtn + vtn+1
xtn) + c2nr2(pbest(n)
n
xt )
(6)
Thus, for a number of iteration, Niter , the rest of the
process ensues as follows:
2) Set the lower and upper bounds for the parameter
constraints based on dimension, d,
3) Obtain the random initial solutions (or nests),
4) For each iteration, until Niter, do
5) Get a cuckoo randomly by Levy Flights
6) Evaluate each solution (or nest) using (5), after
transformation using (1),
7) Obtain the global best value among all nest as Maxfit
8) If Maxfit(t+1) > Maxfit(t)
9) Update the new global best
10) End if
11) Empty a fraction, pa, of the worst nests,
12) Update the new nests using (7)
13) Keep the best solutions
14) Return to 1, until Niter is completed.
D. The use of Cuckoo Search (CS) Algorithm
The Cuckoo Search (CS) Algorithm using Levy flight
[
15
] is applied here for image enhancement and the kernel
function used for finding new solutions is given as:
xt=1 = xtn +
n
Levy( )
(7)
where > 0 is the step size related to the scale of the
problem of interest, in most cases, = 1 is normally used,
and 1 < < 3 is the Levy distribution parameter, while xtn
and xtn=1 are the current and next solutions with dimension,
d. In our work, the CS algorithm was used as follows:
1) Let the number of nests (or different solutions, similar
to particles in PSO) be n, and the dimension of each
particle be d (where d = 4). Let the probability of
discovering an alien egg (or solution) in a nest be pa.
1) For i = 1 to R, do
2) For j = 1 to C, do
3) If G(i; j) TG
4) Increment DO set by 1
5) Else
6) Increment BO set by 1
7) End
8) If F (i; j) TG
9) Increment DE set by 1
10) Else
11) Increment BE set by 1
12) End
13) End
14) End
At the end of Line 14, the algorithm computes the overall
Detailed and Background Variance of both the original and
enhanced image by adding all the counts in DO and BO,
and DE and BE , respectively. In addition, by using a Sobel
detector, the number of edges denoted as NO and NE
respectively for the original and enhanced image are also
considered for evaluating the algorithm’s performance.
BO
BE
DO
DE
NO
NE
For evaluation purpose, four different images (see Figs. 1 –
4) were used for running the PSO and CS algorithm alongside
classical techniques such as the Linear Contrast Stretching
(LCS) and Histogram Equalization (HE) techniques. The
images used for evaluation have various properties relevant
for evaluating these algorithms, such as a variety of both
small and large number of pixels (see Table 1a), different
shades, darkness, and representing different applications, e.g
Fingerprint image for finger print analysis. The parameters
used for the metaheuristic algorithms are provided in Table
1b. Before proceeding, it should be noted that the term
Background (BV) and Detailed Variance (DV) are only
similar terminology-wise to the metrics used in [
3
], but
different in their technical interpretation. Here, BV describes
the number of pixels that belong to the background image
(or noisy component) of the image, while DV describes
the number of pixels that belong to the foreground image
(or true signal component). Both metrics form an effective
measure for evaluating any GI enhancement algorithm. We
provide the enhanced images (see Figs. 1 – 4) outputted
by each algorithm so that readers can make their subjective
evaluation and then proceed to corroborate their judgment
using the corresponding output metrics in Tables 2 – 5. Thus,
by visually analyzing each algorithm’s output (see Figs. 1 –
4) and comparing them with their corresponding objective
measures (see Tables 2 – 5), it can be clearly seen that the
objective measures closely reflect the true outcome of the
enhancement process. Consequently, it can be seen that a
well enhanced image should have lower BV and higher DV
along with more number of edges than its original version.
While noting that these metaheuristic algorithms are highly
statistical in nature, and often converge to solutions close
to the optimal, it might be difficult to conclude which is
better on single runs of the algorithm. However, on the
average, the CS provides better performance. Summarily, this
work adds to the body of evidence supporting the claim
that metaheuristic algorithms possess the potential to replace
subjective and manual methods based on human judgement
in GI enhancement. Future works will provide a thorough,
indepth and objective evaluation of different metatheuristic
algorithms for the GI enhancement problem.