Pattern Recognition Letters 54 (2015) 103108

Contents lists available at ScienceDirect

Pattern Recogn

journal homepage: www.els

I ive

e

L

F 50 Me

a

A

R

A

K

I

C

A

ment

ional

mean

nhan

thew

inten

infrar

to so

nica

1

t

u

d

C

h

a

c

[

b

t

R

d

w

n

e

R

e

o

a

h

0nfrared image enhancement using adapt

nhancement

o Tzer Yuan, Sim Kok Swee, Tso Chih Pingaculty of Engineering and Technology, Multimedia University, Jalan Ayer Keroh Lama, 754

r t i c l e i n f o

rticle history:

eceived 18 February 2014

vailable online 14 November 2014

eywords:

nfrared image

ontrast enhancement

daptive trilateral

a b s t r a c t

A novel contrast enhance

presented. Unlike convent

ground of an image as a

approach which involves e

trilateral is derived from

teristics, namely, contrast,

the experiment on a set of

is evaluated and compared

results show that ATCE sig

entropy.. Introduction

Conventional histogram equalisation (CHE) is considered as one of

he principal image processing algorithms that has been traditionally

sed to improve the image contrast quality through rescaling the

ynamic rangeandhistogramdistributionof agreyscale image [5]. For

HE, a uniform transfer function is often adopted to redistribute the

istogram counts, forcing the histogram distribution to be atten out

nd occupied a wider dynamic range as a means of exaggerating the

ontrast differences between the fore- and background of an image

2,14].

Kim suggested a bi-histogram equalisation model, called the

rightness preserving bi-histogram equalisation (BBHE), to reduce

he aforementioned shortcomings of HE [7]. In 2003, Chen and

amli [4] suggested a recursive approach for BBHE, where the

ecomposition of histogram distribution is performed recursively

ithin each sub-histogram based on their respective mean bright-

ess level. It is called the recursive mean separate histogram

qualisation (RMSHE) method. Although the recursive approach of

MSHE provides a exible means of monitoring the degree of over-

qualisation defect, yet this approach has been overly emphasising

n keeping the mean brightness value. For infrared thermal im-

ges with low image contrast, excessively emphasising on preserving

This paper has been recommended for acceptance by C. Luengo. Corresponding author. Tel.: +60 6 2523480; fax: +60 6 231 6552.E-mail address: kssim@mmu.edu.my, sksbg2003@yahoo.com (S. Kok Swee).

t

e

d

p

p

n

t

i

l

a

w

T

t

v

i

w

t

c

o

s

b

t

w

p

l

t

ttp://dx.doi.org/10.1016/j.patrec.2014.09.011

167-8655/ 2014 Published by Elsevier B.V.ition Letters

evier.com/locate/patrec

trilateral contrast

laka, Malaysia

method called adaptive trilateral contrast enhancement (ATCE) method is

methods which exaggerate the contrast difference between fore- and back-

s to improve image visual quality, ATCE method provides a multifaceted

cement of both image contrast and subtle image details. The designation of

orking principle of ATCEwhich utilises three different types of image charac-

sity and sharpness, to concurrently enhance the visual quality of an image. In

ed thermal images captured under low-light night time environment, ATCE

me existing contrast enhancement methods. The quantitative experimental

ntly surpasses other existing methods on the measure of enhancement by 2014 Published by Elsevier B.V.

he input mean brightness through higher recursive level will in-

vitably lead to inadequate enchantment results. To overcome this

ilemma, a hybrid of bi-histogram and plateau histogrammodelswas

roposed [11].

Vickers [16] suggested a plateau histogram model, known as

lateau histogram equalisation (PHE), as an attempt to reduce the

oise amplication and over-equalisation artefacts during the his-

ogram equalisation process. As the plateau threshold value for PHE

s empirically chosen, the practical usage of PHE method is rather

imited due to the vast differences in nature of greyscale images

nd the requirement of human intervention which may incur un-

anted cognitive bias during the contrast enhancement process.

o resolve this shortcoming, Liang et al. [6] proposed the adap-

ive double plateau histogram equalisation (ADPHE) to derive the

alue of plateau thresholds based on the characteristics of input

mages.

Adaptive histogram equalisation (AHE) is a localised model of HE,

here the image pixels are contrast-enhanced by local transfer func-

ions that derived from dened sub-image plots. Despite the stronger

ontrast enhancement effect offered by AHE, the practical usefulness

f AHE is rather limited as this form is slow and computational inten-

ive; the number of times that the local equalisation process should

e repeated is equivalent to the total amount of pixels available in

he image. Moreover, it may be inappropriate to transform pixels

ithin low-contrast region with a full histogram equalisation map-

ing as the local histogram distribution for highly homogenous (or

ow contrast) region is often severely skewed, thus very susceptible

o over-equalisation artefact [15]. To overcome such deciencies, an

104 L. Tzer Yuan et al. / Pattern Recognition Letters 54 (2015) 103108

F

d

t

p

L

m

j

g

steeper curve denotes a higher degree of inequality, and vice versa.

Meanwhile, a hypothetical straight diagonal line, known as the line

of equality, is often drawn along with the Lorenz curve to represent

a uniform distribution where each subject has the same proportion

as rest of the population. Fig. 1(c) illustrates an example of a typical

Lorenz curve.

With reference to the plotted Lorenz curve in Fig. 1(c), themeasure

of inequality can be computed by calculating ratio between the area

of concavity (area of difference between Lorenz curve and line of

equality) and the area-under-graph (total area under line of equal-

ity). Themeasure of inequality, or more commonly known as the Gini

coecient, can be written as

g =

K1k=0

{cdfe(Ik) cdfc(Ik)}K1k=0

cdfe(Ik)

; k = 0,1,2, . . . ,K 1, (1)

where g represents the Gini coecient; cdfe(Ik) and cdfc(Ik) denotethe CDF for line of equality and Lorenz curve at ordered grey inten-

sity level, Ik, respectively. K is the number of grey intensity levels,

typically set at 256 for greyscale image. Based on Eq. (1), the plateau

threshold values used to re-model the local histogram distribution

can be derived as

Tu = (1 g)hmax + gTl; 0 g 1, (2)

Tl = n/K, (3)where Tu and Tl denote the upper and lower plateau threshold val-improved version of AHE that featured with contrast limitation and

fast computational time, known as the contrast-limited adaptive his-

togram equalisation (CLAHE), was proposed [13].

Several types of conventional contrast enhancement methods

with different nature of working principles have been discussed. In

this paper, the constraints of aforementioned contrast enhancement

methods when applied on low contrast infrared thermal images is

addressed. Therefore, a novel approach is developed as a solution for

these drawbacks.

2. Research and methodology

The proposed adaptive trilateral contrast enhancement (ATCE) is

a novel variation of image contrast enhancement techniques, which

extends the model of conventional histogram equalisation by com-

bining it with the concept of image ltering to provide a multifaceted

contrast enhancement algorithm. The trilateral refers to the three

different types of image characteristics, namely, the image contrast

(IC), image sharpness (IS) and image intensity (II), to concurrently

enhance the visual quality of an image. Such multifaceted enhance-

ment can be achieved by rst manipulating the contrast of image to

improve the overall image visual quality, followed by ne-tuning the

intensity and sharpness of image to further preserve andmagnify the

subtle image details.

2.1. Image contrast (IC) manipulation

For a greyscale image, the image contrast can be dened as the

perceivable grey intensity differences that discriminate objects from

the background of an image [9]. In ATCE, the concept of plateau his-

togram equalisation (PHE) is adopted as the underlying principle for

image contrast manipulation due to its simplicity and exibility, as

elaborated earlier. However, unlike conventional PHE, the image con-

trast manipulation in ATCE utilises two plateau threshold values; an

upper plateau threshold value is deployed to restrain the severity of

over-equalisation and noise amplication artefacts, while an addi-

tional lower plateau threshold value is used to prevent subtle image

details fromoverwhelming by drastic contrast gain. The proposed im-

age contrast manipulation consists of three main stages: image seg-

mentation, local histogram equalisation and image reconstruction.

During the image segmentation process, the input image is di-

vided into a dened number of non-overlapping and equally-sized

sub-images. By regionally isolated the input image, each sub-image

is allowed to be enhanced independently based on their own respec-

tive local histogram distribution, thus capable of achieving stronger

contrast with least inuence from non-adjacent regions with vastly

different nature of grey intensities distribution. Furthermore, such

procedure allows the subtle image details to be well-preserved as

they are less likely to be engulfed during the contrast enhancement

process; the magnitude of contrast gain is ranked proportionally

among the pixels within close proximity, rather than against entire

image.

For histogram equalisation, one of the main concerns is the ef-

fect of over-equalisation due to inequality of histogram distribution.

The inequality of histogram distribution is often referred as a phe-

nomenon where highly dense histogram counts are being populated

within a limited number of grey intensity levels, causing an unde-

sirable drastic brightness change during equalisation process as the

contrast gain values given to each grey intensity level are scaled pro-

portionally according to the density distribution of histogram counts.

However, rather than using an arbitrarily selected threshold to glob-

ally restrict the height of histogram distribution, the proposed image

contrast manipulation attempts to study the nature of such inequal-

ity to adaptively derive an optimal threshold value to remodel the

histogram distribution. A good general measure for analysing the in-equality of distribution is known as the Lorenz curve. uig. 1. Illustration of Lorenz curve. (a) Histogram distribution, (b) sorted histogram

istribution, and (c) the corresponding Lorenz curve.

The Lorenz curve is a statisticalmodel rst developed in economics

o represent the proportionality or inequality of a distribution. The

roportionality of distribution is illustrated by a curved line, or the

orenz curve, where points along the curved line represent the cu-

ulative proportion of distribution assumed by each individual sub-

ect against the cumulative distribution of entire population; the de-

ree of inequality is dened by the curvature of Lorenz curve, with aes, respectively; hmax represents the highest histogram counts; n

L. Tzer Yuan et al. / Pattern Recognition Letters 54 (2015) 103108 105

i

i

t

p

e

d

l

o

s

o

m

t

r

h

w

t

t

u

h

o

l

d

a

w

i

w

h

w

l

a

l

g

i

o

d

p

f

p

w

g

s

s

f

b

t

w

s

p

c

I

r

t

p

l

t

j

2

i

a

a

i

p

p

f

c

h

c

a

w

T

f

t

t

e

o

p

a

i

s

s

o

a

p

p

i

e

i

i

v

i

o

T

I

[

w

r

s

t

f

A

ndicates the number of available pixels within the targeted sub-

mage. The upper plateau threshold is formulated in such a way

hat the greater the measure of inequality, the lower the upper

lateau threshold value would be, thus results in smaller over-

qualisation effect due to lesser amount of regional contrast gained

uring the histogram equalisation, and vice versa. Meanwhile, the

ower plateau threshold serves as a buffer to avoid total elimination

f histogram counts by upper plateau threshold under the worst case

cenario.

Themodicationof the original local histogramdistribution, based

n the derived plateau threshold values, is carried out by limiting the

aximum number of histogram counts for each grey intensity level

o the upper plateau threshold. Such thresholding process can be

epresented as

i(Ik) ={h(Ik),h(Ik) < Tu

Tu,h(Ik) Tu; k = 0,1,2, . . . ,K 1, (4)

here h(Ik) and hi(Ik) denote the original and modied local his-

ogram functions at grey intensity level, Ik, respectively. Those ex-

ra counts, which truncated beyond the upper plateau threshold, are

niformly redistributed among the rest of grey intensity levels with

istogram counts lesser than lower plateau threshold. The purpose

f redistribution is to increase the proportion of minorities, which

argely constituted by the subtle image details, within the histogram

istribution as the given regional contrast is ranked proportionally

ccording to their ratio of distribution; a stronger regional contrast

ill help to prevent subtle details being engulfed during the equal-

sation process. The redistribution of truncated extra counts can be

ritten as

ii(Ik) =hi(Ik),hi(Ik) > Tl

hi(Ik)+cx

kTl,hi(Ik) Tl ; k = 0,1,2, . . . ,K 1, (5)

here hii(Ik) is the redistributed histogram function at grey intensity

evel, Ik; cx is the total amount of truncated extra counts; kTl is the

mount of grey intensity levels with histogram counts lesser than the

ower plateau threshold value.

To conduct the local histogram equalisation, let I be the input

reyscale imagewithKdiscrete grey intensity levels that decomposed

ntoa set of (i j) sub-images,where Ik denotes thegrey intensity levelf sub-image pixels and. The equalisation of modied local histogram

istribution, based on Eq. (5), is carried out by rst computing the

robability distribution function (PDF) and cumulative distribution

unction (CDF) for each local histogram distribution as following

dfs(Ik) =hii(Ik)

n, (6)

cdfs(Ik) =K1i=0

pdfs(Ik), (7)

here pdfs(Ik)and cdfs(Ik)denote the PDF and CDF of sth sub-image atrey intensity level, Ik, respectively; n is the total number of targeted

ub-image pixels. The subscripted symbol s represents the index of

ub-images, with s {1,2,3, . . . , i j}. Next, the local uniform trans-er function, tfs(Ik), which formulated based on the derived CDF cane written as

fs(Ik) = (IK1 I0)cdfs(Ik)+ I0, (8)here IK1 and I0 represent the maximum and minimum grey inten-ity level of the equalised sub-image, respectively. Finally, by map-

ing the original grey intensities to their equalised counterparts, the

ontrast enhanced sub-image, Isout , can be expressed as

sout(Ik) = tfs(Ik), (9)

The third and nal step of image contrast manipulation, the imageeconstruction, is used to eliminate the abrupt intensity change near the boundaries caused by the individual local histogramequalisations

erformed in previous step. The reconstruction algorithm is formu-

ated based on the working principle of bilinear interpolation, where

he weighted average of reference pixel values that derived from ad-

acent sub-images is used to create a set of interpolated pixel values to

smooth-out the blocking effect near the boundaries of sub-images.

.2. Image sharpness (IS) manipulation

In computer vision and image processing, sharpness (or acutance)

s a generic term used to dene the clarity of details within an im-

ge by describing the transition rate of image information at edges;

high acutance results in sharper edge transitions, thus producing

mage details with clearly dened borders, and vice versa. For pro-

osedATCE, the conceptof phase congruency is adoptedasunderlying

rinciple for image sharpness manipulation; phase congruency is a

requency-based feature detection algorithm which utilises the local

oherence characteristics of an image to identify the presence of its

igh frequency components (or image details). By using the phase

ongruency algorithm, the high frequency components of an image

re being extracted and converted into a relative edge-strength index,

hich serves as the foundation for high-boost ltering at later stage.

he proposed image sharpness manipulation consists of two steps:

eature extraction and high-boost ltering.

Image feature is dened as ameasurement function that quanties

he properties and signicant characteristics of an image [3]. Through

he concept of feature extraction, the image characteristics are being

xtracted and formulated into a subset of image domain, often in form

f isolated (corner/interest points) or continuous (curve/edge/ridge)

oints, to serve as the foundation for subsequent image processing

lgorithms. Phase congruency is a frequency-based image process-

ng algorithm which provides a dimensionless index to reect the

trength of high frequency components, such as edge and corner

trength, of an image. This approach attempts to nd the patterns

f phase order of Fourier transform by searching for points within

n image where all the sinusoids of frequency domain are in phase;

hysical evidences have strongly indicated that human eyes respond

ositively towards the image locations with highly ordered phase

nformation [10,12]. Furthermore, one key benet of phase congru-

ncy approach is that the perceived high frequency components are

nvariant towards the contrast of an image, thus capable of provid-

ng highly localised and reliable resulting image features under the

arying illumination conditions [8].

To derive the phase congruency index of an image, let I(x, y) be the

nput image, where Geno and Gono denote the even-symmetrical and

dd-symmetrical lters of log-Gabor at orientation, o, and scale, n.

he outcome of each quadrature pair of lters after convoluting with

(x, y) can be written as

eno(x, y), ono(x, y)] = [I(x, y) Geno, I(x, y) Gono], (10)

here eno(x, y) and ono(x, y) denote the even-symmetrical (complex-

eal) and odd-symmetrical parts (complex-imaginary), respectively;

ymbol represents the convolution operator. Based on Eq. (10),he amplitude, Ano(x, y), and phase, no(x, y), of this response can beurther derived as

no(x, y) =e2no(x, y)+ o2no(x, y), (11)

no = arctan(eno(x, y), ono(x, y)). (12)

With reference to above equations, the mathematical measure ofwo-dimensional (2D) phase congruency that developed by Kovesi

106 L. Tzer Yuan et al. / Pattern Recognition Letters 54 (2015) 103108

F

l

i

b

a

p

w

s

i

i

W

w

t

d

0

m

a

u

t

t

a

d

t

o

a

I

d

i

b

t

adequate reduction of undesirable halo effect commonly generated

by the overshoot and undershoot of high-boost ltering.Fig. 2. Illustration of phase congruency. (a) Original image and (b) map of extracted

edge-strength index.

can be written as

PC(x, y) =

o

n Wo(x, y)Ano(x, y)no(x, y) To

o

n Ano(x, y)+

, (13)

no(x, y) = cos(no(x, y) o(x, y)) |sin(no(x, y) o(x, y))|(14)

where indicates that the enclosed value is self-equivalent for pos-itive value and equivalent to zero otherwise; Wo(x, y) is a weighting

function for frequency spread; is a constant valueadopted topreventdivision by zero;no(x, y) represents amodied phase deviation; To is

the threshold for noise inuence, which estimated by using Rayleigh

distribution function. Fig. 2 shows the map of edge-strength index

after performing phase congruency.

During the high-boost ltering stage, rather than using a single

high-boost kernel to conduct sharpening for entire image, the pro-

posed high-boost ltering is performed by convoluting the input im-

age with a set of high-boost kernels which derived from the relative

edge-strength index previously generated during the feature extrac-

tion stage. By diversifying the coecients of high-boost kernel in

reference to the relative edge-strength index, it allows both ampli-

cation of true edge and suppression of noise to be performed simulta-

neously during the ltering process; higher coecients are assigned

to cells where the true edges are located to intensify the effect of

boosting, while rest of the cells are lled up by lower coecients to

suppress and prevent possible noise amplication. This results in a

more subdued enhancement of image details in comparison to con-

ventional high-boost kernel of equivalent amplication factor. The

generic high-boost kernel, Whb, for proposed image sharpness ma-

nipulation algorithm can be written as

Whb =

1 2 34 A 5

6 7 8

; A = c + 8

i=1|i|, (15)where i denotes the non-central coecients from edge-strengthindex, with i {1,2,3, . . . ,8} ; A represents the centre weighting co-ecient; c indicates the amplication factor. For common practice,

the value of c will always equivalent to one to ensure the sum of all

kernel cells equals to unity to avoid biased amplitude on the output

image.

2.3. Image intensity (II) manipulation

For infrared thermography, rather than representing the amount

of light being reected off, image intensity is associated with the

amount of radiation detected from the object surface and represented

in terms of temperature reading; black colour for low temperature

reading and white colour for high temperature reading, respectively.

For proposed ATCE, the underlying principle of image intensity ma-

nipulation is to perform ltering in range (or intensity) domain of an

2

t

i

i

i

v

s

r

a

[ig. 3. Illustration of high-boost ltering. (a) High-boost ltering with single (dashed-

ine) and combined kernel (dotted-line) and (b) an ideal edge.

mage, as opposed to spatial domainltering commonly implemented

y traditional lters.

Range ltering is a form of nonlinear image ltering which aver-

ges the pixel values based on weighting functions that decay with

hotometric dissimilarity. By performing the range ltering, pixels

ith similar intensity values are being conserved regardless of their

patial locality, thus helps to prevent undesired distortions near the

mage boundaries. Computationally, the complexity of range ltering

s similar to those spatial domain lters and can be written as

r(x, y) = e{I(x,y)Io}

22 , (16)

here Wr(x, y) represents a (m n) range lter, where x and y arehe horizontal and vertical distance from the origin, respectively; enotes the standard deviation or decay factor, which equivalent to

.1; Io is the intensity value of targeted pixel.

Based on Eq. (16), one will soon realise that range ltering is just

erely transforming the colour map of the input image by averaging

way the small differences between pixels with close intensity val-

es. The visual impact caused by range ltering is rather insignicant,

hus renders it almost useless for any practical application. However,

hings are very much different when range ltering is coupled with

nother spatial domain lter; by combining both range and spatial

omain ltering, it allows the enforcement of both spatial and pho-

ometric locality within a single kernel, thus results in an advanced

perator with higher exibility and precision. Fig. 3 demonstrates

n illustrated outcome of combined range and high-boost ltering.

n Fig. 3(a), the solid line denotes the transition at edge, while the

ashed-line and dotted-line represent the outcome after perform-

ng high-boost ltering with standard high-boost kernel and com-

ine kernel, respectively. It can be observed that the combined kernel

ransforms the edge of transition closer to an ideal edge, thus provides.4. Integration of adaptive trilateral contrast enhancement

The nal step of ATCE involves integration of three aforemen-

ioned image characteristics manipulation algorithms, namely the

mage contrast (IC), image sharpness (IS) and image intensity (II),

nto a single multifaceted contrast enhancement method, which is

llustrated in Fig. 4.

Let I be the input greyscale image with Ip denotes the intensity

alue of image I at position p, where p = (px, py) with subscriptedymbol x and y represent the horizontal and vertical position of p,

espectively. The generic equation for ATCE algorithm can be written

s following

Iout]p = 1Wp

qN

Whb(PCq) IS

Wr(|Ip Iq|) II

[Iic]qIC

, (17)

L. Tzer Yuan et al. / Pattern Recognition Letters 54 (2015) 103108 107

w

W

l

t

i

t

i

k

a

s

3

p

l

t

l

i

a

3

a

f

v

t

l

(

t

m

F

(

(

M

E

A

w

grey intensity level in a given sub-block I(u,v); c is a small constant

introduced to avoid division by zero; b1 b2 is the equally-sized sub-blocks; is an additional coecient introduced to control the optimalFig. 4. Overview of adaptive trilateral contrast enhancement procedures.

here [Iout]p denotes the intensity value of output image at positionp;

hb(PCq) represents the high-boost ltering kernel with normalisedocal coherence function as input parameter; Wr (|Ip Iq|) refers tohe range ltering kernel with decay factor, ; [Iic]q indicates themage which is enhanced by image contrast (IS) manipulation;Wp is

he normalisation function to avoid amplitude biased on the output

mage; subscripted symbol q is a designation used to represent the

ernel where convolution is taking place, with q {1,2,3, . . . ,9} for(3 3) kernel which centred at position p. Fig. 5 demonstrates theequential outcome for ATCE algorithm.

. Results and discussions

The scheme of evaluation involves the comparison of evaluated

erformance of ATCE against a collection of conventional global and

ocal contrast enhancement algorithms. An Internet protocol (IP)

ype forward-looking infrared (FLIR) imager, model FLIR A325, is se-ected as the imaging device for acquisition of night time thermal

mages. The dimensions of captured thermal images are typically set

s 320 240 pixels in 8-bit greyscale bitmap (BMP) format.

.1. Image quality assessment

A collection of 85 sets of night-time surveillance thermal image

re processedwith selected contrast enhancementmethods. The per-

ormance of ATCE is compared against four selected methods: con-

entional histogram equalisation (CHE), adaptive double plateau his-

ogramequalisation (ADPHE), bi-histogramequalisationwith plateau

imit (BHEPL) and contrast-limited adaptive histogram equalisation

CLAHE). The quantitative evaluation is being conducted by two of

he standard objective image quality metrics: measure of enhance-

ent by Weber denition (EMEE) and measure of enhancement by

p

t

r

e

d

e

b

o

f

a

m

b

l

Table 1

Parameter settings of contrast enhancement alg

Contrast enhancement methods

Conventional histogram equalisation (CHE)

Adaptive double plateau histogram equalisatio

Bi-histogram equalisation with plateau limit (

Contrast-limit adaptive histogram equalisatio

Adaptive trilateral contrast enhancement (ATCig. 5. Illustration of ATCE. (a) Original image, (b) after image contrast manipulation,

c) after image sharpness manipulation, and (d) after image intensity manipulation

nal output image).

ichelson denition (AME), given as [1]:

MEE = 1b1b2

b2u=1

b1v=1

(Imax(u, v)

Imin(u, v)+ c)

ln Imax(u, v)Imin(u, v)+ c

, (18)

ME = 1b1b2

b2u=1

b1v=1

(Imax(u, v) Imin(u, v)

Imax(u, v)+ Imin(u, v)+ c)

ln Imax(u, v) Imin(u, v)Imax(u, v)+ Imin(u, v)+ c

(19)

here Imax(u,v) and Imin(u,v) represent the maximum and minimumarameters. The purpose of operating on small image regions rather

han the entire image is to reduce the possibility of the evaluated

esult being swayed by the outliers. For both EMEE and AME, a higher

valuated value indicates better degree of contrast enhancement.

Table 1 shows the list of parameter settings used by each of the

iscussed contrast enhancement algorithm. For all discussed contrast

nhancement algorithms, default settings and parameters suggested

y their respective authors are implemented to maintain the delity

f each algorithm, whereas a designation of N/A is shown instead

or parameter-free algorithm.

In Fig. 6, one can easily notice the apparent over-equalisation

rtefact that commonly appeared in all the global contrast enhance-

ent methods, such as CHE, ADPHE, and BHEPL; the human gure is

eing washed-out as huge contrast gains are being assigned to

ow grey intensity levels in order to maintain the uniformity of

orithms.

Parameters

N/A

n (ADPHE) N/A

BHEPL) N/A

n (CLAHE) Contrast limit, = 0.01E) N/A

108 L. Tzer Yuan et al. / Pattern Recognition Letters 54 (2015) 103108

ods such as CHE, ADPHE and BHEPL generally have higher measures

of contrast enhancement compared to local enhancement methods

such as CLAHE. However, such higher rate of contrast enhancement

often comeswith greater risk of over-equalisation defect, as discussed

previously. Meanwhile for CLAHE, a trade-off between measure of

c

i

m

i

f

t

T

4

e

pFig. 6. Illustration of contrast enhancement algorithms. (a) Original image, (b) after

CHE, (c) after ADPHE, (d) after BHEPL, (e) after CLAHE, and (f) after ATCE.

Table 2

Assessments of contrast enhancement algorithms.

Original CHE ADPHE BHEPL CLAHE ATCE

EMEE 0.029 0.120 0.073 0.104 0.069 0.432

AME 0.181 0.110 0.131 0.113 0.127 0.037

histogram distribution. Such a drawback can be easily resolved

by introducing the local contrast enhancements, such as CLAHE

and ATCE, where the segregation of image into smaller contex-

tual regions can effectively reduce the density of histogram counts

and prevent undesirable inuences from non-adjacent pixels with

vastly different characteristics during the enhancement process.

However, the local equalisation process by itself is only capable

of preserving but falls short of enhancing the subtle image details.

Hence a multi-faceted algorithm such as ATCE truly outshines the

rest.

Table 2 shows the averaged EMEE and averaged AME results of all

85 sets of processed images for CHE, ADPHE, BHEPL, CLAHE and ATCE.

From Table 2, one may observe that the global enhancement meth-

e

s

a

l

v

e

a

R

[

[

[

[

[

[ontrast enhancement and image visual quality is made to min-

mise the occurrence of over-equalisation at the expense of lesser but

ore manageable enhancement rate. However, the proposed ATCE

s free of such constraints. The measure of contrast enhancement

or ATCE in terms of averaged EMEE and averaged AME are better

han the rest of the contrast enhancement methods, as shown in

able 2.

. Conclusion

In this study, several conventional and well-established contrast

nhancement methods widely adopted for various types of image

rocessing applications are reviewed. The proposed adaptive trilat-

ral contrast enhancement method is developed based on careful

tudies on the drawbacks of each reviewed conventional methods

nd the inherited characteristics of thermal images captured under

ow-light night time conditions. ATCE attempts to extend the con-

entional HE model by combining it with the concept of high-boost

ltering to simultaneously provide both image contrast and details

nhancement in a singlemulti-faceted algorithm, hencemaking ATCE

simple, exible and elegant algorithm.

eferences

[1] S.S. Agaian, B. Silver, K.A. Panetta, Transform coecient histogram-based image

enhancement algorithms using contrast entropy, IEEE Trans. Image Process. 16(3) (2007) 741758.

[2] H.C. Andrews, A.G. Tescher, R.P. Kruger, Image processing by digital computer,Inst. Electr. Electron. Eng. Spectr. 9 (7) (1972) 2032.

[3] K.R. Castleman, Digital Image Processing, second ed., Prentice Hall, New Jersey,

1996.[4] S.D. Chen, R. Ramli, Contrast enhancement using recursive mean-separate his-

togram equalization for scalable brightness preservation, Inst. Electr. Electron.Eng. Trans. Consum. Electron. 49 (4) (2003) 13011309.

[5] R.C. Gonzalez, R.E. Woods, Digital Image Processing, second ed., Prentice Hall,New Jersey, 2002.

[6] K. Liang, Y. Ma, Y. Xie, B. Zhou, R. Wang, A new adaptive contrast enhancement

algorithm for infrared images based on double plateaus histogram equalisation,Infrared Phys. Technol. 55 (2012) 309315.

[7] Y.T. Kim, Contrast enhancement using brightness preserving bi- histogram equal-ization, IEEE Trans. Consum. Electron. 43 (1) (1997) 18.

[8] P.D. Kovesi, Image features from phase congruency, Videre: J. Comp. Vis. Res. 1(1999) 127.

[9] M. Kumar, Digital Image Processing, Proceedings of the Satellite Remote Sensing

and GIS Applications in Agricultural Meteorology 2003, pp. 81102.[10] M.C. Morrone, D.C. Burr, Feature detection in human vision: a phase-dependent

energy model, Proc. R. Soc. Lond. B 235 (1988) 221245.11] C.H. Ooi, S.P. Kong, H. Ibrahim, Bi-histogram equalization with a plateau limit

for digital image enhancement, IEEE Trans. Consum. Electron. 55 (4) (2009)20722080.

12] A.V. Oppenheim, J.S. Lim, The importance of phase in signals, Proc. IEEE 69 (1981)529541.

13] S.M. Pizer, E.P. Amburn, J.D. Austin, R. Cromartie, A. Geselowitz, T. Greer,

B.T.H. Romeny, J.B. Zimmerman, K. Zuiderveld, Adaptive histogram equal-ization and its variations, Comp. Vis. Graphics Image Process 39 (1987)

355368.14] W.K. Pratt, Digital Image Processing, fourth ed., Wiley Interscience, California,

2007.15] A. Reza, Realization of the contrast limited adaptive histogram equalization

(CLAHE) for real-time image enhancement, J. Very-Large-Scale Integr. Sig. Pro-

cess. 38 (2004) 3544.16] V.E. Vickers, Plateau equalization algorithm for real-time display of high-

quality infrared imagery, Soc. Photo-Opt. Instrum. Eng. 35 (7) (1996)19211926.

Infrared image enhancement using adaptive trilateral contrast enhancement1 Introduction2 Research and methodology2.1 Image contrast (IC) manipulation2.2 Image sharpness (IS) manipulation2.3 Image intensity (II) manipulation2.4 Integration of adaptive trilateral contrast enhancement

3 Results and discussions3.1 Image quality assessment

4 Conclusion References