International Conference on Communication and Signal Processing, April 3-5, 2018, India

Retinal Fundus Image Enhancement using Piecewise

Gamma Corrected Dominant Orientation based
Histogram Equalization
P. Sujith Reddy, Himanshu Singh, Anil Kumar, L. K. Balyan and Heung-No Lee

Abstract—Retinal fundus images play an essential role in the As a standard image modality, fundus camera is usually
diagnosis of retina-related diseases and hence, their quality used to acquire retinal images, showing structure like optical
enhancement is essential for doctors to make a reliable clinical disc, retinal vessels and several others [4]. The changes
diagnosis. This paper presents an advanced retinal fundus image detected in these structures can be As a standard image
enhancement method by employing an efficiently modified and
modality, fundus camera is usually used to acquire retinal
biologically inspired levy-flight firefly optimizer in association
with a novel optimally weighted piecewise gamma corrected indications of a pathological condition associated with
energy redistributed dominant orientation based texture diseases such as glaucoma and diabetic retinopathy, which can
histogram equalization framework for imparting overall quality further be confirmed by performing detailed analysis of these
improvement of retinal fundus images. The key intelligence is to retinal mages [5]. The main intention of this enhancement
utilize a weighted summation of intensity as well as texture based method is to correct the contrast and highlight the retinal
enhancement along with an efficiently defined cost function. The vessels. Therefore the analysis of retinal images is an
cost function has been framed such that more and more intensity important and helpful diagnostic tool. In fact, the analysis of
span can be explored in a positive manner. Rigorous retinal images can also render beneficial to the classification
experimentation by employing the performance evaluation and
of the disease stages, on identifying the underlying problem.
comparison with recently proposed enhancement approaches so
that the explicit outperformance can be underlined. In various domains of engineering and technology, image
processing is in very high demand, both for human vision as
Index Terms—Retinal fundus imaging; piecewise gamma well as machine vision perspective. Some of the retinal images
correction; cuckoo search optimization; quality enhancement. are clinically unacceptable due to eye lesions and imperfect
imaging techniques which include exudates, hemorrhages,
I. INTRODUCTION opacity of refractive media and patient’s eye movement.

DIGITAL imagery and its various forms aggregate the core Irregular blurring, illumination, low contrast and imprecise
basis for digital information era. Day-by-day increasing focus truncate the aspect of retinal images, emerging in loss of
audacious happenings and corresponding revolutionary sensitivity and specificity for diagnostic scope. It may also
improvements in any sphere of science and technology cannot prejudice ophthalmologists’ capability to enact significant eye
be imagined without digital imaging techniques in one form or features or categorize retinal diseases. Retinal images which
the other. It grants much wider range of algorithms to be are of poor quality make it difficult for consequent authentic
applied to the input data and can avert many disputes [1]. segmentation and computer-aided diagnosis of retinal related
Despite of surprising advancements for image capturing diseases, which can be used to automate the disclosure process
devices, still there are various natural as well as artificial and to assist ophthalmologists. Thus, it is significant to
artifacts, which lead to poor quality of the image captured, and overcome the objections associated with poor quality retinal
hence, quality improvement for raw captured images is an images [6]. Wide variety of histogram based and transform-
indispensible part of pre-processing of the images [2]. Retinal domain based techniques have already been available in
fundus imaging provides rich information of pathological literature for general images [5-6]. First, general histogram
changes those are usually with sporadic illumination, low equalization (GHE) approach [7] was initially introduced,
contrast and blur of the details due to the complex imaging thereafter its various variants have been proposed by many
environments [3]. Retinal imaging is an important and researchers. In the same context, requirement of localized
effective tool for screening retinal diseases such as Diabetic processing seems more ambitious and hence various sub-
Retinopathy (DR), Glaucoma, hypertension, stroke, and age- equalization motivated histogram based enhancement
related Macular Degeneration (AMD) and Cardiovascular approaches have also been contemplated. A detailed literature
disease. analysis in this ambience is also available in [5-6]. Momentous
contributions like contrast-limited adaptive HE also dragged
the core attraction of the researchers. Statistical segmentation
P. Sujith Reddy, Himanshu Singh, Anil Kumar and L. K. Balyan are
based sub-equalization like median-mean based sub-image
with Indian Institute of Information Technology Design and Manufacturing,
Jabalpur-482005, India (E-mail: [email protected];{sujithreddy2502; clipped HE (MMSICHE) [8] has also been introduced.
anilkdee; lokendra.balyan}@gmail.com). Afterwards, the averaging histogram equalization (AVGHEQ)
Heung-No Lee is with Gwangju Institute of Science and Technology, [9], HE based optimal profile compression (HEOPC) [10]
Gwangju, South Korea. (E-mail: [email protected]).
method for color image enhancement followed by HE with
978-1-5386-3521-6/18/$31.00 ©2018 IEEE

0124
maximum intensity coverage (HEMIC) [11] were proposed.
^’I  x, y  , ’I  x, y  , ’I  x, y  ,..., ’I  x, y  `
T
GM , (5)
Also, the adaptive gamma correction with weighting 1 2 3 N

distribution (AGCWD) [12] and its productive variations [13- Next, SVD of the gradient map is computed.
17] were proposed for dark images. Eventually, the intensity < U 6V T (6)
nu n
and edge based adaptive unsharp masking filter (IEUMF) [18] Where U  R represents the contribution of each vector to the
based enhancement has been proposed by operating the corresponding singular vector; 6 Rnu2 represents the energy in
unsharp masking filter for edge enhancement followed by the the dominant directions; and V  R 2u 2 represents the orientation.
significant proposal of dominant orientation-based texture In matrix V , the dominant and the subdominant orientations of
histogram equalization (DOTHE) [19] especially of textural the gradient field are represented by the columns v1 and v2
improvement. In this paper, the piecewise gamma correction is respectively. The dominant measure D divides the rough
optimally associated with energy redistributed texture- patches into dominant and non-dominant orientation patches,
orientation dominance framework for transmitting intensity as which can be calculated by:
well as for the texture based quality enhancement approach
D
 6 1,1 - 6  2,2   ,
and is employed for optimal enhancement of dark images. (7)
2
Remaining content is organized as: Section II which deals
with the problem formulation followed by the proposed In the above equation, 6 1,1 and 6  2, 2  are singular values
methodology. Experimentation is discussed in Section III and representing the energy in the dominant direction. The patches
finally, conclusion is drawn in Section IV. having dominant measure lesser than the significance level
threshold D ' contains no dominant orientation as they are only
II. PROPOSED METHODOLOGY pure white noise. The patches which have dominant measure
Parallel band processing is generally required for multiband greater than the significance level threshold D ' are dominantly
images, but for enhancing equivalent color images Hue- oriented. Now, the intensity distribution (histogram) of the
Saturation-Intensity (HSI) model can be applied to decouple texture patches is computed. All the intensities present in the
the chromatic and non-chromatic information content, as [1]: texture patches are required for histogram formation. Further,
Cumulative Density Function (CDF) is used to map the input
> H  m,n  ,S  m,n  ,I  m,n @T TRGB > R  m,n  ,G  m,n  ,B  m,n @ , (1)
HSI T
image histogram into new dynamic range. Locations of various
HSI
Here, TRGB is RGB to HSI transformation process. The color local maxima values have to be observed and accordingly
admissible sub-histograms can be derived. Individual CDFs are
image enhancement can be done by upgrading only the evaluated for all the sub-histograms as:
brightness intensity values, keeping rest (hue and saturation) i

¦h  k ,
values preserved, followed by linear stretching. The gamma 1
c j i  j (8)
compressed interim intensity channel can be evaluated as [1]: Nj
k 0
I gcp  Iin J , J >1,
(2) Here, N j is the net pixel-count in j th sub-histogram
The corresponding gamma expanded interim intensity channel (corresponding to the j th patch). Equalize all the j sub-
can be evaluated as [1]:
histograms autonomously as:
I gex  Iin 1 J , J >1, (3)

I%j I j_ min  I j _ max  I j _ min * c j (i), (9) 
Later on, evaluation for third interim channel can be done by Access the overall equalized image as:
identifying the tile-wise texture dominance followed by
variance based thresholding (for separate identification of Itex _ dom = UI% ,
m
m (10)
smooth as well as non-smooth patches). For this purpose, the
entire image is sectioned into tile-wise collection of several Now, the weighted summation input intensity channel with
5 u 5 (or any odd-ordered) sized patches. These extracted uniformly equalized intensity channel Iˆen can be obtained as- 
patches are divided into smooth or rough by implanting a
§ D · § 1D · § E ·
variance threshold W on each image patch. Further, the rough Iˆen ¨ ¸ I gcp  ¨ ¸ I gex  ¨ ¸ I tex _ dom , (11)
© 1 E ¹ © 1 E ¹ © 1 E ¹
patches are classified into dominant or non-dominant
orientation patches by reckoning their local orientation which Here, unfortunately over-ranging may get resulted, and it
should be minimized efficiently without affecting the resulted
is based on singular value decomposition (SVD) of the
enhancement and hence it can be included as a penalty term in
gradient vectors of the patch. The procedure required for
the cost function framed here, as:
categorizing the rough patch into dominant or non-dominant
§V2 · § nov ·
orientation patches and the method to construct the histogram J H .'V .¨
2
¸ .¨ 1  ¸, (12)
is highlighted below. The local estimate of gradient ’I (m, n)i at © P ¹ © M*N¹

each pixel  m, n  in patch is calculated as: Here, P ,V 2 ,'V 2and H stands for output brightness, contrast,
relative contrast, and output Shannon entropy, respectively for
°­wI  x, y  wI  x, y  ½° an L-bit, M * N image. Here, nov is the count of the normalized
’I  x , y  i ® , ¾, (4)
¯° wx wy ¿° over-ranged pixels, which can be evaluated as:
Collectively for N patches, the gradient map can be framed as: nov ¦^% imn ! 1`,
imn  0 U % (13)

0125
Here, ĥ  i  stands for histogram of the processed image. Cost- Here, entry-wise walk during multiplications can be
function is devised here, so that the relative variance along determined through product operation †. Random exploration
with maximal information restoration can be imparted with follows Levy distributed (having both first as well as second
proper check on relative mean brightness. Biologically moment infinite) random step size, as [20]:
inspired and later on efficiently modified CSOA is employed Lévy : u t O ,  O  1,3@ , (15)
for optimal enhancement for dark images, by efficient This power law step-flight assigned random walk leads to the
exploration followed by generous exploitation in a three- introduction of few new solutions in the proximity of best
dimensional search space so that the required optimal values solution (identified so far), and in this manner local search can
for D , E , and J can be obtained. The efficient parametric be boosted up. In addition, a benevolent share of new solutions
variation for framing search space derived analytically is should be created through far-field randomization, so that the
>D , E ,J @ m > 0,1 ,  0, 5  , 1, 5 @ . Combative psychology of the local deceiving can be avoided and global inspection can be
renewed. Finally, enhanced channel is obtained and hence,
cuckoo bird, and its interesting breeding behavior (more
correspondingly enhanced color image can be derived as:
precisely, its brood parasitism) fascinated various researchers ) ) )
> R  m,n  ,G  m,n  ,B  m,n @T
T
to frame an analogously designed population oriented THSI ª H  m,n  ,S  m,n  ,$I  m,n  º , (16)
RGB
¬ ¼
metaheuristic optimization algorithm. CSOA is highly Here, THSI
RGB
is HSI to RGB transformation process.
perceptible for determining multimodal, multi-objective, and
highly non-linear optimization issues deprived of any kind of III. EXPERIMENTATION AND RESULT ANALYSIS
comprehensive search. Core structure for CSOA and its
problem solving approach in its original form has been already A. Assessment Criterion
detailed in [20]. Following the Levy distributed quasi-random Comparative evaluation is performed qualitatively for
flight; a suitable intelligence has been also introduced, where resultant images [21] and for further quantitative assessment,
the prevailing step has to be decided by keeping “current performance metrics such as brightness (B), contrast/variance
location” and “next-state transition probability” in the mind. (V), entropy (H), sharpness (S), and colorfulness (C) are
This type of step flight pattern is highly compatible and derived here.
beneficial with CSOA behaviour. Simplified analogous
demeanoral modelling has been done by establishing three B. Qualitative Assessments
rules, as already existed in the admissible literature. Levy Reimplementation for various recent state-of-the-art
distributed flight is generally for both local as well as global methodologies (namely, GHE, ADAPHE, AGCWD, and
analysis of the corresponding search space. Levy flight for HEOPC) has been done. Visual results for enhanced images
iterative new solution xt 1 for the i th cuckoo can be forged as: are presented in Fig. 1.
xit 1 xit  w † Lévy  E  , where, w ! 0, (14)

TABLE I

QUANTITATIVE EVALUATION WITH COMPARISON AMONG GHE [7], ADAPHE [3], AGCWD [12], HEOPC [10], AND THE PROPOSED APPROACH USING METRICS
TERMED AS BRIGHTNESS, CONTRAST, ENTROPY, SHARPNESS AND COLORFULNESS.
S. No. INDICES INPUT GHE ADAPHE AGCWD HEOPC OURS
Brightness 0.6368 0.5497 0.6545 0.7273 0.5764 0.6545
Contrast 0.068 0.1223 0.0694 0.0528 0.0889 0.0694
1.
Entropy 5.6271 5.1901 5.7954 5.7838 5.5716 5.912
Sharpness 0.0802 0.1535 0.1904 0.1186 0.102 0.1904
Colorfulness 0.2827 0.2551 0.3082 0.3746 0.2441 0.3899
Brightness 0.6921 0.5462 0.6859 0.764 0.6949 0.6859
Contrast 0.0598 0.1208 0.0709 0.0528 0.0597 0.0709
2.
Entropy 5.6508 4.9697 5.7605 5.7137 5.7504 5.8098
Sharpness 0.102 0.1217 0.199 0.1178 0.1086 0.199
Colorfulness 0.3722 0.275 0.3772 0.4473 96.3018 0.4446
Brightness 0.7191 0.549 0.7026 0.7882 0.6756 0.7026
Contrast 0.0467 0.122 0.0595 0.0398 0.0617 0.0595
3.
Entropy 5.6671 4.9553 5.7758 5.75 5.7236 5.8308
Sharpness 0.0747 0.1204 0.1626 0.0898 0.0909 0.1626
Colorfulness 0.3934 0.2815 0.3881 0.4688 91.8551 0.4609
Brightness 0.6562 0.5466 0.6707 0.7332 0.6637 0.6707
Contrast 0.0744 0.1207 0.0737 0.0658 0.0708 0.0737
4.
Entropy 4.5803 4.0234 4.6757 4.6092 4.6801 4.6893
Sharpness 0.095 0.1029 0.1849 0.1068 0.0972 0.1849
Colorfulness 0.3903 0.3079 0.4081 0.4824 103.0327 0.492

0126
 1a 2a 3a 4a

1b 2b 3b 4b

1c 2c 3c 4c

1d 2d 3d 4d

1e 2e 3e 4e
Fig. 1. Visual evaluation with comparison among 1a-4a: GHE [7]; 1b-4b: ADAPHE [3]; 1c-4c: HEOPC [10]; 1d-4d: AGCWD [12]; and 1e-4e: the proposed
approach.

0127
 [8] K. Singh and R. Kapoor, “Image enhancement via Median-Mean Based
C. Quanitative Assessments Sub-Image-Clipped Histogram Equalization,” Optik -Int. J. Light and
Electron Optics, vol. 125, no. 17, pp. 4646-4651, Sept. 2014.
For explicit quantitative comparison and evaluation, relevant [9] S. C. F. Lin, C. Y. Wong, M. A. Rahman, G. Jiang, S . Liu, N. Kwok, H.
performance metrics have been evaluated and listed in Table I. Shi, Y.H. Yu and T. Wu, “Image enhancement using the averaging
histogram equalization (AVHEQ) approach for contrast improvement
IV. CONCLUSION and brightness Preservation,” Comput. and Elect. Eng., vol. 46, pp. 356-
370, Aug. 2015.
As a concluding remark, it can be explicitly identified that the [10] C. Y. Wong, G. Jiang, M. A. Rahman, S. Liu, S. C. F. Lin, N. Kwok, H.
prospective path will be highly suitable for overall quality Shi, Y. H. Yu and T. Wu, “Histogram equalization and optimal profile
enhancement of retinal fundus images and hence directly helps compression based approach for colour image enhancement,” J. Visual
Commun. and Image Represen., vol. 38, pp. 802-813, July 2016.
the radiologists in various kinds of medical diagnosis as well
[11] C. Y. Wong, S. Liu, S. C. Liu, M. A. Rahman, S. C. F. Lin, G. Jiang, N.
as early detection of symptoms of retinal disorders. Especially Kwok and H. Shi, “Image contrast enhancement using histogram
when followed by segmentation and further by classification equalization with maximum intensity coverage,” J. Modern Optics, vol.
the quality enhanced fundus images outperforms for medical 63, no. 16, pp. 1618-1629, Mar. 2016.
diagnostic purposes. The dominant texture-orientation based [12] S. C. Huang, F. C. Cheng and Y. S. Chiu, "Efficient Contrast
equalization when associated with the recently proposed Enhancement Using Adaptive Gamma Correction with Weighting
Distribution," IEEE Trans. Image Process, vol. 22, no. 3, pp. 1032-1041,
piecewise gamma corrected weighted summation framework, Mar. 2013.
yields highly appreciable results when intuitively governed by [13] H. Singh, A. Kumar, L. K. Balyan, and G. K. Singh,“A novel optimally
highly efficient exploration as well as exploitation following gamma corrected intensity span maximization approach for dark image
the biologically inspired cuckoo search optimization (CSO). enhancement,”IEEE International Conference on Digital Signal
Although the approach is some-how iterative, but the Processing (DSP), London, United Kingdom, 2017, pp. 1-5.
doi:10.1109/ICDSP.2017.8096035.
associated robustness and it’s highly flexible behavior
[14] H. Singh, N. Agrawal, A. Kumar, G. K. Singh and H. N. Lee, "A novel
counter-balances for that. Highly consistent performance gamma correction approach using optimally clipped sub-equalization for
metrics are examined for proper image quality evaluation and dark image enhancement," IEEE International Conference on Digital
accordingly, the outperformance of the proposed framework Signal Processing (DSP), Beijing, China, 2016, pp. 497-501.
can be easily highlighted in addition to the qualitative [15] H. Singh and A. Kumar, "Satellite image enhancement using beta
wavelet based gamma corrected adaptive knee transformation," IEEE
evaluation through visual results. International Conference on Communication and Signal Processing
(ICCSP), Melmaruvathur, India, 2016, pp. 128-132. doi:
REFERENCES 10.1109/ICCSP.2016.7754552
[1] P. Dai, H. Sheng, J. Zhang, L. Li, J. Wu, and M. Fan, “Retinal Fundus [16] H. Singh, A. Kumar, L. K. Balyan and G. K. Singh, "Dark image
Image Enhancement Using the Normalized Convolution and Noise enhancement using optimally compressed and equalized profile based
Removing,” Int. J. Biomed. Imaging, vol. 2016, 2016. parallel gamma correction," 2017 International Conference on
[2] M. Zhou, K. Jin, S. Wang, J. Ye, and D. Qian, “Color Retinal Image Communication and Signal Processing (ICCSP), Chennai, India, 2017,
Enhancement Based on Luminosity and Contrast Adjustment,” IEEE pp. 1299-1303. doi: 10.1109/ICCSP.2017.8286592
Trans. Biomed. Eng., vol. 9294, no. c, pp. 1–1, 2017. [17] H. Singh, A. Kumar, L. K. Balyan and G. K. Singh, "Regionally
[3] K. Jin, M. Zhou, S. Wang, L. Lou, Y. Xu, J. Ye, and D. Qian, equalized and contextually clipped gamma correction approach for dark
“Computer-aided diagnosis based on enhancement of degraded fundus image enhancement," IEEE International Conference on Signal
photographs,” Acta Ophthalmol., no. 1994, pp. 1–7, 2017. Processing and Integrated Networks (SPIN), Noida, India, 2017, pp.
431-436.
[4] L. Xiong, H. Li, and L. Xu, “An enhancement method for color retinal
images based on image formation model,” Comput. Methods Programs [18] S.C.F. Lin, C.Y. Wong, G. Jiang, M.A. Rahman, T.R. Ren, N. Kwok, H.
Biomed., vol. 143, pp. 137–150, 2017. Shi, Y. H. Yu and T. Wu, “Intensity and edge based adaptive unsharp
masking filter for color image enhancement,” Optik–Int. J. for Light and
[5] H. Singh, A. Kumar, L. K. Balyan, and G. K. Singh, “A novel optimally
Electron Optics, vol. 127, no. 1, pp. 407-414, Jan. 2016.
weighted framework of piecewise gamma corrected fractional order
masking for satellite image enhancement,” Comput. Electr. Eng., pp. 1– [19] K. Singh, D. K. Vishwakarma, G. S. Walia, and R. Kapoor, “Contrast
17, 2017. doi 10.1016/j.compeleceng.2017.11.014. enhancement via texture region based histogram equalization,” J. Mod.
Opt., vol. 63, no. 15, pp. 1444–1450, 2016.
[6] H. Singh, A. Kumar, L. K. Balyan, and G. K. Singh, “Swarm
intelligence optimized piecewise gamma corrected histogram [20] X.-S. Yang and S. Deb, “Cuckoo search via Lévy-flights,” inProc.
equalization for dark image enhancement,” Comput. Electr. Eng., pp. 1– World Congr. Nature Biol. Inspired Comput., Dec. 2009, pp. 210–214.
14, 2017. doi:10.1016/j.compeleceng.2017.06.029. [21] J. J. Staal, M. D. Abramoff, M. Niemeijer, M. A. Viergever, B. van
[7] R. C. Gonzalez and R. E. Woods, Digital Image Processing, 4th ed. Ginneken, "Ridge based vessel segmentation in color images of the
Pearson/Prentice-Hall, NY, 2017. retina", IEEE Transactions on Medical Imaging, 2004, vol. 23, pp. 501-
509.

0128