2018 Conference on Information and Communication Technology (CICT'18)

Homomorphically Rectified Tile-wise Equalized

Adaptive Gamma Correction for Histopathological
Color Image Enhancement
Rangu Silpasai, Himanshu Singh, Anil Kumar and L. K. Balyan
Indian Institute of Information Technology Design and Manufacturing, Jabalpur-482005, India
[email protected]; [email protected]; [email protected]; [email protected]

Abstract—In this paper, an efficient approach for quality contextual, textural as well as spectral feature improvement.
improvement of histopathology slides’ images is suggested, along Many methods came into existence for image enhancement.
with proper luminance restoration and color-correction. Here, A plethora of literature is already available in [5, 6, 17-19].
the extracted luminance channel is initially tile-wise equalized In the beginning, general histogram equalization (GHE) [7]
and an efficiently clipped image is acquired for evaluation of the
was introduced. This is one of the methods for enhancement of
required gamma value set for imparting adaptive boosting for the
proper harvesting of the entire intensity span. The core contrast of the image by adjusting the image intensities to
contribution is to harvest the highly adaptive, pixel-wise, make the distribution uniform. Later, various other methods
homomorphic color-correction ratio over the entire three- based on histogram modification and sub-histogram
dimensional color image. In addition, this correction is modification came into existence. Some of them are median-
introduced by intensity level-wise gamma correction. Again, it mean based sub-image clipped HE (MMSICHE) [8] (where,
can be understood that the corresponding gamma value set is sub-equalization and clipping is suggested), averaging
derived from a tile-wise uniformly equalized channel. Hence, this histogram equalization (AVGHEQ) [9] (where, successive
multi-level adaptiveness leads to a new strategy to attain the histogram averaging is suggested), contrast limited adaptive
required objective. Rigorous experimentation is executed by
histogram equalization (CLAHE) [10], adaptive gamma
employing the performance evaluation through standard quality
measures and comparison with pre-existing recently proposed correction with weighting distribution (AGCWD) [11] along
and highly appreciated quality enhancement approaches. with its efficient variants [12-14] are proposed. Later, in the
same sequence, HE based optimal profile compression
Keywords—homomorphic ratio, color correction; gamma (HEOPC) [15] method (which prevents the insufficient
correction; tile-wise equalization; optimal clipping; histopathology enrichment of colorfulness and saturation through a pipeline
images; image quality enhancement. approach), histogram equalization with maximum intensity
coverage (HEMIC) [16] is also proposed along with some
I. INTRODUCTION intelligence. In the same sequence, various other algorithms
Now-a-days, all kinds of biomedical investigations are have been also presented for quality restoration [17-25]. Here,
usually done by analyzing the digitized one-dimensional or these states of the art methods are employed over the test
multidimensional signals. In the same sequence, histology and images, so that the comparative study of these methods with
histopathology are also relying on the digital images acquired the proposed approach can be presented. In this paper, for
from the various kinds of microscopic instruments, at various better restoration and illumination based enhancement,
levels of magnification. Cellular level diagnosis for any initially the patch-wise histogram equalization for the
disease is the prime need of time and hence, by establishing luminance channel is imparted along with suitable histogram
the computer-aided diagnosis through histopathology images, clipping. Thus process, interim channel is itself utilized for
medical systems can become more efficient. This also deciding the adaptive gamma value set for imparting
indicates the importance of the quality and naturalness of the enhancement. Finally, the color coordination and
image. This is most common in images of cells and tissues, corresponding information are improved through pixel-wise
especially in cancer cases as the images of these cells and homomorphic color ratio adaptively.
tissues are microscopic and need large magnification [1-4]. The rest of the content is organized as: In Section II, the
Poor quality images cause a severe difficulty in diagnosis. It is problem formulation and in section III, the proposed
methodology is explained. Experimentation is discussed in
necessary to overcome the challenges associated with poor
Section IV and finally, Section V deals with the conclusion.
acquired images. Conventionally proposed enhancement
techniques do not always give robust and reliable images. Due II. PROPOSED METHODOLOGY
to uneven illumination, low contrast, incorrect focus and Following the RGB color space model, the digitally acquired
blurring the quality of the image is reduced. Image quality histopathology images can be considered as a M × N × 3 sized
enhancement should be identified as enhancement of matrix as:
I M × N ×3 = [ RM × N : GM × N : BM × N ] (1)

978-1-5386-8215-9/18/$31.00 ©2018 IEEE

 γ ( i ) = 1 − CDF ( i ) (7)
By using this gamma value set, the enhanced luminance
channel can be derived as:
γ (i )
Lˆ ( i ) =  L ( i )  (8)
For imparting the homomorphic-ratio correction and thus, for
evaluating the enhanced versions of all the three channels
( )
Rˆ , Gˆ , Bˆ , the information content of the enhanced luminance

channel ( L̂ ) is harvested efficiently. Assuming

the M × N summation matrix SM of the enhanced color
channels and establishing its relation with the enhanced
luminance channel L̂ along() with the ratios
α1 ,α 2 ,α 3 ,Ω R ,ΩG , and Ω B as,
Lˆ ( m, n )
SM ( m, n ) = (9)
α1.Ω R ( m, n ) + α 2 .ΩG ( m, n ) + α 3 .Ω B ( m, n )
This summation matrix ( SM ) when multiplied element-wise
with corresponding color ratio matrices, they yield the
enhanced color channels, as follows:
R̂M × N = Ω R * SM (10)
Fig. 1. Tile-wise symmetric distribution pattern [20].
ĜM × N = ΩG * SM (11)
From this color image, the corresponding luminance channel B̂M × N = Ω B * SM (12)
(L) extracted can be generalized as: By concatenating all three color channels, the enhanced color
LM × N = α1 .RM × N + α 2 .GM × N + α 3 .BM × N (2) image of the size M × N × 3 can be obtained as:
Here, α1 + α 2 + α 3 = 1; and values for α1 ,α 2 ,α 3 lies in the range Iˆ M × N ×3 =  R̂M × N : Gˆ M × N : Bˆ M × N  (13)
(0,1), in general. Following the Rec. ITU-R BT.601-7, and
assigning the values α1 = 0.299 ,α 2 = 0.587 ,α 3 = 0.114 ,
III. EXPERIMENTATION AND RESULT ANALYSIS
accordingly, the eq (3) can be written as;
LM × N = 0.299 * RM × N + 0.587 * GM × N + 0.114 * BM × N (3) A. Assessment Criterion
For accounting the share of each color/channel, color ratio Experimentation and visual comparison with state-of-the-art
matrices (which are used later for homomorphic ratio methods are done along with quantitative evaluations of
correction) are also derived as: several histopathology images [21] using a set of standard
ΩR = R ( R + G + B ) (4) performance metrics which can be expressed as follows.
ΩG = G ( R + G + B ) (5) Brightness (average intensity value) is an attribute of visual
perception in which a source appears to be reflecting or
ΩB = B ( R + G + B ) (6) radiating light. It gives the information about average image
This luminance intensity matrix is sectioned to various brightness level (B) for the considered image, which can be
uniformly sized, non-overlapping tiles (matrices). Regions in numerically expressed as;
the resulted sectioned image can be categorized as [20]: corner 1
Brightness ( B ) =  
M N
I(m,n) (14)
regions ( CRs ) , edge regions ( ERs ) , and inner regions ( IRs ) . In M*N m =1 n =1

other words, the whole luminance matrix can be understood as The term “global image contrast” or average contrast basically
a group of, (i) four CRs; (ii) 2 × ( M + N − 4 ) ERs and the gives the information about the differential spread between the
remaining (iii) ( M − 2 ) × ( N − 2 ) IRs. as presented in Fig. 1. bright and dark areas present in an image. It can be understood
as the variance of the image and indicates the amount of
Individual patch-wise histograms are processed to achieve the
intensity deviation per pixel with respect to mean intensity
uniformly distributed patches’ histogram and contrast-
enhanced equivalent patches, in a parallel manner for all the level (B) of the image. If the gray level values differ from their
patches. Later on, these all enhanced patched are merged to mean then the variance will increase.
obtain the resulting image. Bilinear interpolation is followed The total sum of the intensity dispersions with respect to the
for the avoidance of any kind of artifacts of sudden boundary mean level is known as contrast, and it will be higher for good
changes. Normalized histogram or probability distribution quality enhancement. It can be numerically represented as:
function (PDF) is derived from the processed intensity channel. 1  1 
2

Later, cumulative distribution function (CDF) is extracted from Global Contrast (V ) =  I(u,v)2 −  M * N 
M * N u ,v
I(u,v)  (15)
it, and when subtracted from unity leads to the gamma value  u ,v 
set, as:
 (1a) (2a) (3a) (4a) (5a)

(1b) (2b) (3b) (4b) (5b)

(1c) (2c) (3c) (4c) (5c)

(1d) (2d) (3d) (4d) (5d)

(1e) (2e) (3e) (4e) (5e)

(1f) (2f) (3f) (4f) (5f)

(1g) (2g) (3g) (4g) (5g)

(1h) (2h) (3h) (4h) (5h)

(1i) (2i) (3i) (4i) (5i)

Fig. 2. Visual evaluation with comparison among 1a-5a: input images [26]; 1b-5b: GHE [7]; 1c-5c: MMSICHE [8]; 1d-5d: AVGHEQ [9]; 1e-5e: CLAHE [10]; 1f-
5f: AGCWD [11]; 1g-5g: HEMIC [15]; 1h-5h: HEOPC [16]; and 1i-5i: the proposed approach.
TABLE I. QUANTITATIVE EVALUATION WITH COMPARISON AMONG INPUT IMAGES [26], GHE [7], MMSICHE [8], AVGHEQ [9], CLAHE [10],
AGCWD [11], HEMIC [15], HEOPC [16], AND THE PROPOSED APPROACH USING METRICS LIKE BRIGHTNESS, GLOBAL CONTRAST, AVERAGE
LOCAL CONTRAST, SPATIAL FREQUENCY AND MEAN GRADIENT VALUE.

S.
INDICES INPUT GHE MMSICHE AVGHEQ CLAHE AGCWD HEMIC HEOPC Proposed
No.
Brightness 0.501777 0.499039 0.536901 0.733893 0.54408 0.548094 0.7223 0.76792 0.805885
1. Global Contrast 0.012301 0.08579 0.051292 0.057941 0.068888 0.012622 0.045923 0.034309 0.075573
Avg. Local Contrast 0.004829 0.040162 0.023073 0.019163 0.030414 0.004698 0.013306 0.008906 0.032466
Spatial Frequency 0.064347 0.190537 0.161359 0.144088 0.17163 0.06388 0.130341 0.111511 0.203442
Mean Gradient 0.044115 0.125908 0.096919 0.095147 0.116302 0.039803 0.08821 0.075237 0.110576
Brightness 0.477106 0.500017 0.511016 0.713403 0.534423 0.531225 0.697054 0.732998 0.815489
2. Global Contrast 0.015734 0.085815 0.047405 0.056839 0.071318 0.01697 0.051267 0.042144 0.071116
Avg. Local Contrast 0.006255 0.033916 0.018672 0.01686 0.028526 0.006604 0.014045 0.0104 0.027408
Spatial Frequency 0.065979 0.167099 0.135454 0.12737 0.158624 0.069144 0.121402 0.110115 0.183597
Mean Gradient 0.04423 0.108362 0.082028 0.083809 0.105777 0.042182 0.081227 0.073263 0.097598
Brightness 0.446324 0.499816 0.459925 0.669895 0.51549 0.520898 0.656466 0.686194 0.782641
3. Global Contrast 0.007993 0.085785 0.031112 0.05466 0.052014 0.011883 0.035424 0.025369 0.07153
Avg. Local Contrast 0.004329 0.047849 0.019793 0.02343 0.034064 0.005972 0.012964 0.008848 0.041037
Spatial Frequency 0.060027 0.20276 0.132808 0.15685 0.176377 0.074357 0.126947 0.108138 0.234893
Mean Gradient 0.041992 0.136425 0.083243 0.107472 0.122936 0.049226 0.088535 0.075246 0.140528
Brightness 0.471119 0.499749 0.504869 0.685707 0.529683 0.526774 0.707501 0.725485 0.852929
4. Global Contrast 0.015063 0.085803 0.04879 0.066599 0.07116 0.016141 0.048004 0.044688 0.061459
Avg. Local Contrast 0.005123 0.032261 0.019041 0.017237 0.027507 0.005114 0.011157 0.010059 0.026373
Spatial Frequency 0.064115 0.167491 0.144851 0.133271 0.15856 0.065103 0.114674 0.11119 0.192575
Mean Gradient 0.043278 0.108009 0.085275 0.087039 0.104538 0.040345 0.076755 0.074103 0.087889
Brightness 0.485709 0.500873 0.520699 0.734564 0.540294 0.53706 0.738326 0.74753 0.870837
5. Global Contrast 0.016854 0.086725 0.047296 0.06161 0.060161 0.018193 0.048753 0.04859 0.050492
Avg. Local Contrast 0.00556 0.03181 0.017009 0.014479 0.021692 0.005547 0.0108 0.010496 0.022247
Spatial Frequency 0.061838 0.172913 0.139142 0.118899 0.140357 0.062267 0.106954 0.107502 0.177046
Mean Gradient 0.039321 0.10311 0.076881 0.072773 0.090277 0.03602 0.068024 0.067829 0.075038

It is just a sincere possibility that global variance of the image 1 N M


2
may not be sufficient to analyze it completely and hence, local FC =  I ( m, n ) − I ( m − 1, n ) 
 (20)
M × N n =1 m = 2
or patch-wise contrast for a fixed mean value of that patch can
be evaluated locally and later their mean value can be Accountability for the presence of the edge content can be
identified as Average Local Contrast (ALC). Thus, a window easily identified through the sharpness content of the image
that can be also identified as the gradient of the image, which
of size 3 × 3 can be moved along the image and hence, the
can be evaluated as
pixel-wise equivalent value cij can be identified as follows;
σ 2
Mean Gradient ( MG ) =
1

M × N m,n
( Δm 2 + Δn 2 ) (21)
cw = w
(16)
μw where, Δm = I ( m,n ) − I ( m + 1,n ) and Δn = I ( m,n ) − I ( m,n + 1)
enh enh enh enh

Hence, averaging the local patch-wise contrast over the image symbolizes for the accountability of the local values of gradient
for all pixels leads to the ALC as: content of the image. It is well obvious that the higher value of
1 M N the gradient signifies the more sharpness content of the image.
Average Local Contrast ( ALC ) = 
M × N m =1 n =1
cw ( m, n ) (17)
B. Quantitative Assessments
Here, μ w and σ w2 stands for patch-wise local mean and patch-
wise local contrast, respectively. To analyze the overall For explicit quantitative comparison and evaluation, relevant
activity level of the image, in the terms of spatial repetition, image performance metrics have been evaluated and listed in
spatial frequency measure can be evaluated as: Table I.
Spatial Frequency ( SF ) = (F R
2
+ Fc 2 ) (18)
C. Qualitative Assessments
Here, row frequency ( FR ) and column frequency ( FC ) can be
For explicit analysis, reimplementation for various recent
evaluated as: state-of-the-art methodologies (namely, GHE, MMSICHE,
1 M N
AVGHEQ, CLAHE, AGCWD, HEMIC and HEOPC) is

2
FR =  I ( m, n ) − I ( m, n − 1)  (19)
M × N m =1 n = 2  followed by comparative evaluation. Visual results for all
enhanced images are shown in Fig. 1.
 (a)

(b) (c)

(d) (e)
Fig. 3. Comparative evaluation for performance indices: (a) brightness (B); (b) global contrast (GC); (c) average local contrast (ALC); (d) spatial frequency (SF);
and (e) mean gradient (MG) value by employing various state-of the-art methods over various images.

REFERENCES
IV. CONCLUSION [1] A. Bentaieb and G. Hamarneh, "Adversarial Stain Transfer for
Histopathology Image Analysis," IEEE Transactions on Medical
Finally, the concluding note can be framed by highlighting the Imaging, vol. 37, no. 3, pp. 792-802, 2018. doi:
outperformance of the proposed approach. In addition, it can 10.1109/TMI.2017.2781228
also be noticed that any specific color model conversion is not [2] M. Veta, J. P. W. Pluim, P. J. van Diest and M. A. Viergever, "Breast
employed here. Also, the non-iterative and adaptive behavior Cancer Histopathology Image Analysis: A Review," in IEEE
of the proposed approach fascinates a lot and similar kind of Transactions on Biomedical Engineering, vol. 61, no. 5, pp. 1400-1411,
May 2014. doi: 10.1109/TBME.2014.2303852
strategy can be further explored for other image domains and
[3] Y. Al-Kofahi, W. Lassoued, W. Lee and B. Roysam, "Improved
also for avoiding the display artifacts. A highly relevant set Automatic Detection and Segmentation of Cell Nuclei in Histopathology
of performance metrics is evaluated here for sketching the Images," IEEE Transactions on Biomedical Engineering, vol. 57, no. 4,
quantitative as well as qualitative evaluation and comparison pp. 841-852, 2010. doi: 10.1109/TBME.2009.2035102
for color images from histopathology domain those are [4] C. Mosquera-Lopez, S. Agaian, A. Velez-Hoyos and I. Thompson,
"Computer-Aided Prostate Cancer Diagnosis From Digitized
initially acquired in very poor condition at various Histopathology: A Review on Texture-Based Systems," IEEE Reviews
magnification levels. The same approach can be easily in Biomedical Engineering, vol. 8, pp. 98-113, 2015. doi:
appreciated for several other kinds of biomedical color 10.1109/RBME.2014.2340401
images.
[5] H. Singh, A. Kumar, L. K. Balyan, and G. K. Singh, “A novel optimally [22] H. Singh, A. Kumar, L. K. Balyan and G. K. Singh, "Dark image
weighted framework of piecewise gamma corrected fractional order enhancement using optimally compressed and equalized profile based
masking for satellite image enhancement,” Comput. Electr. Eng., pp. 1– parallel gamma correction," IEEE International Conference on
17, 2017. doi:10.1016/j.compeleceng.2017.11.014. Communication and Signal Processing (ICCSP), CHENNAI, India,
[6] H. Singh, A. Kumar, L. K. Balyan, and G. K. Singh, “Swarm 2017, pp. 1299-1303. doi: 10.1109/ICCSP.2017.8286592
intelligence optimized piecewise gamma corrected histogram [23] H. Singh, A. Kumar and L. K. Balyan,"A Sine-Cosine Optimizer-Based
equalization for dark image enhancement,” Comput. Electr. Eng., pp. 1– Gamma Corrected Adaptive Fractional Differential Masking for Satellite
14, 2017. doi:10.1016/j.compeleceng.2017.06.029. Image Enhancement," in Harmony Search and Nature Inspired
[7] R. C. Gonzalez and R. E. Woods, Digital Image Processing, 4th ed. Optimization Algorithms. Advances in Intelligent Systems and
Pearson/Prentice-Hall, NY, 2017. Computing, vol 741, pp. 633-645 Springer, Singapore, 2019. doi:
10.1007/978-981-13-0761-4_61
[8] K. Singh and R. Kapoor, “Image enhancement via Median-Mean Based
Sub-Image-Clipped Histogram Equalization,” Optik -Int. J. Light and [24] H. Singh, A. Kumar and L. K. Balyan, "Cuckoo Search Optimizer based
Electron Optics, vol. 125, no. 17, pp. 4646-4651, Sept. 2014. Piecewise Gamma Corrected Auto-clipped Tile-wise Equalization for
Satellite Image Enhancement," 14th IEEE India Council International
[9] S. C. F. Lin, C. Y. Wong, M. A. Rahman, G. Jiang, S . Liu, N. Kwok, H.
Conference (INDICON), Roorkee, India, 2017, pp. 1-6. doi:
Shi, Y.H. Yu and T. Wu, “Image enhancement using the averaging 10.1109/INDICON.2017.8487901
histogram equalization (AVHEQ) approach for contrast improvement
and brightness Preservation,” Comput. and Elect. Eng., vol. 46, pp. 356- [25] H. Singh, A. Kumar and L. K. Balyan, "A Levy Flight Firefly Optimizer
370, Aug. 2015. based Piecewise Gamma Corrected Unsharp Masking Framework for
Satellite Image Enhancement," 14th IEEE India Council International
[10] K. Zuiderveld, “Contrast Limited Adaptive Histogram Equalization,”
Conference (INDICON), Roorkee, India, 2017, pp. 1-5. doi:
Chapter VIII.5, Graphics Gems IV. P.S. Heckbert (Eds.), Cambridge, 10.1109/INDICON.2017.8487501
MA, Academic Press, 1994, pp. 474–485.
[26] F. A. Spanhol, L. S. Oliveira, C. Petitjean and L. Heutte, "A Dataset for
[11] S. C. Huang, F. C. Cheng and Y. S. Chiu, "Efficient Contrast Breast Cancer Histopathological Image Classification," IEEE
Enhancement Using Adaptive Gamma Correction with Weighting Transactions on Biomedical Engineering, vol. 63, no. 7, pp. 1455-1462,
Distribution," IEEE Trans. Image Process, vol. 22, no. 3, pp. 1032- 2016. doi: 10.1109/TBME.2015.2496264
1041, Mar. 2013.
[12] H. Singh, A. Kumar, L. K. Balyan and G. K. Singh, "Slantlet filter-
bank-based satellite image enhancement using gamma-corrected knee
transformation," International Journal of Electronics (Taylor & Francis),
vol.105, no. 10, pp. 1695-1715, 2018. doi:
10.1080/00207217.2018.1477199.
[13] H. Singh, N. Agrawal, A. Kumar, G. K. Singh and H. N. Lee, "A novel
gamma correction approach using optimally clipped sub-equalization for
dark image enhancement," IEEE International Conference on Digital
Signal Processing (DSP), Beijing, China, pp. 497-501, 2016. doi:
10.1109/ICDSP.2016.7868607.
[14] H. Singh, A. Kumar, L. K. Balyan, and G. K. Singh,“A novel optimally
gamma corrected intensity span maximization approach for dark image
enhancement,”IEEE International Conference on Digital Signal
Processing (DSP), London, United Kingdom, pp. 1-5, 2017
doi:10.1109/ICDSP.2017.8096035.
[15] C. Y. Wong, G. Jiang, M. A. Rahman, S. Liu, S. C. F. Lin, N. Kwok, H.
Shi, Y. H. Yu and T. Wu, “Histogram equalization and optimal profile
compression based approach for colour image enhancement,” J. Visual
Commun. and Image Represen., vol. 38, pp. 802-813, July 2016.
[16] C. Y. Wong, S. Liu, S. C. Liu, M. A. Rahman, S. C. F. Lin, G. Jiang, N.
Kwok and H. Shi, “Image contrast enhancement using histogram
equalization with maximum intensity coverage,” J. Modern Optics, vol.
63, no. 16, pp. 1618-1629, Mar. 2016.
[17] D. Sheet, H. Garud, A. Suveer, M. Mahadevappa and J. Chatterjee,
"Brightness preserving dynamic fuzzy histogram equalization," IEEE
Trans. Consum. Electron., vol. 56, no. 4, pp. 2475-2480, Nov. 2010.
[18] K. Singh, R. Kapoor and S. K. Sinha, "Enhancement of low exposure
images via recursive histogram equalization algorithms," Optik -Int. J.
Light and Electron Optics, vol. 126, pp. 2619–2625, 2015.
[19] X. Fu, J. Wang, D. Zeng, Y. Huang and X. Ding, "Remote Sensing
Image Enhancement Using Regularized-Histogram Equalization and
DCT," in IEEE Geoscience and Remote Sensing Letters, vol. 12, no. 11,
pp. 2301-2305, Nov. 2015.
[20] H. Singh, A. Kumar, L. K. Balyan and G. K. Singh, "Regionally
equalized and contextually clipped gamma correction approach for dark
image enhancement," IEEE International Conference on Signal
Processing and Integrated Networks (SPIN), Noida, India, 2017, pp.
431-436. doi: 10.1109/SPIN.2017.8049988.
[21] H. Singh and A. Kumar, "Satellite image enhancement using beta
wavelet based gamma corrected adaptive knee transformation," IEEE
International Conference on Communication and Signal Processing
(ICCSP), Melmaruvathur, India, 2016, pp. 128-132. doi:
10.1109/ICCSP.2016.7754552