See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/316705880

Color Retinal Image Enhancement Based on Luminosity and Contrast

Adjustment

Article in IEEE transactions on bio-medical engineering · May 2017

DOI: 10.1109/TBME.2017.2700627

CITATIONS READS
160 1,316

5 authors, including:

Kai Jin Shaoze Wang

Zhejiang University Zhejiang University
53 PUBLICATIONS 797 CITATIONS 14 PUBLICATIONS 313 CITATIONS

SEE PROFILE SEE PROFILE

Dahong Qian
Shanghai Jiao Tong University
83 PUBLICATIONS 1,048 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Computer-aided diagnosis of retinal diseases View project

Ophthalmic biomaterials View project

All content following this page was uploaded by Kai Jin on 29 May 2021.

The user has requested enhancement of the downloaded file.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TBME.2017.2700627, IEEE
Transactions on Biomedical Engineering
TBME-01821-2016.R1 1

Color Retinal Image Enhancement Based on

Luminosity and Contrast Adjustment
Mei Zhou#, Kai Jin#, Shaoze Wang, Juan Ye, and Dahong Qian*, Senior Member, IEEE

 diseases, including diabetic retinopathy, age-related macular

Abstract—Objective: Many common eye diseases and degeneration, and glaucoma [1]. However, some retinal images
cardiovascular diseases can be diagnosed through retinal imaging. are of a clinically unsatisfactory quality due to eye lesions and
However, due to uneven illumination, image blurring, and low imperfect imaging processes [2], [3], such as opacity of
contrast, retinal images with poor quality are not useful for
refractive media, hemorrhages, exudates, and patients’ eye
diagnosis, especially in automated image analyzing systems. Here
we propose a new image enhancement method to improve color movement. Uneven illumination, blurring, incorrect focus, and
retinal image luminosity and contrast. Methods: A luminance gain low contrast reduce the quality of retinal images, resulting in a
matrix, which is obtained by gamma correction of the value loss of sensitivity and specificity for diagnostic purposes, and
channel in the HSV (Hue, Saturation, and Value) color space, is may even impair ophthalmologists' ability to interpret
used to enhance the R, G, and B (Red, Green and Blue) channels, significant eye features or distinguish different retinal diseases
respectively. Contrast is then enhanced in the luminosity channel
[4]. Poor quality retinal images make it difficult for subsequent
of L*a*b* color space by CLAHE (contrast limited adaptive
histogram equalization). Image enhancement by the proposed accurate segmentation and computer-aided diagnosis of retinal
method is compared to other methods by evaluating quality scores diseases [5], which are used to automate the detection process
of the enhanced images. Results: The performance of the method and to assist ophthalmologists. Therefore, it is necessary to
is mainly validated on a dataset of 961 poor quality retinal images. overcome the challenges associated with poor quality retinal
Quality assessment (range 0-1) of image enhancement of this poor images. One effective method is to use image enhancement
dataset indicated that our method improved color retinal image
technology [6] to provide better visibility of the retinal
quality from an average of 0.0404 (standard deviation 0.0291) up
to an average of 0.4565 (standard deviation 0.1000). Conclusion: anatomical structure.
The proposed method is shown to achieve superior image In recent years, many new enhancement methods for retinal
enhancement compared to contrast enhancement in other color images have been proposed to augment the classical histogram
spaces or by other related methods, while simultaneously equalization, including image luminosity and contrast
preserving image naturalness. Significance: This method of color normalization techniques [7], a multi-scale method based on
retinal image enhancement may be employed to assist
the Contourlet transform [8], CLAHE (contrast limited
ophthalmologists in more efficient screening of retinal diseases
and in development of improved automated image analysis for adaptive histogram equalization) [9]-[11], Retinex-based
clinical diagnosis. enhancement algorithm [12], [13], blood vessel enhancement
by multi-scale top-hat transformation and linear stretching with
Index Terms—Contrast enhancement, gamma correction, histogram Gaussian curve fitting [14] or via multi-dictionary
L*a*b* color space, luminosity, retinal image. and sparse coding [15], and some other combinatorial methods.
These existing methods can broadly be divided into three types:
histogram based, filter based, and transformation based. Most
I. INTRODUCTION of these methods focus on enhancing retinal blood vessels to

R ETINAL images are widely used by the ophthalmologists

for early detection and diagnosis of common retinal
achieve better vessel segmentation through increasing the
contrast between blood vessels and the retinal background in
both grayscale and color retinal images. This method is
especially useful for color retinal images, where the green
Manuscript received September 18, 2016; revised December 26, 2016 and
March 14, 2017; accepted April 25, 2017. This work was supported in part by channel of the color retinal image generally displays a high
the National Natural Science Foundation of China under Grant No. 81471748, contrast between the vessels and the background. The enhanced
and the Project supported by Shanghai Key Laboratory of Multidimensional retinal images can lose color information or other important
Information Processing, East China Normal University, China. M. Zhou and K.
Jin are joint first authors. Asterisk indicates corresponding author. image features (e.g., optic disc, macula lutea, and various types
M. Zhou is with the Shanghai Key Laboratory of Multidimensional of lesions), which cannot directly improve the current status of
Information Processing, East China Normal University. diagnosis by ophthalmologists.
K. Jin and J. Ye are with the Department of Ophthalmology, the Second
Affiliated Hospital of Zhejiang University, College of Medicine, Zhejiang
In order to obtain more image details as well as preserve
University. image naturalness which means pleasing perceptual quality
S. Wang is with the Institute of VISI Design, Zhejiang University. without color distortion and over-enhancement, we present a
*D. Qian is with the Institute of Translational Medicine, College of
Medicine, Zhejiang University, Hangzhou 310020, China (correspondence
new enhancement method for color retinal images based on
e-mail: dqian@zju.edu.cn). illumination and contrast adjustment. To avoid the common

0018-9294 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TBME.2017.2700627, IEEE
Transactions on Biomedical Engineering
TBME-01821-2016.R1 2

Fig. 2. Four retinal image samples from our data subset.

Tokyo, Japan) at the Eye Center of Second Affiliated Hospital

of Zhejiang University. A subset of 961 images with overall
poor quality was selected automatically from the 4000 images.
The automatic-selection system was implemented by our
proposed algorithm [18] where three characteristics of the
human visual system -- multi-channel sensation, noticeable
blur, and the contrast sensitivity function -- were utilized to
detect illumination and color distortion, blur, and low contrast
distortion, respectively. The selected images in the subset have
poor assessed scores which are below 0.1; the maximum
possible score is 1. Clinical diagnosis on these retinal images is
usually difficult as they are of very low contrast, blur, and
uneven illumination. In addition, circular masks were drawn
Fig. 1. Flow diagram of our proposed method for color retinal image and the images were cropped to eliminate the area beyond the
enhancement.
region of interest [18]. The preprocessing reduces the size of
the image and hence less number of pixels is investigated in
problem of color distortion, all the processes are performed on
further computation, which can reduce the computational time.
the luminosity channel. The luminance gain matrix, which is
Four representative samples in our data subset are shown in Fig.
obtained by a non-linear transformation of the value channel in
2. (No.104, No.134, No.145, No.854).
the HSV (Hue, Saturation, and Value) color space, is used to
Meanwhile, to verify the effectiveness of the proposed
enhance the R, G, and B (Red, Green and Blue) channels
method on normal dataset and make the result reproducible, the
respectively. Subsequently, contrast is enhanced in the
largest publicly-available Messidor dataset [19] is also used to
luminosity channel of L*a*b* color space [16], [17] by
test. The Messidor dataset contains 1200 retinal images, which
CLAHE, which specializes in improving local contrast and
were acquired at three different ophthalmology departments
avoiding the non-homogeneous regions in retinal images (the
using a non-mydriatic 3CCD camera (Topcon TRC NW6) at
optic disc region is significantly brighter than other regions)
45o FOV with a resolution of 1440*960, 2240*1488 or
and enhances the image uniformly. The method is evaluated on
2304*1536 pixels and stored in TIFF format.
a data subset of poor quality retinal images, as assessed by the
human visual system-based fundus image quality assessment
system [18] from our proprietary datasets, and a B. Luminosity Enhancement
publicly-available dataset. A comparison with contrast Because insufficient or uneven luminance obscures visual
enhancement in different color spaces and other related perception of retinal images, making diagnostic details
methods is also presented. undetectable, it is essential to enhance the luminance effect first.
However, for a color image, the color should not change for any
II. MATERIALS AND METHODS pixel, to prevent image distortion. In general, color retinal
The proposed method includes two steps: luminosity images are stored and viewed using RGB color space. The R, G,
enhancement and contrast enhancement. The flow diagram of and B channels simultaneously contain the luminosity
our proposed method is shown in Fig. 1. information and the color information, which are correlated
with each other. To enhance the luminosity and preserve the
color, the R, G, and B channels should be adjusted by the same
A. Retinal Image Datasets
proportion [20]. Our solution is to obtain a luminance gain
The main color retinal image dataset used for evaluation of matrix G(x, y) which is defined as follow:
the proposed method is selected from a proprietary repository
of retinal images. To collect our proprietary datasets, 4000 r ' ( x, y) g ' ( x, y) b' ( x, y)
images with a resolution of 2144 ×1424 pixels were taken by a    G( x, y) (1)
r ( x, y) g ( x, y) b( x, y)
tabletop TRC-NW8 fundus camera (TopCon Medical Systems,

0018-9294 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TBME.2017.2700627, IEEE
Transactions on Biomedical Engineering
TBME-01821-2016.R1 3

which is also used in our method. The transformation curve of

gamma correction is shown in Fig. 3. Compared with the
identity line when γ = 1, the nonlinearity transformation can
effectively enhance the luminosity. The values in the calculated
luminance gain matrix G(x, y) are always greater than or equal
to 1. The gray level interval [0, 50] in the luminosity channel (V)
is transformed to [0, 122], [100, 150] is transformed to [167,
200], and [200, 255] is transformed to [228, 255]. The
transformation can increase the dynamic range of the low gray
level and compress the high gray level, which can clarify the
details of the retinal image.
Even more importantly, the luminosity enhancement in the
HSV color space can effectively take care of the gamut problem
[23]. Because the pixel value in the luminosity channel (V) is
the maximum value in the R, G, and B channels and the gamma
Fig. 3. Transformation curve of gamma correction and the identity line.
correction does not change the gray level range. The enhanced
values r'(x, y), g'(x, y), and b'(x, y) by multiplying the original R,
where r'(x, y), g'(x, y), and b'(x, y) are the enhanced R, G, and B
G, and B values and G(x, y) respectively cannot go out of
values in the pixel at (x, y) position, and the r(x, y), g(x, y), and
bounds.
b(x, y) are the original R, G, and B values.
To obtain the color-invariant luminance gain matrix, the
color image is transformed into the HSV color space where the C. Contrast Enhancement
luminosity channel (V) is decoupled from the two color The luminosity enhancement above can balance the overall
components, hue (H) and saturation (S). The H and S channels luminance, especially for retinal images with low luminosity.
are irrelevant to luminance, and both are ignored. The Image contrast can also be enhanced to some extent. However,
luminance intensity of a pixel at the (x, y) position is obtained if the retinal image is blurry and has moderate luminosity, there
as the maximum (max) of the R, G, and B values. Therefore, the would be no significant improvement in contrast. To further
luminance gain matrix can be inferred as enhance the contrast of retinal images, the CLAHE method is
applied, which has been verified to be an effective method to
V ' ( x, y) V ' ( x, y) uniformly enhance details of gray retinal images [10], [11].
G( x, y)   (2)
V ( x, y) max(r ( x, y), g ( x, y), b( x, y)) The CLAHE divides the image into small regions called tiles;
the histogram on each tile is equalized so that local contrast is
where V(x, y) is the luminance intensity of a pixel at (x, y) enhanced. This local enhancement technique can result in noise
position, and V'(x, y) is the function of V(x, y), which dominating on individual local regions. Hence, a clip-limitation
determines the effect of luminosity enhancement. We can see strategy is employed to prevent local contrast from hitting the
that the processing can be directly done in the RGB color space, maximum. In our method, the number of tiles is 8×8, and the
which reduces the computational complexity. clip limit is 0.01. Both derive from a public domain
For the V'(x, y), we aim to significantly increase the dynamic implementation of CLAHE in MATLAB. Optimization results
range of the low gray level region, to slightly increase the of the two values showed that higher values generate blocking
moderate gray level region, and to maintain or compress the effects whereas lower values fail to enhance contrast.
high gray level region. Gamma correction [21], [22], a popular For color retinal images, it is better to implement CLAHE on
imaging processing methods, is used to transform luminance the luminosity channel to reduce color distortion. This contrast
nonlinearly. The transformation curve is expressed by enhancement will only operate on the luminosity channel and
can avoid the gamut problem. The L*a*b* color space (also
called CIE LAB) [16], [17] based on the prior master space CIE
w  u (3) 1931 XYZ color space can provide more uniform color
differences in relation to visual differences. It has three
where u  [0,1] denotes the normalized pixel value of the orthogonal dimensions: luminosity, L* (dark-to-bright);
luminosity channel, w is the normalized output, and γ is a chromaticity, a* (green-to-red); and b* (blue-to-yellow). The
constant. The transformation has a simple pointwise operation value of L* extends from 0 (black) to 100 (white), and the
form. Before undergoing the transformation, V(x, y) is values of a* and b* both range from -128 to +127. The L*a*b*
normalized as the input u. V'(x, y) is the reversed normalization color space is able to better separate the luminosity and the
of the output w. The gamma, γ, is typically greater than 1.0 in a color, and its L* component closely matches human perception
display device, and the standard of the National Television of luminosity, which can be used to adjust the contrast. Thus,
System Committee recommends a gamma of 2.2, which can be the retinal image in RGB color space by luminosity
used to effectively recover an overexposed region. Whereas γ < enhancement is converted into L*a*b* color space. The
1 has exactly the opposite effect, the usual setting is 1/2.2, CLAHE is used to enhance the L channel. Then the processed

0018-9294 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TBME.2017.2700627, IEEE
Transactions on Biomedical Engineering
TBME-01821-2016.R1 4

Fig. 5. Corresponding histograms of ROI of the color retinal images in Fig. 4.

The red, green, and blue lines correspond to the R, G, and B channels
respectively.

RAM and elapsed times were reported in seconds.

III. RESULTS AND DISCUSSION

Fig. 4. Procedure of color retinal image enhancement by the proposed method, The performance of the color retinal image enhancement
(a) Original images, (b) luminosity enhancement of original images, and (c)
contrast enhancement of images in 4b. method based on luminosity and contrast adjustment was
mainly analyzed and validated on our proprietary dataset of 961
image in the L*a*b* color space is transformed back into the color retinal images and further on the Messidor dataset.
RGB color space. However, there are no simple formulas for Quality of both original and enhanced images was
conversion between RGB and L*a*b*, because the RGB color quantitatively assessed by our assessment algorithm [18].
space is device-dependent. To obtain device-independent data, Initially, we applied our proposed enhancement method to
the luminosity enhanced image in RGB color space is firstly four retinal images (Fig. 2).The stepwise process of color
transformed to the CIE 1931 XYZ color space and then retinal image enhancement by the proposed method starting
transformed into L*a*b*. The forward transformation process with the original image (Fig. 4a), proceeding to luminosity
of the L channel is described in (4)-(6) as follows. enhancement (Fig. 4b), and culminating with contrast
enhancement (Fig. 4c) is shown for representative images. This
process demonstrates that our method of luminosity
 X  0.412453 0.357580 0.180423  r ' ( x, y ) 
 Y    0.212671 0.715160 0.072169   g ' ( x, y ) (4)
enhancement achieves improvement in the luminance of retinal
      images without color distortion. Enhancement of luminance
 Z  0.019334 0.119193 0.950227  b' ( x, y )  alone results in overall uniform and bright luminance (Fig. 4b),
L  116 f (Y / Yn )  16 (5) but the clarity of the retinal images is still insufficient.
Subsequently applying CLAHE to the L channel of the L*a*b*
  6 
3
 t1 / 3 t   color space results in significant enhancement of the contrast of

f (t )    29  (6) the retinal images (Fig. 4c). The corresponding ROI histograms
2
   t
1 29 4 from the original and enhanced color retinal images in Fig. 4
 3  6  others
 29
verify the effectiveness of our method (Fig. 5). The images
No.134, No.145, and No.854 have low gray-level values in
where X, Y, and Z are the three components in CIE XYZ color whole and narrow dynamic range in the R, G, and B channels,
space, and Xn, Yn, and Zn are the CIE XYZ tristimulus values of especially in the B channel, which means low luminosity and
the reference white [24]. contrast and is consistent with our visual perception. The image
The value of L is proportional to Y in which the green No.104 has moderate gray-level values, but it also has narrow
channel is weighted largely. The green channel exhibits the dynamic range, and the highlight region around the optic disc
highest contrast among the three channels of the RGB color makes some details indistinguishable. After luminosity
space, therefore the L channel is also expected to have higher enhancement, the whole gray-level values in all channels
contrast. Enhancement in the L channel of the L*a*b* color become greater due to the characteristic of luminance gain
space should further improve performance. matrix G(x, y) (Fig. 5b), which indicates better illuminance.
All the operations were implemented using MATLAB And the dynamic range of histograms in the low gray-level
R2013b on a PC with Intel i5 processor at 2.30 GHz and 8 GB interval is also stretched. It is an effective and conventional

0018-9294 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TBME.2017.2700627, IEEE
Transactions on Biomedical Engineering
TBME-01821-2016.R1 5

TABLE I
AVERAGE AND STANDARD DEVIATION OF ASSESSED SCORES OF THE ORIGINAL
AND ENHANCED IMAGES IN THE L*A*B*, HSI, AND HSV COLOR SPACES.
Original
LAB HSI HSV
images
Average score 0.0404 0.4565 0.4372 0.2297
Standard deviation 0.0291 0.1000 0.0981 0.0907

Fig. 6. (a) Original retinal image; luminosity and contrast enhanced images
with contrast enhancement in the (b) L*a*b*, (c) HSI, and (d) HSV color space.
The selected regions in the black rectangle are zoomed in in the bottom row.

Fig. 7. Boxplot of the assessed scores of the original and luminosity and
contrast enhanced 961 images with contrast enhancement in the L*a*b*, HSI, Fig. 8. Enhancement of poor quality color retinal images by different
and HSV color space. enhancement methods. Original images are found in Fig. 2. (a) Image
enhancement was performed (a) by our proposed method based on luminosity
method for contrast enhancement by spreading the histogram of and contrast enhancement, (b) by using CLAHE only in the L channel of CIE
LAB color space, (c) using CLAHE in the R, G, and B channels, and (d) using
the processed image. The luminosity enhancement can improve
the method based on guided image filtering.
the contrast to some extent. The dynamic range of the
histograms in the R, G, and B channels of these processed space (Fig. 6c) has comparable luminance with the enhanced
images is further remarkably broadened by contrast image in the L*a*b* color space (Fig. 6b), it also has increased
enhancement, as indicated by the ROI histograms for the noise and inferior contrast in visual perception compared to the
contrast enhanced images (Fig. 5c), which demonstrates better image enhanced in the L*a*b* color space. Furthermore,
contrast compared to the luminosity enhanced images. The comparison of selected regions demonstrates that blood vessels
analysis result of histograms is consistent with our visual and other important details in the image enhanced in the
perception (Fig. 4). L*a*b* color space have better visibility and clarity than in
The resulting retinal images enhanced by our proposed images enhanced in the HSI and HSV color spaces. To
method display improved visualization of not only blood quantitatively compare the three color spaces, the original and
vessels but also other important anatomical structures of the enhanced images from our 961 image data subset were assessed
retina (i.e., optical disc and macula), and improves the by our proposed quality assessment method. This assessment
prominence of lesions. Moreover, our method preserves the indicated that the quality of images with contrast enhanced in
naturalness of the retinal images while at the same time the L*a*b* color space is superior to those enhanced in the
enhancing critical details, which can assist the HSV and HSI color space, as indicated by higher assessed
ophthalmologists in better retinal image analysis. scores for individual images and for the 961 image data
To verify the effectiveness of contrast enhancement of the subgroup as a whole (Fig. 7, Table I). This demonstrates that
luminosity enhanced image in the L*a*b* color space, we our proposed method is able to achieve the best overall assessed
carried out a comparison of our method with contrast score for enhanced images, and indicates that the L*a*b* color
enhancement of the I and V channels in the HSI and HSV color space has an advantage over the HSV and HSI color spaces for
spaces respectively by the CLAHE approach. The contrast improving image contrast and highlighting retinal vessel
enhancement in the L*a*b color space demonstrates superior characteristics.
luminance and contrast compared to enhancements in the HSI We further carried out a comparison with other related
and HSV color space (representative color retinal images enhancement methods, including a method using CLAHE only
shown in Fig. 6). Although the image enhanced in the HSI color

0018-9294 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TBME.2017.2700627, IEEE
Transactions on Biomedical Engineering
TBME-01821-2016.R1 6

compared to other image enhancement methods.

IV. CONCLUSIONS
Here we present an effective method for color retinal image
enhancement based on luminosity and contrast adjustment.
First, the luminosity of the color retinal image is enhanced by a
luminance gain matrix based on gamma correction, and then
image contrast is enhanced by CLAHE in the L*a*b* color
space. The performance of our proposed method was validated
on two large color retinal image datasets. The results show that,
compared with contrast enhancement in other color spaces and
other methods, our proposed method achieves superior
improvement of color retinal images, especially for those with
initially of poor quality. This method is not only able to
Fig. 9. Boxplot of assessed scores of the original and enhanced images in our enhance important anatomical structures of the retina, but it
proprietary dataset and the Messidor dataset respectively. also preserves the naturalness of the images. This effective
method of color retinal image enhancement will greatly assist
in the L channel of the L*a*b* color space, a method using ophthalmologists in disease diagnosis through retinal image
CLAHE in the R, G, and B channels, and a method based on analysis, and will be greatly beneficial to automated image
guided image filtering [25], in which the input image itself is analysis systems. The clinical evaluation of our method is
used as a guidance image (Fig. 8). In contrast to our proposed currently in progress.
enhancement method based on luminosity and contrast
enhancement (Fig. 8a), the method using CLAHE in only L ACKNOWLEDGMENT
channel of the L*a*b* color space improves contrast for the
The authors would like to acknowledge the Messidor
retinal images with normal illuminance, but it is unable to
database which were provided by the Messidor program
balance luminosity for retinal images with low illuminance (Fig. partners (see http://www.adcis.net/en/DownloadThirdParty/
8b). The image enhancement using CLAHE in the R, G, and B Messidor.html).
channels results in severe color distortion, which may lead to a
clinical misdiagnosis (Fig. 8c). Finally, although the REFERENCES
enhancement method based on guided image filtering can [1] M. D. Abramoff et al., “Retinal imaging and image analysis,” IEEE Rev.
effectively enhance the details of the whole image (Fig. 8d), Biomed. Eng., vol. 3, pp. 169-208, Dec. 2010.
this method results in unnatural looks due to the over [2] A. F. M. Hani and H. A. Nugroho, “Retinal vasculature enhancement
enhancement of details and edges. And the two methods are using independent component analysis,” J. Biomed. Sci. Eng., vol. 2, no. 7,
pp. 543-549, Nov. 2009.
also unable to improve luminosity. Therefore, luminosity [3] J. Paulus et al., “Automated quality assessment of retinal fundus photos,”
enhancement in our method is necessary, especially for poor Int. J. Comput. Assist. Radiol. Surg., vol. 5, no. 6, pp. 557-564, Nov.
quality color retinal images, and contrast enhancement in our 2010.
[4] U. Sevik et al., “Identification of suitable fundus images using automated
method is more effective. quality assessment methods,” J. Biomed. Opt., vol. 19, no. 4, p. 046006,
Moreover, the retinal images in the Messidor dataset were Apr. 2014.
processed by our enhancement method, and were also [5] M. R. K. Mookiah et al., “Computer-aided diagnosis of diabetic
retinopathy: a review,” Comput. Biol. Med., vol. 43, no. 12, pp.
evaluated by the same assessment method (Fig. 9). The 2136-2155, Dec. 2013.
averages of assessed scores of the original and enhanced [6] E. Daniel and J. Anitha, “Optimum green plane masking for the contrast
images are 0.1957 (standard deviation 0.0819) and 0.4610 enhancement of retinal images using enhanced genetic algorithm,” Optik,
vol. 126, no. 18, pp. 1726-1730, Sep. 2015.
(standard deviation 0.1152) respectively. The assessment result [7] M. Foracchia et al., “Luminosity and contrast normalization in retinal
demonstrates that our method can improve the normal retinal images,” Med. Image Anal., vol. 9, no. 3, pp. 179-190, Jun. 2005.
images, too. When the retinal image is collected, it can be [8] P. Feng et al., “Enhancing retinal image by the Contourlet transform,”
Pattern Recogn. Lett., vol. 28, no. 4, pp. 516-522, Mar. 2007.
directly processed by the proposed method without prior [9] E. D. Pisano et al., “Contrast limited adaptive histogram equalization
assessment. By comparing the assessed scores of the two image processing to improve the detection of simulated spiculations in
datasets, the average of the enhanced images in the Messidor dense mammograms,” J. Digit. Imaging, vol. 11, no. 4, pp. 193-200, Nov.
dataset is slightly better in general, but the average of the 1998.
[10] G. S. Ramlugun et al., “Small retinal vessels extraction towards
original images is significantly higher. Despite the good quality proliferative diabetic retinopathy screening,” Expert Syst. Appl., vol. 39,
achieved with enhanced images in both datasets, our method no. 1, pp. 1141-1146, Jan. 2012.
performs better on the poor quality retinal images in our [11] R. GeethaRamani and L. Balasubramanian, “Retinal blood vessel
segmentation employing image processing and data mining techniques
proprietary dataset. for computerized retinal image analysis,” Biocybern. Biomed. Eng., vol.
In summary, the above results demonstrate that our proposed 36, no. 1, pp. 102-118, Jan. 2016.
method works well on both the poor and normal retinal images [12] Y. Zhao et al., “Retinal vessel segmentation: an efficient graph cut
approach with retinex and local phase,” Plos One, vol. 10, no. 4, p.
to improve their luminosity and contrast, and provides e0122332, Apr. 2015.
qualitatively and quantitatively superior enhancement

0018-9294 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TBME.2017.2700627, IEEE
Transactions on Biomedical Engineering
TBME-01821-2016.R1 7

[13] S. Wang et al., “Naturalness preserved enhancement algorithm for [19] E. Decenciere et al., “Feedback on a publicly distributed image database:
non-uniform illumination images,” IEEE T. Image Process., vol. 22, no. 9, the Messidor database,” Image Anal. Stereol., vol. 33, no. 3, pp. 231-234,
pp. 3538-3548, Sep. 2013. Aug. 2014.
[14] M. Liao et al., “Retinal vessel enhancement based on multi-scale top-hat [20] L. Cai et al., “A hue-preserving based algorithm for low-illumination
transformation and histogram fitting stretching,” Opt. Laser Technol., vol. colour image enhancement,” Comput. Appl. Softw., vol. 26, no. 3, pp.
58, pp. 56-62, Jun. 2014. 226-227, Mar. 2009.
[15] B. Chen et al., “Blood vessel enhancement via multi-dictionary and sparse [21] H. W. Guo et al., “Gamma correction for digital fringe projection
coding: application to retinal vessel enhancing,” Neurocomputing, vol. profilometry,” Appl. Optics, vol. 43, no. 14, pp. 2906-2914, May 2004.
200, pp. 110-117, Aug. 2016. [22] S. C. Huang et al., “Efficient contrast enhancement using adaptive gamma
[16] R. Kirifi et al., “CIEL*a*b* color space predictive models for colorimetry correction with weighting distribution,” IEEE T. Image Process., vol. 22,
devices - analysis of perfume quality,” Talanta, vol. 104, pp. 58-66, Jan. no. 3, pp. 1032-1041, Mar. 2013.
2013. [23] S. K. Naik and C. A. Murthy, “Hue-preserving color image enhancement
[17] H. Tang and Y. Zhao, “Edge detection in CIE L*a *b * based on without gamut problem,” IEEE T. Image Process., vol. 12, no. 12, pp.
fractional differential,” J. Image Graph., vol. 18, no. 6, pp. 628-636, Jun. 1591-1598, Dec. 2003.
2013. [24] L. Xu et al., “Moving object detecting using LAB color space,” J.
[18] S. Wang et al., “Human visual system-based fundus image quality Huazhong Univ. Sci. & Tech. (Natural Science Edition), vol. 41, no. I, pp.
assessment of portable fundus camera photographs,” IEEE T. Med. 219-222, Oct. 2013.
Imaging, vol. 35, no. 4, pp. 1046-1055, Apr. 2016. [25] K. M. He et al., “Guided image filtering,” IEEE T. Pattern Anal., vol. 35,
no. 6, pp. 1397-1409, Jun. 2013.

0018-9294 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
View publication stats