Research Article Retinal Fundus Image Enhancement Using the Normalized Convolution …downloads.hindawi.com/journals/ijbi/2016/5075612.pdf · 2019-07-30 · fundus image such as small

Research ArticleRetinal Fundus Image Enhancement Usingthe Normalized Convolution and Noise Removing

Peishan Dai1 Hanwei Sheng1 Jianmei Zhang1 Ling Li1 Jing Wu1 and Min Fan2

1Department of Biomedical Engineering School of Geosciences and Info-Physics Central South University Changsha 410083 China2Department of Education and Law Hunan Womenrsquos University Changsha 410004 China

Correspondence should be addressed to Min Fan dpsgrace163com

Received 22 March 2016 Revised 29 June 2016 Accepted 4 August 2016

Academic Editor Jyh-Cheng Chen

Copyright copy 2016 Peishan Dai et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Retinal fundus image plays an important role in the diagnosis of retinal related diseases The detailed information of the retinalfundus image such as small vessels microaneurysms and exudates may be in low contrast and retinal image enhancementusually gives help to analyze diseases related to retinal fundus image Current image enhancement methods may lead to artificialboundaries abrupt changes in color levels and the loss of image detail In order to avoid these side effects a new retinal fundusimage enhancement method is proposed First the original retinal fundus image was processed by the normalized convolutionalgorithm with a domain transform to obtain an image with the basic information of the background Then the image with thebasic information of the background was fused with the original retinal fundus image to obtain an enhanced fundus image Lastlythe fused image was denoised by a two-stage denoising method including the fourth order PDEs and the relaxed median filter Theretinal image databases including the DRIVE database the STARE database and the DIARETDB1 database were used to evaluateimage enhancement effects The results show that the method can enhance the retinal fundus image prominently And differentfrom some other fundus image enhancement methods the proposed method can directly enhance color images

1 Introduction

Retinal fundus images provide rich information of pathologi-cal changeswhichmay indicate diseases such as arteriosclero-sis diabetes hypertension stroke and cardiovascular disease[1] These images are widely used for diagnosis of relateddisease However the retinal fundus images are usually withlow contrast uneven illumination and blur of the details dueto the complex imaging environmentsThe purpose of retinalfundus image enhancement is to improve the contrast andhighlight the retinal vessels [2]

There aremany image enhancementmethods Histogramequalization (HE) [3] is a popular method to improve imagecontrast but the decreasing of the gray levels may resultin the loss of image details To overcome this deficiencycontrast limited adaptive histogram equalization (CLAHE) isproposed [4] But the CLAHEmethodmay produce artificialboundaries at the region which has an abrupt change in thegray levels Due to the special characteristic enhancement ofretinal fundus image needs specific design Setiawan et al [5]

demonstrated that CLAHE method is suitable for improvingthe retinal image quality Ashiba et al [6] proposed an imageenhancement method based on a wavelet-based homomor-phic filter which can improve image contrast and dynamicrange However it is difficult to find a suitable structuringelement for the morphological filter Oh and Hwang [7] pro-posed an image enhancementmethod based onmorphology-based homomorphic filter and differential evolution algo-rithm But the target image region should be input firstlywhich may limit the scope of its application U Qidwaiand U Qidwai [8] proposed a retinal image enhancementmethod based on the blind deconvolution approach usingmaximum likelihood estimation And this deconvolutionapproach showed a promising result which may be used forclinical application Fraz et al [9] used a 2D Gabor filter toenhance retinal fundus image which could detect the vesselsinmultidirectionHowever it included some sensitive param-eters Bai et al [10] proposed an image enhancement methodbased on multiscale top-hat transformation In this methodthe image details were well enhanced but the contrast is

Hindawi Publishing CorporationInternational Journal of Biomedical ImagingVolume 2016 Article ID 5075612 12 pageshttpdxdoiorg10115520165075612

2 International Journal of Biomedical Imaging

not improved effectively Rampal et al [11] used a complexdiffusion-based shock filter for retinal image smoothingand contrast enhancement And this method outperformedother methods on the DIARETDB1 database Liao et al [12]used a hybrid model of multiscale top-hat transformationand histogram fitting stretching to enhance retinal fundusimage which could enhance the contrast of the retinal imageeffectively and highlight the retinal vessels well But someparameters have to be set cautiously in the part of histogramfitting stretching

In this paper a retinal image enhancement method wasproposed to avoid producing artificial boundaries abruptchanges in color levels and the loss of image detail Firstwe used normalized convolution with a domain transformto obtain an image containing the basic information of theoriginal imageThen the image containing the basic informa-tion was fused with the original image to enhance vessels anddetail of the retinal fundus image Lastly the fused imagewentthrough denoising filters to achieve image enhancement

The rest of this paper is organized as follows In Section 2the proposed method is introduced in detail In Section 3experimental results of three well-known databases are usedto demonstrate the effectiveness of the proposed methodFinally in Section 4 the conclusion is presented

2 Materials and Methods

It is very difficult to enhance retinal fundus images directlydue to the uneven illumination of imaging characteris-tic So different from commonly used image enhancementmethods which enhance retinal fundus image directly theproposed method firstly obtained an image with the basicinformation of the background through some process ofthe original image and then fused this image with theoriginal image to suppress background to achieve the goalof image enhancement As the image enhancement processis realized by background suppression the proposed methodmay reduce occurrence probability of artificial boundariesabrupt changes in color levels and the loss of image detail

We used the normalized convolution with a domaintransform proposed by Gastal and Oliveira [13] to obtain abackground image with the basic information and then fusedthe image with the original image to achieve retinal fundusimage enhancement Lastly the fused image was denoised bythe fourth order PDEs [14] and the relaxed median filter [15]

21 Normalized Convolution with Domain Transform Nor-malized convolution [16] uses neighbor information tomodelthe image pixel of the original image In [13] normalizedconvolution was used to obtain edge-preserving smoothedimages with a domain transform box kernel But if doingsome variation for the parameters the smoothing methodcan produce an image with the basic information of thebackground of the retinal fundus image which does not focuson edge-preservingTheprocessing procedures are as follows

Let 119868 represent the original retinal fundus image and 119901 =(119909119901 119910119901) denotes the position of a pixel in the original image 119868and 119868(119901) = (119903119901 119892119901 119887119901) denotes the pixel value in the 2D RGBcolor image119863(Ω) denotes the domain of119901 that is119901 isin 119863(Ω)

Then the normalized convolution with a domain transform isas follows [13]

119869 (119901) = (1

119870119901

) sum

119902isin119863(Ω)

119868 (119902)119867 (119905 () 119905 ())

119870119901 = sum

119902isin119863(Ω)

119867(119905 () 119905 ())

(1)

where 119870119901 is the normalization factor for 119901 And 119905() is adomain transform

119905 () = 119888119905 (119901) int

119901

0

1 +120575119904

120575119903

3

sum

119896=1

100381610038161003816100381610038161198681015840

119896(119909)10038161003816100381610038161003816119889119909 (2)

where 1198681015840119896(119909) is the derivative of 119896th channel of the retinal

color image and 120575119904 and 120575119903 are standard deviation of typicallyGaussian spatial and range filters [17]

As the domain transform 119888119905(119909) is monotonically increas-ing an efficient moving-average approach [18] is used toperform NC with a box filter 119867(119905() 119905()) The box kernelis defined as follows [13]

119867(119905 () 119905 ()) = 120575 1003816100381610038161003816119905 (119901) minus 119905 (119902)

1003816100381610038161003816 le 119903

119903 = 120575119867radic3

120575119867119894= 120575119867

radic32119873minus119894

radic4119873 minus 1

(3)

where 120575119867119894 is the standard deviation for the kernel used in the119894th iteration119873 is the total number of iterations and 120575119867 is thestandard iteration of the desired kernel

22 Image Fusion In the above section we obtained thebackground image with basic information In order to obtainan enhancement result we fuse the background image withthe original image The fusing result is defined as

119875 = 119868 minus 119886 (119869 minus 119868) (4)

where 119875 is the enhancement result and 119869 and 119868 are the imagewith the basic information of the background (represented as119869(119901) in Section 21) and the original image respectively And119886 is a parameter factor which could determine the contrast ofthe image

23 ImageDenoising Noisemay be amplified in image fusionstep so some noise suppression process should be done toobtain the enhanced image The second order PDEs can dohigh quality denoising yet the methods tend to cause blockyeffects [19] Fourth order PDEs can avoid the blocky effectsThe function is as follows [14]

120597119875

120597119905= minusnabla2[119888 (10038161003816100381610038161003816nabla211987510038161003816100381610038161003816) nabla2119875] (5)

when we get 120597119875120597119905 representing the difference betweenoriginal image and the denoised imaging containing noise

International Journal of Biomedical Imaging 3

information we reduced the noise information to get thedenoising where 119888(|nabla119875|) is as follows [14]

119888 (|nabla119875|) =1

1 + (|nabla119875| 119896)2 (6)

where 119896 was set to be 1 This step can be run more than onetime if necessary

The image processed by the fourth order PDEs may alsoreserve some noise so the relaxed median filter can be usedto smooth the image further [15] Assuming that the imageprocessed by the fourth order PDEs can be represented as119875pde the function is given as follows [15]

119875relax 119898 = 119877119872120572120596 119882(119894119895)

=

119875pde(119894119895) if 119875pde(119894119895) isin [[119882(119894119895)](120572) [119882(119894119895)](120596)]

[119882(119894119895)](119898)otherwise

(7)

where [119882(119894119895)](119898) is the median value of the samples inside thewindow119882(119894119895) And the sliding window119882(119894119895) at the windowlocated at the point is as follows

119882(119894119895) = 119875pde(119894+119903119895+119903) 119903 isin 119882 (8)

3 Results and Discussion

31 Testing Data Sets In this section three public databasesare used to test the proposed method namely the DRIVEdatabase [20] the STARE database [21] and the DIARETDB1database [22] The DRIVE database originally collected byStaal et al [20] contains 40 color retinal images with the sizeof 565 times 584 pixels This database is divided into two setsmdasha training set and a testing set which is captured using the45-degree field-of-view digital fundus camera The STAREdatabase originally collected byHoover et al [21] contains 20color retinal images with the size of 700 times 605 pixels whichis captured using the 35-degree field-of-view digital funduscamera The DIARETDB1 database originally collected byKauppi et al [22] contains 89 color retinal images with thesize of 1500 times 1152 pixels which is captured using the 50-degree field-of-view digital fundus camera

32 Parameter Determination There are several importantparameters directly affecting the enhancement results Thereare two parameters 120590119904 and 120590119903 in (2) of the normalizedconvolution with a domain transform and there is oneparameter 119886 in (4) of the image fusion

In order to determine the parameters 120590119904 and 120590119903 in (2)we changed the values of 120590119904 and 120590119903 to test which values aresuitable for retinal image enhancement As the characteristicof retinal image is relatively stable an original retinal imagewhich is randomly chosen from the DRIVE database isenhanced by the domain transform and the normalizedconvolution algorithm with different values of 120590119904 and 120590119903(120590119904 = 1 10 30 60 100 150 and 120590119903 = 01 02 04 06 1 15)Here 119886 is set as 5 randomly In Figure 1 the parameter 120590119903changes by column and the parameter 120590119904 changes by row

From the bottom right corner of Figure 1 we can see thatas the parameters 120590119904 and 120590119903 increase the background of theretinal image is overenhanced gradually and the vessels in theoptic disk are missing gradually As a compromise we choose120590119904 as 60 and 120590119903 as 04 in this paper

In order to determine the parameter 119886 in (4) we variedthe value of 119886 to test which value is suitable for retinal imageenhancement The original retinal image which is randomlychosen from the DRIVE database is enhanced by the domaintransform and the normalized convolution algorithm withdifferent value of 119886 (119886 = 1 3 5 8 10) Here 120590119904 and 120590119903 areset as 60 and 04 Figure 2(a) is an original retinal image andFigures 2(b)ndash2(f) are the enhanced images with the differentvalue of 119886 (119886 = 1 3 5 8 10) Figures 2(b)-2(c) showed thatsome dim regions of the original image are notwell enhancedIn Figures 2(e)-2(f) the image is overenhanced and the vesselregions are very clear but many noises are also producedTheimage and the vessel regions are efficiently enhanced in Fig-ure 2(d) and fewer noises are produced So the enhancementresult of Figure 2(d) may be a better result so we let 119886 = 5

33 Results of Enhancement Figure 3(a) is the original retinalimage which is randomly chosen from the DRIVE databaseFigure 3(b) is the background image obtained by normalizedconvolution with a domain transform with the parameter 120590119904being 60 and 120590119903 being 04 Figure 3(c) is the image fusingresult with the parameter 119886 being 5 Figure 3(d) is the noiseremoval result Figure 3(e) is the green channel of Figure 3(a)and Figure 3(f) is the green channel of Figure 3(d)

Comparing Figure 3(d) with Figure 3(a) shows that theenhanced image improved image contrast level The detailsof the small vessels are enhanced significantlyThe dim regionsuch as fovea becomes much dimmer

The subregions of the images in Figure 3 are used toshow the denoising effects Figure 4(a) is the subregion ofFigure 3(c) and Figure 4(b) is the subregion of Figure 3(d)Figure 4 shows that the noise produced by image enhance-ment steps is depressed effectively

Figures 5 and 6 show the image enhancement results ofa randomly selected image from the STARE database andthe DIARETDB1 database respectively Figures 5(a)ndash5(f) arethe original image from the STARE database the normalizedconvolution result the fusion result the denoising resultthe green channel of original image and the green channelof denoising result respectively Similarly Figures 6(a)ndash6(f) are the original image from the DIARETDB1 databasethe normalized convolution result the fusion result thedenoising result the green channel of original image and thegreen channel of denoising result respectively These resultsshow that the proposed method has wide applicability

34 Evaluating Image Enhancement Effects by Comparingwith Other Methods There are two main ways to evaluateimage enhancement effects one is subjective evaluation andthe other is objective evaluation The subjective evaluationis done by observing the enhancement by experts or per-sons with image evaluation experience The advantage ofsubjective evaluation is that the evaluation results are more


Figure 1 The enhancement results of the retinal image with different 120590119904 and 120590119903

in line with the purpose of image enhancement but thedisadvantage is that the subjective evaluation may vary fromperson to person and it is a problemhow to quantify the eval-uation The objective evaluation methods can quantitativelyevaluate the image enhancement results but it is difficultfor the indicator of the objective evaluation methods toreflect humanrsquos subject perception So we compare the imageenhancement results with other methods using both thesubjective evaluation and the objective evaluation methodsto evaluate the image enhancement results

As vessel is important information in retinal fundusimage two objective evaluating methods are used to quantifythe quality of the enhanced image One method is used to

quantify the enhancement of the retinal vessels and the otheris used to quantify the enhancement of the whole image

The first objective evaluating method tends to quantifythe enhancement of the retinal vessels The measure is calledthe contrast improvement index [7] which is defined asfollows

CII =119862en119862 (9)

where 119862en and 119862 are the contrast values for the retinal vesselin the enhanced image and the original image respectivelyHere the values of119862en and119862 are calculated without the pixels


(a) (b) (c)

(d) (e) (f)

Figure 2 The enhancement results of the retinal image with different 119886 (a) Original image (b) enhanced image with 119886 = 1 (c) enhancedimage with 119886 = 3 (d) enhanced image with 119886 = 5 (e) enhanced image with 119886 = 8 (f) enhanced image with 119886 = 10

of the black background points outside the pupil Here 119862 isdefined as follows

119862 =

10038161003816100381610038161003816100381610038161003816

119891 minus 119887

119891 + 119887

10038161003816100381610038161003816100381610038161003816

(10)

where119891 and 119887 are the average gray values of the retinal vesselsand the non-vessel regions respectively It is obvious that alarger value of119862means a larger difference between the retinalvessels and the non-vessel regions 119862en can be calculated inthe same way as 119862 and the difference is just the fact that theimage used to produce 119862en is the enhanced image It shouldbe pointed out that the manually segmented retinal vessels ofthe original image should be available using this method

The second objective evaluating method used to quantifythe enhancement of the whole image The measure is calledlinear index of fuzziness [10] which is defined as follows

119903 (119891en) =2

119872119873

119898

sum

119909=1

119899

sum

119910=1

min 119901119909119910 (1 minus 119901119909119910)

119901119909119910 = sin[1205872times 1 minus (

119891 (119909 119910)

119891max)]

(11)

where 119891max is the maximum value of the whole image withsize119872times119873

A larger value of CII shows that the retinal vessels areenhanced better And a smaller value of 119903 shows that thewhole enhanced image is clearer and has less noiseThereforea large value of CII and a small value of 119903 indicate a good imageenhancement method

The image enhancement results of the proposed methodare color images Yet the green channel image is generallyused to evaluate image enhancement results So only thegreen channel of the enhanced image is used in imageevaluation step

The image enhancement results of Figure 3(e) with othermethods such as HE [3] CLAHE [4] and the methodsin [7 9] are shown in Figure 7 Figures 7(a)ndash7(d) are theenhancement result using HE the enhancement result usingCLAHE the enhancement result using method in [7] andthe enhancement result using method in [9] respectivelyTable 1 shows the comparison of CII and 119903 correspondingto different image enhancement results of Figure 3(e) It isshown that the proposed method has a larger value of CIIand a relatively smaller value of 119903 comparing with the othermethods It should be pointed out that before calculating CIIand 119903 of the image shown in Figure 7(d) it is inverted for the


Table 1 The comparison of CII and 119903 with different methods on the image shown in Figure 3(a)

Evaluation measures HE CLAHE Method in [7] Method in [9] The proposed methodCII 23981 30866 41526 54816 69785119903 05406 07402 04453 07094 04523

(a) (b) (c)

(d) (e) (f)

Figure 3 The enhancement process of the chosen retinal image from the DRIVE database (a) Original image (b) normalized convolutionresult (c) fusion result (d) denoising result (e) green channel of original image (f) green channel of denoising result

(a) (b)

Figure 4 The subregion of Figure 3(c) (before denoising) and Figure 3(d) (after denoising) (a) Subregion of Figure 3(c) (b) subregion ofFigure 3(d)

purpose of comparing all the methods at the same conditionthat the vessels appear darker than the background

Three retinal images randomly chosen from the DRIVEdatabase are also used to compare image enhancement resultsshown in Figure 8The first column (Figure 8(a)) is the green

channel of each original image The middle four columns(Figures 8(b)ndash8(e)) are the enhancement results of the HEthe CLAHE and the methods in [7 9] respectively Thelast column is the results of the proposed method From thecomparison of image enhancement results we can see that


(a) (b) (c)

(d) (e) (f)

Figure 5 The image enhancement results of an image selected from the STARE database (a) Original image (b) normalized convolutionresult (c) fusion result (d) denoising result (e) green channel of original image (f) green channel of denoising result

(a) (b) (c)

(d) (e) (f)

Figure 6 The image enhancement results of an image selected from the DIARETDB1 database (a) Original image (b) normalizedconvolution result (c) fusion result (d) denoising result (e) green channel of original image (f) green channel of denoising result


(a) (b)

(c) (d)

Figure 7The enhancement results with different methods (a)The enhancement result using HE (b) the enhancement result using CLAHE(c) the enhancement result using method in [7] (d) the enhancement result using method in [9]

Table 2 The comparison of CII and r with different methods on the images shown in Figure 8

Image Evaluationmeasures HE CLAHE Method in [7] Method in [9] The proposed

method

1st in column (a) CII 23105 23252 38081 22630 59076119903 05391 06264 02012 06401 02267

2nd in column (a) CII 30113 30059 46100 45377 75447119903 05418 05454 02979 07148 02582

3rd in column (a) CII 22450 30392 48968 47682 79369119903 05412 04865 02004 06378 02514

the HE method produces uneven background image whichchanged the illumination distribution of the original imageand the low-contrast region is not enhanced well So manydetails would be missed by using the HEmethod as shown inFigure 8(b) The CLAHE method could enhance the detailsof the retail image but the whole contrast is not improvedsignificantly as shown in Figure 8(c) The results of using themethod in [7] are too dim as shown in Figure 8(d)The image

enhancement results using the method in [9] lose the detailsof the background as shown in Figure 8(e) The values ofCII and 119903 of each image in Figure 8 are shown in Table 2It shows that the retinal vessels are enhanced better and thewhole image is clearer using the proposed method as shownin Figure 8(f)

The image enhancement results for pathological retinalfundus images chosen from the STARE database are shown in


(a) (b) (c) (d) (e) (f)

Figure 8The comparison of image enhancement results of three images in the DRIVE database (a)The original image (b) the enhancementresults of HE (c) the enhancement results of CLAHE (d) the enhancement results using the method in [7] (e) the enhancement results usingthe method in [9] (f) the enhancement results of the proposed method

(a) (b) (c) (d) (e) (f)

Figure 9The comparison of image enhancement results of three images in the STARE database (a)The original image (b) the enhancementresults of HE (c) the enhancement results of CLAHE (d) the enhancement results using method in [7] (e) the enhancement results usingmethod in [9] (f) the enhancement results of the proposed method

Figure 9Thefirst column (Figure 9(a)) is the green channel ofeach original imageThemiddle four columns (Figures 9(b)ndash9(e)) are the enhancement results of the HE the CLAHEand the methods in [7 9] respectively The last column isthe results of the proposed method as shown in Figure 9(f)It shows that the proposed method not only enhances theretinal vessels and makes the whole image clearer but alsoenhances the pathological regions well The values of CII and119903 of each image in Figure 9 are shown in Table 3 It shows that

the proposed method could perform better than the othermethods on the pathological images

The comparison of line charts of CII and 119903 values for theimages of DRIVE database is shown in Figure 10 Figures10(a)-10(b) are the results using the 20 images in the testingset ofDRIVEdatabase And Figures 10(c)-10(d) are the resultsusing all the 40 images of DRIVE database including testingset (20 images) and training set (20 images) In Figures 10(a)ndash10(d) the 119909-axis denotes the sequence number of the image


0 2 4 6 8 10 12 14 16 18 20 220

2

4

6

8

10

12

14

16

HE Method in [9]CLAHE The proposed methodMethod in [7]

Valu

e of C

II

Sequence number of the image

(a)

0 2 4 6 8 10 12 14 16 18 20 2200

02

04

06

08

10

12

Valu

e of r



(b)


0 5 10 15 20 25 30 35 400

10

20

30

40

Valu

e of C

II


(c)


0 5 10 15 20 25 30 35 4000

02

04

06

08

10

12Va

lue o

f r


(d)

Figure 10 Comparison of the line charts of CII and r on the DRIVE database (a) The line charts of CII of the testing set (b) the line chartsof 119903 of the testing set (c) the line charts of CII of the testing and training sets (d) the line charts of r of the testing and training sets

Table 3 The comparison of CII and r with different methods on images shown in Figure 9


method

1st in column (a) CII 12610 19859 30856 13910 46879119903 05189 04203 01781 06664 02845

2nd in column (a) CII 21659 21467 28935 15957 51086119903 05196 05660 02247 06549 04306

3rd in column (a) CII 18862 17994 29650 17290 49442119903 05196 05607 01294 06735 03019



0 2 4 6 8 10 12 14 16 18 20 220

2

4

6

8

10Va

lue o

f CII


(a)

0 2 4 6 8 10 12 14 16 18 20 2200

02

04

06

08

10

12


Valu

e of r


(b)

Figure 11 Comparison of the line charts of CII and 119903 on the STARE database (a) The line charts of CII (b) the line charts of 119903

Table 4 The average running time comparison of different methods on the DRIVE database (including 40 images)

Evaluation methods HE CLAHE Method in [7] Method in [9] The proposed methodTimes 0012 0045 0291 2672 18152

Table 5 The average running time comparison of different methods on the STARE database (including 20 images)


and the 119910-axis denotes the value of CII and 119903 of the differentmethods By the same token Figures 11(a) and 11(b) showthe line charts of the values of CII and 119903 of the differentmethods used to enhance retinal fundus images of the STAREdatabase respectively

It is shown from Figures 10 and 11 that the proposedmethod achieved the largest CII which indicates that theproposed method is the best in enhancing the retinal vesselsAnd the value of 119903 of the proposed method is nearlysecond small comparing with the other methods whichindicates that the whole image is comparatively clearer afterenhancement using the proposed methodThe enhancementperformance of the proposed method on both normal andpathological retinal images is comparatively better

The proposed method shows some merits comparingwith other methods (1)The edges of the vessels change softlyand do not cause artificial edges as there are not abruptchanges in the image containing basic information (2) As theenhanced image is obtained by fusing the image with basicinformation of the background and the original image thereis no main detail losing (3) A proper noise reduction filter isused to depress the noise caused by the fusing step

The running times of different methods were calculatedon a computer with Windows 7 OS a Dual Core 340GHz

CPU 4GB of RAM under Matlab R2012 software Theaverage running time comparisons of different methods onthe DRIVE database and STARE database are shown inTables 4 and 5 respectively It is shown from the two tablesthat the fastest method is the HE method and the mosttime consuming method is the proposed method on boththe DRIVE database and STARE database The proposedmethod needs to do some time consuming processes suchas domain transform convolution the fourth order PDEsand the relaxed median filter to achieve image enhancementwhich makes the proposed method consume much moreaverage running time than all the other methods As runningtime is also a very important factor in clinical application thisis an obvious defect of the proposed method In the future ifwe can find another fast image denoising method suitable forremoving the noise caused by image fusion the running timemay be reduced to some extent

4 Conclusions

In this paper a retinal image enhancement method fusingoriginal image and the image containing the basic informa-tion of the background is presented The proposed methodis tested on the retinal images of the DRIVE database the


STARE database and the DIARETDB1 database respectivelyTwo objective evaluation indexes (CII and r) which couldmeasure both the contrast and clarity of the enhanced imageare used The CII value of the proposed method is distinctlythe largest and the 119903 value is the second smallest in all thecompared methods on the DRIVE and the STARE databasesIt is shown that the proposed method achieves a much betterperformance on enhancing retinal fundus image comparingwith the othermethodsMoreover different fromother imageenhancement methods mentioned above the proposedmethod can handle color images which may be more bene-ficial to diagnosis by the ophthalmologists

Competing Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

This work was financially supported by the National NaturalScience Foundation of China (Grants no 81171420 no61379107 and no 81370913)

References

[1] N Lai ldquoClinical ophthalmology a systematic approachrdquoOptometry and Vision Science vol 81 no 5 p 295 2004

[2] D Marın A Aquino M E Gegundez-Arias and J M BravoldquoA new supervised method for blood vessel segmentation inretinal images by using gray-level andmoment invariants-basedfeaturesrdquo IEEE Transactions on Medical Imaging vol 30 no 1pp 146ndash158 2011

[3] K-Q Huang Q Wang and Z-Y Wu ldquoNatural color imageenhancement and evaluation algorithm based on human visualsystemrdquoComputer Vision and ImageUnderstanding vol 103 no1 pp 52ndash63 2006

[4] A M Reza ldquoRealization of the contrast limited adaptivehistogram equalization (CLAHE) for real-time image enhance-mentrdquo Journal of VLSI Signal Processing Systems for SignalImage and Video Technology vol 38 no 1 pp 35ndash44 2004

[5] A W Setiawan T R Mengko O S Santoso and A BSuksmono ldquoColor retinal image enhancement using CLAHErdquoin Proceedings of the International Conference on ICT for SmartSociety (ICISS rsquo13) pp 1ndash3 Jakarta Indonesia June 2013

[6] H I Ashiba K H Awadalla SM El-Halfawy and F E Abd El-Samie ldquoHomomorphic enhancement of infrared images usingthe additive wavelet transformrdquo Progress in ElectromagneticsResearch C vol 1 pp 123ndash130 2008

[7] J Oh and H Hwang ldquoFeature enhancement of medical imagesusing morphology-based homomorphic filter and differentialevolution algorithmrdquo International Journal of Control Automa-tion and Systems vol 8 no 4 pp 857ndash861 2010

[8] U Qidwai and U Qidwai ldquoBlind Deconvolution for retinalimage enhancementrdquo in Proceedings of the IEEE EMBS Confer-ence on Biomedical Engineering and Sciences (IECBES rsquo10) pp20ndash25 Kuala Lumpur Malaysia December 2010

[9] M M Fraz P Remagnino A Hoppe et al ldquoAn ensembleclassification-based approach applied to retinal blood vessel

segmentationrdquo IEEE Transactions on Biomedical Engineeringvol 59 no 9 pp 2538ndash2548 2012

[10] X Z Bai F G Zhou and B D Xue ldquoImage enhancement usingmulti scale image features extracted by top-hat transformrdquoOptics and Laser Technology vol 44 no 2 pp 328ndash336 2012

[11] H Rampal R K Kumar B Ramanathan and T P Das ldquoCom-plex shock filtering applied to retinal image enhancementrdquoin Proceedings of the World Congress on Medical Physics andBiomedical Engineering (WCMPBE rsquo12) pp 900ndash903 BeijingChina May 2012

[12] M Liao Y-Q Zhao X-H Wang and P-S Dai ldquoRetinal vesselenhancement based on multi-scale top-hat transformation andhistogram fitting stretchingrdquo Optics and Laser Technology vol58 pp 56ndash62 2014

[13] E S L Gastal and M M Oliveira ldquoDomain transform foredge-aware image and video processingrdquo ACM Transactions onGraphics vol 30 no 4 article 69 pp 1244ndash1259 2011

[14] J Rajan K Kannan and M R Kaimal ldquoAn improved hybridmodel for molecular image denoisingrdquo Journal of MathematicalImaging and Vision vol 31 no 1 pp 73ndash79 2008

[15] A B Hamza P L Luque-Escamilla J Martınez-Aroza and RRoman-Roldan ldquoRemoving noise and preserving details withrelaxed median filtersrdquo Journal of Mathematical Imaging andVision vol 11 no 2 pp 161ndash177 1999

[16] H Knutsson and C F Westin ldquoNormalized and differentialconvolution methods for interpolation andfiltering of incom-plete and uncertain datardquo in Proceedings of the IEEE Conferenceon Computer Vision and Pattern Recognition (CVPR rsquo93) pp515ndash523 New York NY USA June 1993

[17] C Tomasi and R Manduchi ldquoBilateral filtering for gray andcolor imagesrdquo in Proceedings of the IEEE 6th InternationalConference on Computer Vision (ICCV rsquo98) pp 839ndash846 IEEEBombay India January 1998

[18] E R Dougherty Digital Image Processing Methods OpticalEngineering CRC Press Boca Raton Fla USA 1994

[19] Y-L You and M Kaveh ldquoFourth-order partial differentialequations for noise removalrdquo IEEE Transactions on ImageProcessing vol 9 no 10 pp 1723ndash1730 2000

[20] J Staal M D Abramoff M Niemeijer M A Viergever andB Van Ginneken ldquoRidge-based vessel segmentation in colorimages of the retinardquo IEEETransactions onMedical Imaging vol23 no 4 pp 501ndash509 2004

[21] A Hoover V Kouznetsova andM Goldbaum ldquoLocating bloodvessels in retinal images by piecewise threshold probing of amatched filter responserdquo IEEETransactions onMedical Imagingvol 19 no 3 pp 203ndash210 2000

[22] T Kauppi V Kalesnykiene J K Kamarainen et alldquoDIARETDB1 diabetic retinopathy database and evaluationprotocolrdquo in Proceedings of the British Machine VisionConference (BMVC rsquo07) pp 1ndash18 Warwick UK September2007

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Active and Passive Electronic Components

Control Scienceand Engineering

Journal of



RotatingMachinery


Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design



Shock and Vibration


Civil EngineeringAdvances in

Acoustics and VibrationAdvances in



Electrical and Computer Engineering

Journal of

Advances inOptoElectronics


Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of


Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Chemical EngineeringInternational Journal of Antennas and

Propagation




Navigation and Observation



DistributedSensor Networks



not improved effectively Rampal et al [11] used a complexdiffusion-based shock filter for retinal image smoothingand contrast enhancement And this method outperformedother methods on the DIARETDB1 database Liao et al [12]used a hybrid model of multiscale top-hat transformationand histogram fitting stretching to enhance retinal fundusimage which could enhance the contrast of the retinal imageeffectively and highlight the retinal vessels well But someparameters have to be set cautiously in the part of histogramfitting stretching

In this paper a retinal image enhancement method wasproposed to avoid producing artificial boundaries abruptchanges in color levels and the loss of image detail Firstwe used normalized convolution with a domain transformto obtain an image containing the basic information of theoriginal imageThen the image containing the basic informa-tion was fused with the original image to enhance vessels anddetail of the retinal fundus image Lastly the fused imagewentthrough denoising filters to achieve image enhancement

The rest of this paper is organized as follows In Section 2the proposed method is introduced in detail In Section 3experimental results of three well-known databases are usedto demonstrate the effectiveness of the proposed methodFinally in Section 4 the conclusion is presented

2 Materials and Methods

It is very difficult to enhance retinal fundus images directlydue to the uneven illumination of imaging characteris-tic So different from commonly used image enhancementmethods which enhance retinal fundus image directly theproposed method firstly obtained an image with the basicinformation of the background through some process ofthe original image and then fused this image with theoriginal image to suppress background to achieve the goalof image enhancement As the image enhancement processis realized by background suppression the proposed methodmay reduce occurrence probability of artificial boundariesabrupt changes in color levels and the loss of image detail

We used the normalized convolution with a domaintransform proposed by Gastal and Oliveira [13] to obtain abackground image with the basic information and then fusedthe image with the original image to achieve retinal fundusimage enhancement Lastly the fused image was denoised bythe fourth order PDEs [14] and the relaxed median filter [15]

21 Normalized Convolution with Domain Transform Nor-malized convolution [16] uses neighbor information tomodelthe image pixel of the original image In [13] normalizedconvolution was used to obtain edge-preserving smoothedimages with a domain transform box kernel But if doingsome variation for the parameters the smoothing methodcan produce an image with the basic information of thebackground of the retinal fundus image which does not focuson edge-preservingTheprocessing procedures are as follows

Let 119868 represent the original retinal fundus image and 119901 =(119909119901 119910119901) denotes the position of a pixel in the original image 119868and 119868(119901) = (119903119901 119892119901 119887119901) denotes the pixel value in the 2D RGBcolor image119863(Ω) denotes the domain of119901 that is119901 isin 119863(Ω)

Then the normalized convolution with a domain transform isas follows [13]

119869 (119901) = (1

119870119901

) sum

119902isin119863(Ω)

119868 (119902)119867 (119905 () 119905 ())

119870119901 = sum

119902isin119863(Ω)

119867(119905 () 119905 ())

(1)

where 119870119901 is the normalization factor for 119901 And 119905() is adomain transform

119905 () = 119888119905 (119901) int

119901

0

1 +120575119904

120575119903

3

sum

119896=1

100381610038161003816100381610038161198681015840

119896(119909)10038161003816100381610038161003816119889119909 (2)

where 1198681015840119896(119909) is the derivative of 119896th channel of the retinal

color image and 120575119904 and 120575119903 are standard deviation of typicallyGaussian spatial and range filters [17]

As the domain transform 119888119905(119909) is monotonically increas-ing an efficient moving-average approach [18] is used toperform NC with a box filter 119867(119905() 119905()) The box kernelis defined as follows [13]

119867(119905 () 119905 ()) = 120575 1003816100381610038161003816119905 (119901) minus 119905 (119902)

1003816100381610038161003816 le 119903

119903 = 120575119867radic3

120575119867119894= 120575119867

radic32119873minus119894

radic4119873 minus 1

(3)

where 120575119867119894 is the standard deviation for the kernel used in the119894th iteration119873 is the total number of iterations and 120575119867 is thestandard iteration of the desired kernel

22 Image Fusion In the above section we obtained thebackground image with basic information In order to obtainan enhancement result we fuse the background image withthe original image The fusing result is defined as

119875 = 119868 minus 119886 (119869 minus 119868) (4)

where 119875 is the enhancement result and 119869 and 119868 are the imagewith the basic information of the background (represented as119869(119901) in Section 21) and the original image respectively And119886 is a parameter factor which could determine the contrast ofthe image

23 ImageDenoising Noisemay be amplified in image fusionstep so some noise suppression process should be done toobtain the enhanced image The second order PDEs can dohigh quality denoising yet the methods tend to cause blockyeffects [19] Fourth order PDEs can avoid the blocky effectsThe function is as follows [14]

120597119875

120597119905= minusnabla2[119888 (10038161003816100381610038161003816nabla211987510038161003816100381610038161003816) nabla2119875] (5)

when we get 120597119875120597119905 representing the difference betweenoriginal image and the denoised imaging containing noise



119888 (|nabla119875|) =1

1 + (|nabla119875| 119896)2 (6)



119875relax 119898 = 119877119872120572120596 119882(119894119895)

=

119875pde(119894119895) if 119875pde(119894119895) isin [[119882(119894119895)](120572) [119882(119894119895)](120596)]

[119882(119894119895)](119898)otherwise

(7)


119882(119894119895) = 119875pde(119894+119903119895+119903) 119903 isin 119882 (8)


















CII =119862en119862 (9)



(a) (b) (c)

(d) (e) (f)



119862 =

10038161003816100381610038161003816100381610038161003816

119891 minus 119887

119891 + 119887

10038161003816100381610038161003816100381610038161003816

(10)



119903 (119891en) =2

119872119873

119898

sum

119909=1

119899

sum

119910=1

min 119901119909119910 (1 minus 119901119909119910)

119901119909119910 = sin[1205872times 1 minus (

119891 (119909 119910)

119891max)]

(11)








(a) (b) (c)

(d) (e) (f)


(a) (b)






(a) (b) (c)

(d) (e) (f)


(a) (b) (c)

(d) (e) (f)



(a) (b)

(c) (d)




method

1st in column (a) CII 23105 23252 38081 22630 59076119903 05391 06264 02012 06401 02267

2nd in column (a) CII 30113 30059 46100 45377 75447119903 05418 05454 02979 07148 02582

3rd in column (a) CII 22450 30392 48968 47682 79369119903 05412 04865 02004 06378 02514





(a) (b) (c) (d) (e) (f)


(a) (b) (c) (d) (e) (f)






0 2 4 6 8 10 12 14 16 18 20 220

2

4

6

8

10

12

14

16


Valu

e of C

II


(a)

0 2 4 6 8 10 12 14 16 18 20 2200

02

04

06

08

10

12

Valu

e of r



(b)


0 5 10 15 20 25 30 35 400

10

20

30

40

Valu

e of C

II


(c)


0 5 10 15 20 25 30 35 4000

02

04

06

08

10

12Va

lue o

f r


(d)




method

1st in column (a) CII 12610 19859 30856 13910 46879119903 05189 04203 01781 06664 02845

2nd in column (a) CII 21659 21467 28935 15957 51086119903 05196 05660 02247 06549 04306

3rd in column (a) CII 18862 17994 29650 17290 49442119903 05196 05607 01294 06735 03019



0 2 4 6 8 10 12 14 16 18 20 220

2

4

6

8

10Va

lue o

f CII


(a)

0 2 4 6 8 10 12 14 16 18 20 2200

02

04

06

08

10

12


Valu

e of r


(b)











4 Conclusions




Competing Interests


Acknowledgments


References


























RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation











119888 (|nabla119875|) =1

1 + (|nabla119875| 119896)2 (6)



119875relax 119898 = 119877119872120572120596 119882(119894119895)

=

119875pde(119894119895) if 119875pde(119894119895) isin [[119882(119894119895)](120572) [119882(119894119895)](120596)]

[119882(119894119895)](119898)otherwise

(7)


119882(119894119895) = 119875pde(119894+119903119895+119903) 119903 isin 119882 (8)


















CII =119862en119862 (9)



(a) (b) (c)

(d) (e) (f)



119862 =

10038161003816100381610038161003816100381610038161003816

119891 minus 119887

119891 + 119887

10038161003816100381610038161003816100381610038161003816

(10)



119903 (119891en) =2

119872119873

119898

sum

119909=1

119899

sum

119910=1

min 119901119909119910 (1 minus 119901119909119910)

119901119909119910 = sin[1205872times 1 minus (

119891 (119909 119910)

119891max)]

(11)








(a) (b) (c)

(d) (e) (f)


(a) (b)






(a) (b) (c)

(d) (e) (f)


(a) (b) (c)

(d) (e) (f)



(a) (b)

(c) (d)




method

1st in column (a) CII 23105 23252 38081 22630 59076119903 05391 06264 02012 06401 02267

2nd in column (a) CII 30113 30059 46100 45377 75447119903 05418 05454 02979 07148 02582

3rd in column (a) CII 22450 30392 48968 47682 79369119903 05412 04865 02004 06378 02514





(a) (b) (c) (d) (e) (f)


(a) (b) (c) (d) (e) (f)






0 2 4 6 8 10 12 14 16 18 20 220

2

4

6

8

10

12

14

16


Valu

e of C

II


(a)

0 2 4 6 8 10 12 14 16 18 20 2200

02

04

06

08

10

12

Valu

e of r



(b)


0 5 10 15 20 25 30 35 400

10

20

30

40

Valu

e of C

II


(c)


0 5 10 15 20 25 30 35 4000

02

04

06

08

10

12Va

lue o

f r


(d)




method

1st in column (a) CII 12610 19859 30856 13910 46879119903 05189 04203 01781 06664 02845

2nd in column (a) CII 21659 21467 28935 15957 51086119903 05196 05660 02247 06549 04306

3rd in column (a) CII 18862 17994 29650 17290 49442119903 05196 05607 01294 06735 03019



0 2 4 6 8 10 12 14 16 18 20 220

2

4

6

8

10Va

lue o

f CII


(a)

0 2 4 6 8 10 12 14 16 18 20 2200

02

04

06

08

10

12


Valu

e of r


(b)











4 Conclusions




Competing Interests


Acknowledgments


References


























RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation















CII =119862en119862 (9)



(a) (b) (c)

(d) (e) (f)



119862 =

10038161003816100381610038161003816100381610038161003816

119891 minus 119887

119891 + 119887

10038161003816100381610038161003816100381610038161003816

(10)



119903 (119891en) =2

119872119873

119898

sum

119909=1

119899

sum

119910=1

min 119901119909119910 (1 minus 119901119909119910)

119901119909119910 = sin[1205872times 1 minus (

119891 (119909 119910)

119891max)]

(11)








(a) (b) (c)

(d) (e) (f)


(a) (b)






(a) (b) (c)

(d) (e) (f)


(a) (b) (c)

(d) (e) (f)



(a) (b)

(c) (d)




method

1st in column (a) CII 23105 23252 38081 22630 59076119903 05391 06264 02012 06401 02267

2nd in column (a) CII 30113 30059 46100 45377 75447119903 05418 05454 02979 07148 02582

3rd in column (a) CII 22450 30392 48968 47682 79369119903 05412 04865 02004 06378 02514





(a) (b) (c) (d) (e) (f)


(a) (b) (c) (d) (e) (f)






0 2 4 6 8 10 12 14 16 18 20 220

2

4

6

8

10

12

14

16


Valu

e of C

II


(a)

0 2 4 6 8 10 12 14 16 18 20 2200

02

04

06

08

10

12

Valu

e of r



(b)


0 5 10 15 20 25 30 35 400

10

20

30

40

Valu

e of C

II


(c)


0 5 10 15 20 25 30 35 4000

02

04

06

08

10

12Va

lue o

f r


(d)




method

1st in column (a) CII 12610 19859 30856 13910 46879119903 05189 04203 01781 06664 02845

2nd in column (a) CII 21659 21467 28935 15957 51086119903 05196 05660 02247 06549 04306

3rd in column (a) CII 18862 17994 29650 17290 49442119903 05196 05607 01294 06735 03019



0 2 4 6 8 10 12 14 16 18 20 220

2

4

6

8

10Va

lue o

f CII


(a)

0 2 4 6 8 10 12 14 16 18 20 2200

02

04

06

08

10

12


Valu

e of r


(b)











4 Conclusions




Competing Interests


Acknowledgments


References


























RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










(a) (b) (c)

(d) (e) (f)



119862 =

10038161003816100381610038161003816100381610038161003816

119891 minus 119887

119891 + 119887

10038161003816100381610038161003816100381610038161003816

(10)



119903 (119891en) =2

119872119873

119898

sum

119909=1

119899

sum

119910=1

min 119901119909119910 (1 minus 119901119909119910)

119901119909119910 = sin[1205872times 1 minus (

119891 (119909 119910)

119891max)]

(11)








(a) (b) (c)

(d) (e) (f)


(a) (b)






(a) (b) (c)

(d) (e) (f)


(a) (b) (c)

(d) (e) (f)



(a) (b)

(c) (d)




method

1st in column (a) CII 23105 23252 38081 22630 59076119903 05391 06264 02012 06401 02267

2nd in column (a) CII 30113 30059 46100 45377 75447119903 05418 05454 02979 07148 02582

3rd in column (a) CII 22450 30392 48968 47682 79369119903 05412 04865 02004 06378 02514





(a) (b) (c) (d) (e) (f)


(a) (b) (c) (d) (e) (f)






0 2 4 6 8 10 12 14 16 18 20 220

2

4

6

8

10

12

14

16


Valu

e of C

II


(a)

0 2 4 6 8 10 12 14 16 18 20 2200

02

04

06

08

10

12

Valu

e of r



(b)


0 5 10 15 20 25 30 35 400

10

20

30

40

Valu

e of C

II


(c)


0 5 10 15 20 25 30 35 4000

02

04

06

08

10

12Va

lue o

f r


(d)




method

1st in column (a) CII 12610 19859 30856 13910 46879119903 05189 04203 01781 06664 02845

2nd in column (a) CII 21659 21467 28935 15957 51086119903 05196 05660 02247 06549 04306

3rd in column (a) CII 18862 17994 29650 17290 49442119903 05196 05607 01294 06735 03019



0 2 4 6 8 10 12 14 16 18 20 220

2

4

6

8

10Va

lue o

f CII


(a)

0 2 4 6 8 10 12 14 16 18 20 2200

02

04

06

08

10

12


Valu

e of r


(b)











4 Conclusions




Competing Interests


Acknowledgments


References


























RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation












(a) (b) (c)

(d) (e) (f)


(a) (b)






(a) (b) (c)

(d) (e) (f)


(a) (b) (c)

(d) (e) (f)



(a) (b)

(c) (d)




method

1st in column (a) CII 23105 23252 38081 22630 59076119903 05391 06264 02012 06401 02267

2nd in column (a) CII 30113 30059 46100 45377 75447119903 05418 05454 02979 07148 02582

3rd in column (a) CII 22450 30392 48968 47682 79369119903 05412 04865 02004 06378 02514





(a) (b) (c) (d) (e) (f)


(a) (b) (c) (d) (e) (f)






0 2 4 6 8 10 12 14 16 18 20 220

2

4

6

8

10

12

14

16


Valu

e of C

II


(a)

0 2 4 6 8 10 12 14 16 18 20 2200

02

04

06

08

10

12

Valu

e of r



(b)


0 5 10 15 20 25 30 35 400

10

20

30

40

Valu

e of C

II


(c)


0 5 10 15 20 25 30 35 4000

02

04

06

08

10

12Va

lue o

f r


(d)




method

1st in column (a) CII 12610 19859 30856 13910 46879119903 05189 04203 01781 06664 02845

2nd in column (a) CII 21659 21467 28935 15957 51086119903 05196 05660 02247 06549 04306

3rd in column (a) CII 18862 17994 29650 17290 49442119903 05196 05607 01294 06735 03019



0 2 4 6 8 10 12 14 16 18 20 220

2

4

6

8

10Va

lue o

f CII


(a)

0 2 4 6 8 10 12 14 16 18 20 2200

02

04

06

08

10

12


Valu

e of r


(b)











4 Conclusions




Competing Interests


Acknowledgments


References


























RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










(a) (b) (c)

(d) (e) (f)


(a) (b) (c)

(d) (e) (f)



(a) (b)

(c) (d)




method

1st in column (a) CII 23105 23252 38081 22630 59076119903 05391 06264 02012 06401 02267

2nd in column (a) CII 30113 30059 46100 45377 75447119903 05418 05454 02979 07148 02582

3rd in column (a) CII 22450 30392 48968 47682 79369119903 05412 04865 02004 06378 02514





(a) (b) (c) (d) (e) (f)


(a) (b) (c) (d) (e) (f)






0 2 4 6 8 10 12 14 16 18 20 220

2

4

6

8

10

12

14

16


Valu

e of C

II


(a)

0 2 4 6 8 10 12 14 16 18 20 2200

02

04

06

08

10

12

Valu

e of r



(b)


0 5 10 15 20 25 30 35 400

10

20

30

40

Valu

e of C

II


(c)


0 5 10 15 20 25 30 35 4000

02

04

06

08

10

12Va

lue o

f r


(d)




method

1st in column (a) CII 12610 19859 30856 13910 46879119903 05189 04203 01781 06664 02845

2nd in column (a) CII 21659 21467 28935 15957 51086119903 05196 05660 02247 06549 04306

3rd in column (a) CII 18862 17994 29650 17290 49442119903 05196 05607 01294 06735 03019



0 2 4 6 8 10 12 14 16 18 20 220

2

4

6

8

10Va

lue o

f CII


(a)

0 2 4 6 8 10 12 14 16 18 20 2200

02

04

06

08

10

12


Valu

e of r


(b)











4 Conclusions




Competing Interests


Acknowledgments


References


























RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










(a) (b)

(c) (d)




method

1st in column (a) CII 23105 23252 38081 22630 59076119903 05391 06264 02012 06401 02267

2nd in column (a) CII 30113 30059 46100 45377 75447119903 05418 05454 02979 07148 02582

3rd in column (a) CII 22450 30392 48968 47682 79369119903 05412 04865 02004 06378 02514





(a) (b) (c) (d) (e) (f)


(a) (b) (c) (d) (e) (f)






0 2 4 6 8 10 12 14 16 18 20 220

2

4

6

8

10

12

14

16


Valu

e of C

II


(a)

0 2 4 6 8 10 12 14 16 18 20 2200

02

04

06

08

10

12

Valu

e of r



(b)


0 5 10 15 20 25 30 35 400

10

20

30

40

Valu

e of C

II


(c)


0 5 10 15 20 25 30 35 4000

02

04

06

08

10

12Va

lue o

f r


(d)




method

1st in column (a) CII 12610 19859 30856 13910 46879119903 05189 04203 01781 06664 02845

2nd in column (a) CII 21659 21467 28935 15957 51086119903 05196 05660 02247 06549 04306

3rd in column (a) CII 18862 17994 29650 17290 49442119903 05196 05607 01294 06735 03019



0 2 4 6 8 10 12 14 16 18 20 220

2

4

6

8

10Va

lue o

f CII


(a)

0 2 4 6 8 10 12 14 16 18 20 2200

02

04

06

08

10

12


Valu

e of r


(b)











4 Conclusions




Competing Interests


Acknowledgments


References


























RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










(a) (b) (c) (d) (e) (f)


(a) (b) (c) (d) (e) (f)






0 2 4 6 8 10 12 14 16 18 20 220

2

4

6

8

10

12

14

16


Valu

e of C

II


(a)

0 2 4 6 8 10 12 14 16 18 20 2200

02

04

06

08

10

12

Valu

e of r



(b)


0 5 10 15 20 25 30 35 400

10

20

30

40

Valu

e of C

II


(c)


0 5 10 15 20 25 30 35 4000

02

04

06

08

10

12Va

lue o

f r


(d)




method

1st in column (a) CII 12610 19859 30856 13910 46879119903 05189 04203 01781 06664 02845

2nd in column (a) CII 21659 21467 28935 15957 51086119903 05196 05660 02247 06549 04306

3rd in column (a) CII 18862 17994 29650 17290 49442119903 05196 05607 01294 06735 03019



0 2 4 6 8 10 12 14 16 18 20 220

2

4

6

8

10Va

lue o

f CII


(a)

0 2 4 6 8 10 12 14 16 18 20 2200

02

04

06

08

10

12


Valu

e of r


(b)











4 Conclusions




Competing Interests


Acknowledgments


References


























RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










0 2 4 6 8 10 12 14 16 18 20 220

2

4

6

8

10

12

14

16


Valu

e of C

II


(a)

0 2 4 6 8 10 12 14 16 18 20 2200

02

04

06

08

10

12

Valu

e of r



(b)


0 5 10 15 20 25 30 35 400

10

20

30

40

Valu

e of C

II


(c)


0 5 10 15 20 25 30 35 4000

02

04

06

08

10

12Va

lue o

f r


(d)




method

1st in column (a) CII 12610 19859 30856 13910 46879119903 05189 04203 01781 06664 02845

2nd in column (a) CII 21659 21467 28935 15957 51086119903 05196 05660 02247 06549 04306

3rd in column (a) CII 18862 17994 29650 17290 49442119903 05196 05607 01294 06735 03019



0 2 4 6 8 10 12 14 16 18 20 220

2

4

6

8

10Va

lue o

f CII


(a)

0 2 4 6 8 10 12 14 16 18 20 2200

02

04

06

08

10

12


Valu

e of r


(b)











4 Conclusions




Competing Interests


Acknowledgments


References


























RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation











0 2 4 6 8 10 12 14 16 18 20 220

2

4

6

8

10Va

lue o

f CII


(a)

0 2 4 6 8 10 12 14 16 18 20 2200

02

04

06

08

10

12


Valu

e of r


(b)











4 Conclusions




Competing Interests


Acknowledgments


References


























RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation











Competing Interests


Acknowledgments


References


























RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation











RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation









Documents

Research Article Retinal Fundus Image Enhancement Using the Normalized Convolution …downloads.hindawi.com/journals/ijbi/2016/5075612.pdf · 2019-07-30 · fundus image such as small