Image Contrast Enhancement based Sub-histogram Equalization Technique without Over-equalization Noise

Transcription

1 World Academy of Sciece, Egieerig ad Techology 5 9 Image Cotrast Ehacemet based Sub-histogram Equalizatio Techique without Over-equalizatio Noise Hyusup Yoo, Yougjoo Ha, ad Hersoo Hah Abstract I order to ehace the cotrast i the regios where the pixels have similar itesities, this paper presets a ew histogram equalizatio scheme. Covetioal global equalizatio schemes over-equalizes these regios so that too bright or dark pixels are resulted ad local equalizatio schemes produce uexpected discotiuities at the boudaries of the blocks. The proposed algorithm segmets the origial histogram ito sub-histograms with referece to brightess level ad equalizes each sub-histogram with the limited extets of equalizatio cosiderig its mea ad variace. The fial image is determied as the weighted sum of the equalized images obtaied by usig the sub-histogram equalizatios. By limitig the maximum ad miimum rages of equalizatio operatios o idividual sub-histograms, the over-equalizatio effect is elimiated. Also the result image does ot miss feature iformatio i low desity histogram regio sice the remaiig these area is applied separatig equalizatio. This paper icludes how to determie the segmetatio poits i the histogram. The proposed algorithm has bee tested with more tha 1 images havig various cotrasts i the images ad the results are compared to the covetioal approaches to show its superiority. Keywords Cotrast Ehacemet, Histogram Equalizatio. Histogram Regio Equalizatio, Equalizatio Noise I. INTRODUCTION ONTRAST ehacemet is acquirig clear image through C brightess itesity value redistributio. I other words, that is ehacig features as stretchig iterval betwee dark ad brightess area. Ehaced image which was result of cotrast ehacemet processig i preprocessig stage will provide clear image to eyes or assist feature extractio processig i computer visio system. They ca be categorized by two approaches, global ad local oes [1-3]. Global approaches improve image quality by extedig dyamic rage of itesity usig the histogram of the whole image. Global histogram equalizatio(ghe) [4,5] is a good example of this approach that is widely used. GHE is achieved by ormalizig the itesity distributio usig its F. Hyusup Yoo is with the Electrical Egieerig Departmet, Soogsil Uiversity, 511 Sado-dog, Dogjak-ku, Seoul , Korea ( beat83@ssu.ac.kr). S. Yougjoo Ha is with the Electrical Egieerig Departmet, Soogsil Uiversity, 511 Sado-dog, Dogjak-ku, Seoul , Korea ( youg@ssu.ac.kr). T. Hersoo Hah is with the Electrical Egieerig Departmet, Soogsil Uiversity,51 1 Sado-dog, Dogjak-ku, Seoul , Korea ( hah@ssu.ac.kr). cumulative distributio fuctio so that the result image may have a uiform distributio of itesity. Sice GHE is basically usig the itesity distributio of the whole image, it may cause washed-out effect by chagig the average itesity to middle level [5-1]. Ad whe equalizatio applyig i the image which has severely crowded distributio i arrow regio the may oise pixel value occurs as over-equalizatio processig. To solve these problems ivolved i global cotrast ehacemet techiques, local histogram equalizatio techiques have bee proposed. AHE (Adaptive Histogram Equalizatio) techique which is oe of basic local histogram equalizatio techiques divides the origial image ito several o-overlapped sub-blocks ad proceeds a histogram equalizatio operatio o idividual sub-blocks. The result image is produced by mergig the sub-blocks usig the bi-liear iterpolatio method [13]. I this case, a discotiuity problem, called a blockig effect, occurs ear the boudaries of the sub-blocks. It occurs sice a local cotrast ehacemet techique uses oly local iformatio iside each sub-block without cosiderig the itesity balace of a whole image [13, 14]. For solvig a blockig problem, sub-blocks are overlapped so that the boudary pixels may be icluded i the eighborig sub-blocks at the same time. But accordig to block size this process eed may calculatio ad memory quatity[14]. I this paper, we propose sub-histogram equalizatio based ew cotrast ehacemet algorithm which are prevetig over-equalizatio ad acquirig global cotrast ehaced effect. They divide optimized umber of histogram based o histogram desity fuctio. Ad divided sub-histogram regio are ehaced by sub-histogram equalizatio. Fially we get result image by mergig each ehaced sub-images. Proposed cotrast ehacemet techique has a advatage that prevet over-equalizatio effect i cocetrated histogram regios ad cotrast ehacemet i sub-histogram regio sice uiformly equalizatio execute i low desity regio. The rest of this paper is orgaized as follows. I Sectio, the advatages ad drawbacks of the covetioal cotrast ehacemet techiques based o global ad local cotrast ehacemet techiques are described. I Sectio 3, desity measurig method for divide histogram regio explaied ad i Sectio 4, sub-histogram equalizatio rage ad result image recovery methods will be explaied. Ad i Sectio 5, proposed algorithm is proved by the experimets where it is 176

2 World Academy of Sciece, Egieerig ad Techology 5 9 compared with those of the covetioal algorithms ad fially Sectio 6 presets the coclusio. II. CONVENTIONAL APPROACHES Global histogram equalizatio techiques [1] acquire the scale factor from the ormalized cumulative distributio of the brightess distributio of the origial image ad multiply this scale factor to the origial image to redistribute the itesity. A image is represeted as a set of data, X = {X (i,j) X(i,j) {X, X 1, X L-1 }}, where each compoet may have oe of L itesity levels. That is, X(i,j) represets the ormalized itesity of the (i,j)th pixel i the image plae ad X k is the k-th itesity level. To equalize the brightess histogram of a give image X, it is ormalized first to obtai the probability desity fuctio (PDF) of the image, represeted by Eq. (1). k px( Xk) = L 1 (1) X 1 ad p ( X ) = 1 k X k k = Where is the total umber of pixels i the iput image, k is the umber that X k appears i X. To obtai a histogram equalizatio fuctio, the cumulative distributio fuctio (CDF) of X k is calculated first from probability desity fuctio usig the followig Eq(). k sk = T( Xk) = px( Xk) () k j = Where k=,1,..,l-1 ad T(X L-1 )=1. S k is the value of the cumulative distributio fuctio at the k-th itesity level i the image. Sice a equalized image has a histogram havig a uiform distributio, the output image may also have a desity fuctio equally distributed over the etire rage. Global cotrast ehacemet scheme geerates differet effects depedig o the characteristics of origial images. I a complex image as show i Fig. 1(a), the bright ad dark areas are emphasized but there is o differece iside the idividual local areas havig similar brightess. I this case, the result image may be evaluated as the image quality is ehaced. However, those examples i Fig. 1 (b) ad (c) where a sigle object is located o simple backgroud shows that the global cotrast ehacemet techiques may deteriorate the image quality. A washed-out effect also appears sice global histogram equalizatio chages the mea brightess of iput image to the middle level. (b) Complex object ad simple backgroud (c) Simple object ad simple backgroud Fig. 1 Results of applyig the histogram equalizatio fuctio to the whole image: left is the origial image ad right is the result image If a image is cosistig of those areas havig similar brightess, the global histogram equalizatio scheme may geerate deteriorated result images havig poor quality. I order to solve this problem ad ehace cotrast i particular regio local histogram which divides a iput image ito sub-blocks havig the same size ad equalizes the brightess distributios of idividual sub-blocks is applied. As doe for the global equalizatio process, Eq.(3) geerates the cumulative distributio fuctio of X k of the i-th block first. k i i i sk = T( Xk) = px( X j) k i j = i L 1 i i Xk T XL 1 px X j 1, ( ) = ( ) = 1 I Eq. (3), i is the umber of pixels i i-th block ad i k represets the umber of pixels havig the k-th level i i-th block. Fig. shows the result image acquired by applyig a block-based histogram equalizatio scheme where the origial Lea's image havig 51x51 pixels is divided by small blocks havig 3x3 pixels. Fig. shows a extreme example of blockig effect geerally occurrig i a block-based cotrast ehacemet techique where the blocks are ot coected smoothly. To reduce this blockig effect, some approaches divide the blocks so that the boudary pixels of the eighborig blocks are overlapped [1-]. But blockig effect ad calculatio amout are depeded o block size ad overlappig amout. (3) (a) Complex object ad backgroud 177

3 World Academy of Sciece, Egieerig ad Techology 5 9 Fig. Result of applyig the histogram equalizatio schemes to idividual blocks differetly I order to miimize losig iformatio lose ad distortio amout ad ehace cotrast i low desity regio, sub-histogram equalizatios were proposed earlier [5-1]. These methods have i commo to preserve mea brightess but umber of sub-histogram regios ad threshold decisio algorithms used were differet. Firstly, BBHE [5] separates the iput histogram ito two parts based o its mea brightess, ad the equalizes the two sub-histograms idepedetly. Ad separatig poit x B is calculated by Eq(4),ad (5). 1 x = rp () r dr (4) P( z) z 1 B xb P r r dr x B = 1 1 Pr r dr xb 1 x xb B r ( ), z < x ( ), z < 1 It is foud that BBHE ca preserve the mea brightess, if the iput histogram has a symmetrical distributio aroud its mea. But every image does't have this property, mea brightess preservig was depeded o iput image. DSIHE[6] is very similar to BBHE, except that the separatig poit x D is selected as the media gray level of the iput image, i.e., x D satisfies Eq (6). x D P().5 r r dr = (6) Also RMSHE(Recursive Mea-Separate Histogram Equalizatio)[7], ad RSIHE(Recursive Sub-Image Histogram Equalizatio)[11] are recursive algorithms of BBHE ad DSIHE. These two recursive methods have improved results comparig with previous methods ad mea brightess was similar to iput image's but equalizatio effect was reduced. Ad they have aother weak poit that has ot flexible iterval of umber of regio ad also its decisio is maual. Ad they caot solve over-equalizatio effect problem i specific image that has high desity distributio arrow rage. As show i Fig. 3, the proposed algorithm aalyze whole histogram desity of the image ad divide histogram regios ad perform sub histogram equalizatio i divided regios fially merges these sub equalized images. B (5) Fig. 3 Block diagram of the proposed algorithm III. HISTOGRAM DENSITY ANALYSIS BASED ON HISTOGRAM DIVISION As a first step i the proposed algorithm, i order to split the etire image, histogram desity based aalysis method is explaied by referrig to Fig.4. Fig. 4 (a) shows image histogram that has rapid variatio ad high desity i specific histogram area. Whe GHE process through this image, the dark area aroud the camerama which has high-desity histogram has a wild iterval as Equalizatio processig. Ad the sky aroud the area also has a large gap ad icreases the cotrast as Equalizatio processig. (Referece Fig. 4(f)) However, as the histogram aalysis of equalized image's light area ad betwee dark area aroud the camerama ad groud area show the brightess iformatio was missed. Ad Fig. 4(g) shows more flat histogram distributio as local histogram equalizatio, but the light area's missig iformatio ad distortig pixel value were icreased. The image that represets over-equalizatio effect has several characteristics that histogram distributio chage quickly ad hardly focus o some histogram areas. Therefore the reaso of over-equalizatio ca be kow by aalyzig histogram CDF fuctio which is stadard of brightess value redistributio. Firstly, characteristics of PDF that has abrupt chage or high desity show CDF's chage is also abrupt. Therefore, the gap betwee redistributio pixels is also wild. As this iterval grows, parts that have see cotiuously with small Differeces are relocated with large iterval. This resulted image is looks discotiued ad damaged. As see i the equalizatio regio that has distributio ragig from high ad low, high distributio areas are relocated with wild itervals ad where as low distributio areas have broade distributio ad lost the focus. These characteristic let lose low distributio area's pixel iformatio ad damage extremely as growig differece betwee high ad low distributio histogram area. Therefore to ehace low distributio histogram area's cotrast ad prevet these distortio effects as equalizatio, some sub-histogram area's equalizatio have to be performed idepedetly. Ad these areas iclude low ad high distributio areas ad rapid chage distributio area. 178

4 World Academy of Sciece, Egieerig ad Techology 5 9 (a) Origial Iput Image (a) Histogram of iput image (b) Histogram of whole image (c) Histogram of ma regio (b) Desity fuctio Fig. 5 (b) Desity fuctio acquired from (a) the example histogram usig a widow size of 6 (d) Histogram of groud regio (e) Histogram of sky regio B. Regio Separatio Based o Desity To divide regio which has differet desity calculated desity fuctio, regio which has similar desity was grouped oe regio with cosiderig eighborhood desity fuctio. High desity area was mai regio because high desity area cause over-equalizatio effects ad rest area was assumed low desity area. Firstly, calculated desity fuctio was sorted descedig order because high desity area is cause over-equalizatio effects. (f) After global equalizatio (g) After local equalizatio Fig. 4 Histogram of the objects i the origial image as show i (a) A. Desity calculatio of histogram regio To divide low ad high desity distributio area, sum of distributio quatity was used based o eighborhood histogram's distributio as defie Eq. (7). Each itesity's desity referece eighborhood brightess based o defied widow size. By referrig to the eighborhood pixel iformatio, histogram desity which ca divide low ad high desity histogram area was calculated without error which caused sudde chage. w + D = f( x ) (7) w i= i above eq(7) represet brightess value, ad ormalized w w brightess value ca have a value of ~ Ad f(x i ) represets frequecy of x i th brightess. For example Fig. 5 shows histogram of iput image (a) ad desity fuctio (b) that was acquired from sum of eighborhood histogram with widow size of 6. Desity fuctio was refereced sixth histogram ad flatted tha origial histogram. i Fig. 6 Labelig the itesity i a descedig order Next, to make group of high desity area, group area was exteded from the ceter poit of the group. Basic widow size was ot suitable to every desity fuctio, so referece histogram rage has to chage. To adjust rage of regio, compared betwee average chage ad mometary chage amout as icreasig widow size(i). If mometary chage amout large tha average chage rate widow size was icreased. Coversely, if its mometary chage rate less tha average chage rate rage extesio operatio was suspeded. Fig. 7 Flowchart of the desity regio expadig algorithm 179

5 World Academy of Sciece, Egieerig ad Techology 5 9 Fially rage extesio was performed based o desity order ad while modifyig the rage of ext order regio desity, the overlappig of two areas are checked. Most of similar desity histogram exists i eighborhood area. Like this overlapped area represet same area i histogram desity based regio groupig, so oly middle value was used. Fig. 8 shows o overlapped regio separatio result with high desity order. Fially, meetig poit betwee approximated Gaussia fuctios is used boudary value for sub histogram equalizatio. Areas are divided as show i Fig. 1 i approximated result as Fig. 9. Fig. 1(a) represet fouded regio i approximated Gaussia regio, ad Fig. 1(b) represet applied fidig regio iformatio i iput image's histogram. Fig. 8 Regios acquired as the result of applyig the desity regio expadig algorithm (a) regio adjustmet usig (b) its applicatio to regio approximated Gaussia fuctios adjustmet i the histogram Fig. 1 Fially obtaied sub-histogram regios IV. EQUALIZATION OF DIVIDED SUB-HISTOGRAM A. Determie rage of sub-histogram Histogram regio was divided by similar desity regios ad the real rage of equalizatio eed to be determied. For example Fig. 8 represets three regios which have high desity, but actually there exist seve regios because aother four regios which were exists i betwee three regios. These four regios which are placed betwee high desity regios ca prevet over-equalizatio effect. But these low desity regio ca be removed without distorted result accordig to aroud high desity regio's chage ad distributio. Ad calculated equalizatio value ad memory size are reduced by reducig umber of uecessary regio. This paper usig Gaussia fuctio approximatio which defies Eq. (9) with regio's mea ad variatio which defied Eq. (8) for remove uecessary regios ad redefie ew regio rage. σ E μ x= S = ( x ) f '( x) (8) 1 ( x μ ) g ( x) = exp( ) (9) σ π σ B. Merge of Cotrast Improved Sub Images Fig. 11 shows each sub equalized image. Sub histogram equalizatio was performed by each regio's brightess values ad each regio's CDF fuctio. Ad each sub histogram equalized image was merged i fial result image. Fig. 11 shows three divided histogram(b), ad each (c),(d),(e) images show sub equalized result which each use uormalized brightess rage of ~46, 47~195, 196~55. Fially result image is obtaied by mergig of (c),(d),ad (e) image. (a) Origial image (b) Segmeted histogram σ i above Eq. (8) represet variatio of selected regio, μ is ceter value of selected regio ad Gaussia distributio was geerated with ceter value μ. Whe two equatios are applied, Gaussia approximatio fuctio ca be obtaied like Fig.9. (c) Result images of regioal histogram equalizatio(~46) (d) Result images of regioal histogram equalizatio(47~195) Fig. 9 The desity fuctio approximated by Gaussia fuctio takig ito accout the selected regios 18

6 World Academy of Sciece, Egieerig ad Techology 5 9 (e) Result images of regioal histogram equalizatio(196~55) (f) Fial result image (a) Origial image (b) Edge image of (a) Fig. 11 Merge of cotrast improved sub images V. EXPERIMENTS The proposed algorithm has bee implemeted usig VC++ i a Petium IV PC havig Widows XP operatig system. Experimets with the local ad global image characteristics of images which are airplae with sky backgroud image ad complex buildig images are performed. I order to evaluate the proposed method previous algorithms image results ad proposed algorithm's result image are compared by represeted features of object's edge ad histogram results. Iitially, previous ad proposed cotrast ehacemet algorithm s extractig edge performace evaluatios are compared usig cay edge detector. Fig. 1 shows result image of comparig edge detectio characteristic with origial image, GHE, RSHIE ad proposed algorithm applyig image. Fig. 1(a) shows a letter writte blurry i the plae as origial iput image which is hard to classify because the cotrast level i plae is low. Fig. 1(b) shows edge characteristic i origial iput image. There oly exist betwee plae ad sky regio without edge of ier plae. Global equalizatio result image i Fig. 1(c) shows as over-equalizatio effect i relatively dark area with ier plai ad bright area with sky regio. Therefore cotrast i GHE result image is gettig poor as ier plae regio was darker ad sky regio was brighter. As result iside of the plae was chaged similar brightess, as show i Fig. 1(d) edge property was disappeared ad icreased rig oise effect which exist i boud betwee plae ad sky tha origial image. Fig. 1(e) shows result image which is applied RSHIE algorithm. Number of RSHIE algorithm's iteratio is ad 4 regios are divided. Over-equalizatio is ot preveted perfectly so blockig effect appears i right sky regio. Ad the edge feature is improved tha origial iput image but the fie feature of ier plae was ot detected as show Fig. 1. Ehaced result image of proposed algorithm i Fig. 1(g) ca do to prevet over-equalizatio as split high desity regio which is a reaso of over-equalizatio ad perform sub equalizatio. Ad also as show Fig. 1(h) detail edge feature of ier plae regio is appeared as flattig low desity regio. Comparig detailed part ad letter writte i the ier plae regio, these features was well represeted i result image which is applied proposed method. (c) Result of global histogram equalizatio (e) Result of RSHIE (d) Edge image of (c) (f) Edge image of (e) (g) Result of proposed method (h) Edge image of (g) Fig. 1 Performace compariso of the proposed algorithm with the covetioal approaches based o edge features Next, proposed algorithm ad previous cotrast ehacemet algorithm are measured by aalyzig of histogram. Experimetal image which is costructed perso ad backgroud is show Fig. 13(a), ad its histogram is represeted i Fig. 13(b). Most of itesity level of backgroud regio is placed aroud 14 level. Whe performig GHE of image which has these property, histogram of uimportat backgroud regio which is takes most part will redistributed across wider regio. Resultat over-equalized backgroud regio make may oise. Also perso regio's cotrast is reduced ad average itesity became lower. Fig. 13(f) show result image which is applied RSHIE algorithm. Number of 181

7 World Academy of Sciece, Egieerig ad Techology 5 9 RSHIE algorithm's iteratio is ad 4 regios are divided. But it caot get large cotrast ehaced effect i the ier perso regio as comparig origial iput image. (a) Origial image (c) Result of global histogram equalizatio (e) Result of RSHIE (g) Result of proposed method (b) Histogram image of (a) (d) Histogram image of (c) (f) Histogram image of (e) (h) Histogram image of (g) Fig. 13 Performace compariso of the proposed algorithm with the covetioal approaches based o histogram aalysis Whe performig 5 sub histograms equalizatio, regio which has low desity ad low cotrast level get fie result which is emphasized ad had high cotrast. These results ca be obtaied by makig oe regio of backgroud regio. Ad this backgroud regio's over-equalizatio is preveted ad other regio's equalizatio is performed evely. which has arrow ad high desity histogram distributio is distortio of regio which has low desity ad wide distributio. To solve these problems, iput images histogram regio was divided based o histogram desity ad sub histogram equalizatio was performed by each regio equalizatio fuctio. Proposed algorithm divides may areas for small equalizatio effect i histogram regio which has arrow ad high desities. Ad other regio's histogram equalizatio is performed ormally. As a result proposed algorithm's superiority ca be proved by improvig feature which has low desity ad broadly distributed without distortio. Our future work ivolves extedig the proposed method to variety of electroic appliaces where cotrast improvemet is ecessary. REFERENCES [1] A. K. Jai, Fudametals of Digital Image Processig, Pretice-Hall, [] J. Y. Kim, L. S. Kim, S. H. Hwag, A Advaced Cotrast Ehacemet Usig Partially Overlapped Sub-Block Histogram Equalizatio, IEEE Trasactios o Circuits ad Systems for Video Techology, Vol. 11, No. 4, pp , 1. [3] C. C. Su. S. J. Rua, M. C. Shie, T. W. Pai, Dyamic Cotrast Ehacemet based o Histogram Specificatio, IEEE Trasactios o Cosumer Electroics, 51(4), pp , 5. [4] J. A. Stark, Adaptive Image Cotrast Ehacemet Usig Geeralizatios of Histogram Equalizatio, IEEE Trasactios o Image Processig, 9(5), pp ,. [5] Y. T. Kim, "Cotrast Ehacemet Usig Brightess Preservig Bi-Histogram Equalizatio," IEEE Trasactios o Cosumer Electroics, 43(1), pp.1-8, [6] Y. Wa, Q. Che, B.-M. Zhag, "Image ehacemet based o equal area dualistic sub-image histogram equalizatio method", IEEE Tras. Cosum. Electro. 45 (1) (1999) [7] S. D. Che, A. Rahma Ramli, Cotrast Ehacemet usig Recursive Mea-Separate Histogram Equalizatio for Scalable Brightess Preservatio, IEEE Trasactios o Cosumer Electroics, 49(4), pp , 3. [8] Soog-Der Che, A. Rahma Ramli, Preservig brightess i histogram equalizatio based cotrast ehacemet techiques, Digital Sigal Processig, 1(5), pp , September 4. [9] Chao Wag ad Zhogfu Ye, Brightess Preservig Histogram Equalizatio with Maximum Etropy: A Variatioal Perspective, IEEE Trasactios o Cosumer Electroics, 51(4), pp , 5. [1] K. S. Sim, C. P. Tso, ad Y. Y. Ta, "Recursive sub-image histogram equalizatio applied to gray scale images ", Patter Recogitio Letters, 8(1), pp , 7. [11] Z. Che, B. R. Abidi. D. L. Page, M. A. Abidi, Gray-Level Groupig (GLG): A Automatic Method for Optimized Image Cotrast Ehacemet-Part I : The Basic Method, IEEE Trasactios o Image Processig, 15(8), pp.9-3, 6. [1] S. M. Pizer, E. P. Ambur, J. D. Austi, R. Cromartie, A. Geselowitz, T. Greer, B. H. Romey, J. B. Zimmerma, K. Zuiderveld, Adaptive Histogram Equalizatio ad Its Variatios, Computer Visio Graphics ad Image Processig, Vol. 39, pp , [13] F. Lamberti, B. Motrucchio, A. Saa, CMBFHE_a ovel cotrast ehacemet techique based o cascaded multistep biomial filterig histogram equalizatio, IEEE Trasactios o Cosumer Electroics, 5(3), pp , 6. [14] Z. Q. Wu, J. A. Ware, I. D. Wilso, J. Zhag, Mechaism aalysis of highly overlapped iterpolatio cotrast ehacemet, IEEE Proceedigs Visio, Image & Sigal Processig, 153(4), pp.51-5, 6. VI. CONCLUSION This paper proposes a ew cotrast ehacemet techique that performs sub histogram equalizatio based o histogram desity. The most importat problem of GHE i some image 18