Image Contrast Enhancement Using Histogram Modification Technique

Transcription

1 Image Cotrast Ehacemet Usig Histogram Modificatio Techique Shekhar R. Suralkar 1, Atul H. Karode 2, Maali S. Rathi 3 1 Asso. Prof. (E&TC Dept.), SSBT s COET, Jalgao, Idia 2 Asst. Prof. (E&TC Dept.), SSBT s COET, Jalgao, Idia 3 Studet, M.E. E&TC (Digital Electroics), SSBT s COET, Jalgao, Idia Abstract Image cotrast ehacemet plays a vital role i digital image processig especially i biomedical applicatios ad secures digital image trasmissio. The objective of ay image ehacemet techique is to improve the characteristics or quality of a image, such that the resultig image is better tha the origial image. To improve the image cotrast, umerous ehacemet techiques have bee proposed. Oe of the covetioal methods adopted is the Histogram Equalizatio (HE) techique. Histogram Equalizatio (HE) has proved to be a simple ad effective image cotrast ehacemet techique. However, it teds to chage the mea brightess of the image to the middle level of the gray-level rage, which is ot desirable i may applicatios. Thus, HE has limitatios sice preservig the iput brightess of the image is required to avoid the geeratio of o-existig artifacts i the output image. To surmout this drawback, Bi-HE methods for brightess preservig ad cotrast ehacemet have bee proposed. Although these methods preserve the iput brightess o the output image with a sigificat cotrast ehacemet, they may produce images which do ot look as atural as the origial images. I order to overcome this drawback, this work proposes a low-complexity histogram modificatio algorithm for cotrast ehacemet ad thereby itroducig specifically desiged terms, the level of cotrast ehacemet ca be adjusted ad thus oise robustess, white/black stretchig ad mea brightess preservatio may easily be icorporated ito the optimizatio. Keywords Cotrast Ehacemet, Histogram Equalizatio, Histogram Modificatio 1. Itroductio Cotrast ehacemet is oe of the most importat issues of image processig ad applicatios. It is believed that it is a fudametal step i Image Segmetatio [1]. Cotrast ehacemet plays a crucial role i image processig applicatios, such as digital photography, medical image aalysis, remote sesig, LCD display processig, ad scietific visualizatio. There are several reasos for a image/video to have poor cotrast: the poor quality of the used imagig device, lack of expertise of the operator, ad the adverse exteral coditios at the time of acquisitio. These effects result i uder-utilizatio of the offered dyamic rage. As a result, such images ad videos may ot reveal all the details i the captured scee, ad may have a washed-out ad uatural look. Cotrast ehacemet targets to elimiate these problems, thereby to obtai a more visually-pleasig or iformative image or both. Typical viewers describe the ehaced images as if a curtai of fog has bee removed from the picture. Image ehacemet techiques ca be broadly categorized ito two groups: direct methods ad idirect methods. Direct methods defie a cotrast measure ad try to improve it. Idirect methods, o the other had, improve the cotrast through exploitig the uder-utilized regios of the dyamic rage without defiig a specific cotrast term. Most methods i the literature fall ito the secod group. Idirect methods ca further be divided ito several subgroups: i) techiques that decompose a image ito high ad low frequecy sigals for maipulatio, e.g., homomorphic filterig, ii) histogram modificatio techiques, ad iii) trasform-based techiques. Out of these three subgroups, the techiques based o histogram of the image received the most attetio due to its straightforward ad ituitive implemetatio qualities. The most commo way to improve the cotrast of a image is to modify its pixel value distributio, or histogram. Cotrast ehacemet techiques i the secod subgroup modify the image through some pixel mappig such that the histogram of the processed image is more spread tha that of the origial image. Techiques i this subgroup either ehace the cotrast globally or locally. If a sigle mappig derived from the image is used the it is a global method; if the eighborhood of each pixel is used to obtai a local mappig fuctio the it is a local method. Usig a sigle global mappig caot (specifically) ehace the local cotrast. Oe of the most popular GCE techiques is histogram equalizatio (HE). HE is widely used for cotrast ehacemet i a variety of applicatios due to its simple fuctio ad effectiveess. However, it teds to chage the brightess of a image ad hece, ot suitable for cosumer electroic products, where preservig the origial brightess is essetial to avoid aoyig artifacts. I additio, HE method teds to itroduce ueccessary visual deterioratio icludig saturatio effect. Preservig the iput brightess of a image ad keepig PSNR i the desired rage are required to avoid the geeratio of o-existig artifacts i the output image. Various methods have bee proposed for limitig the level of ehacemet, most of which are obtaied through modificatios o HE. For example, Mea preservig Bi- Histogram Equalizatio (BBHE) was proposed to overcome the brightess preservatio problems. BBHE seperates the iput image/video histogram ito two parts based o iput mea before equalizig them idepedetly [2]. Later, Equal Area Dualistic Sub-Image Histogram Equalizatio (DSIHE) was proposed to separate the histogram by etropy value [8]. Agai, a extesio of BBHE which was referred to as Miimum Mea Brightess Error Bi-Histogram Equalizatio 1

2 (MMBEBHE) was proposed to provide maximal brightess preservatio [3]. Although they are visually more pleasig tha HE, these techiques caot adjust the level of ehacemet ad are ot robust to oise, which may become a problem whe the histogram has spikes. Also, it should be oted that preservig the brightess does ot imply preservatio of aturaless. Oe recet method proposed by Wag ad Ward [4] suggests modifyig the image histogram by weightig ad thresholdig before histogram equalizatio. The weightig ad thresholdig is performed by clampig the origial histogram at a upper threshold P u ad at a lower threshold P l, ad trasformig all the values betwee the upper ad lower thresholds usig a ormalized power law fuctio with idex r > 0. There are also ucovetioal approaches to the histogram based cotrast ehacemet problem. Gray-level groupig (GLG) [11] is such a algorithm that groups histogram bis ad the redistributes these groups iteratively. Although GLG ca adjust the level of ehacemet ad is robust to histogram spikes, it is maily desiged for still images. Sice gray-level groupig makes hard decisios o groupig histogram bis, ad redistributig the bis depeds o the groupig, mea brightess itesity i a image sequece ca abruptly chage i the same scee. This causes flickerig, which is oe of the most aoyig problems i video ehacemet. Although a fast versio of the algorithm is available, GLG s computatioal complexity is high for most applicatios. Cotrast ehacemet techiques i the first ad third subgroups ofte use multiscale aalysis to decompose the image ito differet bads ad ehace desired global ad local frequecies. These techiques are computatioally complex but eable global ad local cotrast ehacemet at the same time by ehacig the appropriate scales. The aforemetioed cotrast ehacemet techiques perform well o some images but they ca create problems whe a sequece of images is ehaced, or whe the histogram has spikes, or whe a atural lookig ehaced image is strictly required. I additio, computatioal complexity ad cotrollability become a importat issue whe the goal is to desig a cotrast ehacemet algorithm for cosumer products. I order to achieve all these requiremets, this work presets a histogram modificatio based method which aims to better preserve the image quality, preserves better cotrast ad eriches the image details. I summary, our goal i this paper is to obtai a visually pleasig ehacemet method that has lowcomputatioal complexity ad ca be easily implemeted o FPGAs or ASICs ad works well with both video ad still images. Sectio II itroduces the details about the cotrast ehacemet ad mappig fuctios. Sectio III describes the covetioal histogram equalizatio techique. Sectio IV gives the details of the proposed method. I Sectio V, experimetal results ad measures are provided. Fially, Sectio VI cotais cocludig remarks. 2. Cotrast Stretchig Mappig Fuctio Histogram-based cotrast ehacemet techiques utilize the image histogram to obtai a sigle-idexed mappig T[] to modify the pixel values. I HE ad other histogram-based methods, mappig fuctio is obtaied from the histogram or the modified histogram, respectively [5]. HE fids a mappig to obtai a image with a histogram that is as close as possible to a uiform distributio to fully exploit the dyamic rage. A histogram, h[], ca be regarded as a uormalized discrete probability mass fuctio of the pixel itesities. The ormalized histogram p[] of a image gives the approximate probability desity fuctio (PDF) of its pixel itesities. The, the approximate cumulative distributio fuctio (CDF), c[], is obtaied from p[]. The mappig fuctio is a scaled versio of this CDF. HE uses the image histogram to obtai the mappig fuctio; whereas, other histogram-based methods obtai the mappig fuctio via the modified histogram. The mappig fuctio i the discrete form is give as B T[ ] = (2 1) p[ j] (1) j= 0 where B is the umber of bits used to represet the pixel values, ad Є [0, 2 B - 1]. Although the histogram of the processed image will be as uiform as possible, it may ot be exactly uiform because of the discrete ature of the pixel itesities [6]. It is also possible to ehace the cotrast without usig the histogram. Black stretchig ad white stretchig are simple but effective techiques used i cosumer-grade TV sets. Black stretchig makes dark pixels darker, while white stretchig makes bright pixels brighter. This produces more atural lookig black ad white regios; hece, it ehaces the cotrast of the image. Liear black ad white stretchig ca be achieved by the mappig sb, b T[ ] = g[ ], b < < w (2) w + ( w) sw, w where b is the maximum gray-level to be stretched to black ad w is the miimum gray-level to be stretched to white, g[] is ay fuctio mappig the itesities i betwee, ad s b, s w are black ad white stretchig factors both of which are less tha oe. 3. Histogram Equalizatio Techique The mai objective of histogram equalizatio [5], [9] is to obtai a uiform histogram for the output image [15]. The procedure to perform histogram equalizatio techiques is as follows: 1. Fid the ruig sum of the histogram values. 2. Normalize the ruig sum values by dividig by the total umber of pixels. 3. Multiply the ormalized values by maximum gray level ad roud off to the earest iteger value. 4. Map the gray level values usig a oe to oe mappig. For a give digital image G (i,j) with M pixels ad a gray level of [0,L-1].The probability desity fuctio (PDF) is give by u P( k ) = for k = 0,1, K, K,L 1 (3) M where u is the umber of times level k appears i a image. The ruig sum or cumulative desity fuctio (CDF) of histogram values is obtaied by k C( k ) = P( m) for k = 0,1, K, K,L 1 (4) m= 0 2

3 Usig the ruig sum, histogram equalizatio plots a iput level k ito a output level k. k = ( L 1) C( k ) (5) The output level ca be icremet by k = ( L 1) P( k ) (6) The above equatio idicates that distace betwee k ad k + 1has direct relatio with PDF of the iput image at the gray level k. I the Histogram Equalizatio, the pixel values are uiformly distributed with gray level values across their rage. HE over ehaces the backgroud of the image ad ofte produces the improbable effects i the image. A example of histogram equalizatio is show i fig. 1. The origial image ad the equalized image ad their correspodig histograms are show. It ca be clearly see that the equalizatio results i the uiform spreadig of the bis i the origial histogram ad though ehaced, the equalized image may or may ot be as atural lookig as the origial image. Figure 1. Example of HE : The origial image ad its histogram, ad the equalized versios. 4. Proposed Method I this paper, we are implemetig the cotrast ehacemet method proposed i [6]. Here i this sectio, a low-complexity histogram modificatio algorithm is preseted which deals with histogram spikes, performs B&W stretchig, ad adjusts the level of ehacemet adaptively so that the dyamic rage is better utilized while hadlig the oise visibility ad the atural look requiremets Histogram Computatio To deal with histogram spikes i a simple way, istead of smoothig or weightig the iput histogram, oe ca chage the way a histogram is computed. Histogram spikes are created because of a large umber of pixels that have the same gray-level ad these pixels almost always come from smooth areas i the iput image whe they create artifacts/oise i the ehaced image. Hece, histogram computatio ca be modified so as to take pixels that have some level of cotrast with their eighbors ito accout, which will solve the histogram spike problem at the very begiig. It is also possible to relate this practical approach with optimizatio based solutios as follows: For a successful cotrast ehacemet, the histogram should be modified i such a way that the modified histogram, h ~, represets the coditioal probability of a pixel, give that it has a cotrast ~ with its eighbors (deoted by C ). That is, h[ i ] = p[ i C], where p[i C] deotes the probability of a pixel havig graylevel i give the evet C. Performig histogram equalizatio o h ~ rather tha h will ehace the cotrast but ot the oise, sice the former will oly utilize the dyamic rage for pixels that have some level of cotrast with their eighbors. Notig that the histogram modificatio methods preseted i the previous sectio also aim to icrease cotrast but ot the oise visibility, they must modify the histogram i such a way that the modified histogram resembles p [ i C] rather tha p [i]. However, oe ca simply obtai p [ i C] by coutig oly those pixels that have cotrast, rather tha solvig complex optimizatio problems, which i essece correspods to dealig with histogram spikes resultig from smooth area (ocotrast) pixels after computig the histogram i the covetioal way. To obtai the histogram, the local variatio of each pixel ca be used to decide if a pixel has sufficiet cotrast with its eighbors. Oe efficiet way of achievig this for hardware simplicity is to use a horizotal variatio measure by takig advatage of the row-wise pixel processig architecture, which is available i commo video processig hardware platforms. A horizotal oe-lagged differece operatio is a high-pass filter, which will also measure oise. O the other had, a horizotal two-lagged differece operatio is a badpass filter which will atteuate high-frequecy oise sigals. Histogram is created usig pixels with a two-lagged differece that has a magitude larger tha a give threshold (steps 5, 6, 7). The umber of pixels icluded i the histogram is also couted for proper ormalizatio Adjustig the Level of Ehacemet It is possible to adjust the level of histogram equalizatio to achieve atural lookig ehaced images. The modified histogram is a weighted average of the iput histogram ad the uiform histogram, as give i (7). ~ h i + λu 1 λ h = = h i + u (7) 1 + λ 1 + λ 1 + λ The cotributio of the iput histogram i the modified histogram is * κ = 1/(1 + λ) (8) The level of histogram equalizatio should be adjusted depedig o the iput image s cotrast. Low cotrast images have arrow histograms ad with histogram equalizatio, cotourig ad oise ca be created. Therefore, k is computed to measure the iput cotrast usig the aggregated outputs of horizotal two-lagged differece operatio (step 4). Afterwards, k is multiplied by a usercotrolled parameter g, the gk is ormalized to the rage [0, 1] (step 11) to get k *. It is a good practice to limit the maximum cotributio of a histogram, sice this will help with the worst-case artifacts created due to histogram equalizatio. By choosig the maximum value that gk ca take o as a power of two, the ormalizatio step ca be doe usig a bit-shift operatio rather tha a costly divisio. To esure that h i ad u have the same ormalizatio, u is obtaied usig the umber of pixels that are icluded i the histogram (step 12). u mi is used to esure that very low bi regios of the histogram will ot result i very low slope i the mappig fuctio; it will icrease the slope i these regios, resultig i icreased-utilizatio of dyamic rage. 3

4 B&W stretchig is performed usig (9) (step 17) ~ B 1 h = ((1 + λ) I + α I ) ( hi + λu) (9) where I B is a diagoal matrix, I B (i, i) = 1 for i {[0, b] ᴗ [w, 255]} ad the remaiig diagoal elemets are zero. Parameters b, w ad α ca be adapted with the image cotet. b ad w is usually derived from the histogram as the miimum ad maximum itesities. For oise robustess, b should be chose as the miimum gray-level that is bigger tha some predefied umber of pixels itesities, w ca be chose similarly. It is a good practice to impose limits o b ad w. The stretchig parameter should also be adapted with image cotet. For dark images white stretchig ca be favored, while for bright images black stretchig ca be favoured. α may also deped o the iput image s cotrast. ad maximum itesity values i each block, ad averagig them. The MSE value is give by 1 1 i j 1 ( x( m, ) y( i, j) ) MSE = (11) ij m= 0 = 0 where the size of a image is i x j. The origial ad restored images are x(m, ) ad y(i, j). The PSNR is the ratio betwee the maximum possible power of a sigal ad the power of corrupted oise. PSNR = 20 x log 10 (255/sqrt(MSE)) (12) 2 Proposed Histogram Modificatio Algorithm [6] Step 1 : Read iput image, B&W stretch parameters b, w, ad 1/(1 + α), level of ehacemet value g Step 2 : Calculate the histogram usig pixels with a two - lagged differece that has a magitude larger tha a give threshold Step 3 : Aggregate the two-lagged operatio as κ Step 4 : Cout the umber of pixels icluded i the histogram for proper ormalizatio Step 5 : Calculate g, κ which are the cotrol parameters Step 6 : Perform the ormalizatio usig a bit-shift operatio Step 7 : Calculate umi which will icrease the slope i the mappig fuctio, resultig i icreased utilizatio of dyamic rage Step 8 : Obtai u usig u mi ad the umber of pixels that are icluded i the histogram u mi{ cout / 256, umi} Step 9 : Perform cotrast stretchig 5. Experimetal Results ad Discussios I order to test the proposed method, experimets are performed o various images regardless of the image size. The iput image ca be of ay size. Observatio of all these images show that the proposed method outperform well compared to the histogram equalizatio. However, it is usually desired to have some quatitative measures i additio to subjective assessmet i.e. visual iterpretatio. Therefore, to evaluate the image ehacemet ad restoratio performace, the quatitative measures such as Absolute Mea Brightess Error (AMBE), the Discrete Etropy (H), Measure of Ehacemet (EME), Mea Squared Error (MSE) ad Peak Sigal to Noise Ratio (PSNR) are used as the criterio. AMBE is defied as the absolute differece betwee iput ad output mea values. AMBE = E(X) - E (Y) (10) The discrete etropy H is used to measure the cotet of a image, where a higher value idicates a image with richer details. The measure of ehacemet (EME) approximates a average cotrast i the image by dividig image ito ooverlappig blocks, fidig a measure based o miimum Figure 2. Results for image Lady with Hat. Origial Image, Ehaced image obtaied usig HE, Ehaced image obtaied usig the proposed algorithm, Mappigs for ehaced images : blue lie idicates HE mappig, red dotted lie idicates the proposed method ad the dotted lie idicates o chage mappig. The performace of the proposed method is compared with the covetioal histogram equalizatio techique. The proposed algorithm has bee successfully tested o a variety of test images. Oly, a few of the results are show i this paper. Fig. 2, 3, 4, 5, 6 shows the origial image ad their correspodig cotrast ehaced image usig HE ad proposed method. The proposed algorithm is compared with the covetioal histogram equalizatio. I the histogram equalizatio, the pixels are spreadig uiformly. Usually Histogram Equalized resultat images have uatural look that meas HE over ehaces the images. The proposed method reduces the effect caused by the histogram equalizatio. Whe the cotrast of a image is ehaced, as a upshot the oise iduced i the image will be high ad the image looks uatural. Thus, the proposed techique ca be effectively used to improve the cotrast of the images ad filter out the oises i the image ad preserves the edges. 4

5 Figure 3. Results for image Livig Room. Origial Image, Ehaced image obtaied usig HE, Ehaced image obtaied usig the proposed algorithm, Mappigs for ehaced images : blue lie idicates HE mappig, red dotted lie idicates the proposed method ad the dotted lie idicates o chage mappig. Figure 4. Results for image Lea. Origial Image, Ehaced image obtaied usig HE, Ehaced image obtaied usig the proposed algorithm, Mappigs for ehaced images : blue lie idicates HE mappig, red dotted lie idicates the proposed method ad the dotted lie idicates o chage mappig. 5

6 Figure 5. Results for image Hallowee. Origial Image, Ehaced image obtaied usig HE, Ehaced image obtaied usig the proposed algorithm, Mappigs for ehaced images : blue lie idicates HE mappig, red dotted lie idicates the proposed method ad the dotted lie idicates o chage mappig. Figure 6. Results for image Lady. Origial Image, Ehaced image obtaied usig HE, Ehaced image obtaied usig the proposed algorithm, Mappigs for ehaced images : blue lie idicates HE mappig, red dotted lie idicates the proposed method ad the dotted lie idicates o chage mappig. 6

7 Table 1. Quatitative Measuremet Results Image AMBE H EME MSE PSNR HE Prop. Orig. HE Prop. Orig. HE Prop. HE Prop. HE Prop. Lady with Hat Livig Room Lea Hallowee New York Average Computed quatitative measures AMBE, H, EME, MSE ad PSNR listed i Table I supplemet the visual assessmet. Compariso of AMBE values shows that the proposed method outperforms the HE method i all the images, which idicates better preservatio of mea brightess i the image. Compariso of H values show that the proposed method agai outperform HE. Normally, oe would expect HE to give higher discrete etropy value as HE results i more uiform histogram distributio. However, HE results i bi groupig ad this decreases the value. Compariso of EME values show that HE outperforms the proposed method; HE gives higher EME values tha the proposed method. Sice EME measures a form of cotrast, it is o surprise that HE gives the highest value eve though it does ot produce the most visually pleasig image. Also, from PSNR ad MSE values, we ca coclude that these values are better tha the HE method, idicatig less oise i the ehaced image. 6. Coclusio I this paper, we propose a ovel approach of image ehacemet techique usig histogram modificatio algorithm. The proposed framework employs carefully desiged pealty terms to adjust the various aspects of cotrast ehacemet. Hece, the cotrast of the image/video ca be improved without itroducig visual artifacts that decrease the visual quality of a image ad cause it to have a uatural look. To obtai a real-time implemetable algorithm, the proposed method avoids complex calculatios ad memory-badwidth cosumig operatios. The experimetal results show the effectiveess of the algorithm i compariso to covetioal cotrast ehacemet algorithms. Obtaied images are visually pleasig, artifact free, ad atural lookig. The proposed method is applicable to a wide variety of images ad video sequeces. It also offers a level of cotrollability ad adaptivity through which differet levels of cotrast ehacemet, from histogram equalizatio to o cotrast ehacemet, ca be achieved. [4] Q. Wag ad R. K. Ward, Fast image/video cotrast ehacemet based o weighted thresholded histogram equalizatio, IEEE Tras. Cosum. Electro., Vol. 53, No. 2, Pages , May [5] R. C. Gozalez ad R. E.Woods, Digital Image Processig, 2d ed. Eglewood Cliffs, NJ: Pretice-Hall, [6] Tarik Arici, Salih Dikbas, ad Yucel Altubasak, A Histogram Modificatio Framework ad Its Applicatio for Image Cotrast Ehacemet, IEEE Trasactios O Image Processig, Vol. 18, No. 9, pages , September 2009 [7] William K. Pratt, Digital Image Processig, 4 th ed., Wiley Publicatio, 2007 [8] G. de Haa, i Video Processig for Multimedia Systems, Eidhove, The Netherlads, 2000 [9] Mapreet Kaur, Jasdeep Kaur, Jappreet Kaur, Survey of Cotrat Ehacemet Techiques based o Histogram Equalizatio, Iteratioal Joural of Advaced Computer Sciece ad Applicatios, Vol. No. 2, Issue No. 7, Pages , [10] A. Polesel, G. Rampoi, ad V. Mathews, Image ehacemet via adaptive usharp maskig, IEEE Tras. Image Process., vol. 9, o. 3, pp , Mar [11] Z. Y. Che, B. R. Abidi, D. L. Page, ad M. A. Abidi, Gray-level groupig (GLG): A automatic method for optimized image cotrast ehacemet Part I: The basic method, IEEE Tras. Image Process., vol. 15, o. 8, pp , Aug [12] J. A. Stark, Adaptive image cotrast ehacemet usig geeralizatios of histogram equalizatio, IEEE Tras. Image Process., vol. 9, o. 5, pp , May [13] Viay Kumar, Himai Basal, Performace Evaluatio of Cotrast Ehacemet Techiques for Digital Images. Iteratioal Joural of Cosumer Scieces ad Techologies, Vol. 2, Issue 1, March 2011 [14] Ail K. Jai, Fudametals of Digital Image Processig, PHI Publicatio, 2005 [15] Hyusup Yoo, Yougjoo Ha, ad Hersoo Hah, Image Cotrast Ehacemet based Sub-histogram Equalizatio Techique without Over-equalizatio Noise. World Academy of Sciece, Egieerig ad Techology, Refereces [1] H. D Cheg, Huijua Xu, A Novel Fuzzy Logic Approach to Cotrast Ehacemet, Patter Recogit., Vol. 33, No. 5, Pages , May [2] Hojat Yegaeh, Ali Ziaei, Amirhossei Rezaie, A Novel Approach for Cotrast Ehacemet based o Histogram Equalizatio, Proceedigs of It. Cof. o Computer ad Commuicatio Egieerig, [3] David Meotti, Lauret Najma, Jacques Faco, ad Araldo de A. Araújo. Multi-Histogram Equalizatio Methods for Cotrast Ehacemet ad Brightess Preservig, IEEE Trasactios o Cosumer Electroics, Vol. 53, No. 3, pages , August