Contributed PaperManuscript received November 18, 2009Current version published 06 29 2010;Electronic version published 07...
that is quite computationally expensive. Moreover, AHEmay yield unsatisfactory outputs, e.g., images with noiseartefacts a...
where all symbols are as defined in equations (1) to (3). Thecumulative histogram for each RGB component andluminance comp...
The discrete approximation of the continuous pixelbrightness transformation from Equation (18) is, therefore,given by000 0...
andYhigh = min{Yhigh-red, Yhigh-green , Yhigh-red }, (23)whereYhigh-red = min{y|( ) ( ( ) 255) ( (255) (255) )Y R R red ol...
a) ` b) c)d) e) f)Fig. 1. The enhancement results for test image Mountain, a) original image, b) output of the proposed ap...
a) b) c)d) e) f)Fig. 2. The enhancement results for test image Meat, a) original image, b) output of the proposed approach...
a) b) c)d) e) ` f)Fig. 4. The enhancement results for test image Scene, a) original image, b) output of the proposed appro...
[14] J.A. Stark, “Adaptive Image Contrast Enhancement Using Generalizationof Histogram Equalization” IEEE Trans Image Proc...
Upcoming SlideShare
Loading in …5

Colour image enhancement by virtual histogram approach


Published on

For more projects visit @

Published in: Technology, Business
  • Be the first to comment

  • Be the first to like this

No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide

Colour image enhancement by virtual histogram approach

  1. 1. Contributed PaperManuscript received November 18, 2009Current version published 06 29 2010;Electronic version published 07 06 2010. 0098 3063/10/$20.00 © 2010 IEEEColour Image Enhancement by Virtual Histogram ApproachZhengya Xu, Hong Ren Wu, Xinghuo Yu, Fellow, IEEE, Bin Qiu, Senior Member, IEEEAbstract — This paper introduces a new hybrid imageenhancement approach driven by both global and localprocesses on luminance and chrominance components of theimage. This approach, based on the parameter-controlledvirtual histogram distribution method, can enhancesimultaneously the overall contrast and the sharpness of animage. The approach also increases the visibility ofspecified portions or aspects of the image whilst bettermaintaining image colour. The approach was comparedwith other well-known image enhancement techniques. Theexperimental results have shown the superiority of theproposed approach1.Index Terms — Image processing, image enhancement.I. INTRODUCTIONImage enhancement, which transforms digital images toenhance the visual information within, is a primaryoperation for almost all vision and image processing tasks inseveral areas such as computer vision, biomedical imageanalysis, forensic video/image analysis, remote sensing andfault detection [2, 4]. For example, in forensic video/imageanalysis tasks, surveillance videos have quite differentqualities compared with other videos such as the videos forhigh quality entertainment or TV broadcasting. High qualityentertainment or broadcasting videos are produced undercontrolled lighting environment, whereas surveillancevideos for monitoring outdoor scenes are acquired undergreatly varied lighting conditions depending on the weatherand the time of the day. One of the common defects ofsurveillance videos is poor contrast resulting from reducedimage brightness range. A routine examination of thehistograms of the images from the videos reveals that someof the images contain relatively few levels of brightness, andsome of the images have a type of histograms. In the type ofhistograms, a large span of the intensity range at one end isunused while the other end of the intensity scale is crowdedwith high frequency peaks [4], which is typicallyrepresentative of improperly exposed images. The problemis how to approximate or reconstruct information that waslost because of the image having been captured under sub-optimal aperture or exposure conditions. Enhancementtransformation to modify the contrast of an image within adisplay’s dynamic range is, therefore, required in order to1Z. Xu, H.R. Wu and X. Yu are with Royal Melbourne Institute ofTechnology, University, VIC 3001, Australia (;; Qiu is with Faculty of Information Technology, Monash University,VIC 3800, Australia (e-mail: full information contents in the videos, e.g., forforensic investigations.First of all, the basic strategies are briefly reviewed forimage enhancement. Point-operation-based imageenhancement includes contrast stretching, non-linear pointtransformation and histogram modelling [3, 4]. They arezero memory operations that remap a given input grey-levelinto an output grey-level, according to a globaltransformation [2, 4, 9]. Non-linear point transformations,which could be expressed as G(j,k)=[F(j,k)]pwhere F(j,k)represents the original image, G(j,k) represents the outputimage and p is the power law variable, have been shown toimprove visual contrast in some cases whilst clearlyimpairing visual contrast in other cases [2, 3, 15]. Inhistogram modelling [10, 13, 17, 22, 24], the original imageis scaled so that the histogram of the enhanced image isforced to be some desired form such as uniform,exponential, hyperbolic or logarithmic [18, 19]. Thesemethods have the disadvantage of treating the imageglobally only. In order to differentiate between several areasof the image that may require different levels of contrastenhancement, an adaptive histogram modelling techniquewas proposed [16]. Images generated by the adaptivehistogram modelling process, sometimes, is so harsh on theimage visual appearance that an adaptive blurring of thewindow histogram was proposed [14] prior to forming thecumulative histogram as a means of improving the imagequality.Recently, some histogram based approaches, such asdynamic range separate histogram equalization(DRSHE)[6], brightness preserving dynamic histogramequalization (BPDHE)[7] and gain-controllable clippedhistogram equalization (GC-CHE)[8] have been developedin order to overcome some drawbacks of histogramequalization methods.Other classes of methods for image enhancement areapproaches based on the Retinex theory [21, 23], spatialoperations and pseudo-colouring. Spatial operations maysuffer from enhancing excessively the noise in the image orconversely smoothing the image in areas that need topreserve sharp details [3] and these operations are alsoknown to be time consuming. Pseudo-colouring methodsartificially map the grey-scale image to a colour image, withthe disadvantage that extensive interactive trials are requiredto determine an acceptable mapping scheme [2].Local enhancement methods have been developed basedon the gray-level distribution in the neighbourhood of everypixel in a given image. A typical example of localenhancement methods is the adaptive histogram equalization(AHE), which has shown good results in medical imagingapplications. However, AHE uses an enhancement kernel704 IEEE Transactions on Consumer Electronics, Vol. 56, No. 2, May 2010
  2. 2. that is quite computationally expensive. Moreover, AHEmay yield unsatisfactory outputs, e.g., images with noiseartefacts and falsely enhanced shadows [1, 5].Furthermore, all the aforementioned methodologies,except the pseud-colouring, only deal with grey-scale imageenhancement, i.e., they only use luminance component of acolour image for colour image enhancement.With a especial interest in surveillance video/imageprocessing, the proposed colour image enhancement methodis a fast adjustable hybrid approach controlled by a set ofparameters in order to take the advantages of pointoperations and local information driven enhancementtechniques, in making effective use of the entire range ofavailable pixel-values for both colour and luminancecomponents of a colour image.The organization of this paper is as follows. Section IIaddresses the principles of the proposed image enhancementtechnique. Experimental results are presented in Section III.Final conclusions are drawn in Section IV.II. PRINCIPLE OF THE PROPOSED METHODIn surveillance videos/images, the luminance histogram ofa typical natural scene that has been linearly quantised is,more often than not, highly skewed toward the darker levers;a majority of the pixels possess a luminance less than theaverage. In such images, details in the darker regions areoften not perceptible. One means to enhance these types ofimages is a technique called histogram modification, wherethe original image is scaled so that the histogram of theenhanced image follows a desired distribution. Usually, auniform distribution is used to create an image with equallydistributed brightness levels over the entire brightness scale.While histogram equalization applies a transformation thatyields a close-to-uniform histogram for the relativefrequency of the brightness-levels in an image, it onlyenhances the contrast for brightness values close tohistogram maxima, and decreases the contrast nearhistogram minima [2, 12, 15]. If the image analyst isinterested in certain parts or features of an image and thebrightness of the parts or features of the image is not closethe histogram maxima or, even worse, near the histogramminima, which happens often in surveillance video/imageanalysis, histogram equalization is helpless in the requiredcontrast enhancement task. A linear-like or no-linearbrightness stretching is only effective for an image wherethe histogram is narrow [2, 9], which is, unfortunately, notoften the case in practical video/image analysis.In order to meet the above particular practical demandsand stringent requirements for forensic image/videoanalysis, biomedical image analysis and remote sensing, anew colour image enhancement method is proposed asfollows. The proposed enhancement technique is driven byboth global and local processes to achieve not only effectiveimprovement of overall contrast but also the significantenhancement of details in identified features/areas of interestof a colour image. The proposed method also aims atemploying a much less time-consuming enhancementmechanism than those used by the existing methods.Histograms are used to depict image statistics in an imageinterpreted visual format. With histogram, it is easy todetermine certain types of problems in an image, such as ifthe image is properly exposed. Luminance histogram andcomponent histogram both provide useful information aboutthe lighting, contrast, dynamic range, and saturation effectsrelative to the individual colour components [4]. Therefore,in the proposed method the histogram of the enhancedcolour image should not be saturated at one or both ends ofthe dynamic range or at least not bring new significantspikes at the tail ends in order not to introduce new poorexposure like defects in the image. In addition to a full useof the maximum possible dynamic range, colour componentand local information can, certainly, make a contribution tothe contrast enhancement. Ideally, an effective imageenhancement technique devised using colour component andlocal information must not have or introduce very largespikes in the histogram of the enhanced image.Therefore, the intention of the proposed method is to finda monotonic pixel brightness transformation q=T(p) for acolour image such that the desired output histogram can notonly meet specific requirements but also be as uniform aspossible over the whole output brightness scale [9] to fill inthe full range of brightness values.First of all, definitions and notations are presented forintroduction of the proposed enhancement method.Let }1,1|)({ 212,1 NcMccccC denote thepixel coordinates of a colour image, where M and N are theheight and width of the image, respectively. At each pixelcoordinate, Cc , a multivariate value( ) [ ( ), ( ), ( )]TRGB R G Bx x xx c c c c is used to represent the pixelin RGB (Red, Green, Blue) colour space at the currentposition and a multivariate value( ) [ ( ), ( ), ( )]B R B RTYC C Y C Cx x xx c c c c is used to represent thepixel in YCBCR colour space [2, 21]. For each RGB colourchannel, each individual histogram entry is defined,respectively, ashR(i)=card{c | ( )Rx c =i, Cc }, (1)hG(i)=card{c | ( )Gx c =i, Cc }, (2)hB(i)=card{c | ( )Bx c =i, Cc }, (3)where card{•} is the cardinality function, 0 i<K, and K isa scale for a component of the colour image and, usually,256.Second, the colour image is, some times, transformedfrom the RGB colour space or another colour space to theYCBCR colour space necessarily in the proposed imageenhancement [2, 21]. The luminance channel histogram ofan image in the YCBCR colour space is defined ashY(i)=card{c| ( )Yx c =i, Cc }, (4)Z. Xu et al.: Colour Image Enhancement by Virtual Histogram Approach 705
  3. 3. where all symbols are as defined in equations (1) to (3). Thecumulative histogram for each RGB component andluminance component, Y, for the YCBCR colour space aredefined by extending the definition of cumulative histogramfrom grey-scale image, respectively asppiRR ihpH0)()( , (5)ppiGG ihpH0)()( , (6)ppiBB ihpH0)()( (7)andppiYY ihpH0)()( . (8)where the input brightness value is [p0, pk] and p [p0, pk].The cumulative histograms are monotonic no-decreasingfunctions withHR(K)= HG(K)= HB(K)= HY(K)=MN. (9)Based on above definitions, the proposed enhancementmethod is described as follows.Prominent image events, such as objects or a scene suchas edges and contours, originated from local changes inintensity or colour, are highly important for visualperception and interpretation of images. It is thus naturallythat the enhancement of edges has been an important task inimage processing.Compared with the original image, an enhanced imagewith good contrast will have a higher intensity of the edges.Since a histogram of an image contains no information aboutthe spatial arrangement of pixels in the image, luminancehistogram and component histogram do not provide anyinformation about the spatial distribution of the actualcolours in the image. Since we are only interested in how toenhance the edge intensity without regard to its orientation,a linear differential operator, which is a local geometricinformation based operator, widely known as the Laplacian,2221221212212 ),(),(),(ccccccccxxx may be used inorder to enhance edge related area in colour images.In colour images, the scale of brightness is 256represented by 8-bit for most of images, whereas a true RGBcolour space has distinct 224colours, with each colourcomponent pixel represented by 8-bit. Therefore, the edgeinformation is not only described by their luminance but alsoconveyed by their colour. In order to use information fullyfrom both of brightness and colour perception, the Laplacianoperation is applied to each of the RGB channels,respectively.Let)(cLRGB =| )(2cxR |+| )(2cxG|+| )(2cxB |. (10)Based on (10), we can defineSLap={ c| )(cLRGB >Tla, Cc } , (11)where threshold Tla [1, 10] , and the default threshold Tlais set to 3.Hence, SLap is a set of pixel coordinates of an image,i.e., CSLap. Each of the pixels with their coordinates inthe set SLap has a sum of absolute value of output of the pixelprocessed with a Laplace operator 2(10) and the sumvalue is greater than Tla. We definehYv(p)=card{c| ( )Yx c =p, c SLap }, (12)where the input brightness value are [p0, pk] and p [p0, pk].hYv(pi) can be treated as a special density function dependedon the local feature of every pixel and the frequency of thebrightness value of the pixels.If a histogram for an input colour image is hY(p) and theinput brightness value is [p0, pk], H, H1 and H2, are definedas follows:kppiY ihH0)( (13)110)(1kkppiwY ihwH (14)andpkpiYv ihvH0)(2(15)where pk1 and pk10 are in the range of [p0, pk]; hYw(p)= h(p)if p is in the range of (pk10, pk1], otherwise hYw(p)=0; w is aparameter with the default value of 2; v is a parameter withthe default value set to 1. Here, H1(p) is designed to suitspecial enhancement requirements for the imageinterpretation.Using21 HHHHcnas a normalisationcoefficient, a new virtual distribution function is defined as))()()(()(010000ppiYvppiYwppiYnppiihvihwihcihk(16)If M and N are the height and the width of an image,respectively, and the output brightness range is [q0, qk], thedesired output histogram can be approximated with (16) byits corresponding continuous probability density as follows:0 0001( )q pq pkMN ds h s dsq q(17)The left side of Equation (17) is the correspondinguniform probability distribution function. The desired pixelbrightness histogram transformation T is defined as000 0( ) ( )pkpq qq T p h s ds qMN(18)706 IEEE Transactions on Consumer Electronics, Vol. 56, No. 2, May 2010
  4. 4. The discrete approximation of the continuous pixelbrightness transformation from Equation (18) is, therefore,given by000 0( ) ( )pki pq qq T p h i qMN(19)Thus, the quantisation step-size is obtained as follows))()()(()( 000iYviYwinkikipvhpwhphcMNqqphMNqqq (20)On the right-hand side of Equation (20), the second term isused to enhance contrast for a specified range [Pk10, Pk1];whYw(i) = 0, if pi is not in the range of [Pk10, Pk1]; the thirdterm, basically as the first term (input histogram of theimage), is dependent on the image structure, though theparameter v can be adjusted. In most cases, v is fixed as 1,since the enhanced result is not very sensitive to the changeof the v (see Fig.1). Fig1.c and Fig1.d show the results of thetested image enhanced by the proposed method withdifferent values of v. Through (20), it can be clearly seenhow the output interval value between adjacent twobrightness values is produced one by one and how theparameters make contribution to every output brightnesslevel for contrast enhancement since human vision is verysensitive to the interval value q [12, 20]. The defaultvalues of these parameters are: Pk10 = 0, Pk1 = 30 and w = 2.The parameters can be adjusted by an image interpreter tomeet his or her specific requirements. For many cases, theproposed approach with the default values of the parametersworks well without user intervention, as changes of theparameters do not affect the enhanced result very much (seeFig.1). Fig1.b Fig1.c and Fig1.d show the results of thetested image enhanced by the proposed method withdifferent values of its parameters.It is noted that the number of reconstruction levels of theenhanced image must be less than or equal to the number oflevels of original image to provide proper intensity scaleredistribution if all pixels in each quantisation level are to betreated similarly [12].When the contrast of a dark area of an image whosehistogram spans a broad range of the display scale isenhanced, the bright areas of the image may be out of thedisplay range as a result of the above rescaling as defined by(19). When the contrast of a bright area of an image whosehistogram spans a broad range of the display scale isenhanced, the dark areas of the image may also be out of thedisplay range as a result of the above rescaling as defined by(19). Therefore, a hard-limit is needed to map the outputimage pixel values back into the display range [9]. However,the simple hard-limiting method is only suitable for anoutput image with only a few pixels whose brightness valuesare outside [q0, qk].In order to avoid or to greatly reduce the brightness rangeof the output image, a rescaling constraint is employed usingparameter t, which is introduced in the proposed method, tolimit the maxima of q within t, and to smooth theenhancement contrast over the full brightness scale.Consider a patch of light of intensity I + I surrounded by abackground of intensity I and I is actually similar to t forhuman vision. Over a wide range of intensities, it is foundthat the ratio the I/I, called the Weber fraction, is nearlyconstant at a value of about 0.02[2], so the default value of tis set to 3. The rescaling process also relieves further theundesired property of the traditional histogram equalisationtechnique, which tends to reduce contrast near histogramminima.If an image with its histogram basically concentrated in avery bright region, the image can be first inversed. Then, theproposed method is applied and the resultant image isinversed back, to ensure above constraints being met and towork more effectively.After the brightness contrast enhancement in theluminance channel, the output colour image is transformedback to the RGB colour space for display, since almost allhardware generally deliver or display colour via the RGBchannels.Though the limitation of q was applied, the outputresults, unfortunately in some cases, showed that ahistogram/histograms of RGB channels was/were saturatedat one or both ends of the dynamic range. The out-of-rangeoutput pixel values were mapped into the maximum orminimum values of the output scaling range and appeared onthe histogram as significant spikes at the tail ends. Theylooked like what typically occurs in an underexposed oroverexposed image, which is undesirable for improvingvisual quality of an image. In order to remove this defect, amore effective output range boundary control mechanism isfurther applied in the proposed method. The output rangeboundary control mechanism is introduced as follows.The linear mapping of video signal from the RGB colourspace to the YCBCR colour space [2, 21], which is used byvideo and broadcasting television industry, is computed forthe luminance component, Y, by [2, 21]Y=0.299R+0.587G+0.144B (21)where the luminosity (Y) is a function of R, G and B whichare normalised to 1, and denoted as ),,( BGRY .Obviously, according to (23), 0RY, 0GY, and0BY. Hence, ),,( BGRY is a non-decreasing monotonicfunction. In order to determine the new borders of the outputscaling range for the luminance channel (Y) we only need tofind the corresponding Y values to the upper and the lowerbounds of the RGB channels for the colour image. Thevalues can be obtained from the output RGB histograms andthe conversion from the RGB space to the YCBCR space[2,15] is described as follows,Ylow = max{ Ylow-red, Ylow-green, Ylow-blue } (22)Z. Xu et al.: Colour Image Enhancement by Virtual Histogram Approach 707
  5. 5. andYhigh = min{Yhigh-red, Yhigh-green , Yhigh-red }, (23)whereYhigh-red = min{y|( ) ( ( ) 255) ( (255) (255) )Y R R red oldy x x h h sc c } (24)Yhigh-green =min{y|( ) ( ( ) 255) ( (255) (255) )Y G G green oldy x x h h sc c } (25)Yhigh-blue =min{y|( ) ( ( ) 255) ( (255) (255) )Y B B blue oldy x x h h sc c } (26)Ylow-red = max{y|( ) ( ( ) 0) ( (0) (0) )Y R R red oldy x x h h sc c }, (27)Ylow-green = max{y|( ) ( ( ) 0) ( (0) (0) )Y G G green oldy x x h h sc c } (28)andYlow-blue = max{y|( ) ( ( ) 0) ( (0) (0) )Y B B blue oldy x x h h sc c }, (29)where hred-old(i), hgreen-old(i) and hblue-old(i) are RGBhistograms of the original colour image; s=MNs0 is thenumber of the saturated pixels in the image and usually thesame value is taken for both the upper and the lower bounds,with s0 [0.0005, 0.005] . The default s0 is set to 0.002.Using this formulation, the bounds of the new dynamicrange do not depend purely on singular extreme pixels, butcan be based on a representative set of pixels.After the new output scaling range, q [Ylow, Yhigh], isobtained, the transformation (19) is to be redone with thenew output scaling range, and the defect is removed of theimage shown on its histogram as significant spikes at the tailends, therefore the enhanced results show good contrast andmuch better colour maintenance as well(see Fig 1).The number of brightness levels which human can easilydistinguish depends on many factors, such as the averagelocal brightness. Consequently, a display, which avoids thiseffect, will normally provide at least 100 intensity levelswithin its display range [4, 15].Generally speaking, the number of reconstruction levelsof the enhanced image in the proposed method is usuallyless than that of original image to provide proper brightnessscale redistribution since all pixels in each quantisation levelare to be treated similarly. For an original image with 256levels of brightness, if the number of the brightness levels isnot reduced too many, no significant degradation isperceived [4, 15]. However, in some rare cases, if theoriginal image have extremely low dynamic range with onlyfew intensity values, the minimum brightness levels controlwill be applied in the proposed method by adjusting theparameters w and v, in order to ensure that the outputdynamic range is not less than 70% that of the original toavoid over contrast enhancement. The threshold ofbrightness level for applying the control is set to 64[2, 15].Since in some rare cases the last q of the transformationby Equation (20) is very large, a linear contrast stretchtransformation is also applied in the proposed approach toensure full use of the output brightness scale.Sometimes, an enhanced image serves as the input to amachine that traces the outline of the edges, and perhapsmeasures the shape and size of the outline. The imageenhancement emphasizes salient features of the originalimage and simplifies the processing task for a dataextraction machine. In this case, the proposed method canbe applied directly to each of RGB channels, using thecontribution from the colour and the luminance componentsfor contrast enhancement. It shows clearly betterperformance in some of the test images, while introducingfalse colour in others.III. EXPERIMENTSIn the performance evaluation, the proposed method,which works as an automatic enhancement method usingparameters with default values, is compared with fourclassical enhancement methods (linear contrast stretching,contrast reverse, gamma correction and histogramequalization [2, 15, 20] and some recent developedhistogram equalization based methods, such as DRSHE,BPDHE and GC-CHE using test images. The test imagesinclude well-known typical test images including Mountain,Scene, Meat etc, with an image resolution of 500x362 or721x481 or 768x768 or 731x487 pixels.The results enhanced by the proposed method withdifferent values of its parameters are shown in Fig.1b, Fig.1cand Fig.1d. In Fig1b, the tested image Mountain wasenhanced by the proposed method without output rangeboundary control. In Fig1.c, the tested image Mountain wasenhanced by the proposed method with t=2 and v=1. InFig1.d, the tested image Mountain was enhanced by theproposed method with w =3 and v=2. It is observed from theexperimental results shown in Figs. 1-4 that the proposedenhancement algorithm can effectively enhance the overallcontrast and the sharpness of the test images. A significantamount of details that could not be seen in the originalimages has been clearly revealed. For the tested colourimages, better results of the compared techniques, such aslinear contrast stretching, contrast reverse, gammacorrection and histogram equalization, are obtained by firstconverting the image to the Hue, Saturation, Intensity colourspace and then applying the compared techniques to theIntensity component only. However, even this method doesnot fully maintain colour fidelity for the comparedtechniques [11] (see Fig.1f, Fig.3c and Fig.4c), while theproposed technique show much better colour maintenancethan other techniques (see Fig.1c, Fig.1d, Fig.3b andFig.4b). From the output images, these test images wereenhanced by histogram equalization, the histogramequalization with over-enhanced parts of the images andhighlighted blocking artefacts caused by image compression(see Fig.2c and Fig.4c). In terms of revealing the details indark areas of images with a broad low histogram, theperformance of the proposed approach is much better thanthe other image enhancement methods. The details in thedark areas of the test images, Mountain, Light and Scene, aremuch more visible than those in the original image while thecolour of the images are better preserved. The linearstretching failed to make significant enhancement for theseimages.708 IEEE Transactions on Consumer Electronics, Vol. 56, No. 2, May 2010
  6. 6. a) ` b) c)d) e) f)Fig. 1. The enhancement results for test image Mountain, a) original image, b) output of the proposed approach-1, c) output of proposedapproach-2, d) output of proposed approach-3, e) output of modified linear stretching, f) output of histogram equalisation.Z. Xu et al.: Colour Image Enhancement by Virtual Histogram Approach 709
  7. 7. a) b) c)d) e) f)Fig. 2. The enhancement results for test image Meat, a) original image, b) output of the proposed approach, c) output of histogram equalisation, d)output of linear stretching, e) output of contrast reverser, f) output of modified linear stretching.a) b) c)d) e) f)Fig. 3. The enhancement results for test image Light, a) original image, b) output of the proposed approach, c) output of histogram equalisation, d)output of linear stretching, e) output of gamma correction, f) output of GC-CHE.710 IEEE Transactions on Consumer Electronics, Vol. 56, No. 2, May 2010
  8. 8. a) b) c)d) e) ` f)Fig. 4. The enhancement results for test image Scene, a) original image, b) output of the proposed approach, c) output of histogram equalisation, d)output of gamma correction, e) output of PBDHE, f) output of DRSHE.Contrast reverse transfer function [2] and gammacorrection [2, 4] are often helpful in visualizing details indark areas of an image. Therefore, the results of thereverse function and gamma correction are used in theexperiments for comparison (see Fig.2e, Fig.3e andFig.4d). The reverse function, which was clipped at 10%input amplitude levels to maintain the output amplitudewithin the range of unity, was applied to the invertedoriginal image, and then the resultant image was revertedafter the enhancement process. The modified contraststretch was obtained by truncated input scale range at boththe low and the high ends, corresponding to the 1%histogram distribution.IV. CONCLUSIONIn this paper, a new hybrid approach based on a virtualhistogram modification for colour image enhancement isproposed. The novelty of the proposed method is thatcolour image enhancement is based on modification of avirtual histogram distribution, which is a new way tointegrate colour and brightness information extracted fromsalient local features, for global contrast enhancement.The special contributions of the proposed method are theoutput value scaling bounds control and output rangeboundary control for the enhancement mechanism toensure the better maintenance of colour for the enhancedimages. The proposed approach introduces the parametersto increase the visibility of specified features, portion oraspects of the image. If the parameters are set up to defaultvalues, the proposed method will work as an automaticprocess. The proposed approach has a potential for variousapplications to enhance specific categories of images, suchas surveillance videos/images, biomedical images andsatellite images.REFERENCES[1] Cristian Munteanu and Agostinho Rosa, “Gray-Scale ImageEnhancement as an Automatic Process Driven by Evolution,” IEEETransactions on Systems, Man, and Cybernetics—part b: Cybernetics,vol. 34, no. 2, April 2004.[2] William K. Pratt, Digital Image Processing, John Wiley & Sons,2008.[3] G. Ramponi, N. Strobel, S. K. Mitra, and T.-H. Yu, “Nonlinear Un-Sharp Masking Methods for Image Contrast Enhancement,” J.Electron. Imaging, vol. 5, no. 3, pp. 353–366, 1996.[4] R. C. Gonzalez and P. Wintz. Digital Image Processing, 2ndEd.,Prentice Hall, 2002.[5] J. B. Zimmerman, S. M. Pizer, E. V. Staab, J. R. Pery, W. McCartney,and B. Brenton, “An evaluation of the effectiveness of adaptivehistogram equalization for contrast enhancement,” IEEE Trans. Med.Imag., vol. 7, pp. 304–312, Dec. 1988.[6] G. Park, H. Cho, and M. Choi, “A Contrast Enhancement Methodusing Dynamic Range Separate Histogram Equalization,” IEEE Trans.on Consumer Electronics, Vol. 54, No. 4, pp 1981-1987 Nov. 2008.[7] N. S. P. Kong and H. Ibrahim, “Color Image Enhancement UsingBrightness Preserving Dynamic Histogram Equalization,” IEEETrans. on Consumer Electronics, Vol. 54, No. 4, pp 1962-1968 Nov.2008.[8] T. Kim and J. Paik, “Adaptive Contrast Enhancement Using Gain-Controllable Clipped Histogram Equalization,” IEEE Trans. onConsumer Electronics, Vol. 54, No. 4, pp 1803-1810 Nov. 2008.[9] G. Deng, L. W. Cahill, and G. R. Tobin, “The Study of LogarithmicImage Processing Model and Its Application to Image Enhancement,”IEEE Trans. on Image Processing, vol. 4. no. 4, April 1995.[10] C. Soong-Der and A. R. Ramli, "Contrast Enhancement UsingRecursive Mean-Separate Histogram Equalization for ScalableBrightness Preservation," IEEE Trans. on Consumer Electronics, vol.49, no. 4, pp. 1301-1309, 2003.[11] John P. Oakley and Hong Bu, “Correction of Simple Contrast Loss inColour Images,” IEEE Transactions on Image Processing, Vol. 16,No. 2, pp. 511-522, February 2007.[12] Wilhelm Burger and Mark J Burger, Digital Image Processing,Springer, 2008.[13] Soong-Der Chen, Abd. Rahman Ramli, “Minimum Mean BrightnessError Bi-Histogram Equalization in Contrast Enhancement,” IEEETrans. on Consumer Electronics, Vol. 49, No. 4, pp. 1310-1319,November 2003.Z. Xu et al.: Colour Image Enhancement by Virtual Histogram Approach 711
  9. 9. [14] J.A. Stark, “Adaptive Image Contrast Enhancement Using Generalizationof Histogram Equalization” IEEE Trans Image Processing, vol 9, no. 5,889-896, May 2000.[15] Milan Sonka, Vaclav Hlavac and Roger Boyle, “Image processing,Analysis, and Machine Vision,” Brooks/Cole, 2001.[16] S. M. Pizer et al., “Adaptive Histogram Equalisation and its Variations,”Computer Vision, Graphics and Image Processing, 39,3, 355-368, Sept.1987.[17] W. Yu, C. Qian, and Z. Baeomin, "Image Enhancement Based on EqualArea Dualistic Sub-Image Histogram Equalization Method," IEEETransactions on Consumer Electronics, vol. 45, no. 1, pp. 68-75, 1999.[18] E.L. Hall, “Almost Uniform Distribution for Computer ImageEnhancement,” IEEE Trans. Computers, C-23, 2, 207-208, 974.[19] W. Frei, “Image Enhancement by Histogram Equalization,” Graphics andImage Processing, 6, 3, 286-264, 1977.[20] H.R. Wu and K.R. Rao, Digital Video Image Quality and PerceptualCoding, CRC Press, 2006.[21] D. J. Jobson, Z. Rahman, and G. A. Woodell, “Properties and performanceof a center/surround retinex,” IEEE Trans. Image Process., vol. 6, no. 3,pp. 451–462, Mar. 1997.[22] K. Yeong-Taeg, "Contrast Enhancement Using Brightness Preserving Bi-Histogram Equalization," IEEE Transactions on Consumer Electronics,vol. 43, no. 1, pp. 1-8, 1997.[23] Z. Rahman, D. J. Jobson, and G. A. Woodell, “Retinex processing forautomatic image enhancement,” J. Electron. Imag., vol. 13, no. 1, pp. 100–110, Jan. 2004.[24] Tzu-Cheng Jen and Sheng-Jyh Wang, “Generalized HistogramEqualization based on Local Characteristics” in Proceeding of IEEE ICIP2006, pp. 2877-2880.BIOGRAPHIESZhengya Xu received the Master in InformationTechnology from the Vrije Universiteit Brussel, BrusselsBelgium in 1994, and the PhD degree from SwinburneUniversity of Technology, Victoria, Australia in 2000. Hewas a research engineer with Aeronautics ResearchInstitute, Ministry of Aviation and Aerospace Industry,Beijing, China, and a senior research fellow with MonashUniversity, Australia. He works currently as a seniorresearch fellow on Computer & Network Engineering atRMIT University Australia. His research interests include image and videoprocessing, computer vision, biometric pattern recognition, object tracking,software and embedded systems development.Hong Ren Wu received the B.Eng. and M.Eng. fromUniversity of Science and Technology, Beijing (formerlyBeijing University of Iron and Steel Technology), P.R.China, in 1982 and 1985, respectively. He received thePhD in Electrical and Computer Engineering from theUniversity of Wollongong, N.S.W. Australia, in 1990. DrWu worked on academic staff of Chisholm Institute ofTechnology and then Monash University, Melbourne,Australia from April 1990 to January 2005 last as anAssociate Professor in Digital Systems. Dr Wu has been with Royal MelbourneInstitute of Technology, Australia, since February 2005, as Professor of VisualCommunications Engineering and Discipline Head, Computer and NetworkEngineering in School of Electrical and Computer Engineering. His researchinterests include fast DSP algorithms, digital picture compression and qualityassessment, video processing and enhancement, embedded DSP systems andtheir industrial applications. He is a co-editor of the book, Digital Video ImageQuality and Perceptual Coding, (CRC, 2006).Bin Qiu (M’93-SM’00) received the B.Eng degreefrom Beijing Jiaotong University, in 1982, and theMSc and PhD degrees from University ofManchester, Institute of Science and Technology,UK, in 1987 and 1989 respectively.He was a lecturer at Victoria University from1990 to 1995. He joined Monash University in 1995as a senior lecturer, where he is currently anassociate professor. Since 1987, he has participatedand headed research projects in the fields of digital and computercommunication systems and networks, signal processing, imageprocessing and the applications of neuro-fuzzy techniques in the aboveareas. He is also interested in the design and implementation of digitalsystems.Xinghuo Yu (M91-SM98-F08) received the B.Engand M.Eng degrees from the University of Scienceand Technology of China, Hefei China, in 1982 and1984,and the PhD degree from South-EastUniversity, Nanjing China in 1988, respectively. Heis now with RMIT University, Australia, where he isthe Director of RMIT Platform TechnologiesResearch Institute and Professor of InformationSystems Engineering. Prof Yus research interests include variablestructure and nonlinear control, signal processing, complex andintelligent systems. He has published over 300 refereed papers intechnical journals, books and conference proceedings as well as co-edited 10 research books.Prof Yu served as an Associate Editor of IEEE Transactions on Circuitsand Systems Part I (2001-2004) and IEEE Transactions on IndustrialInformatics (2005-2008), and is serving as an Associate Editor of IEEETransactions on Industrial Electronics and several other scholarlyjournals. He has been on the program committees of many internationalconferences, and co-chaired several international conferences. Prof. Yuwas the sole recipient of the 1995 Central Queensland University ViceChancellors Award for Research. He is a Fellow of Institution ofEngineers Australia, and was made Emeritus Professor of CentralQueensland University Australia in 2002 for his long term significantcontributions.712 IEEE Transactions on Consumer Electronics, Vol. 56, No. 2, May 2010