DTM: Dynamic Tone Mapping for Backlight Scaling

Transcription

1 : Dynmic Tone Mpping for Bcklight Scling Abstrct- This pper proposes n pproch for pixel trnsformtion of the displyed imge to increse the potentil energy sving of the bcklight scling method. The proposed pproch tkes dvntge of humn visul system chrcteristics nd tries to minimize the incurred distortion between the perceived brightness vlues of the individul pixels in the inl imge nd those of the bcklight scled imge. This is in contrst to previous bcklight scling pproches where the luminnce vlues of the individul pixels re mtched. Moreover, the proposed pproch bsed on tone mpping is menble to highly efficient hrdwre reliztion nd does not require ny informtion bout the histogrm of the displyed imge, which mkes it desirble for video pplictions. Experimentl results show tht the dynmic tone mpping for bcklight scling method results in bout 35% power sving with n effective distortion rte of 5% nd 55% power sving for 20% distortion rte. This is significntly higher power svings compred to previously reported bcklight dimming pproches. 1 Introduction Current genertion of portble computers nd instruments utilize bcklit Liquid Crystl Displys (LCDs). These displys hve lso ppered in pplictions rnging from medicl equipment to utomobiles, gs pumps nd retil terminls. The smll size nd bttery-powered opertion ssocited with LCD equipped pprtus mndte low component count nd high efficiency for these circuits. Size constrints plce severe limittions on circuit rchitecture nd long bttery life is usully priority. Lptop nd hndheld portble computers offer n excellent exmple. The LCD displys currently vilble require two power sources, bcklight supply nd contrst supply. The disply bcklight is the single lrgest power consumer in typicl portble pprtus, ccounting for lmost 50% of bttery drin with the disply t mximum intensity [1]. As such, every effort must be expended to mximize bcklight efficiency. Study of LCD energy mngement should consider the problem from n interdisciplinry viewpoint. The bcklight presents cscded energy ttenutor to the bttery (cf. Figure 1). Bttery energy is lost in the electricl-to-electricl conversion to high voltge AC for driving the Cold Cthode Florescence Lmp (CCFL). This section of the energy ttenutor is the most efficient; where conversion efficiencies exceeding 90% re possible. The CCFL, lthough the most efficient electricl-tolight converter vilble tody, hs losses exceeding 80%. Additionlly, the opticl trnsmission efficiency of present displys is under 50% for monochrome with color types much lower. The very high DC/AC conversion efficiency highlights some significnt issues. Anything tht improves energy trnsfer in the other ttenutor res will hve greter impct thn further electricl efficiency improvements. Additionl improvements in electricl efficiency, while certinly desirble, re reching the point of diminishing returns. Clerly, overll bcklight efficiency gins must come from lmp nd disply improvements. In [2], Chng et. l. proposed Dynmic bcklight Luminnce Scling technique (DLS) to reduce the energy consumption of the LCD displys. This technique is bsed on key ide tht eye s perception of the light, which is emitted from the LCD pnel, is function of two Energy source Bttery Electricl-To- Electricl conversion DC to AC high voltge converter > 90% efficency Electricl-To- Light conversion Cold Cthode Fluorescent Lmp (CCFL) < 20% efficiency Light-To-Light conversion Diffuser nd LCD disply < 50% efficiency Prsitic pth cpcitnces bsorb some DC/AC converter output Figure 1. Energy conversion pth of the LCD disply component. prmeters, 1) the light intensity of the bcklight nd 2) the trnsmittnce of the LCD pnel. Therefore, by crefully djusting these two prmeters one cn chieve the sme perception in humn eyes t different vlues of the bcklight intensity nd the LCD trnsmittnce. However, since the energy consumption of the bcklight lmp cn be reduced significntly by reducing its intensity, one cn sve energy by simply dimming the bcklight nd then compensting the loss of brightness by djusting the LCD trnsmittnce. However, this pproch suffers from two min drwbcks, ) it mnipultes every pixel on the screen one-by-one, limiting the ppliction of this pproch to still imges or low-frme-rte videos; b) It chieves energy sving t the cost of loss in visul informtion. Reference [3] improved this simple pproch by eliminting the pixel-by-pixel trnsformtion of the displyed imge through minor hrdwre modifictions to the built-in LCD reference driver. These modifictions could implement ny singlebnd gryscle spreding function to improve the brightness nd contrst of the displyed imge extending the pplicbility of the pproch to streming pplictions. More recently, Reference [4] further improved the previous pproches in two spects. First, by using globl histogrm equliztion technique to preserve most of visul informtion, nd second, by modifying the rchitecture of built-in LCD reference driver in order to produce ny piece-wise liner imge trnsformtion function. This pproch hve two min disdvntges, ) the empiricl distortion chrcteriztion curve used in this pproch is dependent on the type of the displyed imge, e.g. lndscpe, portrit, nd fireworks; b) similr to ll of previous pproches this pproch lso requires histogrm informtion of the displyed imge to clculte the imge trnsformtion function. Although ll of the forementioned techniques re effective bcklight scling pproches, but they hve overlooked single most importnt deciding fctor in their optimiztion process; tht is the humn visul system chrcteristic. All of these pproches rely on the luminnce vlues of pixels of the displyed imge s their optimiztion vribles. However, lluminnce vlue of light source is not the sme s its perceived brightness; therefore, the resulting policies will either end up not utilizing the full potentil of the bcklight scling technique or misleding the optimiztion into wrong solution spce. In contrst to forementioned pproches, this pper proposes bcklight scling technique which is bsed on tone reproduction opertor. This opertor will mp the inl imge χ to trnsformed imge χ`, such tht perceived brightness of the imge is preserved while its dynmic rnge is reduced. This reduction in dynmic rnge of the imge will further increse the potentil for bcklight scling nd therefore the energy sving. Moreover, the proposed opertor cn be clculted without ny informtion bout the individul pixels of the displyed imge or its histogrm, further improving the video frme-rte nd power sving gin due to elimintion of ny hrdwre/softwre support for imge histogrm genertion. In the following, the bsic bckground on the photometry, humn visul system, principles of photogrphic tone mpping, nd finlly the TFT LCD rchitecture nd prior work in dynmic bcklight scling will be discussed. Next, in section 3 dynmic tone mpping pproch will be explined. Sections 4 nd 5 will provide the supporting experimentl results nd conclusions of this technique. 2 Bckground nd Preliminries 2.1 Photometric Definitions Although light is form of electromgnetic rdition, the mesurement of luminous intensity from useful light source requires extr informtion bout the reltive sensitivity of the eye to different wvelengths. Photometry is the science of mesuring visible light in units tht re weighted ccording to the sensitivity of the humn eye. The luminous intensity of "white" light source is defined by multiplying the wtts emitted t ech wvelength by the efficiency of

2 tht wvelength in exciting the eye, reltive to the efficiency t 555nm. This efficiency fctor is referred to s the V-lmbd curve [5]. The cndel is the luminous intensity per solid ngle, in given direction of source tht emits monochromtic rdition t wvelength 555nm nd tht hs rdint intensity in tht direction of 1/683 wtt per sterdin[6]. On the other hnd, humen perceive luminnce. It is n pproximte mesure of how bright surfce ppers when one views it from given direction. Luminnce is equl to luminous intensity per squre meter nd is mesured in cndel per squre meter [6]. For exmple, the luminnce of sun, cloudy dull dy, typicl office, good street lighting, nd full moon re 5K, 300, 15, 1.2, nd 0.002cd/m 2, respectively. In the next section we will describe some chrcteristics of humn visul system, which re relevnt to the design of bcklight scling policy. 2.2 Humn Visul system (HVS) When light reches eye, it hits the photoreceptors on the retin, which then send the signl through nerves to the brin, where n imge is formed. The photoreceptors in our retin, nmely rods nd cones, ct s the sensors for the Humn Visul System (HVS). Rods re extremely sensitive to light nd provide chromtic vision t scotopic levels of illumintion (10-6 to 10 cd/m 2 ), tht is why we cnnot see colors in drk surroundings. Cones (which comprise of three distinct types) re less sensitive, but provide color vision t photopic levels of illumintion (0.01 to 10 8 cd/m 2 ). Note tht both rods nd cones re ctive t light levels between 0.01 nd 10 cd/m 2. This rnge is clled the mesoptic rnge. Unfortuntely the mesoptic rnge is the poorest reserched rnge, while it is the rnge tht is mostly encountered for monitors in office environments with subdued lighting. When light hits photoreceptors, n electricl signl trvels by neurons to the brin where n imge is formed. The incoming light cn hve dynmic rnge of nerly 1:10 14, wheres the neurons cn trnsfer signl with dynmic rnge of only bout 1:10 3/2. As result, there is the need for some kind of dpttion mechnism in our vision. This mens tht we first dpt to some luminnce vlue, nd then perceive imges in some dynmic rnge ner this luminnce vlue. One of the most importnt chrcteristics tht chnges with different dpttion levels is the Just Noticeble Difference (JND.) Let L nd L denote the JND nd the dpttion luminnce, respectively. Blckwell [7] showed tht the rtio L/L vries s function of the dpttion level, L nd thus, estblished the reltionship between L nd L to be 2.5 LL ( ) = ( L ) (1) Simply stted, Blckwell s eqution sttes tht if there is ptch of luminnce L +ε where ε L on bckground of luminnce L, it will be discernible, but ptch of luminnce L +ε, where ε < L will not be perceptible to the humn eye. Let us now consider the brightness perception. Brightness is the mgnitude of the subjective senstion which is produced by visible light. Although the rdince cn esily be mesured, the brightness, being subjective metric, cnnot be exctly quntified. Nevertheless, brightness is often pproximted s the logrithm of the luminnce, or the Figure 2. Brightness vs. luminnce chrcteristic of the HVS. luminnce rised to the power of 1/2 to 1/3 depending on the context. More precise, studies hve shown tht there is no one single formul, but rther the brightness-luminnce reltion depends on the dpttion level to the mbient light. In this pper, we will rely on the work of Stevens, which is lso extensively used in the field of computer grphics. Stevens et l. 0 devised the brils units to mesure the subjective vlue of brightness. According to Stevens, one bril equls the senstion of brightness tht is induced in fully drk-dpted eye by brief exposure to 5-degree solid-ngle white trget of 1 micro-lmbert luminnce. 1 Let B denote brightness in brils, L the inl luminnce vlue in lmberts, nd L denote the dpttion luminnce of the eye. Then, where L B = λ L σ = log ( ) L σ (2-) (2-b) λ = 10 L Typicl perceived brightness chrcteristic curves re shown in Figure 2. Note tht the slope of ech curve represents the humn contrst sensitivity tht is the sensitivity of the HVS brightness perception to the chnges in the luminnce. Furthermore, s L is decresed, the humn contrst sensitivity decreses. Finlly, the HVS exhibits higher sensitivity to chnges in luminnce in the drker regions of n imge. Two imges with different luminnce vlues cn result in the sme brightness vlues, nd cn pper to the HVS s being identicl.. Actully, ccording to eqution (2-) we re very poor judges of n bsolute luminnce, ll tht we cn judge is the rtio of luminnces, i.e. the brightness 2.3 Tone reproduction A clssic photogrphic tsk is the mpping of the potentilly high dynmic rnge of rel world luminnces to the low dynmic rnge of the photogrphic print. The rnge of light tht people experience in the rel world is vst. However, the rnge of light one cn reproduce on prints spns t best bout two orders of bsolute dynmic rnge [9]. This discrepncy leds to the tone reproduction problem: how should one mp mesured/sensed scene luminnces to print luminnces nd produce stisfctory picture? The success of photogrphy hs shown tht it is possible to produce imges with limited dynmic rnge tht convey the ppernce of relistic scenes. This is fundmentlly possible becuse the humn eye is sensitive to reltive, rther thn bsolute, luminnce vlues. Consider typicl scene tht poses problem for tone reproduction in photogrphy, room illuminted by window tht looks out on sunlit lndscpe. A humn observer inside the room cn esily see individul objects in the room s well s fetures in the outdoor lndscpe. This is becuse the eye dpts loclly s we scn the different regions of the scene. If we ttempt to photogrph our view, the result is disppointing. Either the window is over exposed nd we cn t see outside, or the interior of the room is underexposed nd looks blck. In 1993, Tumblin et l. [9] introduced this concept to computer grphics community nd proposed primitive tone mpping opertor. Since then gret del of work hs been done on the tone reproduction problem. Generlly speking the tone reproduction literture cn be divided into two min ctegories. The first ctegory which uses globl tone mpping opertor ignores the sptil informtion bout the luminnce of the inl scene nd dopts single non-decresing function s its tone mpping opertor. Reference [10] uses model of brightness perception to derive this mpping opertor. Reference [11] uses globl multiplier to mintin the visibility threshold. In reference [13] nother globl opertor ws proposed bsed on histogrm 1 One lmbert is equl to 3,183 cd/m 2.

3 . LCD component rchitecture b. TFT cell schemtic Figure 3. TFT-LCD screen djustment. This method uses n imge s histogrm to implicitly segment the imge so tht seprte scling fctors cn be used in different luminnce zones. The second ctegory tries to reproduce the visibility of different objects in the scene. This is done through multiple mpping functions which re dopted bsed on locl luminnce informtion of the inl scene. Chiu et l. [12] who used sptilly vrying exposure rmp over the imge ws the first to propose sptilly vrying dynmic rnge reduction opertor. Lter work by Pttnik et l.[14] developed the ultimte still imge opertor bsed on the humn visul system, incorporting color dpttion, locl contrst, nd dynmic rnge. More recently, References [15][16][17] hve proposed more successful tone mpping opertors in seprting the contrst differences tht mtter to vision from those tht do not. The bsic chllenge for sptilly vrying tone mpping opertor is tht it needs to reduce the globl contrst of n imge without ffecting the locl contrst to which humn visul system is sensitive. To ccomplish this, n opertor must segment the high dynmic rnge imge, either explicitly or implicitly, into regions tht humn visul system does not correlte during dynmic rnge reduction. Otherwise, the locl vrying opertors would led into disturbing reverse grdients typiclly seen s hlos round light sources. 2.4 LCD rchitecture nd Bcklight Scling Figure 3 shows the typicl rchitecture of n LCD controller nd pnel. The LCD controller receives the video dt nd genertes proper gryscle i.e., trnsmissivity of the pnel for ech pixel bsed on its pixel vlue. All of the pixels on trnsmissive LCD pnel re illuminted from behind by the bcklight. To the observer, displyed pixel looks bright if its trnsmittnce is high (i.e., it is in the 'on' stte), mening it psses the bcklight. On the other hnd, displyed pixel looks drk if its trnsmittnce is low (i.e., it is in the 'off' stte), mening tht it blocks the bcklight. For color LCD s, different filters re used to generte shdes of three min colors (i.e. red, blue, nd green), nd then color pixels re generted by mixing three sub-pixels together to produce different colors. Ech pixel hs n individul liquid crystl cell, Thin Film Trnsistor (TFT), nd storge cpcitor (cf. Figure 3b). The electricl field of the cpcitor controls the trnsmittnce of the liquid crystl cell. The cpcitor is chrged nd dischrged by the TFT. The gte electrode of the TFT controls the timing for chrging/dischrging of the cpcitor when the pixel is scnned (or ddressed) by the trcer for refreshing its content. The (drin-) source electrode of the TFT controls the mount of chrge. All of the gte electrodes of the pixels on the sme row re driven by single gte driver (clled gte bus line) nd re enbled t the sme time the row is trced. Similrly, single source driver (clled source bus line) drives ll source electrodes of the pixels on the sme column. The source driver supplies the desired voltge level (clled gryscle voltge) ccording to the pixel vlue. In other words, idelly, the pixel vlue trnsmittnce, t(x), is liner function of the gryscle voltge v(x), which is in turn liner function of the pixel vlue X. The trnsfer function of source driver which mps different pixel vlues, X, into different voltge levels, v(x) is clled the gryscle-voltge function. If there re 256 gryscles, then the source driver must be ble to supply 256 different gryscle voltge levels. The source driver mixes different reference voltges to obtin the desired gryscle voltges. Typiclly, these different reference voltges re fixed nd designed s voltge divider. Mthemticlly speking, in trnsmissive TFT-LCD monitor, for pixel with vlue X, the luminnce L(X) of the pixel is: L( X) = b. t( X) (3-) where t(x) is the trnsmissivity of the TFT-LCD cell for pixel vlue X, nd b [0,1] is the (normlized) bcklight illumintion fctor with b representing the mximum bcklight illumintion nd b=0 representing no bcklight. Note tht t(x) is liner mpping from [0,255] domin to [0,1] rnge. In bcklight scled TFT-LCD, b is scled down nd ccordingly t(x) is incresed to chieve the sme imge luminnce. Reference [2] describes two bcklight luminnce dimming techniques. These techniques dim the bcklight nd compenste for the luminnce loss by djusting the gryscle of the imge to increse its brightness or contrst. More precisely, L( X) = β. t( Φ ( X, β )) (3-b) where 0<β 1 is the bcklight scling fctor nd Φ(X,β) is the pixel trnsformtion function. Let x denote the normlized pixel vlue, i.e., ssuming n 8-bit color depth, x=x/255. The uthors of [2] scle the bcklight luminnce by fctor of β while incresing the pixel vlues from x to Φ(x,β) by two mechnisms. Clerly, Φ(x,β)=x denotes the identity pixel trnsformtion function. The bcklight luminnce dimming with brightness compenstion technique uses the following pixel trnsformtion function: Φ ( x, β ) = min(1, x + 1 β ) (4-) wheres the bcklight luminnce dimming with contrst enhncement technique uses this trnsformtion function: x Φ ( x, β ) = min(1, ) (4-b) β In these schemes, the optiml bcklight fctor is determined by the bcklight luminnce dimming policy subject to the given distortion rte. To clculte the distortion rte, n imge histogrm estimtor is required for clculting the sttistics of the input imge. In this pproch there is no considertion for HVS chrcteristic (cf. sec. 2.1) nd only rw imge luminnce vlues re used to chrcterize the imge distortion. Note tht the imge histogrm simply denotes the mrginl distribution function of the imge pixel vlues. Reference [3] proposes different pproch in which the pixel vlues in both drk nd bright regions of the imge re used to enble further dimming of the bcklight. The key ide is to first truncte the imge histogrm on both ends to obtin smller dynmic rnge for the imge pixel vlues nd then to spred out the pixel vlues in this rnge (by pplying n ffine trnsformtion) so s to enble more ggressive bcklight dimming while mintining the contrst fidelity of the imge. This pproch mximizes the number of pixel vlues tht re preserved in order to chieve minimum imge distortion. The min disdvntge of this cost function is tht it trets the gry scle vlues in drk nd white regions of the imge the sme wy, which is in contrst to the perceived brightness chrcteristic of the HVS s shown in Figure 2. More recently, Reference [4] proposed n pproch for imge trnsformtion, bsed on the imge histogrm. In this pproch the dynmic rnge of the inl imge is reduced such tht the incurred imge distortion is no more thn pre-specified vlue. Then, the bcklight scling technique is used to reduce the energy consumption of the LCD pnel. 2.5 Overview of the Proposed Tone Mpping Approch This pper incorportes knowledge from tone reproduction in photogrphy nd computer grphics into bcklight scling techniques to further improve the energy sving potentil of these techniques. This is in turn chieved by Dynmic Tone Mpping () lgorithm. The overview of the lgorithm is s follow.

4 1) Given n inl imge, χ, nd n upper bound on the tolerble imge distortion, we solve the CTM problem to determine the minimum dynmic rnge, R, of luminnce vlues in tonempped imge (cf. Sec. 3). The gol is to chieve the mximum power svings s result of follow-on bcklight dimming. Therefore, this step lso produces the optimum bcklight scling fctor, β. 2) We determine trnsformtion function, ψ, bsed on the brightness vlues of the inl imge nd the minimum dynmic rnge, R of the tone mpped imge. 3) We construct the trnsformed imge by pplying ψ to the inl imge, χ. At the sme time, we dim the bcklight by fctor β. This cn in turn result in significnt energy sving. The dvntges of compred to previous bcklight dimming techniques re: 1) The lgorithm uses perceived brightness of the displyed imge to clculte the mximum dynmic rnge of the imge nd hence the mximum possible energy sving. This should be contrsted to prior works which use the bsolute luminnce vlues of the displyed imge ignoring the HVS model. 2) A powerful, yet simple, tone mpping opertor is developed to reduce the dynmic rnge of the displyed imge for higher energy svings. This tone mpping opertor cn be clculted by using only the verge pixel vlues. This should be contrsted to prior works which cn only reduce the imge dynmic rnge by sturting the pixel vlues either t one end [2] or both ends [3] of the imge histogrm or require the complete informtion bout the imge histogrm to clculte complex trnsformtion function [4]. 3) The lgorithm is very different from previous bcklight scling pproches where the luminnce vlues of the individul pixels in the inl nd bcklight scled imges re mtched. In contrst, mkes use of Steven s results which show tht wht mtters is the brightness mtching between pixel vlues nd tht the brightness is proportionl to the rtio of the individul pixel luminnce to the verge imge luminnce. This hs drmtic effect on the finl bcklight scling policy becuse for one it mkes use of the fct tht the HVS is much more sensitive to the brightness mtching in the drk prts of n imge thn it is to brightness mtching the light prts. 4) The lgorithm llows highly efficient hrdwre reliztion, which mkes it desirble for video ppliction this ws imprcticl for previously proposed pproches due to performnce issues in clcultion of the trnsformtion function ψ. 5) It results in n dditionl power sving of 25% nd 5% compred to [2] nd [4]. Note tht lthough the power sving over HEBS [4] is only 5%, the key dvntge of over HEBS is tht does not require knowledge of the imge histogrm nd therefore results in significntly lower overhed for the bcklight scling. Becuse of this feture, is pplicble to both still imges nd video wheres HEBS cn only be used for still imges becuse it cnnot process video frmes in rel time (unless there is built-in hrdwre support for imge histogrm clcultion, but then the power dissiption overhed of this dditionl hrdwre must lso be ccounted for.) 3 Dynmic Tone Mpping () for Bcklight Scling Let L nd L mx mx denote the mximum luminnce of the inl imge nd the dynmiclly tone-mpped nd bcklight-scled imge, respectively. Moreover, let χ nd χ denote the pixel vlue informtion of the inl nd bcklight scled imges. Then, the perceived imge distortion between imges χ nd χ cn be quntified by function D(χ, χ ). Converse Tone Mpping (CTM) Problem: Given n inl imge χ nd mximum llowble imge distortion D mx, find the tone mpping opertion : [0, ψ L ] [0, L ] such tht L is minimized while mx mx mx D(χ, χ ) D mx (5) where χ ψ( χ ). The forementioned problem is the converse of the tone mpping problem, becuse in the tone mpping problem the gol of optimiztion is to find the mpping opertor Ψ such tht for given mximum disply luminnce, the imge distortion is minimized [9]. In contrst, in the CTM problem, the gol of optimiztion is to find the minimum of mximum luminnce vlue tht gurntees given mximum imge distortion level. Unfortuntely, due to complexity of HVS nd therefore the complexity of the imge distortion function, D, neither CTM problem nor the tone mpping problem hve closed form mthemticl solutions. 3.1 Preservtion of the Perceived Brightness To solve the CTM problem, this pper proposes heuristic pproch bsed on pixel brightness preservtion. The key ide is to mke sure tht the JND in the bcklight scled imge nd tht in the inl imge will be the sme. In other words, the imge perception is preserved, i.e. both imges hve the sme discernible detils. Mthemticlly speking, let L nd L denote the dpttion luminnce for the inl nd the bcklight scled imges. Bsed on eqution (1), the JND for the inl imge is LL ( ) nd the JND for the bcklight scled imge will be LL ( ). Therefore, to preserve the perceptible detils of the imge, one is required to find tone mpping function,ψ, such tht, ( LL ) = ψ ( LL ( )). As simple pproch, one cn ssume vrying scling function in which the scle fctor chnges depending upon locl luminnce. Then, the drker regions of the imge would hve been scled more linerly thn the light regions to compenste the decresing humn contrst sensitivity from drk to light regions of the imge (cf. Sec. 2.2). However, this pproch requires individul pixel mnipultion which is prohibitive in terms of performnce nd energy consumption for online bttery powered pplictions. Therefore, this pper doptsψ, to be constnt scling function ψ ( x) = κ x, where κ cn be clculted using eqution (1) s function of L nd L, κ = ( L ) ( L ) where L nd L could be pproximted by the hlf of the mximum bcklight luminnce before nd fter bcklight scling, i.e. 0.5L nd 0.5 mx mx L. Eqution (6) is n importnt result, becuse it is in contrst with the results in references [2], [3], nd [4], where κ is essentilly set s L / L for luminnce preservtion. mx mx To further fine tune the simple scling function of eqution (6) to cpture the humn contrst sensitivity (cf. Sec. 2.2Figure 2), this pper proposes to use similr functionl form for the trnsformtion function Ψ, s tht of the humn brightness perception function, i.e. (cf. eqn. 2), 5/2 γ ( L, L ) χ ψ ( χ ) = κ( L, L ) (7) L where κ ( L, L ) is simply the luminnce intensity djustment fctor s given by eqution (4) nd γ ( L, L ) is the humn contrst (6)

5 sensitivity chnge between the inl imge nd the bcklight scled imge, tht is σ γ ( L, L ) = (8) σ The motivtion behind introduction of prmeter γ ( L, L ) is to ffect lrge nd smll luminnces differently. More precisely, if only the κ ( L, L ) fctor ws used, in the trnsformed bcklight scled imge the contrst between two pixels would hve been incresed uniformly with respect to tht of the inl imge; however, with introduction of γ ( L, L ), s the contrst between two pixels in the inl imge increses the contrst between sme two pixels in the bcklight scled imge would increse but, grow more slowly for smller pixel luminnce vlues. Therefore, the result would be single tone mpping function which tkes into ccount the sensitivity sturtion of HVS (cf. Figure 2). 3.2 Imge distortion chrcteriztion To del with the complexity of imge distortion function, D, n pproch similr to tht of reference [4] is used. In this pproch first the imge distortion function is chrcterized for set of benchmrk imges s function of the dynmic rnge of the tone-mpped imges. Then, stndrd curve fitting tools is used to generte n empiricl imge distortion curve bsed on this dt. Lter, this empiricl curve is used s the imge distortion function D to find the minimum required dynmic rnge for ny given imge to chieve the mximum imge distortion of D mx fter tone-mpping. We dopted the universl imge qulity index proposed in [18] s our distortion mesure nd used set of benchmrk imges from the USC SIPI Imge Dtbse (USID)[19]. The USID is considered the de fcto benchmrk suite in the signl nd imge processing reserch field [20]. Figure 4 depicts the resulting distortion vlues for these imges when the dynmic rnge of the trnsformed imge is set to twelve different vlues. Figure 7 is subset of benchmrks reported to provide visul reference for the distortion mesure. Next, we used stndrd curve fitting tools provided in MATLAB version 7 relese 14 to find the best verge nd worst-cse globl fits to these distortion vlues. The result is n empiricl curve depicted in Figure 4, which mps trget dynmic rnge of trnsformed imges to the observed distortion vlues. 4 Experimentl Results The CCFL luminnce is complex function of the driving current, mbient temperture, wrm-up time, lmp ge, driving wveform, lmp dimensions, nd reflector design[1]. In our test-bed pltform only the driving current is controllble. Therefore, we model the CCFL luminnce s function of the driving current only nd ignore the other prmeters. Accounting for the sturtion phenomenon in the CCFL light source[1], we use two-piece liner function to chrcterize the power consumption of CCFL s function of normlized luminnce: A. β + C 0 β C lin lin s P ( β ) = bcklight (9) A. β + C C < β 1 st st s Figure 4. Imge distortion vs. Dynmic Rnge Reltionship between the CCFL luminnce nd the driver s power dissiption for the CCFL in LG Philips trnsmissive TFT-LCD LP064V1 [21] is shown in Figure 6. The CCFL illumintion increses monotoniclly s the driving power increses from 0 to 80% of the full driving power. For vlues of driving power higher thn this threshold, the CCFL illumintion strts to sturte. The sturtion phenomenon is due to the fct tht the incresed temperture nd pressure inside the tube dversely impct the efficiency of emitting visible light[1]. After interpoltion, we obtin the following coefficient vlues for the CCFL in LG Philips trnsmissive TFT-LCD LP064V1: C s =0.8234, A lin.9600, C lin = , A st =6.9440, C st = The hydrogented morphous silicon (-Si:H) is commonly used to fbricte the TFT in disply pplictions. For TFT-LCD pnel, the - Si:H TFT power consumption cn be modeled by qudrtic function of pixel vlue x [0,1] [22] 2 P ( x) =. x + bx. + c (10) TFT Pnel We performed the current nd power mesurements on the LG Philips, LP064V1 LCD. During these mesurements we set the CCFL bcklight luminnce to mximum nd displyed full screen rectngle with gryscle level equl to x, i.e. R=G=B=x for ll pixels on the screen. Then, the totl power consumption of the disply is recorded. Next, these power vlues re used to derive the prmeters of eqution (10). The mesurement dt re shown in Figure 5. The regression coefficients re thus determined s: = , b= , nd c= To show the effectiveness of pproch the power sving for different imges from USC SIPI dtbse is reported in tble 1. These power svings re generted for three different vlues of distortion levels. Clerly, by incresing the mximum tolerble distortion level the power sving should increse, which is lso confirmed with listed results. 5 Conclusions In this pper, dynmic tone mpping for bcklight scling with prespecified imge distortion level ws proposed. The proposed pproch ws bsed on mtching of perceived brightness vlues of the individul pixels in the inl imge nd those of the bcklight scled imge. Experimentl results showed the effectiveness of method. In future work lterntive distortion mesures nd histogrms equliztion methods will be evluted. 1 Actul mesurements Qudrtic fit Figure 5. Pixel trnsmittnce, vs. totl power consumption of the TFT-LCD pnel Figure 6. Normlized CCFL luminnce vs. power consumption

6 Originl Dynmic rnge=200 Dynmic rnge = 100 Originl Dynmic rnge=200 Dynmic rnge = 150 Normlized power Distortion=4.3% Power sving =36.19% Distortion0.6% Power sving =45.24% Normlized power Distortion=5% Power sving =35.72% Distortion2% Power sving =46.49% Normlized power Distortion=3.3% Power sving =37.16% Distortion=7.4% Power sving =48.28% Normlized power Distortion=3.6% Power sving =32.21% Figure 7.Smple imges nd their corresponding trnsformed versions Distortion=5.1% Power sving =42.57% Power sving (%) Nme Distortion = 5% Distortion = 10% Distortion = 20% Len Autumn footbll Peppers Greens Pers Onion Trees West Pout Sil Splsh Girl Bboon TreeA HouseA GirlB Testpt Eline Averge Tble 1. Power sving for different distortion levels References [1] J. Willims, A fourth genertion of LCD bcklight technology, Liner Technology Appliction Note 65, Nov [2] N. Chng, I. Choi, H. Shim, DLS: Dynmic Bcklight Luminnce Scling of Liquid Crystl Disply, IEEE Trnsctions on Very Lrge Scle Integrtion Systems, Vol. 12, No. 8, Aug. 2004, pp [3] W-C. Cheng nd M. Pedrm, "Power Minimiztion in Bcklit TFT- LCD by Concurrent Brightness nd Contrst Scling," IEEE Trnsctions on Consumer Electronics, Vol. 50, No. 1, Feb. 2004, pp [4] --, HEBS: Histogrm Equliztion for Bcklight Scling, Proc. of Design Automtion nd Test in Europe, Feb [5] ANSI/IES Nomenclture nd Definitions for Illuminting Engineering, ANSI/IES RP New York, NY: Illuminting Engineering Society of North Americ. [6] Rdiosity: A Progrmmer s Perspective by In Ashdown, October 2002 by Hert Consultnts Limited. (Originlly published by John Wiley & Sons in 1994.) [7] H.R. Blckwell, Contrst thresholds of humn eye, Journl of Opticl Society of Americ, No. 36, Vol. 11, Nov [8] S. S. Stevens, J. C. Stevens, ``Brightness Function: Effects on Adpttion'', Journl of Opticl Society of Amerik, vol. 53, No. 3, pp , Mr [9] J. Tumblin, H. Rushmeier, Tone reproduction for computer generted imges, IEEE Computer Grphics nd Applictions, Vol, 13, No. 6, Nov. 1993, pp [10] J.A.Ferwerd, S.N.Pttnik, P. Shirley, D.P.Greenberg, A model of visul dpttion for relistic imge synthesis, In Proc. SIGGRAPH 1996, ACM SIGGRAPH: Addison Wesley, pp [11] G. Wrd, A contrst-bsed sclefctor for luminnce disply, Grphics Gem IV, P. Heckbert Ed., Chp. VII.2, pp , Cmbridge. [12] K.Chiu, M.Herf, P.Shirley, S.Swmy, C.Wng, nd K.Zimmermn, Sptilly Non-uniform Scling Function for High Contrst Imges, In Proc. Grphics Interfce, MAY 1993, pp [13] L. Wrd, H. Rushmeier, C. Pitko, A visibility mtching tone reproduction opertor for high dynmic rnge scenes, IEEE Trnsctions on Visuliztion nd Computer Grphics, Vol.3, No. 4, Oct. 1997, pp [14] S.N.Pttnik, J.A. Ferwerd, M.D. Firchild, D. Greenberg, A multiscle model of dpttion nd sptil vision for relistic imge disply, In Proc. ACM SIGGRAPH, 1998, pp [15] M.Ashikhmin, A tone mpping lgorithm for high contrst imges, In Proc. of Eurogrphics Workshop on Rendering, 2002, pp [16] P. Choudhury, J. Tumblin, The trilterl filter for high contrst imges nd meshes, In Proc. of the Eurogrphics Symposium on Rendering, 2003, pp [17] F.Durnd, J.Dorsey, Fst bilterl filtering for disply of highdynmic-rnge imges, ACM Trnsctions of Grphics, Vol.21, No.3, 2002, pp [18] Zhou Wng nd Aln C. Bovik, A Universl Imge Qulity Index, IEEE Signl Processing Letters, vol. 9, no. 3, Mrch, [19] A. G. Weber, The USC-SIPI imge dtbse version 5, USC-SIPI Report #315, Oct Also dtbse/dtbse.html. [20] Digitl Imge Processing, Willim K. Prtt, Third Edition, John Wiley & Sons, [21] LG Philips, LP064V1 Liquid Crystl Disply. [22] H. Aoki, Dynmic chrcteriztion of -Si TFT-LCD pixels, HP Lbs 1996 Technicl Reports (HPL-96-19), Februry 21, 1996.