A DISPLAY INDEPENDENT HIGH DYNAMIC RANGE TELEVISION SYSTEM

Transcription

1 A DISPLAY INDEPENDENT HIGH DYNAMIC RANGE TELEVISION SYSTEM T. Borer an A. Cotton BBC R&D, 56 Woo Lane, Lonon, W12 7SB, UK ABSTRACT High Dynamic Range (HDR) television has capture the imagination of the broacast an movie inustries. This paper presents an overview of the BBC s Hybri Log-Gamma solution, esigne to meet the requirements of high ynamic range television. The signal is isplay inepenent an requires no complex mastering metaata. In aition to proviing high quality high ynamic range (HDR) pictures it also elivers a high quality compatible image to legacy stanar ynamic range (SDR) screens an can be mixe, re-size an compresse using stanar tools an equipment. The technical requirements for a high quality HDR television system are presente. Quantisation effects (or baning ) are analyse theoretically an confirme experimentally. It is shown that quantisation effects are comparable or below competing HDR solutions. The psychovisual effects of presenting images on emissive isplays in im environments, system gamma, is shown experimentally an the analysis informs the esign of this HDR system. INTRODUCTION With improvements in technology, television with greater impact, more presence, eeper immersion, a wow factor, or, in short, better pictures, is now possible. Ultra high efinition, UHD, is not just about more pixels, it has the potential to eliver wier colour gamut (WCG), higher frame rates (HFR), an higher ynamic range (HDR). Of these perhaps high ynamic range offers the greatest improvement, an the costs of upgraing to HDR can be relatively low for both prouction an istribution. High ynamic range offers unmistakeably better pictures, across the living room, even on a smaller isplays, an even to those with less than perfect vision. Potentially HDR may be prouce using mostly installe, legacy, stanar ynamic range (SDR) infrastructure, an istribute over largely unchange istribution networks. No woner that, even before the stanars have been finalise, some movie stuios are alreay talking about creating movies in HDR, Ultra HD for home viewing. This paper escribes the signal processing technology require for high ynamic range in the television prouction an istribution chains. It escribes how one solution, the hybri log-gamma approach, provies a isplay inepenent signal that can prouce high quality images, which maintain the irector s artistic intent on a wie range of isplays in iverse viewing environments. So, for example, precisely the same signal may be viewe in a controlle prouction suite, a home cinema, an orinary living room, or on a laptop or mobile evice. Furthermore the signal may be isplaye on a conventional stanar

2 ynamic range isplay to provie a high quality compatible image. The log-gamma HDR signal may be mixe, re-size, compresse, an generally prouce, using conventional tools an equipment. The only specifically high ynamic range equipment neee is cameras an isplays for quality monitoring (signal monitoring may continue to use SDR isplays). No complex mastering metaata is require. Conventional en user istribution techniques may be use (although a 10 bit signal is require). No layere or multichannel coecs are require. Only a single signal is require for both SDR an HDR isplays an expensive multiple graes (for both HDR an SDR) are not necessary. The paper continues by iscussing the meaning of high ynamic range. To unerstan HDR prouction an isplay we nee to look at the television signal chain, which is iscusse next. This then allows us to consier the esign of the camera transfer characteristic (the opto-electronic transfer function (OETF)). Next we iscuss an important psychovisual aspect of HDR TV, the system gamma. Whilst this effect has long been known in television, movies, an the acaemic literature, it assumes an enhance importance for HDR. Base on an unerstaning of system gamma we iscuss the esign of the electro-optic transfer function (EOTF) in the isplay, an how this can allow the isplay of high quality pictures on a iverse range of isplays. Once the EOTF is efine we can analyse the likely effect of quantisation an the performance, in terms of ynamic range, that may be expecte from the system. This is compare to alternative HDR proposals an the theoretical analysis is compare to experimental results. The paper ens with some concluing remarks. DYNAMIC RANGE Dynamic range is the ratio between the whitest whites an blackest blacks in an image. For example printe images have a ynamic range of less than 100:1 (because it is ifficult to make a black ink that reflects less than 1% of incient light). Dynamic range is often measure in stops, which is the logarithm (base 2) of the ratio. So printe images have less than 7 stops of ynamic range. Stanar ynamic range consumer television (8 bit vieo, e.g. DVD, SD an HD DVB) only supports about 6 stops of ynamic range, as iscusse below. Professional SDR vieo (10 bits) supports about 10 stops. But the human eye can see up to about 14 stops (1) of ynamic range in a single image. Higher ynamic range results in an experience closer to reality, an hence of greater impact or immersion. Furthermore higher ynamic range also increases the subjective sharpness of images an so provies a ouble benefit. Some ebate has confuse high ynamic range with high brightness. The two are not the same. You can have high ynamic range in a ark movie environment, with pictures of only 48c/m2. Alternatively you can have stanar ynamic range on very bright screens of hunres, or even thousans, of c/m2. What high brightness oes allow is to see high ynamic range without neeing a very ark viewing environment. It might be thought that higher ynamic range coul be achieve by simply making isplays brighter. But this is analogous to suggesting that you can increase the ynamic range of auio by turning up the volume. With auio, turning up the volume merely emphasises the noise. The same is true for vieo. With vieo the noise is quantisation noise, where the steps between quantisation levels become clearly visible. This is manifest as baning, contouring, or posterisation. An extreme example of baning is shown below.

3 Figure 1:Image Quantisation, left original, right extreme baning The useful ynamic range of vieo is etermine by the ratio between ajacent quantisation levels. If the ajacent quantisation levels iffer in luminance by less than 2% the ifference is probably imperceptible in the image. This threshol of visibility increases at low luminance. The Schreiber curve 1 (shown later) is an approximation to the threshol of visibility. In television systems the luminance represente by the signal is a non-linear function of the signal value. Conventionally this is a gamma curve, exemplifie by ITU-R Rec For example the isplaye luminance L may be the signal V raise to the power gamma, L=V. The ratio between ajacent quantisation levels, also known as the Weber fraction, is given mathematically by: 1 L 1 Weber fraction 1 NV V N L Where N is the number of quantisation levels (220 for 8 bit vieo, 876 for 10 bit), an the algebraic form is given for a conventional gamma curve. Using this equation we may fin the luminance corresponing to the threshol at which baning is visible, which then gives the usable ynamic range. For conventional, 8 bit SDR vieo, with gamma 2.4 an a 5% threshol (allowing for a im isplay), this yiels a ynamic range of only 5.27 stops. This is not a high ynamic range, a bit less than a photographic print, although it can be extene by using techniques such as ither. Overall the conventional isplay gamma curve is not aequate for HDR reprouction an a ifferent non-linearity is require. TELEVISION SIGNAL CHAIN The television signal chain, shown below, intentionally inclues signal non-linearities in both cameras an isplays. The camera non-linearity is known as the opto-electronic transfer function (OETF) an the isplay non-linearity is known as the electro-optic transfer function (EOTF). Colloquially an confusingly in conventional television both transfer functions are known as gamma curves. As is well known the EOTF is not the inverse of the OETF, so overall the signal chain has a non-linear (or system ) opto-optic transfer function (OOTF). The OOTF compensates for the psycho-visual effects viewing pictures on an emissive isplay in im or ark surrounings, an is sometimes known as renering intent (2). This is iscusse in more etail below. 1 Schreiber measure the threshol of contrast visibility experimentally (3). Schreiber s results are broaly consistent with experiments on vieo quantization reporte by Moore (4), an also with the DICOM moel (5), which itself is erive from the Barten moel (6)

4 Originally the OETF, in combination with the CRT isplay EOTF, was esigne to make the effects of camera noise more uniform at ifferent brightnesses. In igital systems the non-linearities help to minimise the visibility of quantisation (or baning ). But, by moifying these non-linearities, it is possible to further reuce the effects of baning an so increase ynamic range. Figure 2: Television signal chain In this conventional moel of the television chain, the relative luminance in the scene is capture by the camera an encoe in the signal. The light from the scene efines the signal. The EOTF then reners the light from scene so that, subjectively, it appears the same as reality. Historically, with CRT isplays, isplay brightness was fairly consistent because CRTs simply coul not be mae very bright. This allowe a single EOTF to be use to rener the signal for all CRT isplays. With the availability of a plethora of bright isplay technologies ifferent EOTFs are neee to ensure that pictures look subjectively the same on isplays of ifferent brightness. The approach escribe in this paper allows the signal to be renere on any isplay (OLED, LCD, local backlight imming, or quantum ot), preserving the irector s artistic intent, without the nee for metaata an without neeing to re-grae for ifferent isplays. That is, the hybri log-gamma approach efines a signal that is inepenent of the isplay. THE HYBRID LOG-GAMMA OPTO-ELECTRONIC TRANSFER FUNCTION (OETF) In the brighter parts an highlights of an image the threshol for perceiving quantisation is approximately constant (know as Weber s law). This implies a logarithmic OETF woul provie the maximum ynamic range for a given bit epth. Proprietary logarithmic OETFs, such as S-Log, Panalog an Log C are, inee, wiely use. But in the low lights it becomes increasingly ifficult to perceive baning. That is, the threshol of visibility for baning becomes higher as the image gets arker. This is known as the De Vries-Rose law. The conventional gamma OETF comes close to matching the De Vries-Rose law, which is perhaps not coinciental since gamma curves were esigne for im CRT isplays. So an ieal OETF woul, perhaps, be logarithmic in the high tones an a gamma law in the low lights, which is essentially the form of the hybri log-gamma OETF. The ynamic range of moern vieo cameras is consierably greater than can be conveye by a vieo signal using a conventional gamma curve (i.e. ITU-Rec 709). In orer to exploit their full ynamic range conventional vieo cameras use a knee characteristic to exten the ynamic range of the signal. The knee characteristic compresses the image highlights to prevent the signal from clipping or being blown out (overexpose). A similar effect is also a characteristic of analogue film use in traitional movie cameras. When a hybri gamma HDR vieo signal is isplaye on a conventional SDR isplay the effect is similar to the use of a igital camera with a knee or using film. It is not surprising therefore, that the hybri gamma vieo signal is highly compatible with conventional SDR isplays, because what you see is very similar to the signal from an SDR camera. Inee the knee

5 characteristic of the hybri gamma characteristic, efine below, is conservative, proviing only 300% overloa. A hybri gamma signal is efine as: r E 0 E 1 E OETF: a ln E b c 1 E where E is proportional to the light intensity etecte in a camera colour channel (R, G, or B), normalize by the reference white level. E is the non-linear, or gamma correcte, signal, where the non-linearity is applie separately to each colour channel. The reference value of E is 0.5, enote r, an correspons to reference white level. Constants, a= , b= , c= , are efine so that the signal value is unity for a (relative) luminance of The hybri log-gamma OETF is shown below alongsie the conventional SDR gamma curve an a knee characteristic. Note that the horizontal axis for the hybri log-gamma curve, as efine above, has been scale to emphasise compatibility with the conventional SDR gamma curve. Furthermore, because the hybri log-gamma signal only escribes the light representing the scene, it is inepenent of the isplay. Consequently, with a suitable EOTF, it may be use with any isplay. Figure 3: Hybri log-gamma an SDR OETFs Figure 4: Hybri log-gamma an SDR OETFs SYSTEM GAMMA AND THE OPTO-OPTIC TRANSFER FUNCTION (OOTF) As is well known, an note above, the light out of a television isplay is not proportional to the light etecte by the camera. The overall system non-linearity, or renering intent (2) is efine by the opto-optic transfer function, or OOTF. Renering intent is neee to

6 compensate for the psychovisual effects of watching an emissive screen in a ark or im environment, which affects the aaptation state (an hence the sensitivity) of the eye. Without renering intent pictures woul look too bright or washe out. Traitionally movies were, an often still are, shot on negative film with a gamma of about 0.6. They were then isplaye from a print with a gamma of between 2.6 an 3.0: This gives movies a system gamma of between 1.6 an 1.8, which is neee because of the ark viewing environment. Conventional SDR television has an OOTF which is also a gamma curve with a system gamma of 1.2. But, for HDR, the brightness of isplays an backgrouns will vary wiely, an the system gamma will nee to vary accoringly. In orer to etermine the necessary system gamma we conucte experiments viewing images with ifferent gammas at ifferent luminances (an with a fixe backgroun luminance). The pictures were erive from HDR linear light images selecte from Mark Fairchil s HDR Photographic Survey. A reference isplay (Dolby PRM4220) an a test isplay (SIM2) were place about 1 metre apart, in a controlle viewing environment (room illumination 10Lux, D65). The reference image, shown at 600c/m 2 on the reference isplay, was chosen by participants from 9 images with ifferent gammas (1.0 to 2.4). This allowe them to choose the artistic effect they preferre. The participants then choose a picture on the test isplay, from 9 ifferent gammas (1.0 to 2.4) that best matche the reference image. The test images were shown at ifferent luminances on the test isplay. The results, illustrate below, provie an estimate of the preferre system gamma, (excluing artistic preferences), at a range of isplay brightnesses from 68 to 5200c/m 2. Whilst only a small number of participants were involve, an further experimental results woul be most welcome, the results are quite consistent an provie a goo guie to the necessary system gamma for ifferent isplay brightness relative to backgroun luminance. Figure 5: Preferre system gamma versus normalise isplay brightness These empirical results may be approximate by the following formula for system gamma, where Y represents luminance.

7 1 Ypeak 1 log 10 5 Ysurroun The results clearly show that the en-to-en system gamma of the HDR TV system has to be ajuste to accommoate isplays of iffering peak luminance. They suggest that, with a backgroun luminance of 10c/m 2, an OLED isplay of aroun 1000 c/m2 woul require a system gamma of aroun 1.4, whilst a brighter LCD of a few thousan c/m2 woul require a system gamma closer 1.5. Whilst these variations in system gamma appear small, they have a significant impact on the subjective appearance of an image. THE HYBRID LOG-GAMMA ELECTRO-OPTIC TRANSFER FUNCTION (EOTF) In orer to specify the complete television system we nee an EOTF as well as the OETF efine above. This maps the relative light representing the scene to the light emitte from the isplay. The EOTF shoul perform this mapping 1) whilst preserving the artistic intent of the programme maker (an proviing a suitable renering intent), 2) allowing for the ynamic range of the isplay from black level to peak white, an 3) minimising quantisation artefacts. The EOTF efine below is similar to the conventional isplay gamma curve, thereby maximising backwar compatibility, whilst also meeting the three preceing requirements; EOTF: Y Y where Y is the luminance of a pixel presente on the isplay, Y s is the relative luminance representing the scene for that pixel, an is the system gamma iscusse above. Parameters, an correspon to similar parameters in ITU-R Rec 1886, which are the traitional contrast an brightness controls respectively. They etermine the peak isplaye luminance an the minimum luminance, i.e. the black level 2. The EOTF maps the linear scene luminance, Y s, to the linear isplay luminance, Y. This iffers from current practice for SDR, which applies the EOTF to each colour component inepenently. But applying the EOTF to each component changes the saturation, an to a lesser extent hue, of the picture. Since the EOTF nees to change with the isplay it must be applie to luminance to avoi inconsistent colours. Scene luminance Ys may be recovere from the signal by first applying the inverse of the OETF to each colour component R', G', an B' to yiel the linear colour components R, G an B. With the same nomenclature as the OETF; Inverse OETF: E exp E E r c 2 s / a b 0 E r r E From the linear colour components the scene luminance may be erive as follows (assuming ITU-R Rec 2020 colorimetry); R G B Y s Scene Luminance: 12 Note that the factor of 12 in the enominator is because the signal normalisation for the OETF yiels a maximum value of 12 for each linear colour component, rather than the 2 LP LB LB where L P is the isplaye luminance for peak white (Y s = 1.0), an L B is the isplaye luminance for black (Y s = 0.0).

8 more conventional value of 1. Having etermine the linear scene luminance the isplaye luminance may be erive from the EOTF, where parameters,, an epen on the isplay an the viewing environment. Given the isplaye luminance we still nee to etermine the iniviual R, G, an B values that shoul be isplaye for each pixel. We obtain these simply by scaling the linear scene colour components as follows: Displaye Colour Components: R B R G G B Y 12Ys Y 12Ys Y 12Y where R, G, B, are the luminances presente on the isplay. Minimising the visibility of quantisation, or baning is an important aspect of the EOTF. We can estimate its visibility by calculating the weber fraction an comparing it to the Schreiber limit, as iscusse above. Doing so an plotting the results yiels the flowing graph for Weber fraction versus isplaye luminance. The EOTF may be use for isplays with ifferent peak luminance an black levels. This example assumes a 10 bit signal, with peak white of 2000 c/m 2, a black level of 0.01 c/m 2, an a system gamma of 1.5. This correspons to the use of bright isplays for programme monitoring an graing, viewe in a ark environment relative to the brightness of the isplay. It represents a ynamic range of 200,000:1 or 17.6 stops, which is more than the ynamic range the eye can perceive in a single image. s Figure 6: Weber fractions versus isplay luminance For comparison this graph also inclues the Schreiber limit, the conventional SDR gamma curve (Rec 1886), an an alternative HDR, a perceptual quantisation curve (PQ 10K) efine in SMPTE ST Baning is likely to be visible when the Weber fraction for an

9 EOTF is above the Schreiber limit. With a 10 bit signal this inicates that for the hybri gamma EOTF baning will, at worst, be at the threshol of visibility across the whole luminance range an similar to or below that of the PQ curve. It also shows that the conventional gamma curve is not aequate, with baning expecte to be visible below 20 c/m 2. Note that this analysis is for a 10 bit signal. With a 12 bit signal, which has been propose for HDR prouction, Weber fractions woul be much lower an baning woul be significantly below the threshol of etectability across the whole luminance range. To confirm the theoretical analysis we performe experiments on the comparative visibility of baning. Highly critical 10 bit, horizontal, shallow ramps, with ajacent patches varying by 1 quantisation level, were compare to a continuous reference 3 using the ITU-R Rec 500 ouble stimulus impairment scale metho. 33 subjects were teste. The test images were isplaye on a Dolby PRM 4220 monitor configure (using a custom internal LUT) to emulate the low tones of a isplay with 2000c/m 2 peak luminance an black level of 0.01c/m 2. Each horizontal grey scale ramp occupie 25% of screen height, an inclue 20 ajacent grey levels each spanning 1/24 th picture with. The experimental results are shown below. In these results -20 is one grae of impairment. The error bars inicate the 95% confience intervals. Figure 7: Baning impairment versus isplaye brightness These results appear to closely corroborate the theoretical analysis. Both the hybri gamma an ST 2084 are less than or about 1 grae of impairment, which is escribe as imperceptible or just perceptible but not annoying at their very worst. Furthermore the hybri gamma EOTF shows marginally more baning than ST 2084 in the region of 1c/m 2, an marginally less baning elsewhere, which is in line with the theoretical analysis. The impairment for conventional SDR gamma (ITU-R Rec 1886) rises from 3 The reference was a carefully ithere 10 bit signal in which baning was unetectable.

10 imperceptible to perceptible but not annoying below 20c/m 2, up to slightly annoying in the region below 1c/m 2, also in line with the theoretical analysis. Overall the hybri gamma curve provies acceptable baning performance at 10 bits for highly critical material, equivalent to that of ST 2084, for a isplay with 17.6 stops of ynamic range. In practice it is highly unlikely that any baning will be visible on naturally occurring scenes. CONCLUSIONS This paper has presente the rationale an esign for a HDR television system. The hybri gamma approach can support a range of isplays of ifferent brightness, without metaata, an so is isplay inepenent. A 10 bit signal is substantially compatible with conventional SDR signals. Shown unprocesse on an SDR isplay the picture is of high quality an so may be use for signal monitoring. This also means prouction can use existing SDR infrastructure, tools an equipment. Only quality monitoring requires an HDR isplay. No metaata is require, thereby simplifying the prouction chain. These features mean SDR prouction may be upgrae to HDR at relatively moest cost. Only a single signal is require for both SDR an HDR isplays an expensive multiple graes (for both HDR an SDR) are not necessary. Consequently layere, or multichannel, coing for en users is unnecessary, thereby simplifying istribution an minimising cost. For a 2000c/m2 HDR isplay, with a black level of 0.01c/m2, i.e stops ynamic range, it is shown, both theoretically an experimentally, that quantisation artefacts ( baning ) will not be visible on real pictures an that baning is comparable, or less, than the competing, more complex, ST 2084 HDR system. Finally note that HDR Rec 2020 signals may be formatte to look like conventional SDR Rec 709 signals, raising the possibility of conventional SDR meia carrying HDR signals in a completely compatible way. This will be the subject of a future paper. REFERENCES 1. Kunkel, T. an Reinhar, E., A reassessment of the simultaneous ynamic range of the human visual system. Proceeings of the 7th Symposium on Applie Perception in Graphics an Visualization. ISBN , pp July Poynton, C., Digital Vieo an HD.Morgan Kaufmann; 2n eition eition (2 Dec 2012). ISBN-13: Schreiber, W. F., Funamentals of Electronic Imaging Systems, Thir Eition: Springer-Verlag, 1992, ISBN: Moore, T.A Digital Vieo: The number of bits per sample require for reference coing of luminance an colour-ifference signals. BBC Research Department Report. BBC RD 1974/ DICOM, DICOM.Digital Imaging an Communications in Meicine (DICOM), Part14: Grayscale Stanar Display Function. National Electrical Manufacturers Association, PS3.14, Barten,P.G. J., Formula for the Contrast Sensitivity of the Human Eye. Proceeings SPIE-IS&T, 5294: , Jan ACKNOWLEDGEMENT We woul like to thank Manish Pinoria of BBC R&D for his work on this project, without which this paper woul not have been possible.