―117― J. Soc. Photogr. Sci. Technol. Japan, Vol.59, No.1, 1996
The Photographic Sensitivity of Electronic Still Cameras
Jack HOLM*
Abstract This paper presents an overview of electronic still camera photographic sensitivity (speed) determination procedures and concepts. The approach used for film speed and exposure determination is described, followed by a review of the work done in ISO TC42/ WG18 toward the development of electronic still and digital camera speed standards. The protocol developed for the determination of saturation based speed is given. The theoretical and experimental work on the signal- to- noise based speed concept is reviewed, including the subjective correlation between EIQ values, midtone log NEQ values, and the noise related subjective image quality. The effect of different pixel pitches is briefly described, and some common film EIQ values are noted in comparison to digital camera values and subjective perceptions. Preliminary protocols for noise based speed determination are provided, and the relation between noise based speed and DQE is noted. Preliminary work on color camera noise measurement is also outlined.
ture it. Consequently, the optical system of the Introduction camera should be set up to deliver the optimum The art and science of photography enable amount of radiation to the sensor. With film humanity to communicate, and record and systems, the primary characteristic used to explore the universe. In some respects the determine the appropriate amount of radiation is phrase "seeing is believing" applies to the view- the film speed. Films with higher speeds are ing of photographic images. Pictures can be more sensitive and therefore capture scenes in manipulated to mislead, but most often the inten- less time. tion of the photographer is to show a perception Film speed designation procedures developed of reality. In photojournalism and the sciences over a number of decades, beginning with the in particular, the intention is frequently to show work of Hurter& Driffield and several astro- reality with greater clarity than can easily be nomical photographers in the 19th century, and seen by the human visual system. Unfortunate- culminating with the work of Jones& Condit1 in ly, the universe which can be photographically the 1940's.* Current film speed designation pro- recorded is limited. One limitation is the num- cedures are under continual review by the Inter- ber of possible locations of the camera; another national Organization for Standardization is the sensitivity of the camera. If the amount (ISO), and are relatively stable. The speed of radiation to be recorded is too small to regis- designations resulting from these procedures are ter, a picture cannot be obtained. This limita- application dependent- it is possible for one film tion applies to photographs of couples by candle- type to have different speeds for different uses. light as well as quasars at the edge of the uni- * Although the paper by Jones& Condit is considered by many verse. In addition to defining the "photographable to be the pivotal paper relating to speed and exposure deter- mination, the work presented therein is based on a broad universe," sensitivity ratings are necessary for foundation of research conducted by a number of important efficient photography. The dynamic range of a photographers and photographic scientists, and has been scene can be quite large, and the full range of a subsequently refined and expanded by a number of individ- modern film or sensor may be required to cap- uals.
* Consultant 7352 East River Road , Rush, NY 14543- 9769, USA ―118― Jack HoLm J. Soc. Photogr. Sci. Technol. Japan
Since speed ratings are designed to be used with square meter, f# is the camera lens effective specific exposure determination systems, they f- number, t is the exposure time in seconds, are related to but different from absolute sensi-
phy, Working Group 18, Electronic still picture plane exposure
The following is a brief overview of the cur- 4. Calculate the speed constant k: rent film speed and exposure determination pro- cedures (more detail can be found in other (3) sources2, 3) 5. Determine the applicable film speed using 1. Calibrate the exposure meter so that when it the equation: is set on a particular exposure index (EI), it † will indicate camera exposure time and lens A statistically average scene is the ensemble average of the luminance distributions of all scenes ever photographed. A f- number settings according to the following number of studies have been conducted to estimate the char- equation : acteristics of a statistically average scene, and the following values are common for pictorial still photography : Scene luminance ratio 160: 1 Focal plane illuminance (or exposure) ratio 80: 1 Scene geometric mean reflectance 14.4% (1) Scene arithmetic mean reflectance 18% Scene highlight reflectance 128% (specular) In equation 1,
Electronic Still Picture Cameras (4) In electronic photography, the situation is
Note: If the film speed is determined using ISO different. The following quotes are from a
standard procedures, including the choice of Href and working draft of the electronic speed standard the constant k (which implies a specific •¢H), the under development: 4 "When an image from an electronic still pic speed value obtained is an ISO speed. -
6. Set the film speed on the exposure meter as ture camera is obtained using an insufficient the exposure index and determine the appro- exposure, proper tone reproduction can gener- ally be maintained by increasing the electronic priate camera exposure settings for each scene photographed. This step can be per- gain, but the image will contain an unacceptable formed automatically by many modern amount of noise. As the exposure is increased, cameras. the gain can be decreased, and therefore the image noise can normally be reduced to an Note: The value set on the exposure meter is the acceptable level. If the exposure is increased exposure index. This value is frequently, but incor- rectly, labeled the ISO [speed], because for normal excessively, the resulting signal in bright areas exposure determination the exposure index should of the image may exceed the maximum signal be equal to the ISO speed. The ISO speed, however, level capacity of the image sensor or camera is a characteristic of the film. The exposure index signal processing. This can cause the image is an index set on an exposure meter to determine highlights to be clipped to form a uniformly camera exposure settings. In some circumstances, bright area, or to bloom into surrounding areas the optimal exposure settings will be obtained using of the image. an exposure index which is different from the ISO speed. This is particularly true for scenes which The maximum exposure level is the exposure are substantially different from the statistically aver- level where typical picture highlights will be age scene, or when unusual film processing tech- clipped as a result of saturating the image sensor niques are employed. Non- standard processing signal capacity or reaching the camera signal techniques can also result in actual film speeds which processing maximum signal level. The mini- are different from the ISO speed. mum exposure level depends on the amount of With film exposure determination, the goal is to noise that can be tolerated in the image. These produce the most pleasing tone reproduction in situations lead to two different ISO speed rat- the final photograph, while maintaining the opti- ings, a saturation signal based rating and a noise mum image quality for other attributes such as based rating. signal- to- noise (S/ N) and detail reproduction. In photographic applications where the scene The behavior of photographic emulsions places illumination level can be controlled, such as in constraints on the film imaging process. studio photography, the photographer normally Compared to digital photography, the number of prefers to use an exposure index value which variables available to the film photographer is provides the best possible image quality. In this small. Film emulsions must be manufactured situation, the saturation signal based ISO speed with a limited number of grain distributions in rating is appropriate. This rating allows the order to assure both appropriate contrast and user to set the camera exposure so that the dynamic range. Film development choices are image highlights are just below the maximum limited by the need to maintain contrast at possible (saturation) camera signal value. specific levels while minimizing grain. The In many photographic applications, it is desir- optical effects found in projection systems, the able to use the lowest exposure possible, in order viewing environment, and the densities achiev- to maximize the depth of field, minimize the able in prints limit the output dynamic range. exposure time, and offer the maximum accept- To a large extent, film systems are physically able latitude for exposure of the image high- and chemically optimized for particular applica- lights. The ISO speed rating that provides an tions prior to exposure; the film exposure deter- acceptably low noise level for typical electronic mination scheme need only take advantage of camera applications is called the noise based this optimization. ISO speed rating of an electronic still picture ―120― Jack HOLM J. Soc. Photogr. Sci. Technol. Japan camera. Its value is based on an objective cor- saturation. relation to subjective judgments of the accepta- Some judgment is required for the determina- bility of various noise levels in exposure series tion of •¢H, but the procedure is still similar to images." the one used for film.•¢H is the expected ratio The approach taken by the ISO committee is of the highlight focal plane exposure to the to describe procedures for the determination of geometric mean focal plane exposure. It is fair- two types of speed ratings for electronic still ly well established that for statistically average picture cameras- saturation speeds and the scenes the highlight focal plane exposure is noise speeds. Recent work has indicated that approximately 9 times the geometric mean focal two noise speeds are important: the noise speed plane exposure, and the shadow focal plane that correlates with the minimum exposure exposure is about 1/ 9 the geometric mean focal required to produce an excellent photograph, plane exposure. and the noise speed that correlates with the A similar situation exists for black and white minimum exposure required to produce an pictorial negative film speed determination, acceptable photograph. However the existence where Href is the shadow exposure and •¢H is 1/ of three, or even two speed values for an elec- 10. Use of the factor 1/ 10 as opposed to the tronic camera could be confusing to some. A statistically expected factor 1/ 9 provides a small proposal put forward by the U. S. delegation to safety margin for metering errors and helps with ISO TC42/ WG18 is to designate the excellent scenes which have large luminance ratios. It is quality noise speed as the ISO speed, with the therefore reasonable to assume that the choice of saturation and acceptable quality noise speeds △H for saturation speed determination should be listed in parentheses to provide a speed latitude. approximately 10. If this proposal is accepted, the electronic cam- In reality, the preliminary value chosen by the era of the future will have a speed designation of ISO TC42/ WG18 committee for •¢H is 9.75. the form: ISO 200 (50- 800). This is because of several effects, which are as If these electronic camera speed ratings are to follows: be compatible with film exposure determination 1. The ceiling imposed by camera saturation is methods, they must be obtained using procedures harder than the floor imposed by fog in developed following the same rationale as was black and white negative films, encouraging used for the development of film speed determi- the use of a larger •¢H. nation procedures. This means that two quan- 2. Flare light tends to bring up the shadow tities must be determined, Href and •¢H, for each focal plane exposure values for scenes with speed. As with film systems, the determination large luminance ratios, thereby decreasing of these quantities is somewhat application the change in the shadow exposure values dependent. In the following discussion, empha- resulting from the large ratios. Flare, how- sis will be placed on the development of speed ever, does not affect the highlights, and rating procedures for pictorial still images, or increases in the scene luminance ratio photographs. directly affect highlight focal plane expo- sures. This effect also encourages the use Saturation Speed of a larger •¢H. The selection of the most appropriate Href and On the other hand: △H for saturation speed determination is 3. The perceptibility of highlight clipping straightforward. In fact the choice of Href is tends to go down as the scene luminance almost trivial, in that the speed values deter- ratio increases. This is because high mined should just prevent clipping of typical luminance ratio scenes tend to have small, image highlights. The critical focal plane expo- specular highlights which are featureless in sure is clearly the saturation exposure (Hsat). images of the scenes which use preferred Hsat can be determined experimentally by tone reproduction. measuring the focal plane opto- electronic con- 4. The existence of multiple speed ratings for version function (OECF)5 and calculating the electronic still picture cameras implies that minimum focal plane exposure that produces intermediate values can be used. Since the The Photographic Sensitivity of Vol.59, No.1, 1996 Electronic Still Cameras ―121―
purpose of the saturation speed is to provide tied together, but the nature of the subject com- a limiting value, it can be chosen to be closer plicates the correlation. For example, Metz, et. to the actual limit than a single, all- purpose al.12 illustrated that images of the same subject speed value. with the same IC can have quite different appear- These effects argue against increasing •¢H too ances, depending on whether the information much over the expected value. limiting factor is the number of allowed gray Saturation Speed Equation levels or the noise. The choice of Hsat for Href and 9.75 for •¢H Most photographic systems, however, are en- results in the following saturation speed equa- gineered so that the number of levels is large tion: enough to be imperceptible. In such systems one would expect a strong correlation between (5) perceived noise related image quality and the IC. This is to some extent the case if all other vari- This equation is used to calculate saturation ables are kept constant, and one approach to the speed ratings from experimentally determined determination of Href for noise speed is to make Hsat values. The U. S. proposal is for the satu- Href the focal plane exposure that produces a ration speed as determined using equation 5 to be specific image IC. Unfortunately, the formula the ISO lower speed limit. for the determination of IC is rather complicated: 8 Noise Speed
The saturation speed concept follows the same line of reasoning as is used for film speed deter- (6) mination. The noise speed concept, however, breaks new ground. Some previous research in In the above equation, Wsis the power spectrum the general area of S/ N metrics was available,6-8 of the signal and WNis the power spectrum of the but a great deal of work was required to corre- noise, both as a function of spatial frequency in late S/ N requirements with speed values. This the x and y coordinate directions (u and v). work was done largely by members of the ISO For film systems, this equation can be simplified TC42/ WG18 committee.9-11 to some extent by considering the isotropic case: The selection of •¢H and in particular Href values for noise speed determination requires the (7) development of an objective metric which is correlatable to subjective impressions of image Equations 6 and 7, however, do not afford a noise. Objective- subjective correlations are practical solution to the problem of determining always tricky, but this one is complicated further Href. They are difficult to apply, and are not by the fact that image noise perceptibility is appropriate for non- linear systems such as picto- affected by gain, spatial frequency, and gray rial films and electronic cameras. A simpler level in addition to the actual amount of noise metric is needed, but one that allows the system
present in the image. The desired objective gain to be a non- linear function of exposure. metric must take all these considerations into However, the form of equations 6 and 7 does account. suggest an approach based on the NEQ, since: Log NEQ and Noise Perceptibility For several decades, an interesting topic of discussion in the imaging science community has been the correlation between subjective graini- (8) ness, the noise level, the S/ N expressed in terms of density over density standard deviation, the In equation 8, Q is the magnitude of the input in signal to noise expressed in terms of the noise quanta, or whatever input unit is used in deter- equivalent quanta (NEQ), and the information mining the gain (g), and MTF is the modulation capacity (IC) of photographic films. It seems transfer function of the imaging system. Note reasonable for these different quantities to be that the middle part of equation 8 is similar to ―122― Jack HOLM J. Soc. Photogr. Sci. Technol. Japan the part of equation 7 in parentheses. DCS 200 is shown in Fig. 1. At this point it is helpful to make some obser- In this example, not only is the response of the vations about pictorial imaging systems. The camera non- linear, but four different overall first is that all but the very darkest tones in gain settings were used to obtain the data acceptable images will have NEQ values much presented. The log NEQ metric defined in equa- larger than one for image areas that constitute a tion 9 appears to be robust against overall gain visual picture element or pixel. The second is changes in addition to being well behaved in that most pictorial imaging systems have rela- other respects. Preliminary results also indicat- tively flat noise power spectra. Some sharpen- ed that it correlates well with subjective impres- ing may increase image quality, but there is a sions of noise in gray test patches. limit to how much is desirable, especially at the Note: The paper printed in the proceedings book cited as middle spatial frequencies where noise is most reference 9 does not contain all the results of the associat- visible. Consideration of equations 7 and 8 com- ed study, although these results were presented at the bined with these observations suggests the fol- conference. lowing form for a noise quality metric that is a Other Factors Affecting Noise Visibility function of exposure in lux- seconds (H), gain, Exposure dependent log NEQ values deal with and the output noise (ƒÐ): the effects of variable gain, but still do not take into account the effects of subjective noise spa- tial frequency and gray level dependence. In (9) order to perform a subjective correlation using photographs of real scenes, it is necessary to In this equation, a is the root-mean-square consider both of these effects. Work to this end fluctuation of the individual pixel output levels was performed,' and it was found that for most about the mean output level (the standard devia- current electronic camera systems the effects of tion) for a constant input exposure. Equation 9 variations in the noise power spectra are small ; implies that noise related subjective image qual- the spectra of cameras and high quality printers ity should correlate with per-pixel log NEQ tend to be relatively flat. Lower quality values. This log NEQ metric is similar to the IC printers do have features in their noise power metric since NEQ>>1 for most usable exposures. spectra which produce visible effects, but the It also allows for non- linear response by con- purpose of this work is to determine a metric for sidering gain to be a function of exposure. The camera measurement. Since printers are avail- base ten log is used for convenience sake, and able with flat spectra, they can be used to pro- differs from the base two log only by a constant. duce the prints for subjective studies. Having speculated on a metric, it is necessary The flatness of the noise power spectra is not to experimentally verify that it is both practical the only frequency dependent effect, however. and applicable. This verification was originally The bandwidth of the noise is also of impor- done using a Kodak DCS 200 monochrome digi- tance, and is determined by the pitch, or number tal camera,9 although other studies have been of pixels per unit length in prints. As the pitch conducted since. A plot of the data from the is increased, more and more of the noise is shift- ed to high frequencies where it is less visible. Consequently, prints made with a higher pitch will appear to have lower noise levels, even if the standard deviation of the digital files remains the same. This situation is analogous to the decrease in variance observed as the size of the scanning aperture (A) is increased in film mi- crodensitometry, as is characterized by the fol- lowing equation:
(10)
Fig. 1 Typical Digital Camera Log NEQ Per-pixel Data. In equation 10, G is a granularity parameter The Photographic Sensitivity of ―123― Vol.59, No.1, 1996 Electronic Still Cameras which has been demonstrated8 to be constant for as follows: most film samples over a wide range of aperture 1. Determine the following imaging system sizes. If G is also assumed to be constant over characteristics: a range of pixel pitches, as will be the case with a. the output level as a function of focal flat noise power spectra, this equation can be plane exposure (the focal plane OECF). rewritten to provide an estimate of the observed b the noise (ƒÐ) as a function of exposure noise for electronic imaging systems: (it may be desirable to normalize this value for a specific pixel pitch using (11) equation 11). c. the gain (slope of the OECF) as a func- where ail, is the observed noise at pitch Pi. In tion of exposure. effect, the MTF of the eye and the viewing d. the log NEQ as a function of exposure distance define a visual sampling "aperture." (using the exposure, gain, and noise as For this analogy to hold, it is necessary that both determined in steps a, b, and c in equa- P1 and P2 be high enough so individual pixels are tion 9). not distinguishable, or greater than about 6 These characteristics can be determined for pixels/ mm. at a viewing distance of 25cm. the camera only, if the printer used does not One fact that is rapidly becoming apparent in add an appreciable amount of noise. electronic photography, however, is that elec- 2. Conduct a four- quadrant tone reproduction tronic camera pixels can be made to have very analysis of each print for which an EIQ good S/N characteristics, but that dramatically value is desired. This analysis is necessary increasing the number of pixels is quite expen- to establish the relation between the focal sive. If, for example, a particular camera noise plane log exposures and the print densities. level needs to be cut in half, it will most likely be 3. Determine the appropriate gray level cheaper to reduce the per pixel noise levels by a weighting function for the output media to factor of two than to increase the number of be used. A typical weighting function for pixels by a factor of four. This situation tends black- and- white prints is shown in Fig. 2. to result in camera pixels being printed at a Most other gray level weighting functions limited range of sizes. In the first pictorial sub- will have similar shapes. jective correlation experiment performed using 4. Select a number of approximately equally the approach presented here, 11all prints were spaced log exposure values and determine made with a bit more than 6 pixels per mm. the corresponding log NEQ and print den- This pitch was determined to be the coarsest sity values. The larger the number of val- acceptable for the reproduction of average pho- ues selected the more accurate the EIQ val- tographic detail. Finer pitches allow for the ue determined will be. If the printer is not reproduction of finer detail with less visible a source of noise, it is not necessary to noise, but require significant increases in the consider its gain in determining the image number of camera pixels for a given print size. log NEQ values. Since the NEQ metric is a Gray Level Dependence and the EIQ Metric S/ N metric, any amplification applied in the Noise perceptibility is also quite dependent on gray level. Fairly high noise levels are almost invisible in dark tones, but relatively small amounts of noise are visible in the middle to light grays. For an objective metric to predict sub- jective impressions of noise in pictorial prints, it must take the gray levels into account. Weight- ed averages of log NEQ values must be taken over the range of log exposures used to capture the image. These averages, which for conve- nience sake will be designated as equivalent image quality (EIQ) values, can be determined Fig. 2 EIQ Gray Level Weighting Function. ―124― Jack HOLM J. Soc. Photogr. Sci. Technol. Japan
printer will be applied equally to both the Table 1 The Relationship EIQ to Perceived Noise signal and the noise, leaving the NEQ value Related Image Quality. unchanged. 5. Determine discrete gray level weighting function coefficients based on the print den- sity values of interest, and normalize the coefficients based on the number of discrete log NEQ values to be averaged (the sum of the weighting coefficients should be unity). Multiply the normalized weights by the cor- responding log NEQ values. Add the results to determine the EIQ for each print. Note: If some of the tones in a print are clipped (the log exposures fall in the saturation region of the camera), the NEQ values will be undefined (the noise and gain will both be zero). In this case, it will be necessary to use an arbitrarily selected log NEQ value with the weights assigned to the clipped tone densities. Experimental work has suggested that a log NEQ value between- 1 and- 2 is appropri- ate; the more negative the value, the sharper the falloff in the EIQ due to clipping. In the work described in the paper" An EIQ- Subjective Image Quality Correlation Study," 11 Fig. 3 Typical Electronic Camera EIQ Curves. the procedure described above was performed for prints from exposure series of three different study were, for practical reasons, limited to four scenes. The tone reproduction of each print images around the optimum. It is possible, how- was digitally optimized so that all prints of the ever, to calculate EIQ values over a much wider same scene had the same, excellent tone repro- range. Fig. 3 shows the results of such calcula- duction. All other factors affecting the per- tions for the three scenes used in the study. In ceived image quality were held constant through these calculations, a value of- 2.0 was used as the use of monochrome prints (to eliminate color the log NEQ for clipped tones. errors) and the same camera and printer (to The mean log exposures that produce the keep the system MTF constant). highest quality images with these scenes vary Subjective evaluation of the prints was perfor- because the geometric mean exposures** and med using both the category scale and paired luminance ranges vary. If a scene has a prepon- comparison techniques. The correlation derance of light tones (a high- key scene) or dark between the EIQ values and both subjective tones (a low- key scene), the peak quality mean methods was extremely high, with the EIQ val- log exposure will be shifted. Large amounts of ues frequently correlating better with the two light tones move the mean log exposure up from subjective ratings than they did with each other. the expected value, and large amounts of dark The prints used in this study were 14 by 21cm, to tones move the mean log exposure down. This produce a diagonal of 25cm. (the standard effect is offset slightly by the fact that the tones viewing distance). As stated previously, the in high key scenes should be printed a bit dark, pixel pitch was slightly higher than 6 pixels per and the tones in low key scenes should be printed mm., thereby producing the largest acceptable a bit light, but the shift in the preferred tone print size for a given number of pixels, and the reproduction is small compared to the shift in the greatest noise sensitivity. The correlation mean log exposure. Also, higher luminance between the EIQ values and the category scale ranges result in lower peak mean log exposures descriptions used in this experiment is provided ** The geometric mean exposure is the antilog of the mea in Table 1. n log The exposure series used for the subjective exposure. The Photographic Sensitivity of Vol.59, No.1, 1996 Electronic Still Cameras ―125―
Table 2 Minimum and Maximum EI Limits Based on Geometric Mean Exposure.
Table 3 Minimum and Maximum EI Limits Based on Scene Illuminance.
because the midtone must be placed at a lower reflector in the scene at the desired midtone log log exposure to allow additional "headroom" for exposure. The midtone log exposure is equal to the extended highlights. Lower luminance the mean log exposure plus 0.1 for a statistically ranges result in higher peak mean log exposures average scene. EI values for incident meters because not as much headroom is needed. The can be calculated from scene illuminance values EIQ values plotted in Fig. 3 behave as the subjec- using equation 1. Scene illuminance values can tive quality would, to the extent that the distribu- be calculated from midtone log exposure values tion of tones in the image is "normal". The if the midtone reflectance values are known. exposure indices required to produce EIQ values The scenes used in the study were measured greater than or equal to 2.0, but without clip- carefully to determine a number of their charac- ping, are also reasonable, and are listed in Table teristics, including the midtone reflectance and 2. mean log luminance. It is therefore possible to In Table 2, the El values are calculated direct- calculate the differences between the EI values ly from the
Table 4 Maximum Acceptable Prints Sizes for Prints objective metrics, and should therefore be from Several 35mm Kodak Films Predicted straightforward, but is complicated by the fact from EIQ= 2.0 Values. that many scenes are different from statistically average, and some are quite different. Also, actual subjective gray level weighting factors are somewhat dependent on the output media and viewing conditions. Ideally, the log NEQ used to specify the Href for noise speed should be chosen to make Href as robust as possible against variations in the scene characteristics and output can be performed by applying the metric to film media. In going back to a single log NEQ val- systems. If the metric produces reliable infor- ue, it is desirable to consider all factors. mation for such a completely different type of In addressing this issue, the ISO committee has system, one can have more confidence that it will determined that the preliminary Href values continue to produce reliable information for the should be the focal plane exposures (Hs/ N) that systems of primary interest. To this end an produce log NEQ values of 3.4 for excellent experiment was performed where the EIQ metric quality, or 2.0 for acceptable quality, and that was used to predict the maximum acceptable these exposures should be placed at a middle enlargement sizes, based on a minimum EIQ gray image tone. In order to simplify things value of 2.0, for several black- and- white Kodak further and eliminate confusion for those who 35mm negative films. 9 are not familiar with the NEQ metric, the desig- These values assume the standard viewing dis- nation of HS/ N is also done by specifying the tance and corresponding "visual effective aper- actual S/ N values, as opposed to the log NEQ ture,"even though the different sized prints values. This means that the reference expo- might be viewed at different distances. This sures are the focal plane exposure that produce assumption is valid because the 25cm. viewing S/ N values of 50 or 10, since the S/ N value is the distance is the average minimum comfortable square root of the antilog of the log NEQ.•õ•õ distance, and is therefore the distance most com- The determination of the other critical value, monly used for critical examination. Photogra- △ H, is straightforward, since it has already been phers with experience using the films listed in established that a middle gray image tone (as Table 4 can confirm that the maximum enlarge- would correspond to an 18% reflectance gray ment sizes predicted are consistent with actual card or Munsell N5 patch) corresponds to a practice. focal plane exposure that is 1.25 times
provided to eliminate confusion as to the exact even if the measured value is less, but it remains definition of S/N applicable: to be seen if this requirement is sufficient to prevent posterization. No cameras are current- (13) ly available that combine very low noise with coarse quantization. In fact, all of the above limitations are avoided if electronic cameras are Speed Rating Limitations well designed. The speed rating procedures described in this Another potential problem with noise speeds paper are the best that the group of experts concerns the effects of different sharpening al- involved could put together. They are probably gorithms and pixel pitches. A number of effects as good as those used for film, although only the are possible: passage of time can truly determine if this is the 1. The optimal amount of sharpening for a case. They do posses a few limitations of which particular rendering of an image depends on the user should be cognizant. First, if a scene is the noise and the pixel pitch. significantly different from statistically average a. As image noise decreases, the optimal in its luminance range or distribution of tones, it amount of sharpening increases, may be necessary to use an EI different from a although not beyond the optimal particular speed rating to produce the desired amount for a noise free image. result. This situation also exists with film sys- b. As pixel pitch increases, the subjective- tems, and has been dealt with by photographers ly optimal amount of sharpening for a number of years, however the behavior of increases, especially at high spatial fre- electronic systems is different than film systems. quencies. The high spatial frequency Different actions are required to compensate for modulation boost compensates for the a few types of scene irregularities. Photogra- falloff of the eye MTF. phers may learn when electronic systems require 2. Equation 11 only approximates the effect of different actions, and what they are. A more different pixel pitches on noise visibility. likely scenario is that the flexibility of digital The exact relationship is more complicated. systems, combined with the possible intelligence a. Sharpening, especially optimal sharpen- of electronic cameras, will allow the camera to ing with low noise and at high spatial deal with unusual conditions automatically. frequencies, can result in significant fea- The proposed upper and lower speed limits could tures in the noise power spectra which facilitate this approach. may reduce the accuracy of equation 11. Several limitations are also unique to the noise b. Increasing the pixel pitch increases the speed concept. The first is that some electronic subjective quality ceiling by increasing cameras may not be able to achieve a S/N of 50 the potential for detail reproduction. at an exposure that is sufficiently below the Since subjective quality is limited by saturation exposure to allow it to be placed at a the weakest image attribute, increasing midtone. A proposal is that such cameras the pixel pitch can shift the limiting would indicate the saturation speed as the pri- factor from the detail reproduction to mary ISO speed, with the upper speed limit in the noise, even though the noise may be parentheses, i.e. ISO 100 (400). In the unlikely less visible at the higher pixel pitch. event that a camera could not even achieve a The above effects are worthy of note, but should midtone S/N of 10, only the saturation speed not interfere with the utility of the speed ratings. would be listed, i.e. ISO 200. At the other It is impossible to consider all factors in a single extreme, it is possible that a digital camera rating, or even multiple ratings. With film sys- could have such a low noise level with respect to tems, speed ratings are designed to be as robust the quantization interval, that at the upper speed as possible. Decades of use testify that this goal limit posterization (contouring) would be vis- has been achieved to a large extent. The elec- ible. The current working draft specifies that tronic still camera speed ratings described in this the noise levels used for S/N calculations be paper are also designed to be robust against greater than or equal to than 0.5 digital levels, factors that cannot be accounted for. ―128― Jack HOLM J. Soc. Photogr. Sci. Technol. Japan
of the luminance required (meter- candles or Ultimate Photographic Speed lux, 3000•‹K) to obtain a satisfactory picture The following section is quoted from Image from a subject in which the highest light is a
Science by Dainty& Shaw. 8 perfectly reflecting matt white surface. "A symposium organized by the Royal Photo - The exposure time is to be 0.02 second and graphic Society on "The Ultimate Sensitivity of the subject is to be imaged originally with a Photography Today and Tomorrow" was held in working aperture of f/ 2. London during December 1960. 14 As part of this There were eight detailed responses to this chal- symposium, a challenge was made to a group of lenge, these coming from Benarie, Berg, Fellgett, scientists- representative of most theoretical and MacAdam, Meyer, Rose, Schade, Vendrovsky practical aspects- to predict the future limits of and Sheberstov..." photographic sensitivity. A set of rules ac- Dainty& Shaw later note that:"Although it is companied this challenge, and since the answers difficult to answer the question of ultimate sensi- provided have become classical in this respect, tivity in terms of conventional speed ratings due these rules are stated here in full: to the necessity of specifying a criterion of 1. The photographic system can be optical, acceptable image quality, the problem is rela- electronic, etc., but it must yield a reason- tively simple in terms of DQE. Without further ably permanent print, say 9•~12cm, of a investigation it is already possible to postulate normal pictorial subject. that ultimate sensitivity corresponds to 100% 2. The system must have a usable latitude of DQE, with individual grains, or receptors, as big 1.2 log exposure units and the final copy as are permissible according to whatever practi- must be capable of a full reflexion density cal criterion is adopted." range up to a maximum of density= 1.2. The noise speed criterion discussed in the
3. The information content of the picture previous section provides the missing piece should not be inferior to that given by a 625 required to answer the ultimate speed challenge. line television system under optimum, The criterion of acceptable image quality is as closed- circuit conditions. A good tele- follows: a log NEQ of 2.0 per pixel at the middle
vision picture of this sort can be taken as the gray image tone, with 6 pixels per mm. on the general guide for quality of the image. print. Since this criterion has now been estab- 4. No practical restrictions are placed whatso- lished, it is possible to determine ultimate speed ever on the sensitized materials and equip- ratings based on pixel size and DQE. These ment used. For example, emulsions can be ratings are determined as follows, and presented imagined of perfectly ordered grains of any in Table 5. desired size and quantum efficiency, and the Photons per lux- second per square micron pigment of each grain can have any desired (Q): covering power. Reciprocity failure, fad- (14) ing, etc., can be ignored. 5. The maximum sensitivity of which the sys- (calculated for a 3100K Planckian radiator with tem is capable should be expressed in terms an IR blocking filter).
Table 5 Upper Noise Speed Limits Resulting From Various Pixel Sizes and DQE's. The Photographic Sensitivity of Vol.59, No.1, 1996 Electronic Still Cameras ―129―
Area (A) of a square pixel with linear dimen- It is possible to obtain many of the values sion L in microns: required for color electronic camera speed deter- mination using this luminance channel. Hsat (15) can be determined, as can g (H). A value can Photons per pixel area A per lux-second (Qp): also be obtained for the luminance channel noise σy (H), but this value will represent only the (16) luminance noise. The chrominance or color Minimum acceptable S/N ratio (NEQmin): difference noise also contributes to noise percep- tion. (17) Work toward determining a noise metric Photons per pixel required to produce NEQmin related to color noise perception is just begin- (Qs/ N): ning, but the initial approach is to model the color noise ƒÐc (H) using the equation: (18)
Lux- seconds required to produce NEQmin (22) (HS/N= 10) where ƒÐ (R- y)(H) and ƒÐ (B- y)(H) are the standard deviations of the R- Y and B- Y color difference channels. Note that equation 22 reduces back (19) to the monochrome form if only a luminance channel is present, thereby making the equation
So the upper noise speed limit, expressed in general. If this model is appropriate, and if terms of DQE, is: values for c1 and c2 can be determined, the devel- opment of noise speed rating procedures for color electronic cameras can be completed. (20) One approach to the determination of the c1 and c2 values is to artificially create test patches with varying noise levels, and try to estimate Color Cameras which patches have the same perceptual noise. So far, all of the material presented in this This approach is facilitated by the fact that all paper has dealt with monochrome images. the patches can be printed at the same gray level, Color images make up the vast majority of picto- and by the fact that all that is needed is a ratio of rial images, but color noise is difficult to define. the noise visibility between the Y, R- Y, and B- The preliminary approach chosen by the ISO Y channels. To this end, the following experi- committee is to define quantities in terms of a ment was performed : luminance (Y) channel whenever appropriate. 1. Sample midtone patch digital files were Such a channel can be constructed from RGB constructed and varying amounts of Gaus- image files by applying the following equation sian noise were added in each of the RGB from the ITU- R BT.709 digital TV standard: color channels. A matrix of combinations was produced with different amounts of (21) noise in each channel combined with
Table 6 Channel Visual Threshold Standard Deviations (8 pixels/ nm.). ―130― Jack HOLM J. Soc. Photogr. Sci. Technol. Japan
different amounts of noise in the other chan- procedures of this type will further the develop- nels. ment of digital photography. 2. The noisy digital files were then printed on a Acknowledgments very low noise color printer (Fuji Pictrogra-
phy) at 8 pixels per mm. Since only ratios I would like to thank all my colleagues on the are needed, the exact tone reproduction and ISO TC42/ WG18 committee for many interest-
pixel pitch characteristics are not important ing discussions and contributions related to the as long as they are reasonable and remain material presented in this paper. In particular, constant. Ms. Sabine Sstrunk has collaborated on many of 3. The visible noise threshold was determined my personal efforts and Mr. Ken Parulski (the for each color channel through visual evalu- ISO project leader) has been instrumental in the ation, and the actual noise present in the development of electronic still picture camera threshold patches was measured. The speed rating procedures. Mr. Tatsuji Kitamoto, results are presented in Table 6. Mr. Andrew Juenger, and Mr. Egbert Buhr have 4. Since the coefficient of the ƒÐY (H) term in also done considerable work, and have provided equation 22 must be unity for the mono- valuable insights on several matters. chrome noise to fall out, the coefficients in References question can be determined by taking the ratio of the Y to the R- Y and B- Y standard 1. L. A. Jones& H. R. Condit,"The Brightness Scale deviation values. This results in the follow- of Exterior Scenes and the Computation of Correct ing values being obtained for c1 and c2: Photographic Exposure," J. Opt. Soc. Amer., 31, p. 651-678 (1941). 2. J. Holm,"Exposure- Speed Relations and Tone Re- (23) production," Proceedings, IS& T's 47th Annual Conference, Vol.II, p.641- 648,Rochester, NY, 1994. 3. T. H. James, ed., The Theory of the Photographic (24) Process, Kodak (1977). 4. ISO 12232: Photography-Electronic still picture These values are preliminary, but hopefully will cameras-Determination of ISO speed, Working Draft be reasonably close to the final values deter- # 7 (preliminary), June 6, 1995. 5. ISO/ CD 14524 (December 20, 1995): Photography- mined. Electronic still picture cameras- Methods for measur- Addendum: Video Cameras ing opto-electronic conversion functions (OECF's) 6. R. Shaw,"The Equivalent Quantum Efficiency of the It is worthy of mention that the noise speed Photographic Process," J. Photogr. Sci., 11, p.199- rating procedures discussed in this paper are for 204 (1963). electronic still picture cameras. Video systems 7. D. Zwick,"Quantitative Studies of Factors Affecting have different noise perceptibility considera- Granularity," Phot. Sci.& Engr., 9: 3, p.145- 148 (1965). tions, in particular the time averaging of the 8. J. C. Dainty& R. Shaw, Image Science, Academic noise by the human visual system. Although Press, London (1974). the approach presented here could also be 9. J. Holm,"A Log NEQ Based Comparison of Several applied to video systems, the log NEQ values Silver Halide and Electronic Pictorial Imaging Sys- required for excellent and acceptable quality are tems," Proceedings, IS& T's 46th Annual Confer- most likely considerably lower. ence, p.35- 37, Cambridge, MA, 1993. 10. J. Holm,"Log NEQ Based Pictorial Print Noise Conclusions Characterization," Proceedings, IS& T's 47th Annual Conference, Vol.II, p.429- 432, Rochester, In the past five years, a great deal of progress NY, 1994. has been made toward the establishment of pro- 11. J. Holm& S. Susstrunk,"An EIQ- Subjective Image cedures for the determination of electronic still Quality Correlation Study," Proceedings, IS& T's picture camera speed ratings. These speed rat- 47th Annual Conference, Vol.II, p.634- 640, Roches- ings will be fully compatible with film speed ter, NY, May 1994. ratings, since the same philosophical approach 12. H. J. Metz, S. Ruchti,& K. Seidel,"Comparison of Image Quality and Information Capacity for was applied. It is the hope of the author that The Photographic Sensitivity of Vol.59, No.1, 1996 Electronic Still Cameras ―131―
Different Model Imaging Systems," J. Photogr. Sci., Photogr. Sci. 9, p.247. (Introduction by G. I. P. 26, p.229- 233 (1978). Levenson; the Rules; Contributions from M. M. 13. L. Stroebel, J. Compton, I. Current,& R. Zakia, Benarie, W. F. Berg, P. B. Fellgett, D. L. MacAdam , Photographic Materials and Processes, Focal Press, R. Meyer, A. Rose, O. H. Schade, K. V. Vendrovsky Boston (1986). and V. I. Sheberstov; Summary by E . W. H. 14. "Towards the Ultimate Speed in Photography," J. Selwyn.)