2.5 ISO

Photographers have long been able to control the brightness of an exposure (i.e., a photo) by carefully choosing an appropriate shutter speed and aperture to accommodate the luminance level (i.e., amount of light) available in a scene.  (Shutter speed, aperture, and luminance are discussed in much greater detail in Chapter 6).  Since light levels in the field can change drastically with different locations, different angles, and different times (even as the sun goes in and out behind a cloud), a photographer needs to be able to juggle these two parameters (shutter speed and aperture) so as to avoid overexposure or underexposure in the different field scenarios.  Of course, doing this effectively requires some skill, and in Chapter 6 we’ll cover the background material needed for developing that skill.
    With the advent of digital imaging, photographers now have a third parameter at their disposal for adjusting exposure: the ISO setting.   Although film photographers have long been able to adjust ISO by changing film types (e.g., taking the ISO 100 film out of the camera and replacing it with a roll of ISO 400 film, for example), in today’s DSLR’s you can now change ISO from one photo to the next by simply turning a dial on your camera before taking the next shot.  Our purpose in this section is to explain how ISO works in modern digital cameras and to illustrate the importance of this feature for producing high-quality images of birds, so that you can make a better informed decision next time you buy a camera.  Techniques for adjusting exposure in the field via ISO settings, shutter speeds, and apertures are deferred to Chapter 6, though the information covered here should serve as a useful foundation for understanding the reasoning behind those techniques when we cover them.

2.5.1 What is ISO?

Just as with the acronym SLR, knowing what the letters in ISO stand for is unimportant.  What’s important to understand is that the ISO setting mimics the effect of different film sensitivities.  Thus, low ISO settings mimic a low-sensitivity grade film, so that a longer exposure is needed to collect enough light to get a bright image.  Conversely, a high ISO setting mimics a high-sensitivity grade film, in which more light would be accumulated per unit time, allowing you to capture a bright image with a faster shutter speed or smaller aperture.  In this way, you can think of the ISO dial on your camera as being a brightness dialif all other settings of your camera are kept constant, increasing the ISO will increase the brightness of the resulting image.  Unfortunately, it can also increase the amount of noise in your image. 
    Exactly how much the noise level will be affected by an increase in ISO depends on a number of factors, but the magnitude of this effect differs markedly between camera models.  Whereas consumer-grade DSLR cameras can, as of today, produce relatively noise-free images at ISO 400 or less, and can often produce acceptably noisy images at ISO’s as high as 800 or maybe 1000, pro-level DLSR’s today can produce images with remarkably little noise at ISO 800 and images with very tolerable amounts of noise at ISO 1600, or even 3200 for the full-frame models with large pixels.  Indeed, whereas the manufacturers appear to be battling each other for larger Megapixel ratings in the consumer segment, they are to some degree battling over better high ISO noise characteristics in the semi-pro and pro markets.  It’s useful to consider the types of
tricks available to manufacturers for improving the ability of their cameras to produce low-noise images at higher ISO’s.

Fig. 2.5.1: How ISO setting affects exposure.
Only the aperture and shutter (at right) affect the amount of light
reaching the sensor.  After photons are converted to electrons by
the silicon matrix of the photosites, the ISO amplifiers artificially
increase the signal, but they also increase any noise that is present

    The first trick to be aware of is the noise-reduction-via-software trick.  Some manufacturers have been accused of trying to hide the low quality of their CMOS sensors by applying a noise reduction filter in the camera, after the image has been captured by the sensor and before it has been written to the memory card.  The problem with doing this is that an aggressive noise reduction filter can end up reducing the detail at the same time that it reduces the noise.  Thus, you may end up with images having very little noise, but also very little detail (or sharpness).  This does a great disservice to the users of that company’s cameras, since it would be much better to allow those users to manually apply noise reduction themselves after downloading the image to their computer, where they can choose the amount of noise reduction to apply, so as to retain the level of detail needed.
    A more
honest approach is for manufacturers to try to improve the sensor’s innate ability to produce low-noise images even when operated at higher ISO settings, without the need for subsequent noise-reduction filters, and this is precisely what the big manufacturers are trying to do.  Before we can survey their current efforts on this front, however, we need to dispel one very popular myth regarding ISO: namely, that higher ISO settings on a digital camera increase the sensor’s sensitivity to light and therefore the amount of light captured, per unit time, by the sensor.  This is a very good analogy to what happens when a film photographer changes to a higher-ISO film in order to collect more light in dim scenes.  Unfortunately, the analogy is wrong.
    In digital cameras, light is collected by photodiodes made of silicon
the same material used to make integrated circuits for computers.  Photons of light striking the (intentionally slightly impure) silicon lattice cause electrons to be jarred free and to flow to metal leads where they are then channeled to and stored in capacitors for eventual release to the further imaging circuitry in the camera.  The quantity of electrons collected from a photodiode in this way correlates very closely with the number of photons striking the pixel.  Since different pixels are constrained so as to receive only red, green, or blue photons of different intensities, the electron counts of different pixels can later be interpolated so as to infer composite color hues for individual pixel elements in the resulting image.  Thus, for example, an overabundance of light waves in the blue part of the spectrum striking a photosite will cause the blue component for that pixel’s hue measurement to be dominant, and the resulting pixel in your image should reflect this by showing a similar hue.
    The key here is that all of this happens irrespective of the ISO setting that you’ve dialed into the camera
at least, the parts of the story involving electrons flowing in direct proportion to the number of photons striking the photosite.  The sensitivity of silicon atoms to photons is in no way affected by the position of your ISO dial; it’s determined entirely by the physical properties of silicon, and by the concentration of impurities within the silicon lattice (called the doping ratio).  Where the ISO dial does have an effect is when the signal is later sent through an analog amplifier circuit, as shown in Fig. 2.5.1.  The ISO setting is, in fact, used to determine the amount of amplification applied to the signalalso know as the gain.  The problem is that the amplification step amplifies both the signal and the noise, so that higher ISO settings will necessarily result in images that are both brighter and have a larger (absolute) amount of noise.
    In order for manufacturers to reduce the amount of noise inherent in high-ISO images produced by their sensors, they therefore need to produce sensors having lower noise even at low ISO settings, since increasing the ISO simply amplifies both the signal and the noise.  This is somewhat of an oversimplification, because what really matters is the signal-to-noise ratio (often written as simply S/N), but for our purposes, it will suffice to consider that the battle for manufacturers is against noise, period, not against some phantom noise source known as
high ISO noise.  In order to examine how manufacturers are battling noise, we need to first consider the sources of noise, which we do next.

2.5.2 Types and Sources of Noise

First, let’s recall what noise is.  It’s the pixels in your image that are the wrong color, or the wrong intensity (brightness).  How does that happen?  There are many possible sources of noise, but two seem to dominate. The first we’ve already mentioned in a previous section: sampling error.  When the pixels are small, and/or the exposure time is very, very short, the actual count of photons striking each photosite may be very small.  Since each photosite in a typical Bayer-type sensor arrangement only registers one of the primary colors (red, green, or blue), the actual pixel colors have to be reconstructed by interpolation (basically, averaging) between the red, green, and blue measurements of the physical photosites corresponding to that pixel.  With small sample sizes (i.e., when the actual number of photons collected is small), that interpolation step can suffer from measurement error, and the actual color that is computed for that pixel may differ to some degree from the correct color for that pixel.  The amount of error in the measurement can vary, and for some pixels there will naturally be more error than in others, due to statistical variation.  The ones with the most error stand out most prominently, and if there are lots of those erroneous pixels, then the image appears noisy.  This type of noise is called photon noise.  When photographing in low light with a camera having a very high pixel density (i.e., very small pixels), photon noise can be a serious problem.
    The other main type of noise is called read noise, and is due to electromagnetic interference (and, to a lesser degree, heat) from the electrical circuitry associated with each photosite, such as the ISO amplifier (mentioned above) and the analog-to-digital converter (ADC), which converts electron counts into digital information (i.e., bits and bytes).  Whereas the pixel density (or, more precisely, photosite size) can be a fairly good predictor of a sensor’s tendency toward sampling error, there is no simple way to guess the amount of read noise that will be produced by a sensor’s electronics, based solely on the manufacturer’s spec.  The amount of read noise produced by a sensor can be affected by a multitude of design decisions made by the manufacturer, so that the only practical way to determine the amount of read noise produced by a particular camera model is to perform controlled tests, in which the camera is operated at different ISO settings and the resulting photos are compared to images from another, well-established model that can serve as a baseline for comparison. 
    These types of comparisons are in fact fairly easy to find on the internet, at least for the major brands.  Whenever a new model is released by a major brand, well-known photo enthusiasts (such as Ken Rockwell or Rob Galbraith) and equipment review sites (such as DPReview.com or FredMiranda.com) will typically put the camera through a number of tests and post the comparisons online, and these very often include ISO performance
i.e., comparison of noise levels at different ISO settings, relative to some well-known and highly popular camera from one of the top brands.  It’s important, however, to read these reviews critically: even if the reviewer or site is a reputable one, skim through their description of their methodology to make sure they’re comparing RAW images that haven’t had any noise reduction filters applied via software.  Comparisons based on JPG/JPEG images are, in my opinion, almost worthless for most purposes, since the JPEG’s produced by most digital cameras have been highly processed by in-camera software.  The noise levels observed in JPEG’s do not reliably indicate the noise characteristics of the camera’s sensor, since manufacturers are free to process the JPEG’s in-camera with aggressive noise reduction filters that both reduce noise and reduce detail/sharpness.  These types of comparisons should always be performed based on RAW images, and should also make use of manufacturer-approved (or supplied) RAW converters, ideally operated with matching, vanilla settings (i.e., no noise-reduction filter).  When I encounter a review that doesn’t say that it was performed according to a protocol like that just outlined, I generally stop reading.
    Note that in the field of astrophotography, another source of noise which doesn’t so much affect bird photography is thermal noise, also sometimes referred to as dark noise, or dark-current noise.  This is noise that results from heat generated by the imaging sensor.  In astrophotography, exposure times can be very long
measuring in the minutes or hours rather than fractions of a second as in bird photography.  Keep this in mind when reading camera reviews: any review published by an astronomy web site or other special-interest group (such as for medical imaging) is likely to be biased in ways that might not apply to bird photography.
    In summary, noise appearing at higher ISO’s isn’t so much caused by the ISO amplification process as revealed by it, meaning that the noise (or the potential for noise) was present at lower ISO’s but was masked by confounding factors such as the use of longer exposures (to collect more photons and reduce sampling error).  Thus, bona-fide technological advances which improve
high ISO noise are really improving the overall noise characteristics of the sensor, which affects all ISO levels, not just high ISO.  Special camera features such as the so-called High ISO Noise Reduction are usually based on in-camera software that both reduces noise and reduces image detail/sharpness.  The most effective way to prevent noise in all settings is to use a camera with large photosites that collect high volumes of photons, so that sampling error, and the resulting noise, don’t occur in the first place.  Minimizing other sources of noise (e.g., electromagnetic interference from camera circuitry) can be accomplished by engineering tricks such as the back-illuminated CMOS technology that we briefly mentioned in section 2.3.4.
    Finally, note that when we delve into postprocessing techniques in Part III of this book, we’ll be classifying types of noise according to a different schema, in which we’ll treat luminance noise (manifested as pixels of the wrong intensity, or brightness) different from chrominance noise (pixels of the wrong color).  Chrominance noise (often called simply chrom noise) can usually be well-controlled in software without reducing image detail/sharpness, whereas algorithms for reducing luminance noise typically obliterate image detail if applied too liberally.  In Chapter 11 we’ll describe in great detail methods for eliminating luminance noise in background regions of an image without affecting detail in the bird.

2.5.3 Preventing Noise Before it Happens

There are several methods worth mentioning at this point, for preventing noise (as opposed to reducing noise after the fact).  Some of these are manufacturing techniques and some are techniques applied by the camera operator (you).  The manufacturing techniques are useful to know about if you’re shopping for a camera, because you can read up on the models you’re considering and try to find out if the manufacturer has employed any of these techniques in the design of that model.  We’ll start with those first.
    As we’ve remarked several times already, the best way to prevent noise in the first place is to collect more photons.  In section 2.3.4 we described the microlenses that are positioned over the individual photosites on the sensor, which channel more of the photons into the photosensitive region of the photosite, resulting in fewer photons being lost (recall that many photons strike regions of the sensor that are not photosensitive, such as the electrical wires connecting each photodiode to the rest of the imaging circuitry).  Recent designs by Canon (and likely others as well) have featured what the manufacturers are calling gapless microlenses.  Early microlens solutions apparently utilized microlenses that didn’t cover the entire area of the photosite.  These newer designs presumably cover all, or nearly all, of each photosite with a microlens, so that nearly 100% of the photons entering the sensor get channeled into a photodiode.  With fewer photons
falling between the cracks, sample sizes at individual photosites should improve, thereby reducing overall photon noise.

Fig. 2.5.2: Gapless microlenses.  Gapped microlenses (A) allow
some light to be lost, because it misses the photosite.  Gapless
microlenses (B) capture nearly all of the incoming light and
channel it toward the photosensitive region of the substrate

    Another recent technological development that was discussed in section 2.3.4 is the use of back-illuminated CMOS, in which the attendant circuitry for each photodiode is moved out of the light path and will therefore no longer interfere with the effective absorption of incident photons.  Exactly how popular this latter technology will become, and the degree to which it is obviated by the more effective use of microlenses, remains to be seen.
    Yet another way to reduce sampling error (i.e., photon noise), in the case of sensors with an overabundance of pixels, is to combine photon counts from several nearby photosites so as to improve samples sizes for photon measurements.  This technique is called binning (or pooling of samples), and has been known for some time, but has remained impractical until just recently, for two reasons.  First, binning drastically reduces the number of pixels in the final image, typically by a factor of two or four, so that it’s really not practical for general-purpose imaging with an image sensor having fewer than about 12 MP (for two-way binning) or 24 MP (for four-way binning).  Second, the use of a standard Bayer pattern (a particular arrangement of red, green, and blue photosites on the sensor) made binning difficult, because same-color photosites aren’t next to each other on the sensor, so binning them tended to blur the image by combining photon counts from pixels some distance apart.  A recent attempt (by camera maker Fujifilm) to resolve the latter issue involved rearranging the colored photosites so as to place same-color photosites next to each other (unlike in the Bayer pattern), so that pairs of adjacent photosites could be combined.  As a result, users now have the choice of operating the camera in a high-resolution (12 MP) mode for bright scenes, or at a lower resolution (6 MP) when light is scarce, with the binning in the low-res mode helping to reduce the photon noise in poorly-lit conditions.
    One issue which I’ve only rarely seen addressed, as of yet, is the so-called full-well capacity of the photodiodes, which is simply the number of electrons that can be stored in the photosite.  The issue is that each photon typically releases one electron in the silicon matrix of the photodiode, so that increasing the light transmission to your sensor will increase the number of arriving photons but may not increase the number of electrons actually counted at the appropriate photosite if the upper limit (the full-well capacity) has been reached.  Additional photons will, presumably, continue to liberate electrons from the (impure) silicon matrix to flow through the induced electrical field of the photodiode, but without sufficient numbers of electron holes (places for those electrons to reside) many electrons may continue along the direction of the field and eventually end up in the well of another photodiode.  In this way, photons striking one photosite can end up being counted at another photosite, resulting in yet another type of noise.  Keep an eye out for reports of advances involving improved well capacities, or that otherwise limit electron travel between photosites, since these may fundamentally improve baseline noise characteristics of image sensors and therefore improve high-ISO performance.  A related avenue for possible future improvement relates to a sensor’s quantum efficiency, or sensitivity to different wavelengths of light.
    One method which deserves special mention here (though we’ll re-iterate this message numerous times throughout this book) for reducing noise, or more precisely, for improving signal-to-noise ratio (S/N), is a technique applied by users of digital cameras rather than by manufacturers.  The method is called exposing to the right (abbreviated ETTR) and results in more bits of information being used to represent your image.  This method can reduce the incidence of certain forms of
noise originating in the digital domain.  The idea is simply to adjust the settings on your camera so as to produce images that are as bright as possible without clipping any highlights.  Because brighter pixels are (on average) represented using more bits in the image file, less information is lost during discretization (analog-to-digital conversion, or ADC) than if you had exposed for a darker image.  During conversion of the RAW image to JPEG, you can then reduce the image brightness back to a more natural-looking level without sacrificing information, since you’ll ideally be working in a 16-bit color space.
This technique and its benefits are discussed in greater detail in Chapter 6.  Here we’ll just note a few things related to the use of ISO to perform ETTR.  First, if you can perform ETTR (exposing to the right) by adjusting only the shutter speed and/or aperture, while keeping ISO very low, you’ll be both maximizing the number of bits allocated to your image’s detail and at the same time minimizing photon noise.  Exposing to the right via shutter speed and aperture (but not by using higher ISO’s) reduces photon noise by allowing the sensor to collect more photons, thereby reducing sampling error and ensuring a more accurate measurement of color information from the incoming light.  Remember, collecting more light allows a more accurate measurement of the individual colors making up that light.  If instead you expose to the right by increasing the ISO setting, you’ll be getting the benefit of better bit utilization (since you’ll have brighter pixels, which are generally allocated more bits in the RAW file than darker pixels), but you won’t be reducing photon noise at all, since you’re not collecting more light (you’re just amplifying the signal after it’s already been measured by the sensor). 
    Just remember that ISO doesn’t affect the sensitivity of the silicon atoms in your sensor to photons; it simply multiplies the photon counts after the photons have already been counted.  When applying the ETTR technique, if you have to do it by increasing the ISO, that’s OK: you’ll still get some benefit due to better bit utilization.  But if you can do it via shutter speed and/or aperture instead of higher ISO, you’ll also be reaping the benefits of lower photon noise.

2.5.4 Other Costs of High ISO

Besides increased noise, there may be other costs to the use of higher ISO settings in a particular camera model, and you should be careful to find out about these before finalizing your assessment of any camera model you’re considering purchasing.  A common cost of high ISO is a reduction in buffer size (sometimes called burst rate).  The buffer is where images are stored between the time they’re captured and the time they’re written to the memory card.  Because images from today’s cameras are typically very large (often well over 10 Megabytes), and because writing to memory cards is typically very slow, the images captured by your camera can take several seconds to be written to the card.  They need some place to be stored before being written to the card, and that place is the buffer.  If you quickly take another photo while the first photo is still being written to the card, it too has to be stored in the buffer.  When shooting action scenes (such as birds in flight), it’s common practice to continuously hold down the shutter-release button so that the camera takes a series of shots in very rapid succession (up to about 10 frames per second for today’s pro camera bodies).  Because today’s memory cards can’t accept images at that speed, larger buffers are needed in today’s cameras for storing the images taken during intense bursts of action, till they can be written to the memory card.  For technical reasons, the use of higher ISO settings sometimes forces the camera to reduce the available buffer size for action shooting.  For example, the camera that I currently use, the Canon EOS 1D Mark III, reduces the available buffer size by 2 to 6 shots when shooting at ISO 640 or greater.  In intense action scenarios, I’ve occasionally run up against the buffer limit, and had to stop shooting (i.e. lost potentially great shots) until the buffer cleared, as previously-captured images were in the process of being written to the memory card.  Though this delay can be reduced somewhat by using faster (and more expensive) memory cards, it’s still worthwhile to investigate whether the camera you’re thinking of buying inflicts a buffer-size penalty for shooting at higher ISO’s, and how significant that penalty is.