Last update: 4/9/09

From the number of emails I receive and posts on various digital camera user groups, it seems that many digital camera users still don't know how the primary function of their camera works. It doesn't help to have representatives from camera companies:

• Make untrue claims ("Our camera doesn't interpolate, you get the real pixels").
• Obscure details in marketing claims ("Our camera produces 12-megapixels").
• Fail to understand basic optic theory ("The smaller sensor size produces 1.5x magnification, making all your lenses longer in focal length").
• Make misleading statements about competitors ("They use a CMOS sensor, which is noisier than a CCD sensor").
• And the list is seemingly endless.

It's important to understand how sensors function if you want to get the best possible results from your camera. So I'll step you through what happens, and the issues associated with a number of practical problems you'll encounter.

What's the Beef?
Digital cameras neither respond to light the same as film, nor does every sensor used work exactly the same. However, the basic concepts used by all digital cameras are similar, and understanding them can help you generate better pictures.

The lens in a camera focuses light on a plane behind the rear element. In 35mm cameras, that plane contains film. In digital cameras, the plane is occupied by a piece of silicon (chip) which I'll refer to as the sensor. Sensors come in two primary types: CCD (charge-coupled device) and CMOS (complimentary metal oxide).

A few words about CCD versus CMOS before we continue: CCD is the older and more mature technology. It is easy to design and produce. CMOS, however, has the advantages of being less expensive to manufacture in quantity and able to contain more complex internal electronics at each photosite. In general it takes more money and a longer time to design a CMOS sensor, but then you get the advantage of lower cost and higher capability. Long term, CMOS is the likely candidate for most sensors. Nevertheless, CCDs are mature and well-known, and they do have a baseline noise production that is lower than CMOS, all else equal. Nikon has shown yet another variation in sensor design: a CMOS variant that uses a different transistor type (the LBCAST sensor used in the D2h and D2hs), so we will continue to see a stream of new acronyms as technology and material use moves inexorably on in the low levels of each sensor. Personally, I don't put much weight into whether a product is CCD or CMOS. I've seen good and bad iterations of both.

On this sensor are an array of light-sensitive spots, called photosites. Photosites are usually square in shape (currently have been two major exceptions that I'll deal with in a moment), and laid out in rows and columns.

The first thing that catches newcomers to digital cameras unaware is this: the light-sensing portion of the photosites--called the photo diode--do not necessarily cover the entire surface area of the sensor. With some sensors the active light gathering area is as little as 25% of the total surface area of the chip. Yes, that means that there is non-light responsive space between adjacent photo diodes. (Pictures you might see of sensors that show adjacent red, blue and green positions are usually of an array of microlenses and filters that lie on top of the actual sensor, and not the photosite and embedded photo diode.)

I said there have been exceptions to the square photosite rule. The most important for Nikon users are:

• The Nikon D1x. The D1x takes two adjacent square photosites and doubles them into a rectangular photosite.
• The Fujifilm SuperCCD (used in the Fujifilm S1, S2, S3, and S5 bodies). The SuperCCD uses quasi-octagonal photosites. Fuji further places the "grid" for the photosites at an angle, though for all practical purposes it still has rows and columns of photosites, just tilted at a 45 degree angle.

You might wonder just how large the individual photosites are. On the original D1, they were 11.8 microns square, which is quite large (though technically, this was a group of fou "binned" photosites that worked together). On a Coolpix 990, they were 3.45 microns square, which is considered small. The tendency has been towards smaller photosites, partly because photo diode efficiencies have gotten better. Thus, current DSLRs tend to have photosites in the 5 to 8 micron square range, while current compact cameras like the Coolpix can have photosites as small as about 2 microns square (some camera phones have even smaller photosites).

Note that it's the area that's important. A 3 micron photosite has 9 square microns of area (of which only a portion may be sensitive to light). An 8 micron photosite has 64 square microns of area, or almost an order of magnitude more area. That turns out to be fairly important for per-pixel noise control, as we'll find out shortly.

top | home

The light-sensing portion of individual photosites don't cover the entire surface area of a chip. In most sensors, they actually cover less than half of the surface area.

Dark Current and Well Overflow

A photosite essentially converts the energy from a light wave into photo-electrons. Light is actually what physicists call a "wave-particle duality." The energy in the light, which is what we're trying to collect, resides in particles called photons. The longer a photosite is exposed to light, the more photons are converted into photo-electrons via the photo diode at each photosite. One photon ~= one photo-electron (you can't gain energy in the transfer). To some degree, photo diode size is directly related to effective ISO sensitivity, as the larger surface area exposes it to more light in any given amount of time than a smaller surface would. You'll note that the larger photosite DSLRs tend to have base ISOs of 200 and the smaller photosite DSLRs tend to have a base ISO of 100, for example.

The physical size of the individual photosites is important beyond effective ISO. The larger the active light gathering surface, the less noise is a problem. That's because every piece of silicon has a baseline level of electron "action" (current). In sensors, this current is usually called Dark Current or Dark Noise (the "dark" in the name implies that the current was formed despite no exposure to light). (There are actually several different underlying types of on-chip noise, but for simplification, I'll just refer to Dark Current in this article.)

Dark Current increases with temperature. This is due to the small gap between the valence and conduction bands within the silicon: the gap is so small that higher temperatures cause more electrons to cross the gap to where they don't belong. Fortunately, it takes really hot temperatures to increase Dark Current to visible, troublesome noise (typically 90 degrees Fahrenheit or higher coupled with long shutter speeds, at least for the smaller Sony sensors used in the Coolpix and most other consumer cameras). At very long shutter speeds (usually 1 second or longer) some of this electron activity can also result in "hot pixels," essentially generated by photosites that prove "sticky" to those wandering electrons due to impurities in the silicon. The longer the shutter speed or higher the temperature, the more likely you'll see some hot pixels in your image.

Every digital camera attempts to deal with dark current by "masking off" a set of photosites so that they don't see light (which is part of the explanation why your 3.34-megapixel camera only produces images with only 3.15 megapixels). Your camera's brains compares the values it sees from photosites that weren't exposed to light to those that were. Dark Current is partially random. So, in the most simplistic form, the camera averages all the values found in the masked off photosites and subtracts that from the values seen by the photosites exposed to light to remove the Dark Current.

Many current production digital cameras go further than that, however. Individual photosites can and do have slightly different responses to light and to current, so many modern cameras do something a bit different on long exposures: they take two pictures with the photosite array, one exposed to light and one not. Then the pattern seen in the exposure without light is subtracted from the one for the exposure exposed to light. (You can do this yourself, by the way. When you take a picture in low light with long shutter speeds, especially in warmer temperatures, put the lens cap on and take another shot at exactly the same shutter speed. In Photoshop you can use the second exposure to remove patterned noise from the first. But make sure that your dark current exposure is taken at the same temperature as the first! I've seen people take their first photo outside at night in the cold, then bring the camera to the warmer indoors while the second dark current exposure is being made. That won't work: the two exposures need to be done at the same temperature.)

At the other end of the spectrum, what happens to a photosite when it contains too many photo-electrons (due to too much exposure to light)? Well, if left to its own devices, the information (electrons) can spill from one photosite to another, corrupting the data in the adjacent site (a concept called "blooming," or well overflow). This is especially true in the physically small photosites of the Sony sensors used in Coolpix models (proximity makes it easier for an electron to escape from its current owner to another). Most sensors have "drain" circuits that attempt to remove excess electrons before they degrade the chip's data too badly, but these circuits are far from perfect, and it's possible to overload them, as well.

I've been speaking about something without really identifying it: the electron well in the photosite. The photo diode in photosite converts the light photons to electrons, but the photosite needs somewhere to store those electrons until the sensor is asked to produce the "value" for each photosite. It does this in an electron well buried in the photosite. Just like the size of the photo diode varies in sensors, so does the size of the electron well. Compact cameras have smaller wells than DSLRs, for example. Let's just use some arbitrary numbers to see why that's important.

Consider a compact camera photosite well that can hold 10,000 electrons versus a DSLR well that can hold 100,000 electrons. If the baseline noise level within the sensor itself is 10 electrons, then the signal to noise ratio in the compact camera has a maximum value of 1000:1 while the DSLR's signal to noise ratio is 10 times better. That's one reason why the compact cameras produce more noise--all else equal--than do DSLRs.

Sensor design is old enough now that we're bumping against physical limits in almost all the areas just mentioned. Photo diodes are getting to be about the maximum size they can be in the photosite without some new technological breakthrough. Baseline noise is about as low as can be mass produced with current materials. Electron well sizes are about maximized for the current photosite sizes. You can make some adjustment to one of the variables, but it tends to make you have to reduce one of the other variables. Thus, we currently see much more emphasis on post processing the sensor data produced and getting cleaner results by taking out problems after the fact.

However, that doesn't mean we're done with breakthroughs, only that it takes a complete rethink of an element to make significant progress. One such thing is to reverse the orientation of the photo diode so that it's on the surface rather than buried further down in the silicon (usually called a "backlit sensor"). Different materials--often more expensive and more difficult to work with--can change the baseline responses. Better microlenses and filtration can get more of the original light to the photo diode itself. We still have plenty of room for improvement, and I think we'll continue to see about the same level of improvements over time for the next decade that we saw in the last.

top | home

The dreaded "purple fringe," usually seen at boundaries of overexposed area (sky) and underexposed (in this case, tree limb). This is a form of axial chromatic aberration most of the time, but it can be due to blooming on CCD-based sensors. Don't confuse this with lateral chromatic aberration, which tends to be more red and blue/green (see right edge of Louvre entrance pyramid, below).

Your Digital Camera Sees in Black and White

It may surprise you to find out that the sensor in your camera reacts to all light with relative equality. Each individual photosite simply collects only the amount of light hitting it and passes that data on; no color information is collected. Thus, a bare sensor is a monochromatic device.

Plenty of ways exist to make monochromatic information into color data. For example, you could split the light coming through the lens to three different sensors, each of which was tuned to react to a certain light spectrum. But most digital cameras use a different method: they place an array of colored filters over the photosites. Two filter arrays are commonly used:

• RGB. This array arrangement usually involves odd-numbered rows of photosites covered by alternating red and green filters, with even-numbered rows covered by alternating green and blue filters. Called the Bayer pattern after the Kodak engineer that invented it, this filter array uses the primary colors of the additive mixing method (used in television and computer monitors). One unique aspect of the Bayer pattern is that each "color" cell has two of each of the alternatives as its neighbors. Most current digital cameras use a Bayer pattern array. This is the method used in the Nikon and Kodak DSLRs. There is one common alternative Bayer arrangement, called the diagonal color pattern, where each row has repeating RGB elements and each row is staggered by one element (i.e., the first row is RGBRGBRGB..., the second row is GBRGRBRBR..., the third row is BRGBRGBRG...), but it is not currently used in any digital camera I know of. There are also some very sophisticated patterns of R, G, and B filtration (and Kodak announced yet another one in early 2007--which adds "no filtration" positions), but again, they aren't currently used as both the demosaicing is more complex and the resolution of current sensors probably not high enough to support them well.
• CYMG. Alternatively, a slightly more complex filter array uses the primary colors in the subtractive process (commonly used in printing technologies) plus green. This is the method used in most Coolpix models beginning with the 900 series (i.e., the 885, 995, 2500, 4500, 5000, and 5700 all use this pattern). CYMG is typically used on sensors that are sensitive to noise at low light levels, as the dyes used to create the CYM colors are lighter than RGB and thus let more light through to hit the photosites.

Each of these methods has advantages and disadvantages. The repeat of the green filter (and addition of a green filter to the subtractive CYM method) is due partly to the fact that our eyes are most sensitive to small changes in green wavelengths. By repeating (or adding) this color in the filter, the accuracy of the luminance data in the critical area where our eyes are most sensitive is slightly improved.

So, each individual photosite has a filter above that limits the spectrum of light it sees. Later in the picture-taking process, the camera integrates the various color information into full-color data for individual pixels (a process usually called interpolation, but more accurately called demosaicing).

But one important point should be made: the color accuracy of your digital camera is significantly influenced by the quality of the filter array that sits on top of the photosites. Imagine, for a moment, a filter array where each red filter was slightly different--you'd have random information in the red channel of your resulting image. A damaged filter would result in inaccurate color information at the damage point.

One thing that isn't immediately apparent about the Bayer pattern filter is that the ultimate resolution of color boundaries varies. Consider a diagonal boundary between a pure red and a pure black object in a scene. Black is defined as the absence of light reaching the sensor, thus the data value would be 0 (for the G and B photosites). That means that only the photosites under the red sensors are getting any useful information! Fortunately, pure red/black and blue/black transitions don't occur as often as you'd think, but it is something to watch out for. (Since no individual color is repeated in a CYMG pattern, all boundaries should render the same, regardless of colors.)

Most sensors these days are built with microlenses that incorporate the filter pattern directly on top of the photosite portions. This microlens layer not only incorporates the Bayer filter pattern just underneath it, but redirects light rays that hit at an angle to hit more perpendicular to the photosites. If light were to hit the photosites at severe angles, not only would the photosite be less likely to get an accurate count of the light hitting it, but adjacent cells would tend to be more influenced by the energy (like Newton's laws suggest, light waves don't tend to change direction unless acted upon by something). All Nikon cameras currently use microlens layers; the Kodak DCS Pro 14n is unusual in that it apparently doesn't.

On top of the microlenses are also another set of filters that take out some of the infrared (IR) light spectrum and provide anti-aliasing (I'll discuss anti-aliasing in the next section). Sensors tend to have the ability to image well out of the visible spectrum and into the infrared (typically they are still highly reactive at 1000nm--I've seen spectral charts for at least one sensor that shows it is actually more responsive to near infrared than visible light). On the Nikon Coolpix models, the 950's IR filter lets in more of the infrared spectrum than does the filter on the 990 (the 995 seems similar to the 990 in my early tests). (The filter that takes out IR is sometimes referred to as a "hot mirror" filter.) On the Nikon DSLRs, the D100 and D2h are notoriously more prone to "IR pollution" than the others, but starting after the D2h it seems that Nikon has tightened the spectrum that gets through to the sensor with each new generation of camera. Current cameras allow very little light outside the visible spectrum to get to the photosites.

One other thing to note about most sensors: they are not particularly sensitive to low wavelengths (ultraviolet, or UV, lives below the visible spectrum). At the blue end of the visible spectrum (~400-500nm) sensors may have less than half the normalized response to light than at the green level (~500-600nm). In the UV spectrum, most sensors are barely reactive. The reduced response of the blue photosites can contribute to a noise problem, though this is well handled by most current cameras. Nevertheless, I do see considerable changes to color neutrality with some digital cameras using UV filters on the lens. The Fujifilm S2 Pro, for example, tends to get purer whites if you use an UV filter, for example. Indeed, the Fujifilm SuperCCD appears to be highly capable of capturing UV wavelengths, as evidenced by Fujifilm's production of specialized UVIR versions of the S3 Pro and S5 Pro. They did that by simply removing the filter over the sensor!

We have one more exception to talk about (sensors have gotten complicated since I first wrote about them in the 1990's). That's the Foveon sensor. Unlike the Bayer-pattern sensors that get their color information by using adjacent photosites tuned to different spectrums, the Foveon sensor uses layers in its photosite design. The primary benefit of this approach is that it gets rid of color aliasing issues that occur when you demosaic Bayer-pattern data, and thus allows you to get rid of (or at least lower the value of) the antialiasing filter over the sensor. The benefit can be described in two words: edge acuity. Another benefit is that there is no guessing about color at any final pixel point, which means that colors are robust and accurate. The primary disadvantage to this approach has to do with noise. Obviously, less light gets to the bottom layer of each photosite than the top layer. Other issues with the current implementation are the 1.7x crop size (a little more aggressive than what Nikon and Canon use in their consumer cameras) and a lower overall pixel count*. Foveon has done a remarkably good job of mitigating the drawbacks while emphasizing the positive in the latest iteration of the sensor (which appears in the Sigma DP1 and SD-14 cameras).

Actually, when I talk about "red," "green," or "blue" filters (or any other color, for that matter), the filter itself may not actually be those colors. Typically those are the color or light that is let through the filter. E.g., a "green" filter lets green spectrum light through to the sensor and removes "blue" and "red" light.

Demosaicing is a word you probably won't find in your dictionary. The filter pattern is a "mosaic" of colors. The routine that deciphers that mosaic performs a de-mosaic action on the data, thus the routine is called demosaicing. The simplest demosaicing routine works this way: (1) record the existing R, G, or B value at each pixel position; (2) invent new G values at each of the R and B photosite positions, often using a multiple pass technique to figure out where edges occur; and (3) fill in the missing R and B values using neighbor sampling techniques.

Hundreds of other, much more sophisticated variants are now used, with most trying to deal with the minor artifacting issues created by the simple routines.

The lower blue response is a problem with incandescent light (the most common indoor lighting type), as that light source doesn't produce much blue spectrum. Indeed, one complaint about some cameras (the Fuji S2 being one of them) is that their blue channel noise levels can be quite high in some incandescent lit situations.

*Yes, I know I'll get grief on that from some Foveon fanatics. Some claim each layer of the Foveon photosite in their megapixel counts, others just simply say that a 5mp Foveon sensor is equivalent in absolute resolution to a 12mp Bayer sensor. This article, however, isn't the place for a discussion of resolution. I'm simply trying to tell you how sensors work. In practical terms, I measure Bayer cameras resolving about 70% of the level of an "equivalent" Foveon sensor (same number of photosites per column and row).

Getting Data Off the Sensor

At this point, we have an array of filtered photosites that respond to different colored light that usually looks something like this:

The data at each of the individual photosites, by the way, is still in analog form (the number of electrons in the well). The method by which that data is retrieved may surprise you, however: in most CCD sensors the values are rapidly shifted one line at a time to the edge of the sensor. This process is called an interline (or row) transfer, and the pathways that the data moves down are one of the reasons why photosites have space between them (to make room for the pathway). While the data is moved off in "rows," it's important to note the short axis is usually the direction that the data is moved (if you're looking at a horizontal image, you'd see these as columns). (CMOS sensors, such as those used in the Canon, recent Nikon, Kodak Pro DSLRs, and a few other cameras are unique, in that the data for each individual sensor can be retrieved directly.)

As the data gets to the edge of the sensor, it is usually first processed to reduce noise, then read by A/D converters (ADC). Now what we have are a series of digital values (8-bit for many Coolpix and consumer cameras, 12-bit to 16-bit for most SLR models). One some recent Nikon DSLRs, the ADC function is actually built into each photosite (the D300 and D3x are examples of this). This tends to reduce "read noise" because the ADC is adjacent to the electron well being counted and thus transmission errors don't come into play.

One common misconception is that bit depth equates to dynamic range (the range of dark to bright light that can be captured). This isn't true. Dynamic range of a camera is determined mostly by the sensor (electron well capacity minus baseline noise determines the maximum range of exposure tolerated; another reason why larger photosites are better than small). If you put a 4-bit, 8-bit, 12-bit, and 16-bit A/D converter on the same chip, the sensor wouldn't respond to low or bright levels of light any differently; you'd only get more or less tonal definition in the conversion.

At this point, we have one-third the data we need for a complete color picture (we need red, green, and blue values at each photosite location, and we have only one of those values from each photosite). Here comes the tricky part: a processor (a Sparc-based computer in many early Coolpix models, dedicated proprietary circuits in most new cameras, called the EXPEED engine by Nikon) looks at blocks of this data and tries to guess the actual RGB color value of each pixel by comparing adjacencies! (The demosaicing I mentioned above. The right-hand column above shows a typical demosaicing algorithm.)

Consider a GB row that was exposed to sky, for example. The blue photosite values across that row might be the same (or change slowly over distance). That would make it relatively simple to guess that the green photosites on that same row should also have very similar blue values. You could simply average the two adjacent blue values to give a green photosite its blue component (and vice versa). This, of course, is a very simple case, but it illustrates how the camera's software has to function: it examines a block of adjacent pixels and uses that data to build the missing two values for each location. Again, this process is often called interpolation, though the software routine that does this is more correctly called demosaicing.

Camera manufacturers are extremely secretive about their demosaicing methods. But given the unclassified data on image processing and the fact that virtually all cameras are pressed for computational power when confronted with huge amounts of data, they all tend to do similar things with near-neighbor lookups. You should know a couple of things about demosaicing (we're about to talk about anti-aliasing, which I promised earlier):

• The process of reconstructing data at a "frequency" (sampling rate) less than the original produces aliasing. What that means is that the reconstructed data may not be a correct record of the original. Let's pretend our sensor is black and white only for a moment. Imagine a series of vertical black lines with white space between them. If each black line fell on one column of photosites and each white space fell on the column of photosites next to it, it should be obvious that we can capture that level of detail perfectly. But what happens if each black line falls partially on a second column of photosites? Instead of recording white, those photosite would record "gray" (part white, part black). Earlier I mentioned that the filter on top of the microlens Bayer pattern took out IR and provided anti-aliasing. Well, what anti-aliasing filtration does is get rid of the highest frequency detail, which would tend to produce the problems we just talked about. Unfortunately, anti-aliasing filters have the net effect of making the level of detail rendered appear "softer" than they would otherwise. But they also insure that the worst artifacts associated with analog-to-digital sampling aren't encoded into the Bayer pattern data. The higher the resolution of the sensor, the less the need for anti-aliasing filtration; the 13.5mp Kodak Pro 14n doesn't have an anti-aliasing filter, for instance. Technically, the D3x shouldn't need one, but it does have one.
• One of those artifacts associated with sampling frequency and demosaicing is that of moire patterns (sometimes called color noise). A moire pattern occurs when aliasing occurs on a highly detailed area. Moire can be partially removed by using complex math (involving what's known as the Nyquist frequency), but there's a real battle going on inside your digital camera between the speed at which images are processed and the amount of data the camera has to deal with. Most cameras saving into JPEG format don't do much, if any, moire processing, and rely more upon the anti-aliasing filter to reduce this artifact before the data is encoded. A few cameras, the Kodak DCS Pro 14n, for example, don't use an anti-aliasing filter. Images from those cameras tend to be slightly sharper and contain more detail than others (especially considering the 14n's 14-megapixel sensor), but this is at the expense of added color noise.
• Sharpness, contrast, and other camera settings may be applied during the demosaicing step or immediately afterwards, depending upon the camera's design. And JPEG compression is yet another variable that enters into the picture. Each additional manipulation of the underlying photosite data gets us a little further from the original information and introduces the potential for artifacts. In essence, by the time the camera is done with all its processing, it is impossible to reconstruct the original data (exception: most modern DSLR cameras have the ability to save the actual photosite data from the camera in a raw file, for later demosaic on your computer).

A couple of other points should be made:

• On Fujifilm's Web site, they make the contention that their SuperCCD does not interpolate to get the 6-megapixels of data the S1 produces from its 3-megapixel sensor (or 12 megapixels from 6 on the S2 Pro, S3 Pro, or S5 Pro). They claim that, because of the angular nature of their array (think of a baseball diamond where each of the bases is a photosite), that they already have the X and Y values that can be used to build the intermediaries (the pitcher's mound in this pattern). Sorry, Fujifilm, but that's still interpolation.
• The D1x is a very unique camera as regards demosaicing. Most interpolation of photosite data is what is called up sampling. This means that you sample existing data to produce additional data (that's what demosaicing does, for example: you have a G value and you look at the other data around it to come up with the R and B values for that G position). You can also down sample, which would be to produce less data than the original contains. The D1x upsamples the short axis and downsamples the long axis to produce its in-camera pictures. NEF conversion products, such as Bibble and versions of Capture later than 3.5, can also produce images that don't downsample the long axis while upsampling the short axis, producing far larger files.) Why Nikon chose to do the in-camera up/down sampling is unclear. One would think that you wouldn't want to downsample sensor data, but the fact that Nikon does so even in their default RAW processing seems to indicate that Nikon knows something about the validity of the data from those split photosites that we don't. It could be, for example, that Nikon downsamples that data to deal with a noise problem. Or it could be that there's a short-cut to the number-crunching that must done to generate the full RGB image if they downsample the long axis. The reason must be a good one, however, because moire and related artifacting is dramatically lessoned if sampling is done on only one axis rather than two.

Dynamic range can also be captured logarithmically, another reason why the number of bits doesn't directly equate to dynamic range.

Fujifilm's Super CCD orientation, while different than traditional CCDs, still requires interpolation, regardless of what Fujifilm claims. Because of the angular layout, however, the interpolation has a slightly different (and arguably better) set of data to work from.

Best Book for Photographers

While I was browsing my bookshelf double-checking some of my material in this report, I pulled out The Manual of Photography Ninth Edition, the highly technical and math-filled volume that defines much of the state-of-the-art. (MoP is highly recommended, by the way. It's one of those books that you pull out and read sections of from time to time when you want to know the underlying theory behind something, like depth of field or fast Fourier transforms.)

Much has been said about when (or whether) digital imaging will pass film in the ability to resolve information. I came upon this interesting passage (the culmination of several paragraphs of theory and math): "It would appear that the digital system has overtaken the photographic process with respect to this [information capacity of images] measure of performance." Note that this says nothing about resolution, only the theoretical amount of "information" contained in an image.