As we've said, probably the most important thing to take into consideration when purchasing digital equipment is output format: the goal in any purchase is to have a finished product that looks as good as possible. Decide what's important to you. Do you need high resolution for enlargements or substantial cropping? Great color saturation for on-screen display, or maybe magazine work? Clarity for paper prints? High contrast for black and white work? All of the above? Once you figure that out, find a camera that gives you the best of what you need in your price range. Sounds simple enough. But how do the main big features of digital cameras-megapixels, sensor size, and image handling-fit into the picture, and how do figure out what you actually need to get the results you want?
Color is, generally speaking, the variable you have the least control over. There is some minor variation in frequency response between sensors and historically different sensor types have reacted differently to different lighting situation, but most current CCD and CMOS sensor are pretty equivalent when it comes to color. Furthermore, most digital images are destined to undergo some sort of software color balancing at some point, anyway, and the final output will depend far more heavily on post processing and the software used to convert the RGB capture to the final (VGA, CMYK, indexed RGB) output format than on the colors in the original capture. Get the best bit depth you can—many current models have 32-bit or better color—and don't worry too much.
Resolution and size are far more interesting; this is where megapixels, sensor format, and the "megapixel myth" come into play. Ignoring lenses for the moment—they can be changed—we often tend to equate megapixels with image quality, but that isn't the whole story. Input resolution and output resolution aren't the same thing.
Our standard for judging photographic equipment and processes is generally to look at how well they hold up to an 8x10 enlargement seen at 1m. A feature looks "sharp" if the individual dots that form it on the paper are smaller than the smallest dot your eye can see. For most people viewing a picture at 1m this falls somewhere in the range of ±5 dots/mm, or a ±200µm dot. The smallest dot you can actually see, however, depends on your eyesight, and this value, the appropriately named the "Circle of Confusion" (CoC), is generally set at around 26µm for 35mm negatives, giving a 184µm dot in a 7x 8x 10 enlargement. For convenience, many photographers use the "Zeiss formula" to calculate CoC. For a given format, the Zeiss CoC is d/1730, where d is the diagonal measurement of the film, in mm. For digital photographers, it is usually more convenient to use the formula (1000*d)/1730 to get the final result in µm, since electronics manufacturers tend to give measurements in microns.
But enough math. Armed with CoC, we can start thinking about sensors. One way to evaluate the range of a given sensor is to compare it's pixel size to the Zeiss CoC for it's format. The smaller the relative pixel size, the more relative detail you'll capture. A current 12.8MP "full-frame" sensor like the one in the new EOS-5D has a pixel size of about 8.2µm. Since this is only about 1/3 the 25.03µm Zeiss CoC for this format, it will stand up to a great deal of enlargement. The 22.5 x 15.0 mm sensor on the EOS 20D, on the other hand, has a pixel size of about 6.4µm, and although this is much smaller than the 5D pixel size, the Zeiss CoC for that format is about 15µm, so the 6.4µm pixel is 42% of the Zeiss CoC. So while the 20D sensor captures smaller detail, the smaller sensor size means that the final output can't be enlarged nearly as much. It can still be enlarged considerably, though. In general, most current "mid-range" cameras are quite capable of producing images suitable for enlargement up to 8x10. Finding out the pixel pitch of a given sensor, however, can be tricky, although some review sites like dpreview have started posting the information for recent models.
At least that's the theory, assuming you have an optically perfect lens and you're exactly focused on a flat subject that lies entirely within the focal plane. In practice, MTF and DOF affect digital in more or less the same way they do film, and you should generally expect the best digitals to perform roughly the same (where detail is concerned) as similar 35mm rigs with quality film. In other words, don't expect your "full frame" digital to consistently give good results for enlargements beyond 8x10, no matter what the CoC calculations might lead you to believe. Unless, of course, you shoot nothing but USAF 1951 charts. At a certain point absolute size matters as much as CoC numbers. Super enlargements will have issues with focus, distortion, spherical and chromatic aberration, and vignetting, just like film, and the smaller your sensor, the sooner, generally speaking, you'll hit that point. Smaller captures need to enlarged more, and smaller lenses tend to be more prone to aberration. The one place this may not hold is in using lenses with undersized sensors. If you're using 35mm lenses with a 2/3 or 3/4 sensor, you may get a little extra mileage: The smaller sensor means that more of the image is captured in the center of the lens, the part with the highest resolving power, and vignetting and blurring in the corners should be mitigated. On the other hand, because more of the image passes through a smaller portion of the lens, the quality of the lens itself becomes a more important factor, and starting with the smaller sensor means you'll have to enlarge the image more to make it useful. and because the amount of light striking each pixel is less, smaller sensor produce much noisier images than larger sensors as well.
Those rules, though, assume that the output format has an extremely high resolution, like traditional photographic paper or a 2400dpi photo printer. For most of us, this isn't the normal case. Most of the places digital pictures end up are fairly low resolution screen and printing devices. In this situation, CoC becomes largely irrelevant. Since the resolution of the media is already worse than most people's eyesight and the bit-depth of the color rendering is almost certainly worse than the camera, the question becomes not "how big do you want to enlarge your image?", but how small you want to shrink it or crop it. If we were to take that same 4368 X 2912 12.8MP image and display it on screen at 72dpi, we would need a 60 x 40" screen. Even at 90dpi, it would still take up an area 47 x 32". Even if we could find a monitor that large, the image quality at would be horrible: so matter how many megapixels it has, no 35mm is really designed to hold up to 33x enlargement.
Here, image format and software design become more important. To effectively display images on screen, we need a couple of things. One is a camera that gives us good files to work with. All manipulation tends to degrade image quality, so we need to start with RAW files, TIFFs, or at the very least lossless JPEGs. In other words, the bigger the file size to image size ratio for a given capture, the better the final output will probably be when we get done cropping, highlighting, color balancing, sharpening edges, RGB indexing and whatever else we plan to do. And we need to give that image to intelligent software, so that we lose as little quality as possible as we make our adjustments. File type and quality, then are the most important concern in choosing a camera, followed closely by a bit-depth at least as good as that of our output device (although, again, most of the time that won't be an issue). If our goal is to produce a 500 x 375 image for a website, the noise and general degradation from repeatedly editing compressed files are going to be far more noticeable than the difference in quality between a 4MP and a 12MP camera. Most entire screens are less than 1MP. A 1024 x 768 screen, for instance, would correspond to a .7MP sensor. Beyond that, we have to use interpolation to drop pixels from our pictures before they can be displayed. And the bigger the original image, the more pixels that have to be dropped, and the more corrupted the displayed image becomes. That doesn't mean a 5MP camera won't produce better results on screen than a 1MP camera; it will. Not all captured pixels make it into the final product; better cameras tend to have better image processors and better color was well; and when you start with more, you can afford to lose more throughout the "photoshopping" process. But if your ultimate goal is on screen display, you'll begin to see rapidly diminishing returns somewhere around 4MP.
As for sensor size, the rule of thumb here is "the bigger, the better." Regardless of resolution (visual or electronic), smaller sensor produce more line noise than larger sensors. To be more accurate, I suppose we should say sensors with smaller pixels produce more line noise than sensors with larger pixels, and a high signal to noise ratio is particularly important to on-screen display. Don't get me wrong, noisy prints are bad, too, but noise is particularly noticeable on screen because of the differences between the illuminated screen image and the reflected light seen when viewing prints. A CRT, in particular, is like a big slide projector noise, which might be mistaken for grain in a print, really pops. If you're going to be doing large-scale cropping or enlargement for on screen display, you'll want to look to your lenses, too: resolution may not be as important as it is for prints, but as with sensor noise, monitors can make even small magenta shadows from chromatic aberration stand out.
