Feature Articles » Displays Of A Different Stripe

A Sharper Image

Any system that dices a picture into a matrix of pixels risks the distortion known as aliasing. A simple example is the jagged rendering of certain characters seen on a computer display. More interesting distortions, such as the moiré pattern, may emerge when one matrix is used to sample data on another, as happens when one looks at a screen door through another screen door. Yet although the human eye filters the world through a matrix, it rarely produces aliasing, and that is why engineers try to copy its powers.

The trick is to find out how the eye ignores extraneous information without overburdening its sampling mechanism. The world’s television standards have been designed according to this biomimetic strategy.

Analog color encoding schemes in both the NTSC color TV and the JPEG/MPEG standards resemble the eye in that they break an image into a luminance channel, which conveys brightness, and two chrominance channels, which convey color. Because both of the chrominance channels are defined by their difference from a third color, it is possible to deduce that third color, which is needed to show a full-color image. The only difference from the human visual system lies in what is inferred. The JPEG/MPEG technologies provide red and blue channels and infer green; the human visual system provides red versus green and yellow versus blue and infers luminance. (NTSC is the National Television System Committee, JPEG is the Joint Photographic Experts Group, and MPEG is the Moving Picture Experts Group.)

In modern digital cameras, the image forms on a charge-coupled device that contains an array of photosites, each of which converts light into electricity. Each photosite—regardless of the color of the color filter array above it—maps to a conventional, full-color, red-green-blue pixel. Algorithms sample the array of photosites to produce a signal that fits a different array—that of the display. The success of the translation depends on the sampling rate.

Sampling a signal can be compared to a person running in the dark over a road with regularly spaced bumps. If he takes small steps, he will detect the shape of the bumps, but if he takes longer steps, he will miss the bumps altogether or misconstrue their form. The general rule is that signal components more than half the sampling frequency will produce aliasing; therefore, the algorithm that assembles and interpolates pixels will work only so long as the channel signals are each less than half the original sampling frequency for their color.

Chromatic aliasing appears as color fringes around the edges of objects. It was a problem in the earliest algorithms that attempted to sample not pixels but subpixels—the elements of which pixels are made. Those early subpixel rendering algorithms used filters based on decimation, a process of dropping information in order to fit high-resolution data into a lower-resolution format. Decimation works moderately well in displaying coarse-grained images in slow video, but when it tries to cram detailed, fast-moving video into a coarse display—say, by displaying NTSC TV resolution on a 240-by-320-pixel color mosaic—it can produce chromatic aliasing. Decimation is commonly used in viewfinders, where the low quality doesn’t matter.

The next step in the development of the art of subpixel rendering was taken by IBM Corp. display engineers in 1988, when they reduced chromatic aliasing with a filter that represented a group of pixel values as a weighted average. In a later development, Microsoft Corp., in Redmond, Wash., designed its ClearType font-rendering technology to minimize aliasing with a system of filters that accounted for the slight spatial displacement of all red subpixels vis-å-vis the green ones and the green ones vis-å-vis the blue ones.

—J.P.