Sunday, July 24, 2005

An Eye on DLP, No. 2

In An Eye on DLP, No. 1, I showed how a postage-stamp-sized array of 1,280 micromirrors across by 720 micromirrors vertically could make a tiny high-definition image that can then be projected onto a large screen. The technology involved is Texas Instruments' Digital Light Processing, or DLP.

The DLP "light engine" uses either three digital micromirror device (DMD) chips, or one:


Either way, red, green and blue images — in the three primary colors, that is — are formed separately and meet at the screen or at the retina of the eye. Each video frame is either spatially (with three DMDs) or temporally (using one DMD and a color wheel) divided into three subframes, one for each color primary. White source light is changed into the appropriate color for each subframe by means of wavelength-subtracting prisms or transparent colored segments of the spinning color wheel.

But, aside from the hue thus optically afforded it, each subframe is in effect a mosaic in black, white, and shades of gray.


An important question is, how many shades of gray?

The simple (much too simple; see below) answer is that consumer DLP displays and projectors divide the luminance range from black to white (ignoring hue) into 28 or 256 steps.

This means that real-world luminance, which actually falls anywhere along the continuous tonal range from black to white, is summarized as 256 discrete shades of gray. Alternately stated, the "grayscale resolution" that is used for consumer-level DLP displays is 256. Yet another way of putting it is that the grayscale's "bit depth" is 8 bits per primary color component: the power of two which 256 is, not surprisingly, corresponds to the number of bits (8) needed to represent all 256 possible integer values from 0 to 255.

An 8-bit bit depth is not ideal; it's an engineering compromise. The problem is that the grayscale steps at lower luminance levels are visually too far apart. Instead of seeming to grade smoothly into one another, they form "bands" or "false contours" on the screen.

For mathematical reasons having to do with the decreasing percentage of step n which each new increment to step n+1 represents, the banding or false contouring problem tends to go away as luminance level rises. But in darker scenes and in dimly lit areas of ordinary scenes, banding or false contouring can be objectionable.


According to "DLP from A to Z," an excellent article by Alen Koebel in the August 2005 issue of Widescreen Review magazine, "Professional [DLP] video sources use 10 bits per component; professional displays should be able to reproduce at least 8,192 (213) gray levels. DLP Cinema projectors, arguably the most advanced manifestations of DLP technology, are currently able to reproduce more than 32,000 levels." (Magazine subscribers may download the WR article in PDF form by clicking on the appropriate hotlink on this page.)

"More than 32,000" equates to 215, or 32,768 levels of gray.

Koebel doesn't make it crystal clear, but I'm guessing that the main reason for the lower grayscale bit depth on consumer (i.e., non-professional) DLPs is that they're 1-chip, not 3-chip. Because of the need to use a color wheel, the time of each video frame has to be divided in three parts — or, rather, at least three parts, depending on how many segments there are on the color wheel. In one place in his article, Koebel does say that "having full gray scale resolution for all three colors ... is not feasible in the time available." I think this is what he means by that.

This problem is only exacerbated by the fact that the color wheel is rotated at a speed high enough to have "four (called 4X), five (5X), or six (6X) sets of RGB filters pass in front of the DMD chip during a [single] video frame." This is done to minimize the "rainbow" artifacts I discussed in An Eye on DLP, No. 1. But it reduces the duration of each primary-color subframe and thereby lowers the number of micromirror on-off cycles that can be squeezed into it. That in turn makes for an upper limit on the gray-scale resolution.


Several strategies are used to offset the banding/contouring problem at low luminance levels which results from not really having enough bit depth or gray-scale resolution. One of these strategies is "dithering."

In spatial dithering, if a pixel needs an "in-between" shade of gray that is not available due to bit-depth limitations, the pixel is formed into an ad hoc unit with adjacent pixels so that, among all the pixels of the unit, the average gray shade is correct.

Temporal dithering averages different shades of gray assigned to a particular pixel from one video frame to the next to get what the eye thinks is an in-between shade.

"Combined, spatial and temporal dithering typically add two or three bits of 'effective' resolution to the gray scale," says the WR article, "at the cost of increased noise and flicker in the image." If we assume an actual gray-scale resolution of 8 bits, dithering can nominally simulate a bit depth of 10 or 11 bits.


The algorithm used for spatial and temporal dithering is much more complex than what I have said actually indicates. Suffice it to say that dithering involves an only semi-intelligent attempt to modify the gray levels assigned to an image's pixels in order to smooth over false contours. (And remember that gray shades translate into colors, by virtue of the fact that each pixel of each red, green, or blue primary hue, before it is tinted such by the optics of the DLP light engine, comes with its own associated "shade of gray.")

On my Samsung 61" rear projector, the dithering algorithm is apparently responsible for a "stippling" or "pebbling" effect I can see in black areas of the screen when looked at up close. For example, when there are black letterboxing bars on the screen, I can see patternless, busily moving "noise" in the renering of black, if I get within inches of the screen — assuming I don't turn brightness down low enough to "swallow" it, that is.

Other people have complained in online forums about single-chip DLPs' temporal dithering in particular. They say it blurs or "pixellates" the trailing edges of brightly colored objects moving rapidly across the screen — sort of like an artifical "tail," where the object is the "comet." I myself have never noticed this effect.


Still, because dithering does produce artifacts, TI has come up with a way to minimize the need for it. It's called dark video enhancement, or DVE.

DVE puts one or two additional segments on the color wheel. In addition to two red, two green, and two blue segments, now a seventh and (depending on the implementation) possibly an eighth segment are introduced for the express purpose of creating what might be called a dark-video subframe (or two subframes) of the entire video frame.

When a red color-wheel segment moves between the light source and the DMD chip, the DMD creates a red subframe. Just the intensity of red is taken into account. When a green segment swings around, the DMD switches to creating a green subframe. When a blue segment is in the proper position, the DMD switches again to take just the intensity of green into account.

So, logically, when a dark-video segment is active, the DMD produces pixels whose intensity corresponds only to the gray-scale information at the low end of the luminance range. The higher luminance levels are ignored so that the gray-scale resolution at the low end can be given a bit depth that has been increased by one or two bits.

I'm not sure what color the dark-video segment actually is, or whether it is clear or possibly a neutral shade of dark gray. I do know (because the WR article tells me so) that "the DMD is "driven only by green-channel data during these segments. Since green contributes nearly 60 percent to the [overall] luminance of each pixel ... this gives a reasonable approximation to having full gray scale resolution for all three colors."

The best I can tell, DVE has been introduced in front projectors only, as of now. I have yet to find a rear projector which has it. (Of course, really expensive front projectors with three DMD chips don't need it, since they lack a color wheel in the first place.)

More on DLP technology in the next installment.

1 comment:

Ian said...

Just happened across your blog via the "Next Blog" link. This looks really cool - I intend to read it all. You should publish as a book when you're done!