In Search of the World’s Oldest Digital Graphics

What would you say if I were to offer you a digital image from the 1740s?  Not an image from back then that just happens to have been digitized, mind you, but an image that was then already digital.  If you’re one of those people who think “digital” means “the convergence of social media, mobile and the web,” or something equally rooted in recent technological developments, you might reject the possibility of 275-year-old digital images out of hand.  But “digital” also has a more basic technical meaning, and on that basis I’d like to argue that natively digital images exist from hundreds of years ago—even thousands of years ago.  To display these antiquated images on a computer screen, or to post them online, we don’t need to digitize them; all we need to do is some superficial reformatting.  Moreover, the people who created them faced some of the same constraints and developed some of the same strategies that still pervade digital graphics today (e.g., how to represent curves or diagonal lines).

This paleodigital content doesn’t generally get acknowledged in histories of the digital image, digital graphics, or digital art, which tend to treat these subjects as synonymous with the computer image, computer graphics, and computer art.  To the best of my knowledge, it’s only the history of “pixel art” that has embraced such precursors to any degree.  But the continuities are pretty striking, and I’d like to see them more fully recognized and theorized than they have been.

One of my signature strategies for verifying the antiquity of supposedly “modern” formats has been to show that older data can be reformatted into them without significant change and then educed in the same way as newer data of the same type.  This is how I’ve tried to make a case for the existence of medieval and early modern audio (“What is Paleospectrophony?”; “How to ‘play back’ a picture of a sound wave”) and motion pictures (“Moon Phase Animations (AD 650-1650)”; “Early Motion Pictures of Eclipses”).  With this blog post, I hope to make a comparable case for the digital image.

So what is a digital image, really?

Let’s start by considering the term “digital” itself.  It’s most often associated with computers, but that’s because computers happen to operate on digital data, and not because digital data as such are inherently tied to computers.  Even though digital data may be uniquely suited to computer processing and certain kinds of electronic storage, that doesn’t define what they are any more than we could meaningfully define alphabetic texts as the type of content handled by Gutenbergian printing presses.  Wikipedia has this to say on the subject:

Digital data, in information theory and information systems, are discrete, discontinuous representations of information or works, as contrasted with continuous, or analog signals which behave in a continuous manner, or represent information using a continuous function.  Although digital representations are the subject matter of discrete mathematics, the information represented can be either discrete, such as numbers and letters, or it can be continuous, such as sounds, images, and other measurements…. Since symbols (for example, alphanumeric characters) are not continuous, representing symbols digitally is rather simpler than conversion of continuous or analog information to digital. Instead of sampling and quantization as in analog-to-digital conversion, such techniques as polling and encoding are used.

And here’s Vangie Beal’s definition of “digital” at Webopedia:

Describes any system based on discontinuous data or events. Computers are digital machines because at their most basic level they can distinguish between just two values, 0 and 1, or off and on. There is no simple way to represent all the values in between, such as 0.25. All data that a computer processes must be encoded digitally, as a series of zeroes and ones.  The opposite of digital is analog. A typical analog device is a clock in which the hands move continuously around the face. Such a clock is capable of indicating every possible time of day. In contrast, a digital clock is capable of representing only a finite number of times (every tenth of a second, for example).

According to these accounts, “digital” information is:

  • Discontinuous, or broken into chunks—not continuous.
  • Made up of discrete values chosen from a limited range of options—not necessarily just two, but not infinitely variable.  I’d add that this applies both to the values assigned to the chunks (such as “0” or “1”) and to the increments that identify the chunks themselves (such as value 1, value 2, value 3….).

And that’s it.  There are no other criteria.  Binary data stored on a computer hard drive may be a “better” example of digital information than the older kinds of image data I’ll be exploring in this article, but only in the same sense that an apple is a “better” example of a fruit than a pumpkin.  Technically, apples and pumpkins are both 100% fruits; we just tend to think of one more readily as a fruit than the other.

It’s worth noting that digital information always assumes a physical form with analog characteristics, even if that aspect gets ignored in practice.  For example, information is encoded into the surface of a CD in the form of “pits” and surrounding “lands”:

Each pit has a complex analog physical structure that can be seen under a microscope.  In terms of accessing the digital data, however, all that matters is which of nine discrete length categories each pit or land falls into, corresponding to the timing at which a laser—taking 4,321,800 one-bit samples per second—encounters transitions between pits and lands (“1”) as opposed to continuations of pits or lands (“0”) as the disc rotates past it.  By the same token, an “exact digital copy” from CD to CD would have pits and lands that fall into the same length categories, but there wouldn’t be any expectation that the specific forms of the pits would be carried over as well.  The physicality of voltages and magnetism is similarly nuanced, and those nuances are similarly irrelevant.

Now let’s consider how the same issues play out with something that’s not ordinarily considered “digital”: an early fifth-century Vergil folio.

In its original form, the image at the bottom consists of continuous information embodied in the infinitely variable spatial placement of pigments.  It can be digitized, and obviously has been—otherwise I couldn’t upload it to my blog—but this requires a lossy process of quantization.  There’s no way to duplicate the pictorial information exactly.  By contrast, the text at the top consists of discrete information.  True, it can be reproduced as a digital facsimile of a continuous analog original, as it has been above, but this is arguably equivalent to the microscopic image of pits and lands on the surface of a CD.  After all, the point here is not the continuous and subtly variable forms of the individual characters, but the distinctions among those characters, just as the point of a CD is not the specific layout of each pit and land, but the lengths between transitions (with only nine possible options).  The information that meaningfully constitutes the Vergil text consists of strings of discrete elements selected from a finite set of possibilities, like this:


We do run into some data compression (QVOQ; for QVOQVE) and corruption (VIRIDIS should be VIRIDES, according to most critical editions), but even so, the text itself can in theory be duplicated losslessly.  Think about what an “exact copy” means in the context of manuscripts: an “A” should be copied recognizably as an “A,” but its precise shape can vary.  We don’t habitually speak of such content as inherently “digital,” but I’d argue that it fits the technical criteria, even though

  • discriminating between letters is typically done by human beings rather than by machines; and
  • there are twenty-odd possible values to differentiate, rather than only two.

Now let’s apply the same distinction, and the same reasoning, to images.  For an image to be “digital,” the information comprising it needs to be broken into discontinuous chunks, each of which has been assigned one out of a finite number of possible values.  For present purposes, I’ll ignore vector graphics and assume that the chunks will be picture elements, known as pixels for short since 1965.  The location of each pixel must itself be identifiable in terms of a constrained set of possible values.  Thus, many mosaics consist of pixels whose values are selected from a finite range of color options, but whose locations are defined in a continuous manner with infinite variability, disqualifying the images as “digital” (in my opinion).  Here’s a typical example (image courtesy of Valdavia):

What’s needed (and lacking from many mosaics) is some kind of regular coordinate system.  Most often this is a Cartesian grid in which each pixel can be clearly and uniquely identified by a pair of x and y coordinates.  In computer graphics, this approach produces what’s known as a raster image, in reference to a five-pointed tool called a raster or rastrum—literally, a “rake”—that was used historically to trace grids for staff-based musical notation (image below courtesy of Steve Sherrill).

A digital raster image can be copied losslessly as long as the locations and values of its pixels can be accurately resolved, just as the Vergil text can be copied losslessly as long as the sequence and values of its letters can be accurately resolved.

Many of the images used in textile production and embroidery fit all the criteria for digital raster images spelled out above.  The physical structure of fabric itself imposes a grid-like image structure—warp yarns correspond to x coordinates, while each successive pick of the weft corresponds to a successive y coordinate—and yarns themselves typically fall into discrete color categories.  It’s no coincidence that the perforated cards of the Jacquard loom, commonly cited as a precursor of computer data storage, contained image data of exactly this kind.  Indeed, things are now coming full circle with the development by artist Pamela Liou of a portable computer-controlled dot-matrix loom to weave patterned cloth much as someone might use a desktop printer.  And there are other intriguing confluences to be found besides.  For example, television (a.k.a. “telectroscope”) pioneer Jan Szczepanik also proposed a system for preparing digital textile patterns photographically using a “‘raster’ or multiplying plate, containing some 435,600 perforations.”  Computer monitors and textile looms may well share some conceptual ancestry.

Digital graphics in black and white

But enough prologue—it’s time we got around to some actual paleodigital content.  For starters, here’s an image of a single leaf from E. A. Posselt, The Jacquard Machine Analyzed and Explained (Philadelphia, 1888), which can be viewed in its original context here.

This is a 32×32-pixel raster image in which each pixel has been assigned a binary value of “black” or “white.”  Each square represents a single bit, and the plate as a whole is essentially a bitmap.  To convert this digital data into a standard modern image format, I keyed it into GNU Octave by hand (0=”black”, 1=”white”) and then used the “imwrite” command to output a TIF.


Here’s the result:

And here’s an enlarged version with each source pixel represented onscreen by a four-pixel square to show the pixelation more clearly:

I don’t see any grounds for denying that this is a digital image from the year 1888.  Do you?

There’s an important distinction to be drawn here between historical images that have merely been digitized using flatbed scanners and other such gadgetry, on the one hand, and historical digital images, on the other—images, that is, that stand out for having already been digital in the deeper historical past.  My instinct is to call images of the latter type born digital, and I think this is reasonably defensible.

Now, it’s true that digital textile patterns were sometimes derived from analog drawings.  In fact, the leaf bitmap shown above was based on a sketch reproduced earlier in the same book:

It could be argued that this sketch was the analog original from which a nineteenth-century digital copy was made.  But such sketches would have been strictly ephemeral, drawn only with the goal of creating digital textile patterns from them.  Treating the textile patterns as “copies” of these sketches would, I think, be a bit like treating movies as “copies” of storyboards.  Besides, the digital version is almost always the only one that survives today; like more recent born-digital items, it’s a self-standing thing that warrants attention on its own terms, and not as a lesser-quality manifestation of something else.

We can contrast born-digital content, either in the usual sense or in the new sense I’m proposing, with the digital facsimile: a digital representation of what a media object looks like (or sounds like), but that doesn’t exploit any natively digital information as such.  The flatbed scans made from the plates in E. A. Posselt’s The Jacquard Machine Analyzed and Explained are digital facsimiles, even if one of them also invites interpretation in natively digital terms.

Next let’s tackle a more elaborate leaf from Thomas Rotherford Ashenhurst, A Practical Treatise on Weaving and Designing of Textile Fabrics, third edition (Huddersfield: J. Broadbent & Co., 1883).  This time the dimensions are 48×48 pixels.  First, here’s a digital facsimile:

And here’s the digital data itself, reformatted for onscreen display (in single-pixel and four-pixel-square versions):

This 68×21-pixel book-marker design from Godey’s Lady’s Book (August 1855), p. 167

—can likewise be rendered in more strictly digital form:

The same goes for this 43×43-pixel doily pattern from the same issue, p. 170

—which comes out like this:

Now let’s reach even further back in time to take in a digital image from circa 1742—the one I mentioned at the start of this essay.  My source for it is Margaretha Helm, Fortgesetzter Kunst- und Fleiss-übender Nadel- auch Laden-Gewirck-Ergötzungen oder des neu-erfundenen Neh und Stick Buchs, reproduced here in digital facsimile courtesy of the Victoria and Albert Museum.

The image is 114×179 pixels—much larger than any we’ve considered so far.  I thought that keying in 114 sequences of 179 binary values would get a bit unwieldy, so I divided the source into 30×30 segments, each divided into thirty-six 5×5 segments, and keyed in each 30×30 block separately using a kind of five-pixel rhythm: 1-1-1-1-1 (pause), 1-0-0-0-0 (pause), 1-1-0-0-0 (pause), and so on, enabling me to keep my place pretty well as I typed.  Meanwhile, I found that keying the data in as “1” and “0” forced me to rely only on my thumb and pinkie with my hand stretched uncomfortably across they keyboard, so in the interest of ergonomics, I ended up keying in “1” and “q” and then replaced all instances of “q” with “0.”

The binary interpretation of the source data was mostly pretty straightforward, but occasionally I ran into dilemmas that resemble problems of data corruption familiar from more recent digital media—instances, that is, where I wasn’t able to resolve a given value as 0 or 1 with absolute confidence.

First, the ink wasn’t applied evenly to the printing plate, so some areas that were presumably meant to be black actually appear white—a phenomenon visible to some extent in the excerpt shown above.  These areas are generally easy to tell apart from “real” white pixels because they tend to turn up in the middle of larger black areas with decently defined edges.  Meanwhile, the black sometimes bleeds through into white pixels, especially border pixels, but usually not enough to cause confusion.  In all these situations, I’ve used my best judgment to infer what was on the printing plate itself.

Second, there are some instances where a pixel clearly has a value of 0 or 1 but this seems to be a mistake.  Here, for example, is a 7×7-pixel block in which the center pixel ought to be white, judging from its place in the design:

However, it’s entirely black—even blacker, in fact, than some of the adjacent pixels that were plainly intended to be black but were poorly inked.  Even presuming this was a mistake, then, it must have been a legitimate eighteenth-century mistake.  In these cases, I’ve keyed the information in as it actually appears rather than as I think it should be.  Of course, I’ve probably made at least one mistake of my own somewhere in the tedious process of data entry.

True ambiguities are rare, but they exist.  Here’s one:

The pixel indicated by the red arrow is split into a white upper part and a black lower part, and this looks too neatly done to be dismissed as mere bleeding or sloppiness.  But each pixel can only be all black or all white, so I’ve resolved this one as black on the assumption that it was meant to be symmetrical with its mirrored counterpart below.  Next (below at left) we have another similar case, except that now the split is vertical rather than horizontal:

Three other repetitions of this part of the pattern exist (one example is given above right for comparison), and in each of them the corresponding pixel is black—so I resolved the split pixel as black too.  And here’s a third ambiguity:

The three pixels in the center of this snippet are a mutually consistent shade of gray.  There’s no way to represent this value in a binary data structure, since it falls somewhere between 0 (black) and 1 (white).  The detail below shows where the ambiguous pixels appear in the overall pattern:

Fortunately, there’s no question about what these three pixels should be to complete the pattern—0, 1, 1—so I’ve gone ahead and keyed in those values for them.  But it’s easy to imagine a truly ambiguous scenario arising along similar lines, in which case we’d have no choice but to choose some method for representing missing values in addition to zeros and ones.

Here’s the result I came up with (which you’ve already seen, but here it is again):

Or if you prefer:

If you’d like to try your hand at entering some similar data and don’t mind the busywork of keying in over twenty thousand individual bits, here’s another plate from the same source.  Or maybe you value your time and would rather try a more automated approach to reformatting.  For example, you might take a digital facsimile, crop it to its borders, rescale the result to the target dimensions, and then choose some threshold intensity value for programmatically interpreting each pixel as black or white.  I tried this with the 32×32-pixel leaf from Posselt, and the results were about as reliable as OCR—which is to say, nice if you’re in a hurry, but you wouldn’t want to rely on it for a critical edition.

Digital graphics in color

The images we’ve considered so far are true bitmaps in which each pixel corresponds to a single bit: black versus white, 0 versus 1, off versus on.  But we can also find plenty of paleodigital images in which pixel values are selected from among three or more values.  Here’s a 32×32-pixel image from Thomas Rotherford Ashenhurst, A Practical Treatise on Weaving and Designing of Textile Fabrics, third edition (Huddersfield: J. Broadbent & Co., 1883), page 207.

There seems to be at least one mistake: the stray gray pixel midway along the right edge should surely be black.  But in any case, we can render the three different pixel values in grayscale—

—or in color, with colors assigned arbitrarily (since none are specified in the source):

Next, here’s a 58×88-pixel color embroidery pattern from Godey’s Lady’s Book (August 1857):
This time, there’s also a color key for the eight symbols:

With this information in hand, we can render the digital data in color:

A few pixels shown as red at upper middle left seem to be in error and should presumably be black, but as usual I’ve chosen to preserve the mistakes as long as there isn’t any ambiguity about them.  It was sometimes hard to distinguish between the symbols for “red” and “green,” but I found that context and repetition helped resolve the uncertainty.

It’s easy to imagine carrying out similar work on the actual punched cards of a Jacquard loom, or on the data left behind by even earlier experiments by Basile Bouchon, Jacques Vaucanson, and others to use perforated media in textile production.  And there’s no reason we couldn’t transcribe textiles themselves into modern digital image formats, for that matter—after all, this kind of analysis is already being done to create viable embroidery patterns and the like from historical source artifacts.  Suddenly it seems all the more fitting that texts and textiles should share the same etymological root: not only can we think of texts as metaphorically “woven” or “plaited,” but textiles can embody digital texts in turn that invite lossless transcription.  Coptic and Pre-Columbian textiles look particularly promising as fields for experimentation, although I think it would be better to transcribe the image data straight from the objects themselves rather than trying to work from photographs, since we’d need to follow and count individual threads.

But if we simply want to pursue born-digital imagery back as far in time as we can, we also have some prime candidates in Sumerian cone mosaics.  These mosaics, estimated to date from 3500-3000 BC, were constructed out of clay cones with their pointed ends inserted into plaster and their flat ends painted black, white, and red.  Several cone mosaics from the Eanna District of Uruk are housed today in the Pergamon Museum in Berlin, and one of these is depicted below (photo by F. Tronchin):

By cross-referencing that photograph with another one that better shows the right-hand side and the contrast between cones—

—I’ve been able to transcribe the sequence of cones with a reasonably high level of confidence (although it would have been preferable to visit the Pergamon Museum itself).  The only areas I had trouble resolving were the very edges, although I also omitted the three unpatterned bottom rows, which don’t line up with the rest.  One cone that should be white based on the repeating pattern actually appears to be red—

—but I’ve maintained my policy of not trying to fix contemporaneous mistakes (presuming that neither the original cones nor the photographs have been “retouched” and that the original arrangement of cones wasn’t compromised during excavation and reassembly).  One other complicating factor is the staggering of cones from row to row, like this:


To reformat this data for onscreen display, I’ve rendered each cone as a 2×2-pixel square.  Each cone is represented below by pixels of the correct color (white, black, or red), while missing cones are represented by gray pixels.

This, I’d argue, is a five-thousand-year-old born-digital graphic.  It’s missing some parts at the top, but the image as a whole consists of repetitions of a single 10×16-cone pattern (with one apparent deviation or “mistake,” as noted above), so as long as we’ve correctly transcribed that, perhaps we’re not missing anything after all.

Now, your gut may be telling you that Sumerian cone mosaic images like this one just couldn’t be properly “digital,” and that I must have lost my way down some conceptual rabbit-hole.  That’s fine, if so—my goal here has been to play devil’s advocate as much as anything else.  But if not “digital,” what would you call them?


Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s