In Spring 1999 I took a course with Mike Bove, "Digital Image Processing for Hard Copy" (MAS 814). For a final project I worked on making nice ASCII graphics, by thinking of letters as a form of halftoning. I started with AA-lib, an excellent ASCII graphics renderer. I went through and understood its algorithms, documented them, fixed a few bugs in their model of Courier, and then augmented AA-lib with a special postscript output that uses bright and dim characters to increase dynamic range.
I have placed my code and my results online. Unfortunately, the images I produced are not directly web-visible; you have to download postscript programs and send them to your printer. An on-screen postscript interpreter is not good enough, you need high resolution.
ASCII graphics are fun. Now that most everyone has access to bitmapped displays and fancy printers there's not much need for reproducing art as characters. But I still like the æsthetic of ASCII art, the coarse retro feel. There are some excellent archives of ASCII art online, such as Christopher Johnson's ASCII Art Collection. Yahoo has an index. I've prepared a few samples that I particularly like.
ASCII art falls into two categories - line art and fill art. All the best ASCII art is drawn by hand; it's hard to automate. But a good program can produce fairly good fill ASCII art. AA-lib, an ASCII graphics renderer, is the best package I know of for doing this. See in particular the BB Demo for what animated ASCII graphics can do. My project was to improve AA lib, in particular to give it a special Postscript output option to let people make better printed ASCII art.
Most of my experiments were with two images. One is "Hive", a picture of myself and a group of folks I work with, photo copyright Webb Chappell. You can see the original image as portable greymap or compressed postscript. The other is "Circle", a radial gradient from 0.0 gray to 1.0 gray (pgm or ps). Finally, I did a final test with the standard Lena image (pgm).
In the old days, the best tool for ASCII graphics was pbmtoascii. This takes a bitmap and converts it into ASCII by examing groups of 1x2 or 2x4 pixels and translating them to a character. For 1x2 pixels, there are four characters possible: a blank, ", o, or M. The results aren't very good. The 2x4 map is more complex, but still looks pretty bad. The actual characters chosen aren't optimal and the restriction of input to a bitmap limits what can be done.
Here are a few examples of the results of pbmtoascii. Note, these are rendered in very small fonts (Courier 3pt or so). A postscript previewer on screen will not do this justice, you need to print them out to really see what's going on.
AA-lib is a much better way to do things. AA-lib works on greyscale input and has a quite sophisticated model of how to convert to characters. The basic idea is to consider ASCII characters as a very funny kind of halftoning, where you want to minimize the error between the original image and the ASCII substitute. A special aspect is that letters don't just have some intensity, they also have a distinctive shape. AA-lib has a nice mechanism for trying to match shapes so that contours are preserved. (The contouring is quite visible in the circle images I tested.)
The only problem is that AA-Lib is 6000 lines of dense, uncommented code. I went through the software and figured it out, here's a sketch of their algorithm:
AA-Lib can be used as a replacment for pbmtoascii (making pgmtoascii). Getting the aspect ratio right this way is tricky, since characters aren't square. I scaled most things vertically by 66%, that should be about right. The results are much better than pbmtoascii, especially using Courier-Bold to make everything darker:
AA-Lib has a very clever feature (inspired by the old PC monochrome text model). Letters don't have to have one intensity - they can be displayed dim or bold. Obviously, with straight text output you can't simulate this. But with postscript output, you can! To take advantage of this, I wrote a postscript output driver for AA-lib (you can get the code below).
I looked at various strategies for simulating "bright" and "dim" in postscript. The most obvious option is to use Courier-Bold instead of Courier for "bright" characters. That works but feels a bit like cheating, and the range isn't very good. A better option is to draw the string, but then to also draw the outline of the characters with a very thin line. The code is like this:
/B { dup gsave false charpath 0.05 setlinewidth stroke grestore show } bind def
The results are a bit unpredictable. A line of width 0.05pt is very small (1/1440 of an inch), below the
rasterizing capability of most printers. Worse, the "charpath" of an
font doesn't follow the nice Type 1 hinting that the "show" of a font
does. But it happens to look good for
my printer, especially at small sizes like Courier 3pt.
I also tried simulating overstrike by printing the character near to on top of itself
/B { dup gsave -0.04 0 rmoveto show grestore 0.04 0 rmoveto show -0.04 0 rmoveto } bind def
But that didn't work very well - it is also unpredictable, but in this
case wasn't much darker and led to some odd banding.
The obvious thing to try for dim is the opposite of bold: show the character, then draw the outline in white on top.
/D { dup dup gsave gsave show grestore 1 setgray false charpath 0.05 setlinewidth stroke 0 setgray grestore stringwidth rmoveto } bind def
This works OK, but makes the characters look a bit broken. It's also
quite slow for the printer (as is the charpath bold). One can just
fill the character with a light grey:
/D { 0.75 setgray show 0 setgray } bind def
This looks quite good, but feels like cheating.
Personally, I like the charpath solutions best, since they are the most "pure". Multiple Master fonts are really the right way to do this kind of thing, since they have code to say things like "draw this font, only heavier".
AA-Lib does its magic by calculating a table of substitutions based on a bitmap model of the font. If you look at the best circle output, you see that the tone scale is not very uniform - there are obvious bands where things are too light or dark. By tweaking the font model, I fixed these problems.
|
When I extracted the original font table (at right), I discovered a bad off-by-one bug in the courier model that comes with AA-lib. The problem is quite visible in the single pixels below 'h' and 'i' - one pixel from the next character to the right has been carried over! Fixing that problem, the results of AA-lib are much nicer. |
 |
|
Fixing this problem in the font model, AA-lib does much better. However, the circle still shows some tone scale problems; AA-lib thought that some characters were darker than they really were, or lighter. To fix this, I went in and hacked the font table by hand. I changed the models of 'c', 'K', 'w', 'g', 'o', and 'x', and removed '4' entirely (because I couldn't fix it). The corrected table is at right, and the results, particularly on the circle, are quite clear. |  |
Putting all these improvements together: the use of AA-lib, the use of bold/dim, and the font table fixes, the result is quite impressive. These demonstrations are on the classic Lena image. The AA-lib version really looks like a person!
AA-Lib isn't just restricted to ASCII: ISO-Latin-1 and graphic control characters can also be used. And with my code, any postscript font can be used. I wanted to do a Cyrillic-graphics version of this work, but figuring out how to make postscript use a Cyrillic font defeated me. (There are some good free Cyrillic Courier fonts online). Doing ASCII graphics with a proportional font would be lots of fun, although it would make the encoding process quite tricky. Finally, there's some room for entropy in the font table - several characters are often a good match. These extra bits could be used for æsthetic purposes, or to stuff some data into the ASCII image.
I wrote a program that can read in PNM images and write out postscript files, with the full suite of AA-lib options. If you set the environment variable AAMODE to 0, it'll also write out plain text. Feel free to play with my code. It's a hack, not polished or particularly well explained, but it worked well enough for me. If you do something interesting with it, please let me know. I'll help someone if they want to clean this up and release it as an add-on for AA-lib and/or netpbm.
| Nelson Minar | Created: May 13, 1999 |
| <nelson@media.mit.edu> | Updated: May 13, 1999 |