Starting Friday, 2007-01-05, I spent about 24 hours making a TrueType font from a handwriting sample from a previous night — my first experience making a TrueType font. Much to my surprise, the result was good enough to use at smallish sizes. So I thought I'd write about the experience, in case other people wanted to try something similar; related pages include Tom 7's fontmaking page and Ben Sittler's announcement of some fonts of his own. 2008 update: pandaba recommends the MyFirstFont.com tutorial, which looks pretty awesome, although it involves proprietary software.
I finished up on Saturday morning, and you can download the resulting font at the oilpencil font page.
Apparently this same service — turning your handwriting into a TrueType font — is available for US$9 from a fully-automated web site called Fontifier, by a company called "Human-Computer Interface", apparently by a person named David Johnson-Davies.
For the most part, I created the glyphs by the following process (illustrated here with capital R):
This was a bad way to do things, because it took me a lot longer to turn the page into good character outlines than it would have taken to draw them from scratch in Inkscape or even Fontforge, which is what I did for some of the later characters. However, the pencil is much faster for experimenting with shapes and weights and so forth than Inkscape, at least for me with my tiny low-quality optical mouse.
Alvin "DRAFT/MATIC No. DM03" 0.3mm mechanical pencil.
I like this pencil a lot, but I think that's mostly because it's an 0.3mm mechanical pencil. You used to be able to get 0.3mm pencils and lead at office-supply stores that sold drafting supplies, but I had to go to an art store to find it. It cost US$12, plus another $8 for the lead. With it I can handwrite legible text in my paper notebook at 5-7 lines per inch, 3-4 words per horizontal inch, which is somewhere around twice the information density of "standard" textual computer printouts (66 lines of 80 columns per A4-size page). Sadly, though, a nice pencil is not enough to compensate for my lack of artistic expertise.
Konica Minolta DiMAGE X50 5-megapixel digital camera.
This belongs to Beatrice. It's a lot more portable than a flatbed scanner, and I used it to photograph the page at about 470 dpi. Unfortunately, we don't have RAW-processing software for it, so we lose a bit of its potential resolution for monochrome applications like this. This model of camera currently (2007-01-07) costs about US$110 on eBay. (Auctions to check after they close for pricing: 1 2 3 4 5 6)
The GIMP, version 2.2.8. Current is 2.2.13. An image editing program similar to Adobe Photoshop.
Current is 1.7. Turns pixel graphics into outlines.
Current is 0.44.1. A vector graphics editor, similar to Adobe Illustrator, that uses SVG as its native file format. I find editing shapes in Inkscape noticeably easier than editing them in Fontforge.
Current is 20061220. An outline font editor similar to Fontographer. I haven't used any other font editors in the past, but man, Fontforge is awesome! It looks like the other programs that do similar things are Fontographer, Fontlab, Metafont, Ikarus/FontMaster, and TypeTool, more or less in order of popularity.
An optical mouse.
My laptop has a "nipple mouse" rather than a real mouse, and it's very difficult to control. I've finally gotten used to it, but it's very slow, especially for freehand-drawing kinds of things. A tiny no-brand USB optical mouse helps a lot here.
Here's some of the stuff I learned doing things this way.
I spent most of my time interfacing one thing to another, not drawing letters. This was partly because I was learning things, partly because some things were buggy, partly because I couldn't figure out how to script them. Mostly, I think it was because I didn't know there was a better way.
I find drawing with a pencil a lot easier than drawing on a computer, even with the optical mouse. But eventually, I found drawing in Inkscape easier than going through the whole paper path. Still, drawing in Inkscape went better with some sketches on paper first.
I photographed the page at about 450-500 dpi, and a typical source letter was 80x80 pixels, with typical letter features being 5 or 10 pixels across. This ought to be enough resolution for the task, and actually I should have been able to get away with considerably less; however, Tom 7 recommends 300xN pixels per scanned letter, and that probably would have worked better. In fact, though, the image quality gave me problems from beginning to end — but not because of resolution. A reduced-size version of the image, with one pixel for every 5x5 pixels in the original image, is shown below.
For one thing, the picture was slightly blurry, possibly out of focus, although I didn't notice this until too late. The biggest problems were lighting variation and noise.
I probably should have used sunlight or intense off-camera light, but I didn't. And the camera was fairly close to the page, so the difference in distance-to-flash between different parts of the page was fairly substantial, and the whole page was toward the top of the camera's dynamic range. The whites in the middle of the page were 254 (on a scale of 0 to 255), while whites toward the edge of the page were only around 180. The difference between "white" and "pencil lead" was only around 30 counts at best, and in many places was only 5 or 10 counts. So the spurious variation in brightness across the page was 2-25 times larger than the variation representing the pencil marks.
The flash illumination also caused some trouble by causing specular reflections off the pencil lead in places, putting bright white in the middle of some of the letters.
There were also (at least) two sources of noise — camera sensor noise and the surface texture of the paper, which resulted in brightness variations under the oblique point-source lighting conditions of the on-camera flash. I don't know how to separate the two, but different parts of the page had the following ranges of apparently random local variation: 177-187, 189-197, 199-208, 218-225, 229-232, 233-238, 244-247, 253-254.
In general, light helps with noise and blurring, by turning down the gain on the sensor, using a smaller aperture, and using a shorter exposure time.
Shorter exposure times reduce blurring because things don't have as much time to move or vibrate; smaller apertures reduce blurring because they increase the depth of field, reducing the necessity of correct focus; and lower gain (expressed in terms of ISO) reduces noise originating at the sensor.
According to the EXIF information, my original photo used ISO equivalent of 160, exposure time of 5.6ms, and an aperture of f/6.7; a new photo of the page that I took today with the same camera, without the flash, of the same page, in cloud-attenuated sunlight used ISO equivalent of 50, exposure time of 2.0ms, and an aperture of f/2.8. (I'm assuming these numbers are correct.) f/2.8 (which means the aperture diameter is the focal length, f, divided by 2.8) represents a larger aperture than f/6.7, with a diameter 2.4 times as great, and thus with 5.7 times the area — but with a PSF (point spread function) dilated by 2.4 times larger. The other two changes represent about a 9x increase in the amount of incident light required to reach a given pixel value; the larger aperture cuts it down to about 1.6x.
So when I retook this photo in diffuse sunlight instead of with the flash, the lighting field was much flatter, any sensor noise was reduced by more than a factor of 3, movement of the book or my hand probably would have caused less than half as much blur (I'm assuming that the exposure time wasn't significantly longer than the on time of the strobe), and any being out of focus would have caused more than twice as much blur.
The upshot is that in this version of the photo, the lighter parts of the page vary from the darker ones only by about 40 counts, the random noise on the white paper is much less, and nearly all the pencil marks are darker than nearly all of the paper. Compare the original flash shot with the sunlit version:
Thresholding is the process of taking a picture with lots of colors in it (shades of gray, in this case) and turning it into a picture with just black pixels, representing ink, and white pixels, representing paper — no gray pixels.
This is where things started to get hard.
In this version of the reduced-size image, pixels with a brightness between 230 and 240 have been turned white. There's a big white band running around the page in areas where the paper was mostly between 230 and 240, although it's dotted with specks where there was random local variation outside that range. Even with this severe reduction in size, many specks remain, showing how strong the random noise is. Some of the pencil marks in this area are only 15 counts darker than the white paper in their area, although others are as much as 30 counts darker. So it takes some work to find a threshold, even for a single letter, that includes the pencil marks of that letter, but not any of the white paper in the vicinity.
Perhaps more important for the quality of the final product, notice how the black pencil shapes in the white band vary in thickness. They were all drawn with the same pencil lead, and they're all pretty much the same thickness in reality --- about 0.3mm, although there are parts of some letters that are thinner. The aspect ratio of the crossbar of the capital A is about 6:1. But in this thresholded picture, it looks like it's about 2.5:1. Consistently, the letter parts toward the outside (dimmer) part of the white band appear wider, and the ones toward the inside (brighter) parts appear brighter. Why?
Apparently, because they were blurry in the original photo. If they had been clear in the original photo, there would be a one-pixel-wide line of partly-penciled-in pixels separating a pool of pure pencil pixels from a pool of pure paper pixels. But instead there's a gradual slope from one to the other. If you run an edge-detection algorithm like Sobel's on a blurry image like that, it doesn't detect hard edges anywhere, but it detects soft but very wide "edges" occupying the entire volume of the letter, with perhaps a single-pixel-wide line of non-edge in the middle.
So if you threshold some part of this image with a fixed threshold, not only will you get noise, but also, the width of the lines will depend on what the threshold is. So some letters will look much "heavier" than others. This effect is noticeable in the resulting font, since I haven't taken the time to finish cleaning up the mess.
So, in the GIMP, I decided to flatten out the lighting field as well as I could before thresholding. I duplicated the layer (Layer -> Duplicate), did a Gaussian blur (Filters -> Blur -> Gaussian Blur...) with a fairly large radius, like 128-256 pixels. The idea was to make a layer with the "paper color" of the area near each pixel, approximately representing the lighting field. Then, to find the difference between each "paper color pixel" and the actual image pixel, I opened the Layers dialog (Dialogs -> Layers) and set the opacity of the new layer to 50% and the "Mode" to "Subtract".
This resulted in a much "flatter" image, with a more even paper brightness over the entire image, which made it possible to set more sensible thresholds. It's not perfectly flat — paper that's close to lots of pencil markings, or close to the edge of the paper, gets a brightness boost — but it's a big improvement. Unfortunately, it also lost one bit of precision from the original pixel values, which is important when some of the pencil pixels only differ from the paper around them by 5 or 10 counts. (I think it may have lost more, but I can't prove it.) And it has some funky behavior with the edge of the image — maybe it's treating it as black? I'm not sure.
Here you can see the results; while without flattening, no single threshold was even close to adequate for even two successive lines of text, after flattening, results are much more consistent. But another problem remains.
Because of the combination of thick, blurry letter-edges, and random noise that's almost as bright as the pencil "signal", any threshold picks up all kinds of random shapes attached to the edges of letters, even if it avoids picking out speckles all over the paper.
At first, I was so stupid that I set my thresholds with the "Curves" dialog box, by drawing a curve that was 0 until it hit the threshold, then jumped right up to 1. This was painful and time-consuming, partly because there's some kind of built-in limit on how fast the curve can rise, so sometimes I had to go through multiple iterations. After about 20 minutes of this, I noticed that Tools -> Color Tools also had a "Threshold..." option. That made thresholding a lot easier! Be less stupid than me. I'm still too stupid to find the GIMP option that does the same thing as 'pgmnorm', though &mdash turning the brightest color into white, the darkest color into black, and linearly stretching everything in between. I can manually do it with the "Levels" dialog box.
In general, this strategy of thresholding each pixel by its difference from other pixels nearby instead of by its raw value is called "adaptive thresholding", and I probably should have used some existing adaptive-thresholding code; but I thought doing it by hand would be faster in this case.
Around this time, I started working with small parts of the image instead of the whole thing. It wasn't just that setting a single threshold for the whole image was impossible (even after flattening); it was also that the GIMP was using lots of memory and all the operations were kind of slow, and once I started importing paths into Inkscape (see below), it got way slower.
I also learned that the "Threshold..." dialog can apply to just the current selection; this allowed me to set slightly different thresholds for each letter, which helped a fair bit with the later letters.
As mentioned previously, due to the blurry letter edges, lots of noise sticks to the sides of letters. I haven't found a way to get potrace (see below) to ignore chunks of noise smaller than a certain size when they're stuck on the sides of bigger shapes like this. It has a "-t" option (short for "turdsize") for ignoring stray pixels off by themselves, but setting that to 10 or so didn't have any visible effect on letter outlines.
xv (wow, I sure do miss xv, but I stopped using shareware a while back) had a filter called "oilpaint", which replaced each pixel with the most common pixel in its neighborhood. It occurred to me that something similar, used on the output of thresholding, would clean up some of that poop stuck to the sides of letters.
It turns out that the GIMP does have something called "Oilify" (Filters -> Artistic -> Oilify), and while what it does is not documented, it seems to do more or less the right thing.
I used stand-alone potrace to vectorize the paths. I used this tiny Makefile to invoke it:
%.pnm: %.png pngtopnm < $< > $@ %.svg: %.pnm potrace --svg < $< > $@
There's not much more to say about potrace, except that this was a dumb idea and I should have used one of the integrated potrace interfaces in Inkscape and Fontforge (see Better Ways to Do It).
Most of the time that I spent on this project was time I spent shuttling data from one program to the next. Most of the rest was time I spent cleaning up paths.
Basically, once you load an SVG file into Inkscape, you can edit the "paths" in it. Each letter is outlined by "paths" — one that goes around the outside, and maybe more that make holes in the middle. Each path is made of "nodes" connected by curves and lines. The "F2" key lets you click on a letter and see all the nodes.
Pretty much anything in Inkscape can be undone with Ctrl-Z and redone with Ctrl-Y. I use Ctrl-Z several times a minute on average in Inkscape.
If you potrace several letters at once, initially all the paths will be combined into one single path object. There's an option on the "Path" menu to "Break Apart" this object into its component paths, which you can then manipulate individually — for example, selecting some of them, deleting them, cutting and pasting them into a new document, or just moving them outside of the area where they started. You'll have to use the "Path ->Combine" option to put them back together in order for holes in the middle of letters to work right.
The letters out of potrace will generally have a lot of ugly knobs and things on them. The "Path -> Simplify" option can help a lot with this by removing extra nodes, and you can remove more of them yourself by clicking on the node and hitting the "Delete" key.
You may have to adjust the control points attached to the remaining path nodes to get the shape of the curve back, but usually you can get by with only one control node on each side of each 90-degree turn in a curve (so two for a right-angle bend, three for a half-circle, four for a circle), plus a node at each corner. As produced by potrace, Inkscape thinks of the nodes as "cusp" nodes, which can have sharp angles at them, but none of them actually will have sharp angles. If you're moving the control points attached to one of these nodes, you will often want to hold down the Shift key to move both control points in sync, to avoid introducing a sharp point.
The distance between the control point and the node controls the "stiffness" of the curve on that side of that node. Nodes whose control points are very close to them are only introducing small local jogs in the curve. Usually I delete them. If you delete a lot of nodes from a curve, you'll probably want to stiffen the curve at the remaining control points so that it remains curved.
If a node has very different stiffnesses on its two sides, the abrupt change in curvature will be conspicuous and possibly ugly. If you want to avoid this automatically, you can Ctrl-click on the node until it says it's become a "symmetric" node in the statusbar at bottom. A "smooth" node is constrained to have its two control points be in line with each other, so there's no sharp angle, but abrupt changes of curvature are allowed.
A smooth control point (or a cusp that's currently configured like one) will be an inflection point in the curve if its control points are on opposite sides of the curve. If that's not what you want, adjust them until they're on the same side of the curve. You may want to zoom in ("+" key) for this.
For a curve segment that's relatively evenly curved — neither flat in the middle nor flat toward the ends — the control points should be about as far apart as they are from their nodes, so they're near the one-third and two-thirds points along the curve.
Fewer control points are generally better than more. In the rare cases where you do need to add a control point to a potrace-output curve, click on the node at one end of the segment, shift-click on the other, then click the "insert new node" button. The new node starts out as a "smooth" node, so you may have to Ctrl-click it into submission. Also beware — all three nodes stay selected, which is handy if you want to add more nodes (just click the button again) but annoying in the normal case where you just want to move your newly created node.
Because all of this curve editing was so painfully slow, I imported a lot of the lowercase letters and digits without cleaning them up at all. Much to my surprise, at the sizes I usually have them on my screen, I can barely see all the pixel poop on, say, "h", or "j", "q", or "9", or especially "w", "1", "2", and "8".
Tom 7 knows roughly a thousand times as much as I do about making fonts, and his page on making fonts with Fontographer explains many more fine points. Most of the details are applicable to Fontforge as well, down to things like Ctrl-K to open the metrics window.
Once I had things looking OK in Inkscape, I would do Ctrl-C to copy, Ctrl-N to open a new drawing, Ctrl-V to paste, possibly resize the results (I think the size and position of the letterform in the drawing controls the initial size and position of the letterform in the character cell in Fontforge, but I'm not sure), then Ctrl-S to save the result as EPS &mdash I have to select the "EPS" format from the menu near the bottom of the "Save as" dialog box. (It isn't sufficient to type "uni0041.eps" in the filename box; that will result in an SVG file named "uni0041.eps.svg" if it's set to save as SVG.) Then there's some annoying "EPS save options" dialog box that comes up. My version of Inkscape has a couple of bugs — it can't reimport the EPS it exports, and if you close the window that saved the first EPS, later EPS saves fail silently (although they do print error messages in your xterm if you launched Inkscape from an xterm.)
For some reason, much of the time, the letter would paste into the new document sideways and backwards by default. There are toolbar buttons conveniently placed to fix this.
Closing a drawing that has been saved only as an EPS requires dealing with an annoying confirmation dialog box; if you pay enough attention to it, you might notice that the error message is different than usual if an EPS save silently failed.
(I really wish there was a "save selection as EPS file" command. Maybe there is and I'm just too dumb to find it. There is an online tutorial for creating extensions to Inkscape in Python so it should be easy to add.)
All of this makes exporting letterforms from Inkscape into EPS time-consuming and maddening. Fortunately, importing them into Fontforge is marginally less time-consuming and maddening.
"Import" is on the "File" menu in Fontforge, and the dialog box that comes up has a list of file formats available, one of which is "EPS". You can select a filename by typing the beginning of it, but only if you type fast.
It's probably better to use the "EPS Template" feature of Fontforge (see below). But, either way, once you have the letterforms loaded into Fontforge, you have to move and resize them into position, and possibly also rotate them. For no reason I could figure out, about a third of my letterforms got loaded in rotated about 135 degrees counterclockwise.
Moving and resizing the outlines in Fontforge, once you've imported them, is fairly straightforward, if boring. There's a "resize" tool in the Fontforge toolbox; you drag from the fixed-point of the resize in a direction that indicates what kind of resizing you want to do. Normal resizing is up and right for larger, or down and left for smaller, and holding down "shift" constrains it to screw up the aspect ratio only by integer ratios. If some nodes are selected (highlighted in yellow), only they are affected; otherwise all the nodes and also the character width are affected, and sometimes the character width ends up someplace far away.
I started out manually kerning character pairs, but eventually I gave up and kerned by classes instead. Since none of my software pays any attention to the kerns, this was wasted effort for me, and I don't know how bad or good a job I did.
It turns out that the Fontforge .sfd file is a text file that can be sensibly version-controlled --- changes tend to be limited to a few lines at a time. After the first time or two that Fontforge crashed on me, I decided to keep this project in a source-control system so I could go back to older versions if Fontforge wrote something to the file that caused it to crash at startup.
This was as simple as
darcs init darcs add oilpencil.sfd darcs record
and then periodically doing darcs diff -u and darcs record.
I considered checking all the original source photos and parts of them into the project, but they are much too large. I have, however, included the whole web site in the darcs repository in order to make it easier to manage, and this did save me from scrozzling a file once.
I did this all in much more difficult ways than I needed to. Be less dumb — here's several ways to make things easier next time.
If your input image isn't as screwed up as mine, Inkscape is capable of doing thresholding and running potrace itself; there's a tutorial on tracing with Inkscape on the project web site, and the same tutorial is also included in the Inkscape package itself, on the "Tutorials" submenu of the "Help" menu.
Fontforge is also capable of running potrace itself ("Import" on the "File" menu to read in the graphic, then "Auto Trace" on the "Element" menu; it'll use either potrace or the "autotrace" program, whichever one is installed) if you have already thresholded it, but doing it in Inkscape has a couple of advantages. Fontforge can't import color images in a useful way (even color images that contain only colors that look pretty gray), so you have to convert the image to gray before trying to load it into Fontforge; and unsurprisingly it doesn't support JPEG. Also, Fontforge doesn't seem to be very smart about thresholding them; Inkscape has a handy dialog box for playing with the thresholding so you can get it right.
However, Fontforge does have the ability to import and autotrace a whole bunch of characters at once, which might matter more. And if you have a JPEG file of a letter that you want to turn into a two-color JPEG, you can do it with these two commands (assuming, of course, that you have netpbm and ImageMagick installed):
jpegtopnm < capital-r/photographed.jpg | ppmtopgm | pgmnorm > tmp.pgm convert -monochrome tmp.pgm tmp.png
If I were again importing a bunch of outlines created in Inkscape and saved in EPS into Fontforge, I would use the "EPS Template" feature of Fontforge's file import. This feature allows you to import a bunch of EPS files named uni0041.eps, uni0042.eps, and so on, as Unicode characters U+0041, U+0042, etc.
If, for whatever reason, I import characters from physical objects such as notebooks in the future, I will use a better scan. An actual scanner would be great, but even with a camera, I could have done a lot better:
It would be straightforward to write a script to extract blocks of pixels from predetermined places in an image, saving each one to a specially named file, so that they could all be potraced and imported into Fontforge at once, without any further human interaction. With specially-marked paper, such as standard blue-lined graph paper, it might even be straightforward to undo perspective distortion from the photo-taking process.
Sketching stuff out on paper to see how it looks, fine. But for the final letterforms, it's probably quicker to use Inkscape to draw the thing than to use Inkscape to clean up the results of scanning. A lot of the punctuation in oilpencil was drawn this way, because it would have been just too much trouble to clean up all the pixel poop.
For line drawings, Inkscape has this awesome feature called "outset", which lets you take a path and make a bigger path around it, just by hitting Ctrl-). So you can draw a "stroke" with the stroke-drawing tool, adjust it to the right shape, convert it to a path (on the Path menu, I think it's Stroke to Path) and "outset" it a few times.
Inkscape is apparently somewhat scriptable now. Fontforge is definitely scriptable. Adaptive thresholding is definitely scriptable (unless you have a totally hopeless image like I did). This could have eased the workflow considerably.
I ended up with about the top 30% of the character cell completely unused. This leads to a variety of problems in practice: characters substituted from other fonts (accented letters, smart quotes, em dashes) look out of place, multifont text looks awful, I have to set my browser's default font size to 20 points to get readable text (which makes text in other fonts huge by default), bullets in bulleted lists are ginormous, and liquid web page layouts end up with far too much space. And all the extra leading (line spacing) makes it hard to get stuff on the screen.Kragen Sitaker <email@example.com>