Permalink
Switch branches/tags
Nothing to show
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
188 lines (142 sloc) 6.66 KB

Random project: custom wallpapers from handwritten notes

I tend to have a bunch of loose pages of notes lying around, and a few days ago Joyce scanned them in so I could store digital copies. Some fairly aimless scripting later, I ended up with a bunch of generated backgrounds (thousands, but these are some of the good ones):

image

image

image

image

image

Extracting note images

Joyce used a ScanSnap, which is a two-sided sheet feed scanner that produces PDFs. The first step was to pull the individual pages out using ImageMagick, inverting each in the process:

$ convert -density 600 -trim input.pdf \
          -negate \
          -quality 95 output-%03d.jpg

My notes are awkward to work with digitally because I don't use the whole page in a linear way. So each output file looks something like this:

image

I opened a bunch of them in the GIMP and selected/rotated/cropped to get correctly-oriented note snippets like this:

image

The last change was to convert black to alpha and blur the images a bit. Blurring hides the fact that the scanner dropped the images to monochrome, which loses some detail for handwriting.

$ convert output-001.jpg -fuzz 10% -transparent black -gaussian-blur 10x2 \
          output-transparent-001.png

You might not be able to see this in github, but if you open the image separately you can see white text against the browser's default PNG background:

image

I ended up with 126 of these from 104 scanned images (some of which were blank; I'd guess there were about 80 actual pages of stuff).

Compositing the background

Almost everything about these is random: the background photo, which note images go where, the opacity, blurring, etc. It was a lot easier to automate this and throw away 95% of the outputs than it would have been to lay it out by hand.

The basic idea is to composite the transparent images onto a darkened photo using the plus blend mode. This uses a convert sub-context:

# composite a single note onto the background
$ convert photo.jpg \
          -resize 3840x2160 \
          -evaluate multiply 0.5 \
          \( note.png -resize 30% -geometry +400+600 \) \
          -compose plus -composite \
          +repage output.png

ImageMagick is concatenative, so if we want more than one note we can just add more arguments, in this case \( ... \) -compose plus -composite. The generator script generates the command-line arguments by repeatedly adding stuff to a bash array.

Parameter space

At first I just randomly generated each parameter on an individual basis, but this tended to produce a bunch of wallpapers that looked more or less the same; if you generate enough identically-distributed random numbers, you'll converge to the mean. To avoid this, I started by generating some higher-order parameters up front:

#!/bin/bash
# Usage: ./wallpaper > out.png

# Distribution parameters
opacity_floor=$(( 30 + $RANDOM % 50 ))
size_floor=$(( 4 + $RANDOM % 4 ))
size_scale=$(( 8 + $RANDOM % 16 ))

x_quantum=$(( 1 << $RANDOM % 8 ))
y_quantum=$(( 1 << $RANDOM % 8 ))

max_blur=$(( $RANDOM % 24 + 1 ))

declare -a composite_args
for f in `ls transparent-pngs/* | shuf | ni r.$(( 10 + $RANDOM % 39 ))`; do
  # single-image compositing code...
done

convert `find photos -name '*.jpg' | shuf | head -n1` \
        -resize 3840x2160 \
        -evaluate multiply 0.$(( 10 + $RANDOM % 80 )) \
        "${composite_args[@]}" +repage png:-

Just a heads-up about the ni r.$((stuff)) in there: this is a quick way to randomly select some fraction of input rows. I did it this way because I wanted the process to scale up with the number of images available, which in hindsight isn't a great idea. A better approach would be to replace this with just head -n $(( 20 + $RANDOM % 60 )) or similar; then you've got a more stable control over the density of notes you're mixing in.

Single-image compositing

The contents of the for-loop generate a bunch of variates from the distributions we set up in the header. Each image has a few variables:

  • XY position, quantized to (x_quantum, y_quantum)
  • Orientation
  • Opacity, at least opacity_floor
  • Size as a percentage of the original 600DPI, between size_floor and size_floor + size_scale
  • Blur sigma, up to max_blur (which can be 0)

Here's the code:

# bash uses int-only math, so /q*q is a way to quantize something
px=$(( $RANDOM % 3600 / xq * xq ))
py=$(( $RANDOM % 3600 / yq * yq ))

opacity=0.$(( opacity_floor + $RANDOM % (90 - opacity_floor) ))
size=$(( size_floor + $RANDOM % size_scale ))
blur=0x$(( $RANDOM % max_blur ))

# this is a hacky way to get a biased distribution (in hindsight I should have
# parameterized it, but this configuration seemed reliably inoffensive)
case $(( $RANDOM % 64 )) in
1|2) rotation=90 ;;
3)   rotation=-90 ;;
4|5) rotation=180 ;;
*)   rotation=0 ;;
esac

composite_args+=( \
  \( $f -resize $size% \
        -rotate $rotation \
        -geometry +$px+$py \
        -channel a -evaluate multiply $opacity +channel \
        -bordercolor none -border 100x100 \
        -gaussian-blur $blur \) \
  -compose plus -composite )

It's important to expand the virtual canvas of images before you blur them; otherwise you get blurs that run off the edges and create sharp lines.

Also note the hack to get floating-point values; it's totally possible to use integers to do this as long as they're all in the same order of magnitude.

Rendering a bunch of these

This is a pretty CPU-bound process for a number of reasons, but not all of it is parallelized by default. This means there's some speedup to be had by using xargs:

$ seq 1000 | xargs -P12 -I{} sh -c 'echo {}; ./wallpaper > results/{}.png'

Each wallpaper takes a few minutes of CPU time on the server.

Flying blind: examples of fail

Throwing random parameters at a problem like this is a high-recall, low-precision strategy, which is ok because it's quick to go through the images and throw out the ones I don't like (or more accurately, rescue the few that I do). Here are some that didn't make the cut:

image

image

image

image