A computer has mastered Sumi-e, a style of oriental ink painting that conveys a scene with just a few, delicate brush strokes.
The challenge Ning Xie and colleagues at the Tokyo Institute of Technology faced was to get a computer to generate natural-looking brush strokes in a repeatable manner.
Previously, researchers have tackled the problem using a so-called dynamic programming approach, but doing so was computationally expensive and only produced a stroke for a specific shape.
That’s a problem if you want to turn a snapshot of a flower vase on the table into a Sumi-e painting with the touch of a button and do it all again moments later with a picture of your child in a sandbox.
So, Xie’s team used a technique called reinforcement learning, whereby the computer was rewarded when it made nice, smooth lines for a variety of strokes but not when it stroked poorly.
In this way, the computer learned a library full of brush strokes and now calls on them at any time to create a Sumi-e painting.
The only catch is that humans manually draw contours on a photo that the computer then fills in with strokes.
"The results ... we think are of good quality," Xie and colleagues conclude.
— Via Technology Review