Colorizing old black-and-white photos with Photoshop has been a popular subject on the Internet over the past few years, as skilled retouchers use their time and skills to offer a new view of vintage images. In the future, though, software may be able colorize B&W photos with the click of a button.
Automatically colorizing black-and-white photos is a problem that computer science researchers are hard at work on. In a new paper titled “Colorful Image Colorization,” UC Berkeley computer vision PhD student Richard Zhang and his team share how they’re using a “convolutional neural network” to create automatic colorizations that can often fool humans.
Their new system is full automatic — other past systems have required a great deal of use input — and is trained on over 1 million color photos. Compared to previous systems, this new one creates colorized photos that are more vibrant and realistic.
To test the system they created, the scientists found human test subjects and showed them two versions of photos — one would be the real color photo, and the other would be one that artificially colorized by the software. The humans were asked to figure out which was which.
20% of the photo pairs actually fooled the humans, meaning the colorized versions were guessed as the real color photos — this “fool rate” is much higher than prior research in this area.
Here are some sample photos that were automatically colorized by the system. This first set contains photos by renowned landscape photographer Ansel Adams (left is the original Adams shot, and right is the colorized one):
Here’s another before-and-after using photos by photographer Henri Cartier-Bresson:
While the results aren’t perfect, it’s a big step in the right direction. Researchers are working on making the results indistinguishable from actual color photos.
Perhaps one day we’ll find an Auto-Colorize feature built into Photoshop that turns a monochrome photo into a color one with a simple click of your mouse.