New App Makes It Easier to Colorize Old Photos
The software combines human input and a sophisticated neural network to make historical images pop
/https://tf-cmsv2-smithsonianmag-media.s3.amazonaws.com/filer/6b/0d/6b0da437-5ee0-4108-8f77-6964bef50c0d/einstein_color.png)
In recent years, colorizing historical photos has become something of a trend. Though scholars have long debated whether images should be colorized, many argue that it helps bring history alive in modern times.
But painting over images in Photoshop is no small task and scientists have long worked to find new and faster ways to add that extra visual bling. Now, a colorizing app could help speed up the process, using a combination of artificial intelligence and a human artist, Andrew Liszewski reports for Gizmodo.
The app, called Interactive Deep Colorization, is the work of Richard Zhang and his team at the University of California, Berkeley. In 2016 the researchers posted a paper on the preprint server arXiv, demonstrating the use of convolutional neural networks—a type of neural network known for its ability in object identification—for colorizing images. The results were overall impressive, but the machine would occasionally make unusual color choices and lacked an artist's eye.
Now, in its latest update, the researchers have tweaked the system so a person can add in their own two cents, placing color-swatches throughout the photo that the neural network then uses as the basis for its color choices.
As Liszewski reports, the app reduces the time colorization takes from hours to minutes and creates much more realistic images. In order to test the new system, the researchers gave 28 users a short two-minute training session before giving each of them 10 images to colorize, Zhang writes.
Though the images were originally in color, the team converted them to black and white for the project so the researchers could “ground truth” the colorization choices. While the accuracy of the colors may be off in the user-assisted images (i.e. a green telephone could be rendered red upon user suggestions), the user-assisted images look better overall than those colorized by Zhang’s earlier automatic algorithm and automatic colorization algorithms created by other teams.
There may come a day when such capabilities are built into imaging editing software like Photoshop, but until then Zhang has released the app for free on Github.
Zhang, however, does not address the controversial question of whether black and white photos should be colorized. That’s something that artists, archivists and historians will likely continue to wrestle with—no matter how simple the process becomes.
