Colorizing old photos is a long process, but using a “convolutional neural network,” Richard Zhang, a UC Berkeley computer vision PhD student, has developed a program with a much higher success rate than earlier attempts.
Convolutional neural networks (CNNs) are advanced image-recognition programs. They use multiple layers of overlapping input regions to create a better representation of the original than earlier technology allowed. Zhang took the idea of using CNNs a bit further and trained the artificial intelligence program by using over a million color photos.
The results? The computer-colored images fooled people in a test about twenty percent of the time. Of course, that means most of the time the images aren’t as good as photos colorized manually, but twenty percent is significantly better than previous attempts at building color conversion software.
Colorizing black and white photos is a tough task to give a computer because often things like a t-shirt could be any number of different colors. According to Zhang’s research, black-and-white images still offer plenty of clues using texture and shape of objects that are always (or almost always) a certain color, like green grass and a blue sky. “Our goal is not necessarily to recover the actual ground truth color, but rather to produce a plausible colorization that could potentially fool a human observer,” he says.
While the program still has a way to go, Algorithmia shows early promise that computers could actually be capable of quite a bit with the right training. Regardless, the new program is fun to play around with — the early version is available to colorize any image from a URL and even download the results.
