Never seen before
Once the neural network has learned these relationships as best it can, it can be asked to apply the learnings onto a neverbefore-seen black and white image (e.g. one of George Nepia from 1924). It looks through the new image for any features it discovered in the training phase and then for each pixel it provides its best guess for the red, blue and green values – the output of which is a colour image.
Jones was happy with how the photos turned out. Often on the vast majority of these types of projects it’s trial and error, he said, essentially fine-tuning a wide range of parameters to get things to work the way you want. In this case he felt he got a bit lucky, as it worked pretty well early on.
There’s always room for improvement, though, he acknowledged, as in some cases the model doesn’t quite understand what colour to apply, which at one stage saw Nepia sporting a head of bright orange hair.
The solution is for him to train the model on more data – instead of giving neural network around 1200 images to learn, which he said is quite low, you would look to train it on tens or even hundreds of thousands of images.
While Jones doesn’t have any further projects specifically lined up, he said if he finds the time he would like to play around with the idea of ‘enhancing’ old photos as well.
He said that would involve obtaining high definition pictures and artificially ‘age’ them, then get a model to learn the relationship between the aged versions and HD versions. That would mean potentially taking the colourised All Blacks photos and making them look like they were taken on a modern camera.
Attempting to colourise video clips is also something he’s had an initial attempt at doing, and something which he said could certainly bring old sporting video to life for newer generations to enjoy and appreciate.