Colorizing black and white photos with deep learning (opens in new tab)

Well, as a professional photo editor, you know that if the original b+w photo captures an image with say, a 50% grey value, you don't know if the original color was bright red, or closer to that 50% gray. Bright red has a much higher chroma value, but chroma isn't recorded in a b+w photo - color saturation is lost. Easily demonstrated by making an image in your favorite image editing program that's just straight up red, then changing the image mode to grayscale, then asking someone else entirely to guess the original color.

That, and I think the style is to mimic hand tinting photographs, where you would paint right on a b+w photo. The colors would look "faded" because whatever was used to tint the photograph needed to have a transparent medium, for the information of the photograph itself to shine through. That, and there's just so many colors you could use when hand-tinting. Back to our 50% gray. What if that was... bright yellow? You can't tint "bright yellow" onto a 50% gray area of a photograph. Yellow is highly transparent, and the grey would be too powerful to let its chroma value shine through.

Sure, the colors will never be exact, because we don't know what the original colors were. But that doesn't matter in any material way.

True for all AI; neural networks are doing amazing things, but the output is a synthesis, it's a complex interpolation of it's training inputs that may seem "good" or reliable, but it is never to be taken as truth or fact, and it can be arbitrarily wrong with unbounded errors.

> I'm also a little confused as to why colorization always aim to restore color to the equivalent of a faded color negative, with muted tonality and grain. Human logic is funny.

This isn't a human logic problem. Normally colorizations don't affect tonality and grain much, they are putting color splashes on top of a B/W image. This is true of hand-painted colorization, as well as the digital colorization here. You can't get rid of grain or adjust tone by adding color.

One can adjust tone and grain, but then you're doing more than colorizing, and going even further down the road of "interpretation" you're concerned about.

In this particular case, the author did mention "A more diverse dataset makes the pictures brownish". Brown is the average color in natural photos, so minimizing error tends to make things browish. That is separate from leaving faded tone & grain in tact, but it's a second reason why AI based colorization will tend toward muted color.

I always assumed that if you tried to use "full color" it would look weird, since the photos themselves usually are quite faded and grainy.

As a photographer, collector and enthusiast of vintage prints and photos and amateur historian, I deeply appreciate the aesthetic of colorized photos and understand the motivation to reproduce and master it. There is more to art than truth.

The artist is not the transcriber of the world, he is its rival. - L'Intemporel (Third volume of 'The Metamorphosis of the Gods'), André Malraux (1957)

Many movies set in the past, or that have flashbacks to the past, will often mute the colors. Most modern movies muck around with the colors in post production, too. The worst is when they go for the blue/orange palette.

That means the luminance values can be completely different from what they actually were, and skew the color choices. On the other hand, this might be something that a good ML algorithm can detect and compensate for.

https://en.wikipedia.org/wiki/List_of_black-and-white_films_...

At least to the extent that GANs work, it works. They will alternate between the observed colours based on the noise vector. They do not simply converge to averages, because the discriminator easily recognizes brown apples as fakes.

I have an urge to teach it simple logic. Instead of making it brown, it selects the color with the highest probability from a range of colors. However, I haven't come across a deep learning implementation like this to mimic.

Using something akin to a variational autoencoder would solve this problem, because it learns a distributional approximation rather than a single point estimate of the color, and then the random noise vector input allows one to sample from this output distribution. Similarly, Mixture Density Networks allow you to model a distribution and then sample from it.

Alternatively, use a discrete colour space and consider colours as categorical data not implying any ordinal scale.

I dunno, does it look more like a red apple, a green apple or a yellow apple?

[1] http://colah.github.io/

[2] https://iamtrask.github.io/

[3] https://distill.pub

[4] https://experiments.withgoogle.com/ai

These types of techniques should let you explore the hyperparameter space much more quickly (and cheaply!), but I agree - having money to burn on EC2 (or access to powerful GPUs) will still be a major factor in tuning models.

[1] https://twitter.com/karpathy/status/913619934575390720

An example implementation would look something like edges2cats https://affinelayer.com/pixsrv/