Surely you’ve all seen a black and white photo of your great grandparents. Have you ever wondered what they would look like in color? Well, thanks to the Artificial Intelligence of ColouriseSG the magic of little color can bring us a little closer to the moment those pictures were taken.
The process is very simple. Go to web site ColouriseSG press ‘Try it yourself’ and select a photo. After a few seconds you will get a colored version. You can download the generated image or the comparison of both (the application does not save the photos you upload).
There are other coloring tools but perhaps ColouriseSG is the best. Although this is not an exact representation of the moment the photo was captured. The creators warn: “The coloring is an active research field and our model is not perfect. It works well in some photos but in some others not”.
To color the black and white photo, ColouriseSG uses a deep learning technique
When a person colors a black and white photo, he faces two phases that require a lot of time and skill, which can be summarized as follows:
- Research: study in depth what is the historical, geographical and cultural context of the photo to obtain the appropriate colors
- Colored: use software tools such as Photoshop
A computer program does these two tasks a little differently:
- Identification: identify the objects in a black and white photo, and determine a plausible color based on the images you have analyzed in the past
- Colored: colorize the image in black and white
To color the black and white photo, ColouriseSG uses a deep learning technique known as GAN (Generative Adversarial Network). This process consists of two phases:
- Generator: with many mathematical parameters (> 20 million). Try to predict the color values in different pixels in a black and white image, based on the characteristics of the image.
- Discriminator: try to identify if the generated colors are photo-realistic compared to the original color image.
The model is learning until the generator can predict the colors that the discriminator cannot effectively distinguish as false. In this case he trained using more than 500,000 old images.