Technical details

We have developed a novel Artificial Intelligence (AI)-based technique for the automatic colorization of black and white images. Historical photography consists of black and white images due to the limitation of the technology at the time. By adding color to the photographs, we can make them not only seem more realistic, but make it easier to interpret them. Up until now, the colorization of images has required significant user interaction to obtain good results, in which the user has to manually indicate which areas are of which color. In contrast, in this work we construct a novel AI model that is able to automatically learn the colors of images for a wide variety of scenes, achieving high-quality colorizations without requiring any user interaction. In our model, global features representing the type of the scene are jointly learned with local features representing the texture of different areas in the image in order to effectively colorize images. The global features provide information at an image level such as whether or not the image was taken indoors, while the local features represent local objects and textures such as grass, brick walls, or human skin. We validate our approach with a user study and compare it against the state-of-the-art in which we show significant improvements. We also show how our approach can realistically color 100 year old photographs and believe it can provide people with a chance to more intuitively understand the past.

Movies

Exhibition content

We will show a real-time demo of the colorization of digital black-and-white images.