Aim : to guess how an image would look like if it had been taken by another imaging tool (example: a human body scan taken by different kinds of medical scanners) Application : to convert a greyscale image into a color one, based on other examples of color images Method : to learn the correlation between texture and color in the training set Difficulties : multiple possibilities locally, and spatial coherency
Example : we forget the original colors of (a part of) the "Mona Lisa" painting by Leonardo da Vinci, and recolor it automatically thanks to (a part of) another painting by the same author, namely "Madonna of the yarnwinder":
Another example : colorisation of a picture taken in Iceland thanks to another picture of the same area (but of a different moutain):
+
=
Yet another example : colorisation of a zebra picture thanks to another zebra picture:
+
=
A more challenging example : colorisation of a picture of Charlie Chaplin thanks to a set of several very dissimilar pictures:
Guillaume Charpiat, Ilja Bezrukov, Yasemin Altun, Matthias Hofmann and Bernhard Schölkopf, Machine Learning Methods for Automatic Image Colorization, chapter of the book Computational Photography: Methods and Applications, R. Lukac Editor, CRC Press. [bibtex]