| Popis: |
Colorization is the process of converting a black and white image - lot of different shades of grey - into a realistic colour image. Most of the published works uses data sets with colour images and turns these into greyscale at training stage. CNN are quite good in classification task and can easily recognize skies, trees and foliage, faces and persons, and usually colorization works well for these elements, but fails to give accurate colour to a car, mostly because if a car is highlighted in an image, it’s a sport car and all sport cars are red.! Improvement is possible by using a pre-trained network for colorization and further strengthen the training process by adding a palette of relevant (salient) colours to the monochrome image. Our palette generation emphasizes salient colours aside from memorable colours (sky blue, leaf green), the latter are generally well recreated by most pre-trained CNNs devoted to colorization. When the network is finally used, the user defines some hints (i.e. small colour palette) prior to the colorization tasks. Unlike manual colorization these hints are provided once for a collection of images, making the technique suitable for movie colorization. |
