Improving CNN-based colorization of B&W photographs
Sanae Boutarfass, Bernard Besserer · 2020
Colorization is the process of converting a black and white image - lot of different shades of gray - into a realistic colour image. Our contribution highlights the fact that the CNNs used for colorization (and specifically the data set used for their training) are not adequate to colourize legacy B&W pictures. In fact the data sets are exclusively composed of colour images, and turning these colour images into greyscale pushes the CNN to learn the contribution of each colour to the resulting luminance. Anyway, CNN are quite good in classification task and can easily recognize skies, trees and foliage, faces and persons. The colorization works well for these elements, but it fails even on memorable objects such as flags and well-known monuments albeit their representation is present in the data sets. We make improvement by using a pretrained network and add hints in the colorization process. Global hints (a colour palette given to the CNN conjointly to the image) still leave a high degree of automation, and these hints could be provided once for a collection of images related to the same theme, which is very suitable for movie colorization. Adding manual hints by scribbling colours onto the B&W image leads to even better results but needs user interaction.