An Efficient Automatic Digital Image Colorization Method Using Convolutional Neural Network

conference contribution

posted on 2020-09-15, 06:39 authored by Gaurav SubediGaurav Subedi

Colors are the elements of human visual

perception that portray the vivid liveliness of the cre-

ation. This perception of color is so diverse that there

is no perfect color combination for a given artifact.

Colorization is a generic term used to describe a com-

puterized process for adding color to black and white

pictures, or motion graphics. Digital Image Colorization

is the colorization process applied to still image artifacts.

Although people are controversial about the artistic

value of colorization, it is no doubt that colorization of

monochrome artifacts enhances the visual effects.

Colorization is basically a mapping between the inten-

sity values and the chrominance values with no ’correct’

but plausible solution. Inspired by the current trends in

deep learning, we propose a colorization framework that

utilizes both the local and global features for colorization.

Global features include the color rarity of each color

class in the quantized ab plane calculated over the

whole dataset. We have also implemented a custom

loss function suitable for this purpose. We tackle the

colorization by treating it as a multi-class classification

by determining the probabilities of colors in the color

gamut and combining the probabilities to a specific

color. We have successfully implemented the method

to various types of images ranging from legacy b&w

images to modern images (having colored versions of

their grayscale counterparts).

History