Do We Always Need AI for Image Colorization?

Abstract

Monochrome image (re-)colorization is a problem where human guidance is considered indispensable (even if AI is used). We challenge this approach by presenting a method for converting grayscale images into credible color counterparts with no priors provided/learned on image content, semantics, etc. In contrast to our recent works (focusing on images from non-visual domains) this paper discusses mainly colorization of real-world images for which certain coloristic expectations exist. We argue (and show on selected examples) that a simple technique combining non-deterministic heuristics and randomized flood-fill can deliver colored outputs which are visually convincing and (occasionally) similar to the ground-truth images.