Grayscale Image Colorization with GAN and CycleGAN in Different Image Domain

TL;DR

This work addresses grayscale image colorization using GAN-based approaches. It first reproduces a conditional GAN baseline and a GAN variant, then introduces a conditional CycleGAN that leverages grayscale conditioning and cycle-consistency to improve multi-modal colorization. Evaluations across LSUN bedroom, face, and comic datasets show that the conditional CycleGAN delivers more plausible colorizations for human faces and comics, with improved training stability compared to the baselines, though color diversity can be limited for unseen content. The approach operates in the YUV color space by predicting UV channels from the grayscale Y, enabling straightforward fusion into full color images and demonstrating cross-domain generalization with practical implications for colorizing historical footage, comics, and portraits.

Abstract

Automatic colorization of grayscale image has been a challenging task. Previous research have applied supervised methods in conquering this problem [ 1]. In this paper, we reproduces a GAN-based coloring model, and experiments one of its variant. We also proposed a CycleGAN based model and experiments those methods on various datasets. The result shows that the proposed CycleGAN model does well in human-face coloring and comic coloring, but lack the ability to diverse colorization.

Figures (7)

• Figure 1: The network structure of the baseline model.
• Figure 2: coloring conditional CycleGAN model structure
• Figure 3: Labeled Face dataset coloring
• Figure 4: Conditional cycleGAN colorized comic
• Figure 5: Conditional cycleGAN colorized comic 2