Grey is the new RGB: How good is GAN-based image colorization for image compression?

Abstract

GAN-based image colorization techniques are capable of producing highly realistic color in real-time. Subjective assessment of these approaches has demonstrated that humans are unable to differentiate between a true RGB image and a colorized image. In this work, we evaluate the fidelity of such colorization and for the first time analyze the GAN-based image colorization scheme in the context of image compression. Our analysis shows that the palette (set of colors) recommended by the GAN-based framework is very limited even for highly realistic interactive colorization. We propose two novel methods of automatic palette generation that allows for the GAN-based framework to be useful for image compression. We demonstrate that provided true colors at a few pixel locations, GAN-based approach results in good spread of color to other image regions. Subjective analysis on a number of public datasets shows that the current system has low fidelity but performs better than JPEG at low data rate regimes.