Title:	AESTHETIC GRADING: COLOR CORRECTION VIA NEURAL NETWORKS
Author(s):	John Kundert-Gibbs
ISBN:	978-989-8533-80-7
Editors:	Ajith P. Abraham, Jörg Roth and Guo Chao Peng
Year:	2018
Edition:	Single
Keywords:	Color Correction, Generative Adversarial Neural Network
Type:	Full Paper
First Page:	89
Last Page:	96
Language:	English
Cover:
Full Contents:	click to dowload
Paper Abstract:	Color grading and correction for film and video is a significant and complex aesthetic and technical task that requires a trained operator and a good deal of time. We compare two deep neural network frameworks—a classification network and a conditional generative adversarial network, or cGAN—for quality of their output as potential automated solutions to color correction. Results are very good for both networks, though each has its own problem areas. The classification network has issues with generalizing due to the need to collect and especially to label all data being used to train it. The cGAN on the other hand can use unlabeled data, which is much easier to collect. The cGAN, however, directly changes image pixels, introducing potential damage to the images, so multiple adjustments to the network need to be made to create high quality output. We find that the data labeling issue for the classification network is a less tractable problem than the image correction and continuity issues discovered with the cGAN method, which have direct solutions. Thus we conclude the cGAN is the more promising network with which to automate color correction and grading.
	Go Back