Wiener Loss: A Strong Correlative Loss Applied to Conditional GAN for Color Prediction

Abstract

Colorization is a task to generate plausible color for a given grayscale image, where the target in input always have variable color styles. Due to the ill-posed inverse problem in colorization, the generated color image easily suffers from the phenomenon of color cast, where unexpected particular color affects the generated image. To cope with such problem, this paper proposes a novel loss function, called Wiener Loss, to constrain the training of colorization network. Concretely, we adopt conditional generative network for training. This paper uses the colorized image from generator and ground truth corporately to calculate their relevance defined as Wiener Loss and feeds this loss back into generator network for training. The experiments demonstrate that our method generates better results compared with its counterpart.