DDGAN: Double Discriminators GAN for Accurate Image Colorization

The purpose of image colorization is to map reasonable colors for grayscale images. Although more and more deep neural networks have been proposed and shown good performance, the color images generated usually have the phenomenon of color crossing. In this paper, we propose DDGAN, a Generative Adversarial Networks (GANs) with double auto-encoding discriminators to solve this problem. The purpose of adding a discriminator is to use the complementary statistical properties of KL divergence and reverse KL divergence to avoid mode collapse and improve training stability. The auto-encoding structure is used for pixel level adversarial training to obtain color images with accurate edges. Experiments show that DDGAN has a better performance in the correctness of color migration and the suppression of excessive smoothing of color at the boundary.