Semantic image synthesis is to render foreground (object) given as a text description into a given source image. This has a wide range of applications such as intelligent image manipulation, and is helpful to those who are not good at painting. We propose a generative adversarial network having a pair of discriminators with different architectures, called Paired-D CAN, for semantic image synthesis where the two discriminators make different judgments: one for foreground synthesis and the other for background synthesis. The generator of paired-D GAN has the encoder-decoder architecture with skip-connections and synthesizes an image matching the given text description while preserving other parts of the source image. The two discriminators judge foreground and background of the synthesized image separately to meet an input text description and a source image. The paired-D CAN is trained using the effective adversarial learning process in a simultaneous three-player minimax game. Experimental results on the Caltech-200 bird dataset and the Oxford-102 flower dataset show that Paired-GAN is capable of semantically synthesizing images to match an input text description while retaining the background in a source image against the state-of-the-art methods.
展开▼
