Semantic layouts based Image synthesizing, which has benefited from thesuccess of Generative Adversarial Network (GAN), has drawn much attention inthese days. How to enhance the synthesis image equality while keeping thestochasticity of the GAN is still a challenge. We propose a novel denoisingframework to handle this problem. The overlapped objects generation is anotherchallenging task when synthesizing images from a semantic layout to a realisticRGB photo. To overcome this deficiency, we include a one-hot semantic label mapto force the generator paying more attention on the overlapped objectsgeneration. Furthermore, we improve the loss function of the discriminator byconsidering perturb loss and cascade layer loss to guide the generationprocess. We applied our methods on the Cityscapes, Facades and NYU datasets anddemonstrate the image generation ability of our model.
展开▼
