基于多角度约束与谱归一化的图像修复算法

IMAGE INPAINTING ALGORITHM BASED ON MULTI-ANGLE CONSTRAINT AND SPECTRAL NORMALIZATION

针对图像修复任务过于困难的问题,采用基于生成对抗网络的双判别器模型,通过增设局部判别器追踪图像局部缺失区域信息,有效提升了修复准确性。但模型在产生合理语义性信息方面并不乐观。为此,提出Multi-Angle GAN模型。在双判别器模型基础上增设分类器和Vgg19特征提取网络,分别向生成网络提供类别、风格和内容损失约束。针对GANs判别器设计存在的训练不稳定问题,向判别器设计中引入谱归一化和Wasserstein距离。在CelebA、Places2数据集上进行大量实验,结果表明,Multi-Angle GAN较之前方法在PSNR和SSIM上分别提升0.6~0.8 dB和0.02~0.05。

Aiming at the problem that image inpainting task is too difficult,the double discriminator model based on generative adversarial network is used to track the image local missing regions information by adding a local discriminator,which improves the inpainting accuracy effectively.However,the model does not perform well at generating semantically plausible information.To this end,we propose a multi-angle GAN model.A classifier and a Vgg19 feature extraction network were added based on double discriminator model,respectively providing category,style and content loss constraints to the generation network.Aiming at the training instability problem of GANs discriminator design,the spectral normalization and the Wasserstein distance were applied in the discriminator design of GANs.A lot of experiments were carried out on the CelebA,Places2 datasets.The results show that the multi-angle GAN is improved by 0.6~0.8 dB and 0.02~0.05 respectively compared with the previous methods on PSNR and SSIM.