基于级联生成对抗网络的人脸图像修复

Face Image Inpainting Using Cascaded Generative Adversarial Networks

人脸图像修复技术为近年来图像处理领域的研究热点.该文提出一种基于级联生成对抗网络的人脸图像修复方法,从生成器、判别器、损失函数三个方面进行改良.生成器采用由粗到精的级联式模型,并结合密集连接模块使所修复区域更加精细;判别器采用局部与全局特征相融合的双重判别式模型以提升判别准确性;损失函数采用最小化重构损失和对抗网络损失相结合以获得更好训练效果.基于CelebA数据集的实验显示,该方法可实现面部区域丢失50%以上的人脸图像修复,在客观评价指标PSNR和SSIM上,较现有方法分别提高了1.1~7.5 dB和0.02~0.15.从主观效果来看,该方法修复的人脸图像拥有更丰富的细节、更显自然.

Face image inpainting is a hot topic of image processing research in recent years. This paper proposes a face image restoration method based on cascade generative adversarial network. In this method, the generator employs a cascading structure consisting of a coarse network and a refinement network and adopts dense connections to recover more details of the missing face area;the discriminator uses a dual discriminant model combining local and global features to improve the discriminant accuracy;the loss function consists of reconstruction loss and generative adversarial loss for better training performance. Experiments on CelebA dataset show that the proposed method can restore facial image with more than 50% missing area. The objective evaluation index PSNR and SSIM are 1.1 dB to 7.5 dB and 0.02 to 0.15 higher respectively compared with state of the arts. For subjective evaluation, the restored face images look more detailed and natural.