ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2019, Vol. 56 ›› Issue (11): 2485-2493.doi: 10.7544/issn1000-1239.2019.20180656

• 图形图像 • 上一篇    下一篇

一种基于局部属性生成对抗网络的人脸修复算法

蒋斌,刘虹雨,杨超,涂文轩,赵子龙   

  1. (湖南大学信息科学与工程学院 长沙 410082) (jiangbin@hnu.edu.cn)
  • 出版日期: 2019-11-12
  • 基金资助: 
    国家自然科学基金项目(61702176);湖南省自然科学基金项目(2017JJ3038)

A Face Inpainting Algorithm with Local Attribute Generative Adversarial Networks

Jiang Bin, Liu Hongyu, Yang Chao, Tu Wenxuan, Zhao Zilong   

  1. (College of Computer Science and Electronic Engineering, Hunan University, Changsha 410082)
  • Online: 2019-11-12

摘要: 最近对神经网络模型的研究在图像修复任务中显示出巨大的潜力,其核心任务是理解图像语义信息并重建缺失的图像内容.这些研究可以生成语义和内容上合理的结构和纹理,但通常会导致与孔洞周围区域不一致的扭曲结构或模糊纹理,特别是人脸图像修复问题.人脸图像修复工作经常需要为包含大量外观元素以及局部属性的缺失区域(例如眼睛或嘴巴)生成语义上的新内容,这些缺失区域往往具有独特的属性和语义信息从而导致生成内容不合理.为了解决以上问题,提出了一个有效的深度神经网络模型,模型的生成器结合全连接卷积和U-net网络的优越特性,同时提出局部属性辨别器使修复内容具有创新性的同时也能够使整体与局部保持语义一致性.模型不仅提升了对于人脸图像整体语义信息的感知能力,同时也基于局部属性能够有效地修复人脸关键部位,通过在CelebA数据集上的实验证明了该模型能够有效地修复人脸缺失部分并且能够生成新颖的修复内容.

关键词: 神经网络, 人脸修复, 局部属性辨别器, 全链接卷积, U-net

Abstract: Recent researches in neural network models have shown great potential in image inpainting task, which focuses on understanding image semantic information and reconstructs missing image content. These researches can generate visually reasonable image structures and textures, however, they usually produce distorted structures or blurry textures that are inconsistent with the surrounding areas, especially for the face inpainting task. The face inpainting task is often necessary to gene the advantages of fully connected convolution and U-net network, and the model proposes locally attributes discriminator to make the inpainted contents more innovative and is able to keep the global and local semantic consistency. The model not only improves the perception of the overall semantic information of the face image, but also restores the key parts of the face based on the local attributes. Experiments on the CelebA dataset have shown that our model can effectively deal with face image repair problems and generate novel results.

Key words: neural network, face inpainting, local attributes discriminator, fully connected convolution, U-net

中图分类号: