东北大学学报(自然科学版) ›› 2024, Vol. 45 ›› Issue (9): 1227-1234.DOI: 10.12068/j.issn.1005-3026.2024.09.002

• 信息与控制 • 上一篇    

基于去噪扩散概率模型的人脸图像修复模型

刘纪红(), 黄熙雄   

  1. 东北大学 信息科学与工程学院,辽宁 沈阳 110819
  • 收稿日期:2023-05-08 出版日期:2024-09-15 发布日期:2024-12-16
  • 通讯作者: 刘纪红

Face Inpainting Model Based on Denoising Diffusion Probability Models

Ji-hong LIU(), Xi-xiong HUANG   

  1. School of Information Science & Engineering,Northeastern University,Shenyang 110819,China. cn
  • Received:2023-05-08 Online:2024-09-15 Published:2024-12-16
  • Contact: Ji-hong LIU
  • About author:LIU Ji-hong,E-mail:liujihong@ise.neu.edu.

摘要:

针对使用主流人脸图像修复模型在修复图像后,出现图像质量欠佳、修复边缘模糊,且模型复杂、训练困难的问题,提出了一种基于去噪扩散概率模型的人脸图像修复模型.通过使用Guided?diffusion中的U-Net网络结构,并在网络中引入快速傅里叶卷积来改进去噪扩散概率模型,最后在CelebA-HQ高清人脸图像数据集上进行模型的训练与结果评估.实验结果表明,改进后的去噪扩散概率模型在修复随机掩码的人脸图像时,修复结果与原图的PSNR(峰值信噪比)可以达到25.01,SSIM(结构相似性)可以达到0.886,优于改进前的去噪扩散概率模型与现有的基于生成对抗网络的人脸图像修复模型.

关键词: 深度学习, 人脸图像修复, 去噪扩散概率模型, 快速傅里叶卷积, U-Net网络

Abstract:

A face inpainting model based on the denoising diffusion probability model is proposed aiming at the problems of poor image quality, blurred repair edges, complex model, and difficult training of the mainstream face inpainting model after image inpainting. By improving the denoising diffusion probability model, the U-Net network structure in Guided?diffusion is adopted. The fast Fourier convolution is introduced into the network, and then the model is trained and tested on the CelebA-HQ high?definition face image dataset. The experimental results show that the improved denoising diffusion probability model can achieve a PSNR of 25.01 and a SSIM of 0.886 compared to the original image, when inpainting face images with random mask, both of which are better than the model before improvement and the existing face image inpainting model based on generative adversarial networks.

Key words: deep learning, face inpainting, denoising diffusion probability models, fast Fourier convolution, U-Net network

中图分类号: