原文 标 题 : Bringing Old Photos Back to Life

Original author: Ziyu Wan

Dried fish and duck duck

Writing in the front

In GitHub Hot Trends Vol.046, HG introduced a Microsoft open source AI tool called Bringing- old-photos-back-to-Life, which can bring broken and Old Photos to Life and earned nearly 3k stars this week. Enjoy It ~ is a post on Hacker News by Ziyu Wan, a member of the project team

An overview of the

We are going to use deep learning to repair old photos that have deteriorated. Different from the repair task of traditional supervised learning, aging of real photos is more complex, and the domain gap between synthetic images and real old photos makes network generalization impossible. Therefore, we propose a novel triple domain translation network using real photos and a large number of synthetic image pairs. Specifically, we trained two variational self-encoders (VAE) to convert old and clean photos into two potential Spaces respectively. Transitions between two potential Spaces are learned by synthesizing paired data. Because domain gap is closed in a compact potential space, this translation network can be well generalized to real photos.

In addition, to solve the multiple degradation problems mixed in the old photos, we designed a global branch with partial nonlocal blocks for structural defects such as scratches and dust spots, and a local branch for non-structural defects such as noise and blur. The two branches fuse in potential space, improving the ability to recover old photos from multiple degradation problems. This method is superior to existing methods in visual quality of photo restoration.

The architecture overview

  1. First, we train two variational autoencoders VAE: VAE1 for real photo R ∈ R and synthetic image X ∈ X, and narrow their domain gap by training an adversarial discriminator jointly. Clean image Y ∈ Y is trained by VAE2. VAE is used to transfer images to compact potential Spaces.
  2. We learned about mappings that restore a corrupted image to a clean photo with partial non-local blocks in a compact potential space.

Contrast figure

More details on fixes

Close up

Team member Ziyu Wan’s introduction of Bringing old-photos-back-to-life has been translated. If you are interested, you can go and have a look. Project address:

Github.com/microsoft/B…

Finally, welcome you to join HelloGitHub “translation dance” series, let your talent dance! Share excellent articles with more people. Requirements:

  • Usually browse GitHub, open source, programming, programmers and other English information and articles
  • I want to share my excellent English articles with more people
  • Translation is accurate but not straight or machine translation
  • Ensure that at least one high quality article is translated or corrected per month
  • Understand Markdown and layout rules
  • Contact me (note: translation)