A group of students worked together on the 'Bringing Photos Back To Life' project, which restores old photographs that have gone through severe degradation. Using layers of technology, the project developed a solution that learns the visuals through supervised learning techniques and artificial intelligence.
It measures the domain gap between the old photos and synthetic images through a novel triplet domain translation. Designing a global branch with partially nonlocal block targeting as well as the structured defects including dust, scratches, and measuring noise and blurriness. The method trains two VAE, which are variational auto encoders to two latent spaces using the paired data of old and clean photos to tackle the degradations in the old photos.
Image Credit: Bringing Old Photos Back to Life, Bo Zhang