Supervisors: Florian Kleber, Christian Stippel
Status: open

Problem Statement

Historical images, such as digitized photographs, manuscripts, maps, or scientific records, often suffer from degradation due to aging, storage conditions, or limitations in early imaging technologies. Common degradations include noise, fading, blur, physical damage, and low resolution. These artifacts hinder the readability, analysis, and digital preservation of historically valuable content. With the increasing availability of digitized historical archives, there is a growing need for automated restoration methods that preserve visual fidelity while enhancing interpretability.

Deep learning offers state-of-the-art solutions for restoring degraded images by learning transformations from corrupted to clean representations. However, restoration of historical images presents unique challenges, including limited ground truth data, domain variability, and the importance of maintaining authenticity. This work explores how deep learning methods can be adapted and applied to the restoration of such historical visual data.

Goal

The main goal of this course is to implement and compare deep learning-based models for the restoration of historical images.  Various architectures should be applied —including convolutional networks, transformer-based models, and autoencoders—to tasks such as denoising, deblurring, super-resolution, and inpainting. The restored outputs will be evaluated with standard metrics and domain-specific criteria. Also a web application that allows users to upload and restore their own images using the trained models should be developed.

Workflow

• Literature review on deep learning methods for image restoration, with emphasis on historical data
• Selection and implementation of multiple models (e.g., DnCNN, UNet, SwinIR, Restormer)
• Preparation of synthetic and real-world datasets, including historical scans
• Training, testing, and evaluation of models
• Development of a web-based interface for interactive restoration (e.g., using Gradio or Streamlit)
• Documentation and presentation of results

Requirements

• Python
• Knowledge of deep learning (PyTorch)
• Background in computer vision and image processing
• Basic experience with web development frameworks