Supervisor: Simon Brenner
Before continuing to read we strongly recommend to let this play in the background: https://www.youtube.com/watch?v=-bTpp8PQSog
Uncountable ancient manuscripts lie in archives, libraries, monastaries and haunted tombs around the world. Unfortunately, the centuries took their toll: due to unsuitable storage conditions, fire- or water damage, clumsy librarians or the recycling of parchment (“palimpsests”) many writings are now lost forever.. or are they??
Multispectral Imaging (MSI) has proven a powerful method to recover such lost texts. For the last 10 years, the Computer Vision Lab has participated in a series of interdisciplinary research projects on the analysis of ancient manuscripts, where we have been imaging manuscripts around Europe and on Mount Sinai and worked on image processing and computer vision methods for text enhancement.
Now that you are totally excited about the secrets of ancient manuscripts, here are some possible topics for practica, bachelor- and diploma thesis:
1. Ruling Line Extraction from 3D Scans
In early manuscripts, the ruling lines guiding the scribe were not made with ink/pigments, but merely engraved into the parchment with a hard tool (“hard point ruling”). Those ruling lines are not visible in conventional photographs (Fig. 2) or multispectral images, but in high-resolution 3D reconstructions (Fig. 3). The shown example was acquired using Photometric Stereo . The goal of this project is to develop robust algorithms to extract ruling schemes from 3D reconstructions of manuscript pages.
2. Experiments with Generative Adversarial Networks (GANs)
Image-to-image translation is a prominent application example of GANs. This even works when you don’t have paired training data: https://arxiv.org/abs/1703.10593. This approach can be used to transform a photo to a Van Gogh painting or what have you..
Now the question is: can you use a GAN for manuscript enhancement? Possible scenarios:
-) Multispectral layers –> readable document (the standard problem)
-) False-color image –> natural colored image
-) Intact manuscript –> degraded manuscript (generate “synthetic training” data for other enhancement approaches)
3. A study on the contribution of individual wavelebands to text reconstruction performance
The question is: can Multispectral Imaging be simplified by omitting certain wavebands that do not significantly enhance the restoration result?
This would enable a faster acquisition process (which is important as manuscripts typically have to be imaged on-site), more efficient processing and cheaper hardware.
Based on data acquired in past research projects, a systematic study must be carried out, testing the contribution of individual wavelengths to the final enhancement results. The study should be evaluated on different types of manuscripts and different enhancement methods.
A common way to post-process multispectral image is to perform a statistical decomposition of some sort (e.g. PCA). Then you can select some components that visualize certain properties of the data and combine them to expressive pseudocolor images (like the right side of Fig. 1 for example). Currently, this is mostly done by hand. This topic is about establishing guidelines or automated procedures for the creation of visually pleasing pseudocolor images.
5. A viewer/editor for multispectral images and material analysis data
While we at the CVL were concerned with the enhancement of degraded text based on MSI, our colleagues at the Academy of Fine Arts Vienna performed chemical meterial analysis of the same manuscripts via X-Ray Fluorescence (XRF), Fourier Transform Infra Red Spectroscopy (FTIR) or Raman Spectroscopy. Now the plan is to build a common database that is connecting the two worlds. This would require a nice front-end enabling the viewing and editing of MSI and specriscopic measurements in a common coordinate frame.
– navigating through spectral layers
– intuitive composition of pseudocolor images
– inserting/editing of material-analysis sample points
– adding/displaying spectroscopy data
The tool can be implemented as a web application (e.g. with React JS) and/or as a plugin for the awesome nomacs image viewer (C++).
If you are interested in one of the topics, write to firstname.lastname@example.org