Project ‘Glashelder’: how AI assists in metadating 62.000 photographic plates

In ‘Glashelder’ (Crystal clear) people and machines collaborate to enrich nearly 62.000 digitised glass plates with quality metadata. Glass plates are among the oldest photographic materials used mainly in the 19th and 20th centuries. They were used to reproduce photos or to project images, for example in an educational context. This early photographic medium is historically relevant, but also fragile and hard to access.

Research project ‘Glashelder’ examines how existing Visual Language Models and other AI-techniques can assist humans in enriching these visual collections with metadata. The glass plates provide a testing ground for analysing how processes of intelligent automation, learning and valorisation can be deployed in heritage organisations. An ethical and legal framework provides guidance to tackle the myriads of challenges that come with the use of AI.

The challenge: creating a rich, substantive description

How can the expansive collection of digitised glass plates receive substantive metadata, and how can it be disclosed to the public? The GIVE-project digitised nearly 62.000 glass plates housed in the Boekentoren. From topographical images and educational content from the Arts Faculty to an ethnographical cluster: the collection contains a broad and diverse range of images.

Even though the glass plates received a limited technical description, those metadata alone are not detailed enough to be useful for the public or for research purposes. Even more, it is currently difficult for the collection managers to get an overview of the nature and diversity of the images. Manually enriching this collection with metadata is practically impossible. The sheer scale and required expertise results in considerable challenges.

The ambition to reach a complete appraisal and quality metadata for the entire collection, however, comes ever closer thanks to artificial intelligence. The size and typological diversity of the glass plates make them the ideal case study for testing methodologies and techniques involving AI in a heritage setting.

A hybrid approach: a collaboration between humans and machines.

Project ‘Glashelder’ uses various AI-techniques to enrich the sizable glass plate collection with metadata. We approach the project in three steps: intelligent automation, learning and valorisation. This is a hybrid method: both humans and machines collaborate and alternate with each other. In doing so, we can process human feedback to train the AI-models and to improve their results. We consistently choose to use open-source models. Not only can we efficiently refine and adapt these models to the needs of our use cases, but we also guarantee the accessibility, transparency, and security of the generated data.

At the end of project ‘Glashelder’, we hope to have:

• Enriched the beautiful and diverse collection of glass plates with detailed metadata, making them as accessible as possible for anyone.
• Developed expertise in the use of artificial intelligence in heritage projects, and to integrate this knowledge in the daily operations of the Boekentoren.
• Analysed the various AI-techniques that are reusable for the broader heritage sector for metadating images.
• Encouraged a critical reflection on the collaboration between humans and machines in the heritage field.