An example of image recognition: detecting comics in the archive
History and artificial intelligence seem like polar opposites: old versus new, dusty versus trendy, past versus future. But that is merely an illusion, as PhD student Bas Vercruysse of the Ghent Centre for Digital Humanities knows. His research focuses on using AI methods in various disciplines, including historical research. For his colleagues specialising in the evolution of comics, he uses computer vision to automatically detect comic strips in old journals. And that saves a lot of time.
Comics in old magazines
My grandfather was an avid comic book collector, and I inherited his wonderful collection as a child. I spent hours in Zonnedorp, to then have adventures with Suske and Wiske in Antwerp, eventually putting myself in Marcel Kiekeboe’s street, somewhere between Fernand Goegebuer and Leon Van der Neffe.
Belgian comic book culture is known worldwide and has a long tradition, which has gone through a lot of evolutions. My colleagues are doing historical research on this, and I am using AI to help them.
First, we had to build a large collection of comics. We started from the early 20th century: back then, short comics often appeared in magazines (think of it as the Dag Allemaal of yesteryear), just as they are still published in newspapers today.
All those old magazines are kept in the warehouse of the Royal Library of Belgium (KBR). But how do you start looking for pages of comics in such a massive amount of material? Fortunately, a whole bunch of magazines were recently digitised, allowing us to browse them digitally in advance. That helped us get started, but the entire collection still contained more than 500 000 pages! 🤯 There had to be a more efficient way to search them. Could AI provide a solution?
Automatic comic stirp recognition
As a researcher today, you no longer have to spend time on boring repetitive tasks, such as flipping through magazines looking for the right pages. Computers and AI can do that kind of work for you.
We used the AI tool YOLO, ‘You Only Look Once’, to automatically detect comics using computer vision. YOLO is an object recognition system: via a specific type of neural networks – so-called convolutional neural networks – it learns to recognise objects on images.
The system looks for the main characteristics of an object. For example: what makes a cat a cat? To recognise a cat, we often look at its ears, whiskers, and fur. This AI system does something similar, but on a much more abstract level. As a result, it often performs even better than humans!
YOLO can be used for a variety of tasks. My colleagues and I taught it specifically to recognise comics, pictures, and cartoons – all the visual materials in a magazine.
Recognising an image 70 times faster thanks to AI
AI
As a pilot project, we unleashed our model on 80 000 magazine pages. The model detected around 4 000 comics among these pages. The model not only succeeded in accurately recognising comics, but also did so improbably fast:
- In five and a half hours, all 80 000 pages were processed.
- That works out at 0.25 seconds per page.
Human
For comparison, I asked ChatGPT how much time it would take a human to turn a page, read it and classify the images:
- 17 to 35 seconds per page, depending on the interface and the person’s speed and experience
- For 80 000 pages, this is between 378 hours (fastest scenario) and 778 hours (slowest scenario)
- AI completes this job at least 70 times faster than humans!
“ Our AI model managed to accurately recognise comics in magazines and did so improbably fast. ”
AI frees up time for researchers
At this stage of the project, the choice to implement AI assistance was quickly made; it saves a lot of time that researchers can better invest in interpreting the data.
In the future, we may expand that support. While we now mainly use artificial intelligence to automate tasks, new applications such as ChatGPT Vision offer additional possibilities, for example to interpret images as well.
AI thus offers many new possibilities for historical research. History and AI turn out to be quite compatible after all!
Find out more
The basics of AI
AI systems learn from experience, meaning they get better at their tasks over time. How does AI exactly work? We’ll explain some of the basics.
Bas Vercruysse
Bas Vercruysse is a historian and took a course in applied AI after his studies. At the moment, he is working as a PhD student at the Ghent Centre for Digital Humanities where he is trying to insert computer vision and large language models into the fascinating but sometimes somewhat archaic world of historians. He focuses on making visual and textual historical sources accessible, with applications ranging from detecting humour comics in old journals to supporting research on climate data.
Related courses
SAIAR Summer Studio: From latent to physical space
Zomerschool - Kortrijk - Howest Hogeschool, SAIARlab
HPC Café
Seminarie - Online, Karlsruhe - KIT, HammerHAI, Training geselecteerd in de Belgische AI Factory Antenna (AIFA)
