Coverserver for UAntwerpen

October 2023 - May 2024

For Anet, the IT service behind the University of Antwerp, I trained an AI model that can distinguish between good and bad cover photos. As a result, hundreds of cover photos are checked daily in about ten libraries.

This model was trained on a dataset of 20,000 photos, of which 12,000 were good and 8,000 were bad. I used a pretrained model 'Resnet50' as a base and further trained it with the dataset.

UA.title 1

The catalog of the Anet network contains approximately 2,500,000 records, including those from the city of Antwerp library and the KMSKA. After training the model, I integrated it, which was a challenge due to the complexity of the codebase, especially since I had no experience with the framework used, Django.

Additionally, I also developed a REST API that enables communication between the front end and the AI model.

UA.title 2

All predictions generated by the model are stored in the database. Cover photos can be annotated with remarks such as 'too much white space' or 'blurry', with the aim of training a better model in the future.

For the developers, I also created two screens where the predictions are visually displayed.

UA.title 3

Want to collaborate?

Do not hesitate to get in touch, for a collaboration or an introduction.