Fast and scalable image auto-tagging

Inside Invenio, the web-based integrated system for handling digital libraries developed at CERN, there is a media module, enabling users to upload photos and videos. Especially in CDS, the Invenio instance used at CERN, people use this digital library to upload pictures of official events that took place at CERN. However, so far, there was no way of tagging what’s inside these photos. This project is meant to solve the problem of tagging persons in a photo in an easy and fast way. First, by implementing a complete tagging interface that allows the user to square parts of the photo, resize them, move them and give them a name. Second, by running face detection so that squares already appear on faces and the user just has to fill the title field. Finally, by running a face recognition system that learned from previous tags created by users. In this report, we will show how we implemented the tagging interface, how we improved the existing face detector to make it more efficient, which face detection methods we used and how we combined them to have a fully working framework. Finally, we will show the results we obtained and the study we did on the parameters to choose for the different algorithms.