Stichting AMBI
Research University of Utrecht

Research University Of Utrecht

Tackling the abuse of images

A man spreads nude pictures of his ex on the internet. A video of happy slapping continues to
appear for months on various websites and loverboys use compromising photos (and the possibility to spread them on the internet) of their victims as blackmail.

"Would it not be a good idea if you could search and find these kind of images of victims on the internet and then remove them? With this question in mind Peter Nuyts chairman of the AMBI Foundation contacted the knowledge centre of Bètawetenschappen at the University Utrecht.
"Is this already possible? And if not, what problems must be solved to achieve this?”

The Citizens Committee against injustice and the Office for Education & Training: ‘Open Ogen’ were also interested in these questions. And this project.

The original problem was too big to answer at once, but ict student Yannick van Welzen has
made a start with researching some of the questions. The Game and Media Technology master's student has researched how well current facial recognition software works and which factors affect the performance of this kind of software.

Van Welzen has researched the performance of three free available programs. This software
works well with photos in a controlled environment, but score worse in uncontrolled environments. To achieve a good facial recognition, it is important that the pictures have a good exposure. And the persons face looks towards the camera. People who wear glasses or a hat, are more difficult to recognized. Unfortunately these optimal conditions are not guaranteed with
photos of victims that circulate the Internet.

Van Welzen has also built an application for this research. This application uses a training
set of 50 pictures of random people. From this training set the program creates an "average face". The software then calculates how much a reference picture differs from this "average face". Faces from a database, which have a similar difference as the reference image, are much like the face of the victim and are therefore possible the same person.

Van Welzen concludes that the performance of the facial recognition software go up if a photo of the person that has to be detected is added to the training set. Additionally, the performance go up of the program has multiple reference pictures of that person.

This is a first step towards developing the software the AMBI foundation needs. In the meantime it will be necessary to do more research. Including improved software performance in uncontrolled environments and the use of that software on the Internet.

Article copied from the newsletter of the knowledge centre of Bètawetenschappen, Utrecht
University (March 24, 2011)