A fusion method for blurring faces on platforms using belief functions
Résumé
In this paper, we propose a global face detection system based on information fusion. Such a system is necessary to make non-identifiable the faces that are visible on platform videos, as required by French legislation related to public video usage and storage. Our system relies on efficient state-of-the-art face detectors, which find the face positions on images. Specifically, it performs first a pixel-wise combination of their outputs in order to take advantage of the potential complementarities of these detectors, which use different image features and different classification procedures. Then, for each pixel of the image to treat and using the result of the merging, a decision is made whether it should be blurred or not. The combination step is grounded on a now well-established framework for reasoning under uncertainty called the Dempster-Shafer theory of belief functions. In this step, detector outputs are converted into a common representation known as belief function using a calibration procedure, and then are merged using so-called Dempster's rule of combination. Our approach is tested on a classical face detection dataset from the literature, showing good performances.
Domaines
Informatique [cs]
Origine : Fichiers produits par l'(les) auteur(s)