An Evidential Pixel-Based Face Blurring Approach

Blurring faces on images may be required for anonymity reasons. This may be achieved using face detectors that return boxes potentially containing faces. The most direct way to exploit these detectors is to combine them in order to obtain a more efficient face detection system, producing more accurate boxes. However, contrary to detection, blurring is actually a decision problem situated rather at the pixel level than the box level. Accordingly, we propose in this paper a face blurring system based on face detectors, which operates at the pixel-level. First, for each pixel, detector outputs are converted into a common representation known as belief function using a calibration procedure. Then, calibrated outputs are combined using Dempster's rule. This pixel-based approach does not have some shortcomings of a state-of-the-art box-based approach, and shows better performances on a classical face dataset.