A method for anonymizing a face based on a generative adversarial network and a feedback loop to control the intensity of anonymization.

Abstract

A method for anonymizing a face, referred to as the original face, said method consisting, from image data representative of the original face, of generating (100) a synthetic representation of the original face, referred to as the "embedding" vector, of detecting (200) facial landmarks of the face, of generating (300) segmentation masks for each determined region of the face, of extracting (400) demographic characteristics of the face, of extracting (500) descriptive attributes of the face, of extracting (600) the direction of gaze and the orientation of the face, of combining (700) all the information to generate a realistic and unidentifiable anonymized face; said method further consisting of adding noise (110) to the "embedding" vector and comparing (120) a similarity score measured at each iteration to a predetermined parameterized target similarity score corresponding to the desired intensity of anonymization. (Figure 2)

Images