Protecting Your Faces: MeshFaces Generation and Removal via High-Order Relation-Preserving CycleGAN

Abstract

Protecting person's face photos from being misused has been an important issue as the rapid development of ubiquitous face sensors. MeshFaces provide a simple and inexpensive way to protect facial photos and have been widely used in China. This paper treats MeshFace generation and removal as a dual learning problem and proposes a high-order relation-preserving CycleGAN framework to solve this problem. First, dual transformations between the distributions of MeshFaces and clean faces in pixel space are learned under the CycleGAN framework, which can efficiently utilize unpaired data. Then, a novel High-order Relation-preserving (HR) loss is imposed on CycleGAN to recover the finer texture details and generate much sharper images. Different from the L1 and L2 losses that result in image smoothness and blurry, the HR loss can better capture the appearance variation of MeshFaces and hence facilitates removal. Moreover, Identity Preserving loss is proposed to preserve both global and local identity information. Experimental results on three databases demonstrate that our approach is highly effective for MeshFace generation and removal.