A GAN-Based Framework For High-Fidelity Face Swapping

Abstract

In recent years, face manipulation has attracted much attention which has both positive and negative effects for us. Face swapping is a face manipulation technique that modifies the identity information while preserving the attribute information of the face. It is necessary for us to know the advanced methods for high-quality face swapping and generate high-quality face swapping images to train the forgery detection algorithm. Currently, the face swapping algorithm faces many challenges, such as poor quality of the generated images or poor generalization of the algorithm, and sometimes the attribute features of the source image are also transferred to the target identity, resulting in the loss of attribute information of the generated images. In this paper, we propose a GAN-based face swapping framework which can obtain high-fidelity face swapping images. Firstly, this framework uses two separate encoders to decouple the identity and attribute feature of face images. Then a feature fusion module is used to coupled the identity and attribute feature of the face. Specifically, we treat the face swapping problem as a style transfer problem to solve the feature fusion problem, furthermore, an attention mechanism is introduced to guide the feature fusion process. The experimental results show that our method is effective.