Manipulation de visage réel par un modèle génératif profond.

Abstract

In the last few years, the development and improvement of artificial intelligence techniques have drawn a lot of attention to face manipulation. Due to its importance and usefulness in several fields such as : cinema, video games..., researchers have developed new techniques to manipulate faces in several ways (style transfer, facial expression change, facial expression transfer...). The main objective of our work is to design an intelligent system that is capable of manipulating a face and more precisely we are interested in the transfer of facial expressions from one face to another. The proposed system starts by detecting the facial expression of an input face, which will then be transferred to a synthetically generated face. This facial expression detection is a crucial step in our system, since its result will be the input to the face generation system. We explored two types of generators for this generation, a pre-trained stylegan2 and a conditional stylegan2-ADA with training performed by us on a random selection of a portion of the FER-2013 dataset. The facial expression detection system was trained on the FER2013 dataset and obtained a score of 65%, while the StyleGAN2-ADA generator was trained on a modified FER2013 dataset. Although the result is encouraging, a longer training period will significantly improve the quality of the generated faces.