PhD defense Gwilherm Lesné: Image Attribute Editing and Film Grain Rendering using Deep Neural Networks

Original title: Édition d’Attributs d’Images et Rendu de Grain Argentique à l’aide de Réseaux de Neurones Profonds

Jury

• Julie Delon, Professeure, Université Paris Cité (MAP5) (Rapportrice)
• Julien Rabin, Maı̂tre de conférences, ENSICAEN (GREYC) (Rapporteur)
• Pierre Hellier, Chaire de professeur junior, Université de Rennes (IRISA) (Examinateur)
• Céline Hudelot, Professeure, Centrale-Supélec (MICS) (Examinatrice)
• Yann Gousseau, Professeur, Télécom Paris (LTCI)  (Directeur de thèse)
• Saïd Ladjal, Professeur, Télécom Paris (LTCI) (Co-Directeur de thèse)
• Alasdair Newson, Maı̂tre de conférences, Sorbonne Université (ISIR) (Co-Directeur de thèse)
• Henri Maître, Professeur émérite, Télécom Paris (LTCI) (Membre invité)

Abstract

This thesis introduces works related to image editing while providing background on existing methods to accomplish such tasks. More specifically, two methods are presented, one on attribute editing and the other on film grain rendering.

The first method presented consists in a neural network where attributes of face images can be edited by modifying a single value, in the manner of a slider.

The particularity of our approach lies in the fact that each attribute can be edited independently from the others.

The second method described in this thesis consists in another neural network. It adds film grain to a digital image. To do this, the size of the grain is fed to our method so that it can render the different aspects of film grain. It is also possible to obtain several outputs for the same grain size.