Pixel-wise Conditioning of Generative Adversarial Networks

Cyprien Ruffino (1, 2, 3, 4) , Romain Hérault (2, 1, 3, 4) , Eric Laloy (5) , Gilles Gasso (1, 2, 3, 4)

1 LITIS - Laboratoire d'Informatique, de Traitement de l'Information et des Systèmes
2 DocApp - LITIS - Equipe Apprentissage
3 NU - Normandie Université
4 INSA Rouen Normandie - Institut national des sciences appliquées Rouen Normandie
5 Belgian Center for Nuclear Energy

Generative Adversarial Networks (GANs) have proven successful for unsupervised image generation. Several works extended GANs to image inpainting by conditioning the generation with parts of the image one wants to reconstruct. However, these methods have limitations in settings where only a small subset of the image pixels is known beforehand. In this paper, we study the effectiveness of conditioning GANs by adding an explicit regularization term to enforce pixel-wise conditions when very few pixel values are provided. In addition, we also investigate the influence of this regularization term on the quality of the generated images and the satisfaction of the conditions. Conducted experiments on MNIST and FashionMNIST show evidence that this regularization term allows for controlling the trade-off between quality of the generated images and constraint satisfaction.

Domaines: Intelligence artificielle [cs.AI]; Vision par ordinateur et reconnaissance de formes [cs.CV]; Machine Learning [stat.ML]

Cyprien Ruffino, Romain Hérault, Eric Laloy, Gilles Gasso. Pixel-wise Conditioning of Generative Adversarial Networks. European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning (ESANN), Apr 2019, Brugge, Belgium. ⟨hal-02347732⟩