Arthur Leclaire (CMLA, ENS Cachan)

Texture synthesis with FRAME Models using Neural Networks Filters

vendredi 5 mai 2017, 11h00 - 12h00

Salle de réunion, espace Turing

In this talk, I will present the paper « Learning FRAME Models Using CNN Filters » by Y. Lu, S.C. Zhu and Y.N. Wu (2015).
In 1998, Zhu and Mumford have proposed the FRAME model (Filters, Random Fields and Maximum Entropy) for texture synthesis. This model takes the form of an exponential distribution which depends on the reponses of the image to a set of filters (linear or non-linear). The parameters of this model can be adapted to an exemplar texture image with a (time-consuming) maximum-likelihood procedure relying on Monte-Carlo simulation. However, one key advantage of this model is to provide a guarantee to respect the average values of the responses to the selected filters.
After a brief introduction to exponential models, I will explain how Lu et al. [1] have adapted the FRAME model to responses of an image to a neural network. I will also discuss the connection and differences with the texture synthesis method designed by Gatys et al. [2].