Efficient domain adaptation for painting theme recognition

Abstract

In this paper we approach the problem of scene recognition in paintings. We tackle this task with the aid of Convolutional Neural Networks and a large database consisting of around 80,000 paintings. The main purpose is to identify an efficient method to enlarge the database by domain transfer from photographic content to artistic content. Thus, we discuss the practical capabilities of a recent method of domain transfer from photographs to paintings while augmenting the employed database and aid the learning of difficult styles. We propose a set of improvements to increase the feasibility of the domain transfer in the context of large databases.