Abstract

The development of computer graphics technologies has been bringing realism to computer generated multimedia data, e.g., scenes, human characters and other objects, making them achieve a very high quality level. However, these synthetic objects may be used to create situations which may not be present in real world, hence raising the demand of having advance tools for differentiating between real and artificial data. Indeed, since 2005 the research community on multimedia forensics has started to develop methods to identify computer generated multimedia data, focusing mainly on images. However, most of them do not achieved very good performances on the problem of identifying CG characters. The objective of this doctoral study is to develop efficient techniques to distinguish between computer generated and natural human faces. We focused our study on geometric-based forensic techniques, which exploit the structure of the face and its shape, proposing methods both for image and video forensics. Firstly, we proposed a method to differentiate between computer generated and photographic human faces in photos. Based on the estimation of the face asymmetry, a given photo is classified as computer generated or not. Secondly, we introduced a method to distinguish between computer generated and natural faces based on facial expressions analysis. In particular, small variations of the facial shape models corresponding to the same expression are used as evidence of synthetic characters. Finally, by exploiting the differences between face models over time, we can identify synthetic animations since their models are usually recreated or performed in patterns, comparing to the models of natural animations.