rObust statistics-basEd lIte Learning – OEIL

From an original image from a digital camera (OIDC), this image can be either modified globally for purposes of improvement of the quality of the image, or modified locally, this mainly for an ill-intentioned objective. The objective of this project is to be able to separate the manipulations of global images of the manipulations of local images to detect manipulations of images ill-intentioned. The main idea of this project leans on the statistical analysis on small-sized sets. Indeed, completely in the opposite of the law of large numbers (classic approach in statistics), we shall develop our approach on analyses of small numbers. We justify this approach of the fact that for the detection of falsified images we have no original images and that we are not sensible to have big picture archives containing all the manipulations of possible images. As a result, for every image, thus it is necessary to us to learn a characteristic statistical contents of the image, its acquisition and the treatments which they underwent from the contents of this one only, then to use this analysis to detect the most finely possible if a zone of the image does not contain the same characteristics. If it is the case, then the image will be supposed to be falsified and we could possibly be capable of locating this zone. This type of approach is confronted with three main difficulties:
i) The learning of global statistical characteristics which is not sensitive to the local modifications (robustness face to face of the contamination of the statistical laws),
ii) the characterization statistics of zones of small size, that is a statistical description based on a small number of information (robustness face to face of the size of the sample),
iii) the comparison of a global statistical criterion learnt with a lot of information possibly contaminated in a local statistical criterion possibly less contaminated but learnt with a small number of information (robustness face to face of the unhomogeneity of the information).

It is the reason why we have decided to name the project "“robust statistics-based lite learning", that is in French "Statistiques robustes pour l’apprentissage léger".

The project consists of four parts:

1) Robust statistical characterization of phenomena in the contamination.
2) Analysis of the meaning of a statistical criterion according to the size of a block of pixels.
3) Make decision in the uncertain from not homogeneous information.
4) Characterize each of the possible and/or known modifications.