Research Seminar Series Abstracts

- Basile Van Hoorick

Context has a large impact on our interpretation of a picture, and cropping photos is a very simple but potentially drastic way of changing their meaning. In this work, we construct a neural network that can trace image patches back to their original position relative to the centre of the lens. Experiments suggest that this localization mechanism uses lens imperfections and colour information, notably chromatic aberration and vignetting, as its biggest clues. Since the method is designed to work ’in the wild’, both training and inference do not rely on camera model information or other metadata. We then leverage this novel network to detect crops based on deviations of densely sampled patches from their expected positions. Hence, we successfully demonstrate that our system advances the state of the art in image forensics for soft tampering, breaking the common assumption that image regions are translationally invariant, and enabling opportunities to reveal faulty photojournalism among other applications.