Evaluation of random field models in multi-modal unsupervised tampering localization
Paweł Korus, Jiwu Huang · 2016
While it is commonly known that successful forensic detectors should combine clues from various forensic features, unsupervised multi-modal tampering localization is still an open problem. State-of-the-art fusion methods perform simple pixel-wise combination of the input tampering maps. In this study, we show that pixel-wise combination is sub-optimal and successful fusion needs to model dependencies between neighboring pixels and exploit the content of the tampered image. We evaluate two methods based on conditional random fields and demonstrate that they can exploit image content and precisely delineate the shape of the forgery. In contrast to existing methods based on explicit image segmentation, such an approach does not suffer from subtle object removal forgeries where meaningful segments do not exist. We also demonstrate that existing performance measures are insufficient to accurately assess tampering localization performance. Further work in this direction is needed.