A multi-scale noise-resistant feature adaptation approach for image tampering localization over Facebook

摘要

This work introduces an approach to localize the tampered region among the images from social media platforms. We propose a joint model to integrate the predictions from a set of features, each of which represents the inherent relation among the pixels within a certain distance to detect the forgery. Within a fixed distance, the feature is adapted from a few basic statistics through a stacked Autoencoder to a proper version in a noise-resistant manner, so that it will be more robust to detect the tampering when the forgery gone through some common social media platform operations. The classifier is trained using a standalone dataset from a benchmarking but is pre-processed properly to simulate its possible imperfections when spreading over the Internet. The approach was tested on images from Facebook, with results showing an encouraging improvement from the prior arts.