Single-image splicing localization through autoencoder-based anomaly detection

We propose a new method for single-image splicing localization. Lacking any side information, such as training data or prior knowledge on the source camera and the image history, we cast the problem as an anomaly detection task. Expressive local features, extracted from the noise residual of the image, feed an autoencoder which generates an implicit model of the data. By iterating discriminative feature labeling and autoencoding, the implicit model fits eventually the pristine data, while the spliced region is recognized as anomalous. Experiments on a suitable test set of spliced images show that the proposed method outperforms the previous state-of-the-art. In addition, it exhibits a good robustness against typical social net post-processing, showing promises for real-world applications.