A parallel attention mechanism for image manipulation detection and localization

Existing image manipulation detection and localization methods tend to detect the trail of manipulation and achieve decent performance. So far, however, there has been little discussion about the category imbalance. In this paper we propose a parallel attention mechanism based network to localize tampered regions, which tends to have better generalization, while it possesses higher model capacity. In addition, we designed a trainable parameter constrained shifted-window dual attention module to strengthen the manipulation features. Finally, to settle the difficulty of category imbalance, we conduct a category-based normalization loss function, which allows the model to pay more attention to the manipulated regions and further improves the generalization capability. Extensive experimental results demonstrate that the proposed approach can effectively reconcile the weights of different categories during training and produce state-of-the-art performance in various benchmark datasets.