DAE-Net: Dual Attention Mechanism and Edge Supervision Network for Image Manipulation Detection and Localization

Image manipulation detection and localization (IMDL) aim to identify tampered regions from a suspicious image. Although existing works have achieved impressive results, they fail to capture sufficient tampering artifacts resulting in inaccurate localization of tampered regions. In addition, the problem of poor robustness to postprocessing methods also occurs frequently. To address these problems, this article presents a dual attention mechanism and edge supervision network (DAE-Net) to locate forged regions. Specifically, DAE-Net first extracts multiscale semantic features from RGB images and captures global and local inconsistent noise features from noise maps. Subsequently, to mine the complementary relationship among different domains and obtain more discriminative feature representations, a dual attention mechanism (DAM) module is designed to integrate semantic features and noise features. Then, we exploit a mask generation module (MGM) consisting of fusion upsampling module (FUM) to generate tampered regions in a progressive decoding manner. FUM fully fuses multiscale semantic features to enhance tampering traces. Finally, an edge supervision module (ESM) is introduced to capture subtle edge information to optimize the boundary of tampered regions. The experimental results on five public datasets demonstrate that DAE-Net outperforms the state-of-the-art (SoTA) methods in terms of higher localization accuracy and stronger robustness under widely used evaluation metrics.
$F_{1}$
/AUC of the proposed DAE-Net reach 0.494/0.805, 0.871/0.984, 0.972/0.973, 0.330/0.741, and 0.338/0.826 on CASIA, NIST16, Columbia, Coverage, and IMD2020, respectively.