SEINet: Semantic-Edge Interaction Network for Image Manipulation Localization

Ye Zhu, Na Qi, Yingchun Guo, Bin Li

Abstract: Previous manipulation detection methods usually utilize semantic to detect suspected regions but the edge is abridged, which poses a greater challenge in manipulation location. We propose a novel dual framework named Semantic-Edge Interaction Network (SEINet) for locating the manipulated regions, including splicing, copy-move and removal. The dual streams and Cross Interaction (CI) pattern aim to extract semantic and edge features under the supervision of semantic and edge Ground-Truth, respectively. In addition, we propose a Bidirectional Fusion Module (BFM) to incorporate the dual stream feature maps with the decoder of U-net. Extensive experiments, which are evaluated on Synthetic, CASIA and NIST16 datasets, prove that the proposed SEINet can locate the manipulated regions more accurate than state-of-the-art methods, and is more robustness to noise, blur, and JPEG recompression attacks.