Résumé: Current research on image tampering localization focuses on finding region features that distinguish manipulated pixels from non-manipulated pixels. As tampering with a specific area of a given image inevitably leaves cues in the boundary between the tampered region and its surroundings, how to utilize sufficient region and boundary features also matters for image manipulation localization. In this paper, we propose a unified network (called RB-Net), which is a two-branch network (i.e., region module and boundary module) to learn region and boundary features separately. Then the fusion module is implemented to integrate the region features from the region module and the edge features from the boundary module, respectively. Particularly, to identify unnatural boundary traces, we propose edge gate components deployed on different layers of the region module to activate manipulated boundary information from the rich region features. Quantitative and qualitative experiments on four benchmark datasets demonstrate that RB-Net can accurately locate the tampered regions and achieve the best results relative to other state-of-the-art methods.

Langue: Anglais