Skip to content
Book Open access

BiPro-Net: Boundary-Aware Dual-Prototype Network for Robust Image Inpainting Detection

Nov 2026 · Proceedings of the 34th ACM International Conference on Multimedia · 0 citations · 46 references

Abstract

Image inpainting detection remains a major challenge in visual forensics, especially as modern tampering techniques produce increasingly realistic results, making the already subtle traces even harder to capture. Existing methods exploit diverse forensic cues but often rely on local feature aggregation or implicit binary feature separation, without explicitly constructing image-conditioned references for both pristine and tampered regions. Consequently, they may struggle to capture manipulation-specific patterns and suffer from severe feature confusion along ambiguous tampering boundaries. To address these limitations, we propose BiPro-Net, a Boundary-Aware Dual-Prototype Network for robust image inpainting detection. Its core component, a Dual-Prototype Memory (DPM) module, leverages cross-attention to dynamically construct image-conditioned memory tokens for both authentic background and tampered foreground, enabling explicit dual-reference discrimination. To further resolve feature confusion in transitional regions, we introduce a Prototype-Anchored Contrastive Loss, which constructs foreground and background class anchors during training to selectively optimize hard pixels along tampering boundaries and sharpen the decision boundary. Additionally, we design an Uncertainty-Aware Dual-Role mechanism guided by predicted uncertainty maps. It reduces the contribution of unreliable evidence when refining class-conditioned contexts while applying a learned spatial gate to emphasize ambiguous regions. Extensive experiments across multiple benchmarks demonstrate that BiPro-Net achieves leading overall performance in both detection accuracy and localization precision.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.