A robust image tampering positioning method based on consistent guided learning
By employing a dual-stream structure and a consistency-guided learning framework, the problem of insufficient robustness of image tampering localization models under post-processing operations is solved, enabling effective tampering localization under different conditions and improving the model's robustness and generalization ability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV OF POSTS & TELECOMM
- Filing Date
- 2026-02-11
- Publication Date
- 2026-06-02
AI Technical Summary
Existing image tampering localization models are not robust enough to post-processing operations. Directly introducing enhancement strategies will weaken the model's ability to learn tampering traces, leading to performance degradation.
A localization network with a master-slave dual-stream structure is adopted, combined with a consistency-guided learning framework. Through alternating training of the target network and the source network, the target network captures fine features without post-processing, and the source network fits the response of the target network. Multi-scale feature fusion is performed using an improved cross-attention and feature pyramid structure, and the network is optimized through localization and contrastive loss.
It effectively improves the model's robustness to post-processing operations, maintains the model's good generalization performance under different datasets and conditions, and achieves effective tamper localization.
Smart Images
Figure CN122135186A_ABST