EdgeAdapter for Image Forgery Localization

The rapid development of digital image editing technology has enhanced the user experience, but it has also exacerbated the risk of the spread of false information. While existing visual foundation models possess strong semantic understanding capabilities, they lack the ability to focus on underlying texture information. Aiming at the above problems, this study proposes Edge Adapter, a parameter-efficient finetuning framework based on edge feature enhancement. This work innovatively constructs a dual-stream feature fusion architecture, and through a lightweight adapter module, it enables cross-scale dynamic interaction between the image edge features and the deep semantic representations of the pre-trained model. This framework adopts the Parameter-Efficient Fine-Tuning (PEFT) strategy. Under the condition of completely freezing the parameters of the backbone network, the learnable edge adapter module guides the network to focus on local abnormal patterns, achieving the adaptive fusion of semantic context and texture features. Cross-dataset experiments show that our method has achieved significant improvements compared with the latest models in benchmarks such as NIST16. The F1 score and IoU score for tampering localization have increased by 3 % and 1 % respectively. By only fine-tuning 2.99 M trainable parameters, it can effectively avoid the problem of knowledge forgetting in fullparameter fine-tuning, providing an innovative solution for the image tampering localization task in terms of both accuracy and computational efficiency.