Image splicing localization method driven by device difference feature guidance

Xiaofeng Wang

doi:10.1016/j.patcog.2025.112002

Return

Image splicing localization method driven by device difference feature guidance

2025-07-08

PRE

DOI：10.1016/j.patcog.2025.112002

Original

Save

Abstract

En 中文

The increasing popularity of image editing tools poses a significant threat to the originality and authenticity of image content. The most common form of attack on image content is splicing forgery, which involves combining multiple donor images from different imaging devices to create a composite spliced image. In this study, we propose an image splicing region localization method based on device difference guidance. Its core idea lies in taking the physical characteristics of imaging devices as forensic clues, and determining the tampered image regions through effective extraction and analysis of these characteristics. The proposed method is a two-stage multi-scale network model named device difference feature guidance network (DDFG-Net), including the device difference guidance stage and the abnormal region localization stage. In the device difference guidance stage, a dual-stream Vision-Transformer network is adopted for camera source identification to extract the physical characteristics of different imaging devices. Immediately afterwards, the inconsistent features in the image are obtained by using the physical characteristics of the device, and the semantic information is decoupled from the inconsistent features introduced by splicing forgery to avoid the interference of semantic information on the forensic features. In the abnormal region location stage, we proposed the Channel-Wise up-sample and Spatial-Wise down-sample(CWU-SWD) and Cross-Scale Concatenator(CSC). The CWU-SWD module is used to reduce the loss of feature information when the scale changes and ensure the integrity of the information, thereby enabling the network to extract richer subtle features introduced by forgery. The CSC module is used for dynamic selection and fusion of multi-scale features, enabling the network to adapt to tampered areas of different scales. A large number of experimental comparisons on four different datasets have proved the excellent performance of DDFG-Net. Cross-dataset experiments and robustness tests further demonstrated the significant generalization and robustness of DDFG-Net. Furthermore, the ablation experiment further confirmed the rationality and effectiveness of each module in DDFG-Net.