Image tampering localization network based on multi-class attention and progressive subtraction

Image tamper localization is an important research topic in the field of computer vision, which aims at identifying and localizing human-modified regions in images. In this paper, we propose a new image tampering localization network, which is named MAPS-Net. It combines the advantages of efficient multi-scale attention, shift operation, and progressive subtraction, which not only improves the sensitivity and generalization to novel data tampering behaviors but also significantly reduces the computation time. MAPS-Net consists of upper and lower branches, which are the fake edge-enhancing branch and the interfering factors-weakening branch. The fake edge-enhancing branch uses an efficient multi-scale edge residual module to enhance the expressiveness of the features, while the interfering factors-weakening branch uses progressive subtraction to weaken the interference of image content fluctuations in capturing general tampering behaviors. Finally, the features of both branches are fused with a position attention mechanism via a shift operation to capture the spatial relationships between different views. Experiments conducted on several publicly available datasets show that MAPS-Net outperforms existing mainstream models in both image tampering detection and localization, especially in image tampering localization in real scenes. Code is available at: https://github.com/dklive1999/MAPS-Net.