Region Aware Framework for Constrained Image Splicing Detection and Localization

Constrained image splicing detection and localization (CISDL) is a newly proposed image forensics task. It provides two images to detect whether they have undergone splicing manipulation and localize the spliced regions on both images. Recent studies have made significant progress, but some limitations remain. Existing methods extract global features from images and compute feature correlation on them, introducing a large amount of background redundant information, which may interfere with the feature correlation results. Moreover, they calculate correlation matching on high-level coarse features, which results in the loss of many fine texture details, making it difficult to localize small manipulated regions. We propose a novel Region Aware framework that leverages a Region Aware module to perceive highly correlated regions in two images, reducing the impact of background information on feature correlation and final localization performance. In addition, we introduce the linear transformer for calculating feature correlation, which effectively models the relationship between features while addressing the high computational complexity issue of the vanilla transformer. As a result, our model can perform correlation matching on higher-resolution fine-grained features, enhancing its ability to detect and localize small splicing regions. The experimental results demonstrate that our method outperforms existing state-of-the-art methods on several benchmark datasets, proving its effectiveness.