Image splicing region localization with adaptive multi-feature filtration

Abstract

With the proliferation of retouching software, the detrimental effects of malicious image tampering are becoming more prevalent. In this study, we propose an adaptive method based on multi-feature filtration for detecting and localizing manipulated regions within images. In the proposed method, we introduce a noise artifact feature that leverages inconsistencies in noise levels and an ELA weighted feature that capitalizes on compression quality inconsistencies. Then we establish an adaptive feature filtration algorithm to generate forensic features using point cloud-based data representation. We employ the Alpha-shape trend connectivity method to aggregate forensic features for precise localization of image tampering regions. Experimental results on public datasets demonstrate that our approach exhibits superior detection performance and robustness against downsampling, gamma correction, and low-quality images from open social networks.