SPDNet: Spliced Person Detection Network Using Two-Channel Four-Stream CNN Architecture

Image splicing, particularly inserting individuals into scenes, presents significant challenges for detecting image forgery. This paper introduces SPDNet, a two-stage deep learning framework for detecting spliced persons. The first stage segments all person instances using a PointRend-based model, while the second stage analyzes each instance using a two-channel, four-stream architecture that fuses RGB and SRM noise features via EfficientNet-B4 backbones and Squeeze-and-Excitation modules. We evaluated SPDNet on the DSO-1 and OIS datasets as well as our newly introduced SPD Evaluation dataset, comparing it against eight state-of-the-art image splicing detection models. Experimental results show that SPDNet outperforms existing methods in both accuracy and localization of spliced individuals, demonstrating strong potential for person-level image forensics.