Image Inpainting Forensics Algorithm Based on Dual-Domain Encoder-Decoder Network

Image inpainting can fill regions of images with plausible content and can also be used to remove specific objects and leave only faint traces in inpainted images, which pose serious security issues. At present, there are relatively few forensic works on image inpainting. Moreover, there is a problem of poor generalization. Therefore, This paper proposes a dual-domain encoder-decoder network (DDEDNet) based on different input types, which is a two-branch network. The first branch is a spatial domain-based encoder network (S-Encoder) used to capture the tampering traces left by image inpainting in the spatial domain; the second branch is an encoder network based on the frequency domain (F-Encoder), which is used to mine the subtle artifacts left in the frequency domain. Then a cross-modal attention fusion module (CMAF) is used to fuse the features of the two encoder networks to obtain rich fused features. Finally, attention-gated (AG) skip connections are utilized to improve localization performance by properly incorporating multi-scale features in the decoder. Experimental results show that in the face of data sets with both deep inpainting and traditional schemes, DDEDNet can locate the inpainting area more accurately, effectively resist JPEG compression and Gaussian noise attacks, and performs better generalization.