A Multi-level Feature Enhancement Network for Image Splicing Localization

Abstract

Most current neural network-based splicing localization methods are based on subtle telltales from inter-pixel differences. But for recompressed and downsampled data, these artifacts are weakened. In this paper, we propose a novel multi-level feature enhancement network (MFENet) to enhance the features. Tampering with an image not only destroys the consistency of the inherent high-frequency noise in host images, but also is performed post-processing operations to weaken this discrepancy. Therefore, based on the high-pass filtered image residuals, we combine the detection evidence of post-processing operations to complete splicing forensic task. For the purpose of enhancing the distinguishability of features in the residual domain, we use bilinear pooling to fuse low-level manipulation features and residuals. In order to improve the consistency between the ground truth and the splicing localization result, we integrate global attention modules to minimize the intra-class variance by measuring the similarity of features. Finally, we propose a multi-scale training generation strategy to train our network, which provides local and global information for the input and pays more attention to the overall localization during gradient feedback. The experimental results show that our method achieves better performance than other state-of-the-art methods.