Image Tampering Localization based on Pyramid Multi-Scale Pooling Module

Abstract

In this paper, we proposed a new two-stream image forgery localization method based on pyramid multi-scale pooling module. One stream learns the differences in JPEG compression rate between original regions and tampered regions by processed through ELA. In the second stream, the network learns the global image manipulation features from the input image RGB values. Next the fusion stage fuses multiple tampering features from the dual-channel feature extraction network to generate the shallow feature map, Coordinate Attention is applied to refine the predicted mask and further improve the localization accuracy. And then through the multi-scale pyramid pooling module, so the secondary fusion is carried out. Finally, the fused feature map restores the original image size by using UpSampling for Pixel level classification prediction. The proposed net can effectively decrease incorrect prediction since it makes better use of the contextual spatial information in images. And most of all, from two stages extraction fuses lowlevel and high-level information to refine the global manipulation features. The experimental results on several public datasets show that the proposed model outperforms some state-of-the-art methods.