Strong robust copy-move forgery detection network based on layer-by-layer decoupling refinement

This paper proposes an all-encompassing methodology called Strong Robust Copy-Move Forgery Detection Network based on Layer-by-Layer Decoupling Refinement (DRNet) which concentrates on detecting a pair of structurally complete similar areas (the source and the tampered area) in the copy-move forgery image by fully extracting the semantically irrelevant shallow information. The DRNet consists of two interacting modules: the Coarse Similarity Area Detection (CD) module and the Shallow Suppression Similarity Area Detection (SD) module. Specifically, the CD module is leveraged to obtain a coarse locating of similar target areas which also work as prior knowledge to guide the detection of the SD module. The SD module fully mines the suppressed information at the shallow layer of the network through layer-by-layer decoupling and uses it as a supplement to refine the coarse detection from the CD module. In addition, we propose a High-Order Self-Correlation Scheme (HS) by dealing with the problem of introducing noise during the process of utilizing the shallow feature to avoid false alarms and improve the robustness. The designed experiments are conducted on USC-ISI CMFD, CASIA CMFD, and CoMoFoD public datasets and the pixel-level F1 score tested by DRnet is improved by 2.27%, 3.82%, and 4.60% respectively than State-of-the-Art in CMFD.