RA-Net: A Deep Learning Approach Based on Residual Structure and Attention Mechanism for Image Copy-Move Forgery Detection

To reduce the difficulty of image forensics on forgery images, in this paper, we present an efficient end-to-end deep learning approach using Residual Structure and Attention Mechanism (RA-Net) for image copy-move forgery detection (CMFD). The RA-Net can locate the forged areas and corresponding genuine areas, and it is composed of two modules, Residual Feature Extraction module (RFEM) and Feature Matching & Up-sampling module (FMUM). RFEM is designed to extract deep feature maps, which enriches the combination of gradient information and attention mechanism that focuses the attention of RA-Net to the forged areas. The FMUM assists RA-Net is used to detect copy-move forgery areas and return the previous output to the size of the input image for analysis and visualization of the results. Furthermore, we create a RANet-CMFD dataset for the training, the way to generate RANet-CMFD dataset could help solve the problem of not having enough dataset in some research areas. Otherwise, comparison results show that our model can achieve satisfied performance on CoMoFoD dataset at the pixel level, and performs superior than the compared methods.