Pyramid Copy-move Forgery Detection Using Adversarial Optimized Self Deep Matching Network

In this paper, we propose a pyramid copy-move forgery detection framework based on multiscale Conditional Random Field (CRF) using an adversarial optimized self deep matching network. We present a patch-level adversarial optimization scheme to optimize a pre-trained self deep matching network, in order to approximate the distribution of ground truth masks more accurately. The optimized network is adopted as the basic self deep matching network in pyramid copy-move forgery detection. Coarse-to-fine images in pyramid form are put into the adversarial optimized self deep matching network to generate a set of score maps which contain rich multiscale information. Multiscale CRF is designed based on average score unary potentials and pairwise potentials with multiscale information kernels, to adequately explore multiscale information. Extensive experiments on publicly available datasets demonstrate the state-of-the-art performance of the proposed method.