Locating splicing forgery by fully convolutional networks and conditional random field

Abstract

To expose and locate splicing forgery, hand-crafted features are often utilized to discern tampered area in a synthesized image. However, given a spliced picture without prior knowledge, it is difficult to tell which feature will be effective to expose forgery. In addition, a certain hand-crafted feature can only handle one kind of splicing forgery. To address these issues, a method based on using deep neural networks and conditional random field is proposed in this paper. It is achieved by training three different fully convolutional networks (FCNs) and a condition random field (CRF). Each FCN is specialized to deal with different scales of image contents. CRF adaptively combines detection results from these neural networks. Then the trained FCNs–CRF can be used to perform image authentication, yielding pixel-to-pixel forgery prediction. Our FCNs–CRF framework achieves improved performance comparing to existing methods relying on hand-crafted features.