An Improved Splicing Localization Method by Fully Convolutional Networks

Liu and Pun proposed a method based on fully convolutional network (FCN) and conditional random field (CRF) to locate spliced regions in synthesized images from different source images. However, their work has two drawbacks: 1) FCN often smooths detailed structures and ignores small objects and 2) CRF is employed as a standalone post-processing step disconnected from the FCN. Therefore, an improved method is proposed in this paper to overcome these two drawbacks. For the first drawback, region proposal network is introduced into the FCN to enhance the learning of object regions. For the second one, the use of CRF is changed to make the whole network an end-to-end learning system. Moreover, the proposed method uses three FCNs (FCN8, FCN16, and FCN32) with different upsampling layers, and all the three FCNs are initialized from VGG-16 network. Experimental results on three publicly available datasets (DVMM dataset, CASIA v1.0 dataset, and CASIA v2.0 dataset) demonstrate that the proposed method can achieve a better performance than the state-of-the-art methods including some conventional methods and some deep learning-based methods.