C2R Net: The Coarse to Refined Network for Image Forgery Detection

Yang Wei, Xiu-Li Bi, Bin Xiao

Abstract: Image splicing forgery means to extract a portion from one image and then place it to another image for merging a new image. Distinguishing a splicing forgery image is a challenging task. In this paper, we propose a deep learning based method to detect splicing image. The proposed network includes two convolutional neural networks (CNNs): the coarse CNN and the refined CNN, which extracts the differences between image itself and splicing regions from patch descriptors of different scales. Unlike previous detection methods that always rely on one property difference detection between the splicing image and the original image, the proposed detection method learns various intrinsic property differences between the splicing images and the original images by CNNs with different scales. For decreasing the complexity of computational time, we further propose image-level CNN to replace the previous patch-level CNN for fast computation. Experimental results show that the proposed detection method is better than the previous detection methods, especially the testing dataset is the real-world dataset.