Learn with diversity and from harder samples: Improving the generalization of CNN-Based detection of computer-generated images

Abstract Advanced computer graphics rendering software tools can now produce computer-generated (CG) images with increasingly high level of photorealism. This makes it more and more difficult to distinguish natural images (NIs) from CG images by naked human eyes. For this forensic problem, recently some CNN(convolutional neural network)-based methods have been proposed. However, researchers rarely pay attention to the blind detection (or generalization) problem, i.e., no training sample is available from “unknown” computer graphics rendering tools that we may encounter during the testing phase. We observe that detector performance decreases, sometimes drastically, in this challenging but realistic setting. To study this challenging problem, we first collect four high-quality CG image datasets, which will be appropriately released to facilitate the relevant research. Then, we design a novel two-branch network with different initializations in the first layer to capture diverse features. Moreover, we introduce a gradient-based method to construct harder negative samples and conduct enhanced training to further improve the generalization of CNN-based detectors. Experimental results demonstrate the effectiveness of our method in improving the performance for the challenging task of “blind” detection of CG images.