Distinguishing computer-generated images from photographic images using two-stream convolutional neural network

Abstract

Today’s advanced multimedia tools allow us to create photorealistic computer graphic images, effortlessly. There are various fields such as the film industry, virtual reality, video games where computer-generated (CG) images are used widely. CG images can also be misused in many ways. Therefore, there is a need of distinguishing CG images from real photographic (PG) images. This paper proposes a method to distinguish CG images from PG images using a two-stream convolutional neural network (CNN). In the proposed method, the first stream takes the advantage of the knowledge learned by the pre-trained VGG-19 network, and then this knowledge is transferred to learn the distinct features of CG and PG images. Here, we propose a second stream, that preprocesses the images using three high-pass filters which aim to help the network to focus on noise-based distinct features of CG and PG images. Finally, we propose an ensemble model to merge the outcomes of the proposed two streams. Comparative and self-analysis experiments demonstrates that the proposed method outperforms the state-of-the-art methods in terms of classification accuracy. The experimental results also show that the proposed method performs satisfactorily under the additive white Gaussian noise postprocessing operation in the images.