Distinguish computer generated and digital images: A CNN solution
Ming Ge He · Concurrency and Computation Practice and Experience · 2018
Summary The development of computer graphics has promoted the creation of computer generated images (CG) to a degree of unrivaled realism. It is of benefit to some industries like games and movies, which is aiming at making photo‐realistic images. At the same time, it brought attacks on many vision systems. An artist can modify one fake image using knowledge based on computer graphics to deceive most people, turning this into a very dangerous weapon. It is of great importance to differentiate a photo‐realistic computer generated image from a real photograph (PG). This problem can be modeled as a binary classification problem. Given one image, we just predict a label like “CG” or “PG” on it. To address this classification problem, we propose a method based on convolutional network through transfer learning. We choose VGG and ResNet as our base network structure and develop different models. Current state‐of‐the‐art approaches rely on hand‐crafted feature while we adopt a power convolutional network as an alternative and achieve the state‐of‐the‐art performance. In comparison, our method is end to end and more stable.