Color Patterns And Enhanced Texture Learning For Detecting Computer-Generated Images

• Published in: Computer journal Vol. 67; no. 6; pp. 2303 - 2316
• Main Authors: Xu, Qiang, Xu, Dongmei, Wang, Hao, Yuan, Jianye, Wang, Zhe
• Format: Journal Article
• Language: English
• Published: Oxford University Press 24.06.2024
• Subjects: Computer-Generated (CG); enhanced texture; natural photographs; neural network; color patterns
• ISSN: 0010-4620; 1460-2067
• DOI: 10.1093/comjnl/bxae007

Detection of computer-generated (CG) images can reveal the authenticity and originality of digital images. However, recent cutting-edge image generation methods make it very difficult to distinguish CG images from natural photographs. In this paper, a novel method based on color patterns and enhanced texture learning is proposed to tackle this problem. We designed and implemented the backbone network with a separation-fusion learning strategy by constructing a multi-branch neural network. The luminance and chrominance patterns in dual-color spaces (RGB and YCbCr) are leveraged to achieve a robust representation of image differences. A channel-spatial attention module and a global texture enhancement module are also integrated into a backbone network to enhance the learning of inherent traces. Experiments on several commonly used benchmark datasets and a newly constructed dataset with more realistic and diverse images demonstrate that the proposed algorithm outperforms state-of-the-art competitors by a large margin.