Robust Texture-Aware Computer-Generated Image Forensic: Benchmark and Algorithm.

Area of Science:

• Computer Vision
• Digital Forensics
• Machine Learning

Background:

• Computer-generated (CG) images are becoming indistinguishable from photographic (PG) images due to advanced rendering and generative adversarial networks.
• Existing datasets for CG image forensics are outdated, limited in scope, and lack diversity.
• Current forensic algorithms often overlook subtle texture differences, relying solely on global visual features.

Purpose of the Study:

• To address the limitations of existing CG image forensic datasets and algorithms.
• To introduce a large-scale, diverse, and unbiased benchmark for CG image forensics.
• To propose a novel texture-aware network for enhanced CG image detection.

Main Methods:

• Development of the Large-Scale CG images Benchmark (LSCGB) containing 71,168 CG and 71,168 PG images.
• Collection of diverse CG images from various scenes and rendering techniques, alongside varied PG images.
• Implementation of a texture-aware network that strengthens texture information and explores feature channel relationships via Gram matrices.

Main Results:

• The LSCGB benchmark is significantly larger and more diverse than previous datasets.
• The proposed texture-aware network effectively captures finer texture differences between CG and PG images.
• Experimental results show the proposed method surpasses existing CG image forensic techniques.

Conclusions:

• The developed benchmark and texture-aware network offer a significant advancement in CG image forensics.
• The findings pave the way for more robust detection of synthetic media.
• The LSCGB benchmark is publicly available to facilitate further research.