Background
When an image is maliciously edited or partially destroyed, traditional forensics can only detect the tampered area, but not restore what was lost. Our research goes one step further: self-recovery. By embedding hidden representations of the original image into itself. Our research can automatically reconstruct tampered regions with high fidelity. This approach enhances media integrity verification, supports forensic evidence restoration, and provides a foundation for resilient digital watermarking.
This research explores two complementary directions:
Traditional Fragile Watermarking-Based Self-Recovery (传统脆弱水印方法的自恢复): Embeds fragile watermark into the host image. When tampering occurs, these embedded codes are extracted to reconstruct the altered areas. This approach offers high authenticity assurance and high-quality restoration, suitable for forensic and legal use.
Deep Learning-Based Self-Recovery (基于深度学习的自恢复): Leverages convolutional or generative models to predict and regenerate tampered regions directly from learned representations. This approach offers stronger robustness against compression, noise, and blurring, making it more suitable for real-world applications.
Qiyuan Zhang, Xiaochen Yuan, Chan-Tong Lam, Zheng Xing, Guoheng Huang
Journal of King Saud University - Computer and Information Sciences, vol. 35, 2023, p. 101795
Tampering localization and self-recovery using block labeling and adaptive significance
Qiyuan Zhang, Xiaochen Yuan, Tong Liu, Chan-Tong Lam, Guoheng Huang, Di Lin, Ping Li
Expert Systems with Applications, vol. 226, 2023, p. 120228