Revisiting fragile watermarking recovery: a deep learning-based image restoration framework
Signal, Image and Video Processing, cilt.20, sa.13, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 20 Sayı: 13
- Basım Tarihi: 2026
- Doi Numarası: 10.1007/s11760-026-05734-3
- Dergi Adı: Signal, Image and Video Processing
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Aerospace Database, Compendex, INSPEC, zbMATH, Technology Collection (ProQuest)
- Anahtar Kelimeler: Deep learning, Digital forensics, Fragile watermarking, Image authentication, Image restoration, Tamper recovery
- Kayseri Üniversitesi Adresli: Evet
Özet
Fragile watermarking schemes enable tamper detection and partial recovery, yet their reconstruction quality remains fundamentally limited by coarse block-level approximations. This work treats fragile watermarking recovery as a learned image restoration problem and proposes a modular deep learning (DL) framework that operates as a post-processing stage over classical recovery outputs. This framework feeds a two-channel input to a deep restoration network, using two outputs already available in the classical pipeline. One channel is a binary manipulation map (MAP) identifying damaged blocks at the watermarking grid resolution. The other is a coarsely recovered image (REC) derived from the embedded block-level payload. Restoration then refines the classical REC. Explicit MAP conditioning is architecture-dependent and is not the sole source of the observed gain. Across 8790 attacked samples from three public datasets (MS COCO, FFHQ, NWPU VHR-10) and 30 attack types, all six architectures improve peak signal-to-noise ratio (PSNR) by at least + 3.14 dB. The reference for this gain is the implemented classical baseline on the same test set. MAP-UNet, the best model, achieves 38.85 dB PSNR and a structural similarity index (SSIM) of 0.9544 on this fixed set.