EUVIMEDEUVIMED entdecken
Crossref

Self-Supervised Semantic Prior Generative Adversarial Network for Heritage Image Restoration

Suma N., Kusuma Kumari B M.

Journal of Innovative Image Processing · 08.04.2026 · 8 (2) · 439-462

DOI: 10.36548/jiip.2026.2.001 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Innovative Image Processing, online erschienen am 08.04.2026. Verfasst von Suma N. und 1 weiteren Autor:innen. Die Arbeit nennt 25 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The heritage images are subjected to large and irregular missing regions due to long-term degradation. In order to repair the gaps, a filling process is used which creates visible seams. High computational expense and prolonged inference time are involved in diffusion and transformer models. A framework for restoration based on GAN is built to handle the irregular missing regions while preserving delicate brush strokes and structural lines with faster speed. The proposed framework makes use of the Adaptive Semantic Prior Injection (ASPI), Topology-Aware Semantic Refinement (TASR) and Entropy-Guided Self-Supervision (EGSS). ASPI contextualizes guidance while TASR provides line and edge continuity, and EGSS emphasizes uncertain regions during training. SSPL learns semantics from masked images with no labels and is used to aid the performance of irregular restoration. The Kaggle Damaged Paintings dataset is used to train the model with an image resolution of 256×256 for 50 epochs. Masked WikiArt images are utilized for self-supervised prior learning. The Indian Heritage Monuments pictures are validated for model performance. The model recorded a Peak Signal-to-Noise Ratio (PSNR) of 32.6 dB, a Structural Similarity Index Measure (SSIM) of 0.92, and a Fréchet Inception Distance (FID) of 18.7. According to the results, diffusion and transformer-based baselines are consistently outperformed. Addditionally, more texture continuity is gained when compared to exemplar based inpainting and Conditional GAN (cGAN) using partial convolution-based models. The ablation study shows that TASR mainly promotes structural fidelity while ASPI and EGSS provide additional benefits. The architecture remains efficient with 34.8 million parameters, 108.5 Giga Floating Point Operations (GFLOPs), and an inference time of 0.118 seconds per image Despite the improvements, the model remains unable to manage thin cracks and the edges of the restored region can hardly be seen. Additionally, significant omissions may not completely correspond to the original in the final restoration. Future studies will focus on training at greater resolutions and applying diffusion priors appropriately.

Zitieren

APA

N., S., & M., K. K. B. (2026). Self-Supervised Semantic Prior Generative Adversarial Network for Heritage Image Restoration. Journal of Innovative Image Processing, 8(2), 439-462. https://doi.org/10.36548/jiip.2026.2.001

Vancouver

N. S, M. KKB. Self-Supervised Semantic Prior Generative Adversarial Network for Heritage Image Restoration. Journal of Innovative Image Processing. 2026;8(2):439-462. doi:10.36548/jiip.2026.2.001

BibTeX
@article{N2026,
  author = {N., Suma and M., Kusuma Kumari B},
  title = {Self-Supervised Semantic Prior Generative Adversarial Network for Heritage Image Restoration},
  journal = {Journal of Innovative Image Processing},
  year = {2026},
  volume = {8},
  number = {2},
  pages = {439--462},
  doi = {10.36548/jiip.2026.2.001}
}
Kostenlos starten →