Deep Learning Based Digital Artwork Restoration Using ResNet-GAN for Cultural Heritage Preservation
Abstract
The research suggests a new way to restore digital art using a GAN (Generative Adversarial Network) working together with a pretrained Residual Network and a modified version of U-Net architecture. The proposed system attempts to recover excellent images from images that are damaged, even without having to use masks. The encoder in the generator uses ResNet, which allows it to pick useful features from the degraded photos, and the decoder sorts them back into their original form. The model is trained and evaluated on a dataset of artwork images collected from open-source publicly available sources. To simulate real-world degradation, synthetic damage such as missing regions, noise, and structural distortions is introduced during training. It was the PSNR, SSIM, MSE, and FID methods that helped test the quality of restoration. The model achieved a PSNR of 31.85 dB, being better than the existing methods. Further evaluation showed SSIM scores of 0.57 and FID scores of 345. This process enhances the way restorations are done and keeps the original features, so it can greatly help with India’s heritage preservation and its cultural assets.References
[1]Basu, A., Paul, S., Ghosh, S., & Das, S. (2023). Digital restoration of cultural heritage with data-driven computing: A survey. IEEE Access.
[2] Jia, Y., Yu, W., & Zhao, L. (2024). Generative adversarial networks with texture recovery and physical constraints for remote sensing image dehazing. Scientific Reports, 14, 31426.
[3] Bano, A., & Salamat, M. (2025). Image Processing Methods: A Systematic Literature Review of Classical and Modern Approaches.
[4] Kniaz, V. V., et al. (2026). Generative inpainting of partially destroyed frescoes.
[5] Fan, Y., Hong, C., Zeng, G., & Liu, L. (2024). A deep convolutional encoder–decoder–restorer architecture for image deblurring 56, 27.
[6] Liu, Y., Chen, X., & Zhang, W. (2024). A survey on image-to-image translation with generative adversarial networks. IEEE Access, 12, 45678–45702.
[7] Zhu, J. Y., Park, T., Isola, P., & Efros, A. A. (2017). Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE International Conference on Computer Vision (ICCV) (pp. 2223–2232).
[8] Dhariwal, P., & Nichol, A. (2021). Diffusion models beat GANs on image synthesis. Advances in Neural Information Processing Systems, 34, 8780–8794.
[9] Wang, Z., Zheng, H., He, P., Chen, W., & Zhou, M. (2022). Diffusion-GAN: Training GANs with diffusion. arXiv preprint arXiv:2206.02262.
[10] Zhao, W., Bai, L., Rao, Y., Zhou, J., & Lu, J. (2023). UniPC: A unified predictor-corrector framework for fast sampling of diffusion models. arXiv preprint arXiv:2302.04867.
[11] Wang, R., Li, C., Wang, J., Liu, G., Zhang, H.-Y., Zeng, B., & Liu, S. (2021). UPHDR-GAN: Generative adversarial network for high dynamic range imaging with unpaired data. arXiv preprint arXiv:2102.01850.
Published
Issue
Section
License
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.
