APPLYING DEEP NEURAL NETWORKS FOR AUTOMATED RESTORATION OF DIGITALLY DEGRADED CULTURAL ARTIFACTS

Abstract [English]

The conservation and repair of culturally important artifacts have become a crucial issue due to the fast digitization of heritage collections. Nevertheless, the noise, blur, compression artifacts, and environmental degradation are some of the common reasons as to why digital forms of artifacts are degraded. The research paper proposes a deep learning representation of automated restoration of culturally damaged artifacts through digitally corrupted systems. The presented solution develops restoration as an inverse problem, and combines convolutional neural networks (CNNs), generative adversarial networks (GANs), and attention models based on transformers to successfully produce high-quality images on the basis of the damaged inputs. A multiset of paintings, manuscripts, sculptures and historical records is utilized in this application and pre-processed with normalization and augmentation methods to improve generalization of the models. The model is extracted through the composite loss function which involves the incorporation of pixel-wise reconstruction loss, perceptual loss, and adversarial loss to maintain structural fidelity and small artistic features. Experimental analysis is performed on the basis of such standard measurements as Peak Signal-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Fréchet Inception Distance (FID). Findings show that the hybrid architecture is far more successful in terms of quality results when compared to the current state-of-the-art techniques in terms of restoring texture, color consistency, and finer detail features. The framework also provides a scalable and efficient digital heritage conservation solution, as well as promotes AI-mediated cultural preservation in the future.