Abstract
Cultural heritage continues to be endangered by environmental, social, and political issues requiring new digital preservation methods. This paper proposes a deep learning framework consisting of an Feature Pyramid Network (FPN) and an Modified Builder Optimization Algorithm (MBOA) whose purpose is to classify, restore, and synthesize culturally significant artwork. Using the WikiArt dataset, the model represents the cultural styles of many artistic traditions by fully utilizing FPN's multi-scale architecture and identifying both fine scale textures as well as global stylistic semantics. The MBOA optimizes hyperparameters through an innovative bio-inspired search strategy, which improves convergence and model performance and improves training efficiency. The framework achieves style classification accuracy of 96.7% in terms of preference, a PSNR of 32.41 dB and SSIM of 0.937 for restoration, and an Fréchet Inception Distance (FID) score of 14.3 for generation, therefore outperforming state-of-the-art CNNs and generative models. The results illustrated the framework's powerful reconstruction capabilities in restoring degraded works of art, or generating new works inspired by cultural aesthetics. By combining a Feature Pyramid Network, Modified Builder Optimization Algorithm, and Neural Style Transfer, this work further expands the capabilities and role of Generative AI for use in digital preservation. This approach supports not only archival fidelity and quality but forms a creative interaction, enhancing the capabilities of networks and reinterpreting global heritage assets in the digital age as a scalable tool for museums, teachers, and cultural institutions.