Deep Learning-Based Restoration of Historic Building Images with Blockchain-Driven Maintenance Data Management

Abstract

Ancient buildings, steeped in millennia of human history, stand as precious legacies requiring meticulous preservation and restoration. With this premise, this study proposes a restoration method for damaged ancient structures based on a deep neural network. A curated dataset of ancient building images—including diverse damage patterns and material textures-was established to train and validate the model, ensuring robustness across varied architectural styles. The core network architecture is an enhanced U-Net, in which all convolutional layers are replaced with partial convolution layers. The sliding window of partial convolutional layers exclusively performs convolution operations in the relevant region of the image, enabling accurate prediction of building structures with irregular apertures. To further enhance feature extraction and generalization, a dual-transfer learning mechanism is introduced: structural features learned from large-scale architectural datasets are transferred to fine-tune restoration on heritage-specific samples. Consequently, the semantic richness of restoration results is reconstructed. Simultaneously, in accordance with the functional requisites of the ancient building maintenance data storage and sharing system, a dedicated blockchain-driven data storage platform is designed. This platform caters to diverse stakeholders, including lay users, professional restorers, museums, and governmental bodies. The platform employs Hyperledger Fabric technology and a traditional database to ensure secure and reliable data management. Model performance is quantitatively assessed using standard image restoration metrics—Learned Perceptual Image Patch Similarity (LPIPS = 0.725), Peak Signal-to-Noise Ratio (PSNR = 26.16), and Structural Similarity Index Measure (SSIM = 0.57)—demonstrating competitive reconstruction quality. It is worth noting that the proposed building restoration method demonstrates significant potential for cost savings, while also achieving a 25% improvement in efficiency. This potential is particularly relevant within the context of restoring old buildings.