Cultural relic image restoration using two-stage transformer-CNN framework

Cultural relic image restoration presents unique challenges due to irregular damage and historically specific textures, which standard deep learning methods struggle to address. This paper proposes a novel two-stage Transformer-CNN framework tailored for this task. The first stage leverages a Transformer to capture global structural dependencies from low-resolution priors, generating coherent coarse proposals. The second stage employs a specialized CNN to refine fine-grained textures from these proposals, optimized by a compound perceptual loss function. Validated on a new large-scale dataset of 88,000 East Asian cultural relic images, our approach demonstrates state-of-the-art performance. A key contribution is the generation of diversified restoration outputs, providing conservators with multiple valid references for decision-making. This work establishes an effective paradigm for digital heritage conservation that balances global structural integrity with local texture fidelity.