OraGAN: a deep learning based model for restoring Oracle Bone Script Images

Oracle bone script, China’s earliest known mature writing system, epitomizes ancient Chinese civilization’s origins. However, many characters are severely damaged by temporal degradation and natural erosion, complicating their interpretation. Thus, restoring oracle bone script images is imperative. Given the limited availability of individual character datasets for oracle bone script, we developed the Colored Oracle Bone Script Dataset (RGB-Oracle) and propose the Oracle Bone Script Image Restoration Model (OraGAN), designed to restore character features and textures. OraGAN integrates a Feature Alignment (FA) module and a Multi-Scale Restoration (MSR) module. The FA module aligns damaged images with reference features, while MSR restores content at multiple scales, preserving fine details. These modules leverage glyph and contextual texture information from reference images to accurately reconstruct oracle bone scripts. Experimental results demonstrate that the proposed restoration network outperforms existing methods in texture detail restoration, recovering the original appearance.