MSR-Net: a multi-scale skeleton restoration network for oracle bone inscription rubbing denoising

Abstract Oracle bone inscriptions (OBIs) suffer from over three thousand years of weathering, resulting in complex noise (dense white areas, abnormal edges, broken strokes) that severely hinders automatic deciphering. This paper proposes MSR-Net, a multi-scale skeleton restoration network tailored for OBI rubbings. Its core innovation is the Multi-scale Skeleton Kernel (MSK) block, comprising a Lightweight Channel-wise Self-Attention Block (LCSAB) for efficient global context modeling and a Glyph Structural Network Block (GSNB) to implement multi-scale skeleton supervision in a coarse-to-fine manner. An edge consistency loss further enforces glyph fidelity. On the OBIMD dataset, MSR-Net achieves 12.95 dB PSNR and 0.821 SSIM, surpassing the state-of-the-art OBIFormer by 0.80 dB and 0.032. On cross-domain datasets (Oracle-P15K, Oracle-241), it shows favorable generalization, improving downstream ResNet accuracy by 47%–54%. MSR-Net offers an efficient, structure-preserving solution for digital restoration of OBI rubbings and a new paradigm for cultural heritage image restoration.