OBIFormer: A Fast Attentive Denoising Framework for Oracle Bone Inscriptions

TL;DR

This paper presents OBIFormer, a fast attentive denoising framework for Oracle Bone Inscriptions that integrates channel-wise self-attention, glyph extraction, and selective kernel feature fusion to reconstruct denoised glyph Images while preserving stroke structure. The model adopts a U-shaped encoder-decoder with specialized OFBs that fuse reconstruction and glyph information, yielding state-of-the-art PSNR/SSIM on Oracle-50K and RCRN and demonstrating promising generalization to the real-world OBC306 dataset. Additional contributions include a jointly optimized loss for image and skeleton outputs, comprehensive ablations, and evidence that denoising enhances downstream OBI recognition across multiple ResNet backbones. The work achieves strong practical impact by offering a computationally efficient denoising approach that supports automatic OBI recognition and provides code for reproducibility.

Abstract

Oracle bone inscriptions (OBIs) are the earliest known form of Chinese characters and serve as a valuable resource for research in anthropology and archaeology. However, most excavated fragments are severely degraded due to thousands of years of natural weathering, corrosion, and man-made destruction, making automatic OBI recognition extremely challenging. Previous methods either focus on pixel-level information or utilize vanilla transformers for glyph-based OBI denoising, which leads to tremendous computational overhead. Therefore, this paper proposes a fast attentive denoising framework for oracle bone inscriptions, i.e., OBIFormer. It leverages channel-wise self-attention, glyph extraction, and selective kernel feature fusion to reconstruct denoised images precisely while being computationally efficient. Our OBIFormer achieves state-of-the-art denoising performance for PSNR and SSIM metrics on synthetic and original OBI datasets. Furthermore, comprehensive experiments on a real oracle dataset demonstrate the great potential of our OBIFormer in assisting automatic OBI recognition. The code will be made available at https://github.com/LJHolyGround/OBIFormer.

Figures (12)

• Figure 1: Four types of noise in real rubbings. The red rectangle indicates the corresponding noise. (a) Stroke-broken, (b) Bone-cracked, (c) Abnormal edges, (d) Dense white regions.
• Figure 2: Our model achieves state-of-the-art performance on the OBI denoising task while being computationally efficient.
• Figure 3: Examples of oracle character images in different OBI datasets: (a) Oracle-50K han2020self, (b) HWOBC li2020hwobc, (c) EVOBI wang2022study, (d) OBC306 huang2019obc306, (e) OBI125 yue2022dynamic, and (f) EVOBC dataset guan2024open. The zoomed-in images are different structural variations of the same character.
• Figure 4: The overall architecture of our OBIFormer. (a) OBIFormer block (OFB) that injects glyph information into the denoising backbone, (b) Glyph structural network block (GSNB) that extracts glyph features, (c) Channel-wise self-attention block (CSAB) that generates channel-wise self-attention effectively and efficiently, (d) Selective kernel feature fusion (SKFF) module that aggregates reconstruction features and glyph features.
• Figure 5: Qualitative comparisons of baseline methods and our OBIFormer on Oracle-50K dataset han2020self.