A Cross-Font Image Retrieval Network for Recognizing Undeciphered Oracle Bone Inscriptions

TL;DR

Problem: deciphering undeciphered Oracle Bone Inscriptions (OBI) is hindered by glyph evolution and limited labeled data. Approach: CFIRN, a Siamese ConvNeXt-based cross-font image retrieval network, uses Multiscale Feature Integration (MFI) and a Multiscale Refinement Classifier (MRC) to align OBI images with deciphered gallery-font scripts; the training objective combines cross-entropy, KL divergence, and triplet losses as $L = CEL + KL + \alpha TL$, with $\alpha = 5$, and $TL_i = \max(\|V_i - V_p\|_2 - \|V_i - V_n\|_2 + M, 0)$, where $M = 0.3$. Contributions: first unified cross-font framework for OBI decipherment; achieves state-of-the-art cross-font retrieval on three datasets (OBI-BI, OBI-BSI, OBI-CS); validates usefulness on real undeciphered OBI data. Significance: enables scalable paleographic decipherment by linking undeciphered inscriptions to deciphered forms via historical-font intermediaries, supporting scholars and digital archives.

Abstract

Oracle Bone Inscription (OBI) is the earliest mature writing system in China, which represents a crucial stage in the development of hieroglyphs. Nevertheless, the substantial quantity of undeciphered OBI characters remains a significant challenge for scholars, while conventional methods of ancient script research are both time-consuming and labor-intensive. In this paper, we propose a cross-font image retrieval network (CFIRN) to decipher OBI characters by establishing associations between OBI characters and other script forms, simulating the interpretive behavior of paleography scholars. Concretely, our network employs a siamese framework to extract deep features from character images of various fonts, fully exploring structure clues with different resolutions by multiscale feature integration (MFI) module and multiscale refinement classifier (MRC). Extensive experiments on three challenging cross-font image retrieval datasets demonstrate that, given undeciphered OBI characters, our CFIRN can effectively achieve accurate matches with characters from other gallery fonts, thereby facilitating the deciphering.

Figures (6)

• Figure 1: (a) Over half of OBI characters remain undeciphered. (b) The goal of deciphering is to match OBI characters to corresponding modern Chinese characters. (c) Utilizing historical font intermediaries to overcome glyph gaps and enable effective decipherment of undeciphered OBI characters.
• Figure 2: (a) Taking the OBI Branch as an example, high-quality feature extraction and class prediction are performed for the input image $\mathbf{OI^{1}}$. (b) The siamese network architecture of the proposed CFIRN. (c) The loss functions utilized in CFIRN include Cross-Entropy for improving classification accuracy, KL Divergence for aligning feature distributions, and Triplet Loss for optimizing the feature space.
• Figure 3: Illustration of the multiscale feature integration (MFI) module.
• Figure 4: Top-5 retrieval results and corresponding modern Chinese characters on the OBI-BI dataset.
• Figure 5: Top-1 retrieval results and corresponding modern Chinese characters on the OBI-BSI dataset and the OBI-CS dataset.