TLDR: This research introduces a novel clustering-based feature representation learning method to improve the automated detection of Oracle Bone Inscriptions (OBIs) from rubbing images. By leveraging a clean Oracle Bones Character (OBC) font library as prior knowledge, the method guides deep learning models to learn more discriminative features, effectively separating authentic inscriptions from noise and cracks. Experiments on mainstream detection frameworks demonstrate significant improvements in accuracy and robustness, offering a generalizable approach for integrating expert knowledge into ancient script analysis.
Oracle Bone Inscriptions (OBIs) are an invaluable window into ancient Chinese civilization, representing the earliest mature writing system in East Asia. These inscriptions, often preserved through rubbing images, are crucial for historical research, character decoding, and studies in various fields like traditional Chinese philosophy and astronomy. However, the automated detection of OBIs from these rubbing images is a significant challenge for digital archaeology.
Understanding Oracle Bone Inscriptions and the Detection Challenge
The primary difficulties in automatically detecting OBIs stem from various degradation factors present in the rubbing images. These include substantial background noise and cracks that often share similar textural properties with the inscriptions themselves. This makes it incredibly hard for conventional detection networks to accurately distinguish between authentic characters and interference. While modern deep learning has improved efficiency over manual methods, these unique challenges persist.
A Novel Approach: Clustering for Clearer Features
To overcome these limitations, researchers have proposed a new method called clustering-based feature space representation learning. This innovative approach leverages the Oracle Bones Character (OBC) font library dataset as a source of ‘prior knowledge’. The core idea is that real OBIs and non-OBIs (like noise and cracks) should occupy distinct areas in the model’s internal feature space. By using the clean, expert-curated OBC font library as a reference, the model is guided to learn more distinctive feature representations, making it better at telling characters apart from artifacts.
How the Method Works
The method integrates both the OBI detection dataset (rubbing images) and the OBC font library dataset during the training process. Images from both datasets are fed into a feature extraction network. The system then identifies ‘sample features’ (potential OBIs from rubbing images), ‘negative features’ (noise/cracks from rubbing images), and ‘positive features’ (clean characters from the font library). Using a clustering algorithm, the negative and positive features are grouped. A specialized loss function, derived from these clustering results, is then introduced. This function works to pull the features of actual OBIs closer to the clean, positive features from the font library, while simultaneously pushing them further away from the negative features (noise and cracks). This ‘contrastive learning’ helps the model to create a clearer separation between what is an inscription and what is not, directly improving its ability to discriminate.
Demonstrated Improvements
The effectiveness of this new method was validated through extensive experiments on two OBI detection datasets. It was applied to three widely used detection frameworks: Faster R-CNN, DETR, and Sparse R-CNN. Across all tested frameworks, the method consistently showed significant improvements in detection accuracy and robustness. Visualizations of the feature space confirmed that the approach successfully enhances character feature discrimination, making the model more resilient against various forms of interference. While there was a slight increase in training time and memory usage, these were deemed acceptable given the performance gains.
Also Read:
- Unlocking Ancient Chinese: A New Digital Character List and Corpus for Language Models
- Enhancing Text Line Detection in Historical Manuscripts with Limited Data
Looking Ahead
This research makes several important contributions to the field of automated OBI analysis and the digital preservation of ancient writing systems. It offers a novel way to combine traditional expert knowledge, in the form of font libraries, with modern deep learning techniques to achieve superior detection performance. The idea of using font libraries as prior knowledge to guide network learning is generalizable and holds promise for other character detection tasks, especially as character databases and font libraries continue to advance.
For more in-depth information, you can read the full research paper here.
