TL;DR
This paper introduces AI-powered tools for recognizing Oracle Bone Inscriptions (OBI) and rejoining fragmented Oracle Bones (OB). They developed a novel rejoining algorithm achieving 98.39% accuracy and an OBI recognition system with 80.9% accuracy, alongside annotation software and a large-scale, character-level annotated OBI dataset.
What Problem Does This Paper Address?
Oracle Bone Inscriptions (OBI) are crucial for understanding ancient Chinese history and the evolution of Chinese characters. However, studying them is challenging because many Oracle Bones are found in fragments, requiring laborious manual rejoining. Furthermore, recognizing the inscriptions themselves is difficult due to variations in character shape, size, orientation, and image artifacts like cracks and moires. Existing OBI research often relies on manual efforts, which are time-consuming and require extensive domain expertise. This paper aims to automate and improve these processes using Artificial Intelligence.
Methods
The authors propose a two-pronged AI approach:
- OB Fragments Rejoining: They developed a "red-green-line" algorithm. This method utilizes manually marked "red lines" along the fragmented borders and "green lines" as tangents to ensure smooth joins. The red lines are converted into numerical "time series" data, allowing for similarity measures like their novel "tolerance difference" (T-Diff) to compare potential fragment matches. This approach is designed to be robust to imperfections in the fragmented borders.
- OBI Recognition: For character localization within rubbing images, they employ deep learning-based scene text detection algorithms, specifically mentioning EAST. For recognition, they introduce a novel deep learning-based template matching algorithm. This Siamese network-based approach learns the similarity between OBI character crops from rubbings and template images from an OBI font library, effectively handling the variations in handwritten characters.
Additionally, they developed:
- Annotation Software: A tool to facilitate fine-grained, character-level annotations of OBIs on rubbing images. This overcomes the limitation of existing image-level annotations.
- Filtering Software: A tool to help OBI scholars quickly filter out unpromising fragment rejoining candidates.
Datasets
- OB-Rejoin: A newly built real-world dataset containing 1,000 OB fragments, with borders manually annotated by OBI scholars.
- OracleBone-8000: A large-scale, character-level annotated dataset comprising 7,824 OBI rubbing images and 128,770 annotated character crops. This dataset is noted as being imbalanced and sparse, providing a unique benchmark.
Key Insights and Ideas
- Bridging the Gap: The paper effectively bridges the gap between traditional, manual OBI research and modern AI techniques, offering practical tools for domain experts.
- Data is Key: The creation of the OracleBone-8000 dataset with character-level annotations is a significant contribution, enabling the application of advanced deep learning models for OBI recognition.
- Robustness in Imperfect Data: The T-Diff similarity measure for time series data is a clever way to handle noisy and incomplete boundary information in fragment rejoining.
- Leveraging Existing Strengths: The use of established scene text detection algorithms (like EAST) for localization demonstrates a practical approach to adapting state-of-the-art computer vision techniques to a specialized domain.
Key Results
- OB Rejoining: The "red-green-line" algorithm with T-Diff achieved a top-10 accuracy of 98.39% on the OB-Rejoin dataset.
- OBI Recognition: The novel deep learning template matching algorithm achieved an overall accuracy of 80.9%.
- Character Localization: Using the EAST algorithm, they obtained an overall recall of 88.8% and a precision of 90.7% on the OracleBone-8000 dataset.
Limitations and Future Work
- Annotation Effort: While they developed annotation software, creating the OracleBone-8000 dataset still required significant manual effort from OBI professors and students.
- Recognition Accuracy: The 80.9% accuracy for OBI recognition, while good, still leaves room for improvement, especially given the inherent challenges of OBI data.
- Future Work: The paper implies that the developed algorithms and software tools are ready for use by domain experts, suggesting that further refinement and application of these tools will be the next steps. The availability of their resources on GitHub also points towards community-driven improvements.
Summary updated