HCSU: A Dataset and Benchmark for Fine-Grained Historical Calligraphy Style Understanding (V1.0)

ECCV 2026

Yinsheng Yao^*, Yan Liu^*, Chen Ye^†

^* Equal contribution. ^† Corresponding author.

Abstract: Addressing the "Knowledgeable but unperceptive" dilemma where existing Large Vision-Language Models (LVLMs) possess historical knowledge but lack fine-grained calligraphy style perception, we introduce HCSU—the first large-scale dataset and evaluation benchmark specifically tailored for fine-grained historical calligraphy style understanding. HCSU contains 39,307 meticulously annotated high-definition Chinese character images, accompanied by expert-level hierarchical aesthetic descriptions. Through a pioneering rigorous data processing pipeline, we successfully decouple authentic ink manuscripts (Tie) from stone rubbings (Bei), thoroughly resolving the "modality aliasing" problem that has long plagued the digital cultural heritage field.

Figure 1: The panoramic data experience of HCSU. Beyond single images, we provide technical dimensions of calligraphy expression (Ink Style, Stroke Style, Structure) and expert-level descriptions of aesthetic spirit, enabling highly interpretable fine-grained visual reasoning.

📊 1. Dataset Overview

Traditional OCR or Document AI datasets typically focus only on "content recognition." HCSU marks a paradigm shift by focusing the core task on fine-grained perception of artistic style. To eliminate interference caused by different physical media in historical archives, HCSU implements strict domain partitioning during its construction.

Core Dataset Statistics

• Total Scale: 39,307 high-resolution character images (normalized to 256×256) subjected to rigorous manual verification.
• Diversity Coverage: Spans 10 major historical dynasties (from the Three Kingdoms to the Modern era), covering 49 calligraphic masters who serve as milestones in Chinese art history.
• Script Classification: Full coverage of the 5 major scripts: Seal (Zhuanshu), Clerical (Lishu), Regular (Kaishu), Semi-cursive (Xingshu), and Cursive (Caoshu).
• Three Sub-domains:
  • 🖌️ Ink Manuscripts (Tie): 3,780 images, highly preserving authentic ink dynamics (e.g., ink wash, "flying white" textures).
  • 🪨 Stone Rubbings (Bei): 3,240 images, extracting pure geometric structures and stroke tension.
  • 📜 Wild Data: 32,287 images, raw character images from calligraphy works without the following data processing pipeline.
• Archive Passwords:
  • bei.zip: eWZNA9hzJoyMHtQwVh7QnBtN7
  • tie.zip: eWZNA9hzJoyMHtQwVh7QnBtN7
  • wild.zip.001 to wild.zip.006: To request the password, email yechen@tongji.edu.cn with your institution, professional title, and a brief description of the intended use.
• Annotations: Full metadata for the released Bei/Tie archives is provided in bei_annotations.json and tie_annotations.json.
• Example Subsets: bei_samples/ and tie_samples/ are provided uncompressed for quick preview.

Repository File Tree

.
├── README.md
├── bei.zip
├── tie.zip
├── wild.zip.001
├── wild.zip.002
├── wild.zip.003
├── wild.zip.004
├── wild.zip.005
├── wild.zip.006
├── bei_annotations.json
├── tie_annotations.json
├── bei_samples/
├── tie_samples/
└── assets/

Figure 2: Comprehensive statistics of the HCSU dataset. Displays the distribution of script types across the three domains (Wild, Bei, Tie) and a stacked chart of character counts by dynasty.

---

👁️ 2. Data Experience: Beyond "Flattened Labels"

Existing calligraphy datasets (such as MCCD, CalliNet) usually provide only "flattened labels" like "Author" or "Character ID," which cannot support deep visual interpretation by large models. HCSU introduces a multi-dimensional, interpretable hierarchical annotation framework that bridges low-level visual features with high-level aesthetic judgments.

Each instance in HCSU provides the following rich metadata experience:

1. Semantic and Contextual Ground Truth: Character ID (char), Calligrapher (author), Dynasty (dynasty), Source Medium (source_type), and preservation quality.
2. Fine-grained Visual Attribute Decomposition:
   • ink_style: Describes ink density, moisture, permeation, and "flying white" (Feibai) textures.
   • stroke_style: Micro-level brushwork features (e.g., "center-tip brushwork," "hidden tip," "angular turns").
   • character_structure: Overall spatial layout and geometric proportions (e.g., "tight center and loose exterior," "dynamic balance between tilted and upright").
3. Expert Natural Language Description: Professional aesthetic commentary (under 50 words) written by calligraphy connoisseurs, serving as the Ground Truth for generation tasks.

Figure 3: Decoupling style and content. HCSU requires models to identify consistent artistic signatures across visually distinct instances (e.g., a and c belong to the same calligrapher) while distinguishing instances with identical content but different artistic expressions (e.g., a and b).

---

⚙️ 3. Strict Data Processing Pipeline

To build an evaluation benchmark with absolute structural consistency, free from complex background interference (e.g., paper aging, mounting borders, stone cracks), we designed a multi-stage rigorous data processing pipeline. This is the core competency that allows HCSU to preserve "authentic ink charm" without loss.

Phase 1: Domain-Adaptive Polarity Normalization and Visual Heuristic Correction

• Metadata-driven Polarity Inversion: Stone rubbings (Bei) typically appear as white characters on a black background. We perform a global polarity inversion ($255 - I_{gray}$) to unify them into a standard "black on white" configuration, followed by Otsu's method to obtain an initial binary map.
• Visual Heuristic Error Correction: Metadata in historical archives is often inaccurate. We check the average pixel intensity of four $N \times N$ corner regions. If the intensity is below a strict threshold $\tau_{corner}$ (indicating a dark background), the heuristic algorithm overrides the metadata and automatically executes bit inversion, ensuring zero-human-intervention polarity accuracy.

Phase 2: Geometric Lossless Normalization

Traditional direct resizing leads to severe anisotropic deformation, destroying the critical structural tension of calligraphy.

1. We first embed the original image into the center of a new $S \times S$ pure white canvas ($S$ is the maximum of the original width and height).
2. We use the high-quality Lanczos interpolation algorithm to downsample the square image to the target $256 \times 256$ resolution. This operation effectively suppresses aliasing effects and perfectly maintains the original aspect ratio and stroke sharpness.

Phase 3: Morphological Denoising and "Authentic Ink" Appearance Reconstruction

While binary maps provide clean structural priors, they discard ink textures (like Feibai) crucial for expert appreciation.

1. We apply 8-connected component analysis (CCA) to the binary mask, filtering out independent noise spots smaller than 50 pixels (e.g., paper grain or stone cracks).
2. The cleaned semantic mask is re-mapped back onto the geometrically aligned original high-resolution RGB image.
3. All background pixels are replaced with pure white ([255, 255, 255]). This mechanism perfectly strips away complex background interference while preserving 100% of the authentic ink density, stroke texture, and writing pressure variations.

Figure 4: Demonstration of the HCSU data construction pipeline. Top row: Background stripping and ink reconstruction for Ink Manuscripts (Tie); Bottom row: Polarity inversion and denoising for Stone Rubbings (Bei).

---

🏆 Evaluation Protocols

Based on the processed high-quality data, HCSU proposes two rigorous evaluation tasks (the test set employs strict Anti-Leakage sampling rules, ensuring that target and reference images never contain the same character):

1. Fine-grained Style Discrimination (8-way Candidate Selection): Tests whether the model can correctly pair styles among 1 correct style and 7 strong distractors, relying solely on micro-level brushwork and structural features.
2. Explainable Aesthetic Reasoning (Style Description Generation): Given a 2-shot prompt, the model is required to generate a calligraphic critique under 50 words using professional terminology. We use a dual-track evaluation framework consisting of BERTScore and LLM-as-a-Judge (evaluating terminology accuracy and detail richness).

Evaluation code and Prompt templates can be found in our open-source repository eva.

---

📝 Citation

If you use this dataset in your research or are inspired by our data processing pipeline, please consider citing our paper:

@inproceedings{hcsu2026,
  title={HCSU: A Dataset and Benchmark for Fine-Grained Historical Calligraphy Style Understanding},
  author={Yao, Yinsheng and Liu, Yan and Ye, Chen},
  booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
  year={2026}
}

📄 License

• Code: Apache 2.0 License.
• Dataset (Images and Annotations): Licensed under CC BY-NC 4.0. For academic research only; unauthorized commercial use is strictly prohibited.