MCCD: Multi-Attribute Chinese Calligraphy Dataset
- MCCD is a large-scale isolated Chinese calligraphy character dataset annotated with character identity, script style, dynasty, and calligrapher.
- It was curated from authoritative sources with 329,715 images spanning 7,765 characters, supporting classification, attribution, and multi-task learning.
- Benchmark results reveal high accuracy for style classification and highlight challenges in calligrapher identification due to significant intra-class variation.
Searching arXiv for the cited MCCD paper and closely related Chinese calligraphy datasets/frameworks. The Multi-Attribute Chinese Calligraphy Character Dataset (MCCD) is a large-scale dataset of isolated Chinese calligraphy character images annotated with character identity and three additional historical-stylistic attributes: script style, dynasty, and calligrapher. Introduced in 2025, it contains 329,715 image samples spanning 7,765 character categories, 10 script styles, 15 dynasties, and 142 calligraphers, with the explicit aim of supporting calligraphic character recognition, script style classification, dynasty attribution, calligrapher identification, and multi-task learning on culturally grounded visual data (Zhao et al., 9 Jul 2025).
1. Definition and attribute scope
MCCD is defined at the level of isolated character images rather than full-page manuscripts. All images are single-character crops extracted directly from original calligraphic works, and the dataset is intended for classification-oriented study of calligraphic form and metadata rather than page-layout parsing or sequence transcription (Zhao et al., 9 Jul 2025).
Its core annotation axes are four categorical variables: character identity, script style, dynasty, and calligrapher. The character space comprises 7,765 categories. The style space comprises 10 types: Oracle Bone, Bronze, Stamp, Lithography, Wooden, Seal, Clerical, Cursive, Semi-Cursive, and Regular. The dynasty subset covers 15 historical periods, with explicitly mentioned examples including Shang, Zhou, Qin, Han, Sui, Tang, Song, Yuan, and Qing, extending to the modern era. The calligrapher axis contains 142 individuals (Zhao et al., 9 Jul 2025).
A common misconception is that MCCD is a uniformly dense four-attribute corpus in which every image carries all labels simultaneously. The published description instead states that missing attributes are labeled as null during collection, and that three attribute-based subsets are re-extracted under stricter completeness constraints. The same limitation appears in multi-task use: only 67,628 samples could be used for four-task learning. This indicates that “multi-attribute” in MCCD refers to the dataset-level availability of multiple annotation axes, not to a guarantee that every raw sample has every label (Zhao et al., 9 Jul 2025).
This design makes MCCD especially suitable for studying entanglement among content, style, period, and authorship. At the same time, it remains narrower than page-level corpora with layout or transcription supervision, because no bounding boxes, reading order labels, or stroke-level annotations are described for the released benchmark specification.
2. Provenance, curation, and labeling protocol
MCCD is built from two authoritative online calligraphy platforms: ZiTong (\begin{CJK*}{UTF8}{gbsn}字统\end{CJK*}), with 258,830 images, and ShuFaTuJi (\begin{CJK*}{UTF8}{gbsn}书法图集\end{CJK*}), with 70,885 images (Zhao et al., 9 Jul 2025). The sources provide high-resolution, authentic calligraphic images together with metadata such as character, script style, calligrapher, and dynasty. The dataset authors did not perform primary page segmentation themselves; the source websites had already extracted isolated-character crops.
Collection proceeds through a custom web crawler designed to maximize acquisition while preserving image-label alignment. For each sample, the pipeline attempts to retrieve character code or text, style label, dynasty label, and calligrapher label, assigning null when attributes are missing. Cleaning then merges the two sources and applies manual inspection to remove images containing two characters in one crop, garbled or incorrect label mappings, and blurred or low-quality images. Two annotators each spent about 50 hours on review and filtering (Zhao et al., 9 Jul 2025).
The label criteria are explicitly selective. Only samples with reliable attribute labels are retained in the attribute-based subsets. Dynasty categories are merged when historical periods are adjacent and calligraphic style differences are minor. Calligraphers with too few samples or uncertain attribution are removed, while some over-represented calligraphers are subsampled to reduce extreme imbalance. Style labels follow the classification provided by ZiTong, with minor normalization (Zhao et al., 9 Jul 2025).
The calligrapher subset is not an indiscriminate crawl of all attributed works. It is constructed from well-known, widely recognized calligraphers so as to preserve cultural and artistic representativeness while maintaining sufficient sample size per individual. Examples explicitly mentioned in the dataset description include Huai Su, Mi Fu, Wu Rui, Deng Shiru, Zhi Yong, and Li Ye (Zhao et al., 9 Jul 2025).
3. Corpus composition, subsets, and storage format
MCCD consists of a full corpus plus three structured subsets aligned to the three non-character attributes. The subset construction is central to how the dataset is meant to be used.
| Dataset | Images / classes | Train / Test |
|---|---|---|
| MCCD | 329,715 / 7,765 characters | 234,225 / 95,460 |
| MCCD-Style | 258,830 / 10 styles | 181,186 / 77,644 |
| MCCD-Dynasty | 258,830 / 15 dynasties | 181,187 / 77,643 |
| MCCD-Calligrapher | 92,122 / 142 calligraphers | 64,544 / 27,578 |
Within each category folder, images are randomly split 70% for training and 30% for testing. To improve I/O efficiency, the training and testing sets are also stored in LMDB format. The published organization is category-per-folder: character folders for MCCD, 10 style folders for MCCD-Style, 15 dynasty folders for MCCD-Dynasty, and 142 calligrapher folders for MCCD-Calligrapher (Zhao et al., 9 Jul 2025).
The class distributions are constrained but still nonuniform. In MCCD-Style, each style has at least 10,000 images and 80% of styles have more than 20,000 images. In MCCD-Dynasty, each dynasty has at least 2,000 images and 9 dynasties have more than 10,000 images. In MCCD-Calligrapher, each calligrapher has at least 100 images and 36 calligraphers have more than 1,000 images (Zhao et al., 9 Jul 2025).
The image files are stored as PNG with original colors preserved. Aspect ratios are mostly in , with mean $0.9241$ and standard deviation $0.2988$, although the overall range extends from below $0.2$ to above $3.0$. Training-time preprocessing resizes inputs to and applies RandAugment with severity , depth , and augmentation_set = "all". No binarization is reported for MCCD; this distinguishes it from some earlier calligraphy synthesis datasets that normalize to binary images (Zhao et al., 9 Jul 2025).
The dataset is released at https://github.com/SCUT-DLVCLab/MCCD. The paper body does not state a license, so licensing details must be checked in the repository itself (Zhao et al., 9 Jul 2025).
4. Benchmark tasks, protocol, and empirical difficulty
The benchmark suite covers four single-task problems: character recognition on MCCD, script style classification on MCCD-Style, dynasty classification on MCCD-Dynasty, and calligrapher identification on MCCD-Calligrapher. The reported models are ResNet50, Vision Transformer, and Swin Transformer v2; character recognition additionally includes HierCode and CCR-CLIP. Training uses AdamW with initial learning rate $0.01$, MultiStepLR halving the learning rate every 10 epochs, 100 epochs, batch size $128$, and a single NVIDIA GeForce RTX 3090 GPU (Zhao et al., 9 Jul 2025).
The reported metrics are Top-1 Accuracy, Top-5 Accuracy, and Macro Accuracy, where
$0.9241$0
This metric is especially informative under class imbalance because it weights classes equally rather than by frequency (Zhao et al., 9 Jul 2025).
| Task | Best reported model | Top-1 / Top-5 / MacroAcc |
|---|---|---|
| Character recognition | Swin Transformer | 79.141% / 93.919% / 75.949% |
| Style classification | ResNet50 | 95.405% / 99.995% / 94.579% |
| Dynasty classification | ResNet50 | 88.469% / 98.881% / 79.477% |
| Calligrapher identification | ResNet50 | 67.676% / 88.209% / 60.265% |
Several empirical patterns are emphasized in the benchmark analysis. Style recognition is comparatively easy, with nearly perfect Top-5 accuracy, whereas calligrapher identification is the hardest task. Character recognition sits between them: it is more difficult than style or dynasty classification because the same character can vary radically across scripts, while different characters may become visually similar in cursive or ancient forms. The large gap between Top-1 and Top-5 accuracy on character, dynasty, and calligrapher tasks is treated as evidence of genuine ambiguity rather than mere optimization failure (Zhao et al., 9 Jul 2025).
The benchmark also shows that radical-oriented methods that perform well on regular printed or handwritten Chinese do not transfer cleanly to calligraphy. HierCode reaches 60.225% Top-1 and 46.650% MacroAcc, while CCR-CLIP reaches 67.670% Top-1 and 63.730% MacroAcc on MCCD. The reported explanation is that large intra-class variation across script styles and the breakdown of typical radical structures in cursive and ancient scripts undermine methods that assume stable radical decomposition (Zhao et al., 9 Jul 2025).
5. Position within the Chinese calligraphy dataset landscape
MCCD occupies a distinct position relative to earlier and adjacent datasets. CCS-4, introduced for Chinese calligraphy synthesis with specified style, is a paired image-to-image translation benchmark linking standard font images to calligraphic renderings. It contains four subsets of about 7,000 images each, one per calligrapher/style, uses $0.9241$1 binary images, and enforces train/validation splits with no overlap in characters. Its structure is explicitly paired and style-specific, but it does not provide the style–dynasty–calligrapher annotation combination that defines MCCD (Lyu et al., 2017).
A 2020 calligraphy image release associated with CalligraphyGAN reports 138,499 images of Chinese calligraphy characters written by 19 calligraphers from the Internet, spanning 7,328 different Chinese characters. However, that paper does not name the dataset MCCD, does not use the phrase “multi-attribute,” and from a modeling perspective uses character identity as the only explicit conditioning label. This makes it a large calligraphy image dataset, but not a direct precursor of MCCD’s attribute-centered design (Zhuo et al., 2020).
Later resources broaden the comparison in different directions. CalliReader uses a single-character dataset of 742,975 stylized calligraphy character images covering 6,763 GB2312 characters, plus a page-level dataset of 10,549 artworks with manually labeled character bounding boxes and contents, and a benchmark with style, layout, author, and semantic interpretation tasks. UniCalli introduces a corpus of over 8,000 classical Chinese calligraphy works, with more than 4,000 annotated and more than 150,000 character instances, focusing on column-level generation and recognition. Chronicles-OCR is a cross-temporal benchmark of 2,800 strictly balanced images across the Seven Chinese Scripts, targeting VLLM evaluation rather than isolated-character attribute classification (Luo et al., 9 Mar 2025, Xu et al., 15 Oct 2025, Li et al., 12 May 2026).
This comparison suggests a useful taxonomy. MCCD is primarily an isolated-character, multi-attribute recognition dataset. CCS-4 is a paired synthesis benchmark. The CalligraphyGAN dataset is a large character image collection used mainly for character-conditioned generation. CalliReader and UniCalli are page- or column-level resources with richer structural supervision. Chronicles-OCR is cross-temporal and document-centric. MCCD’s contribution lies in making script style, dynasty, and calligrapher explicit, sizable, and benchmarked at the single-character level.
6. Research uses, misconceptions, and future development
The authors position MCCD as a resource for calligraphic character recognition, script style classification, dynasty classification, calligrapher identification, multi-task learning, and evolutionary studies of Chinese characters (Zhao et al., 9 Jul 2025). Because the labels are simultaneously linguistic, stylistic, and historical, the dataset is also relevant to digital archives, search, indexing, provenance analysis, and quantitative study of character-form evolution across time.
A second misconception is that “multi-attribute” here implies the kind of structural attribute coding used in open-set Chinese character recognition, such as Pinyin, layout structure, stroke count, Cangjie, Zhengma, Wubi, or Four-Corner. In fact, MCCD’s attribute axes are script style, dynasty, and calligrapher, not pronunciation- or decomposition-based descriptors. The 2018 multi-typed attribute literature shows that such structural codes can be powerful for zero-shot and few-shot recognition of unseen characters, especially under open-set conditions (He et al., 2018). A plausible implication is that MCCD could support future work that combines historical-stylistic labels with structural attribute codes, thereby linking author/style attribution to open-set character recognition.
The limitations stated or implied in the dataset description are substantial. Coverage is restricted to materials available on ZiTong and ShuFaTuJi. Not all images have complete attributes. Class imbalance persists, especially for dynasty and calligrapher labels. Image quality varies because the corpus mixes rubbings, inscriptions, manuscripts, prints, and lithography, and segmentation quality is inherited from the source websites. No bounding boxes for components, no stroke-level annotations, and no richer physical metadata such as material, tool, layout, region, or purpose are included in the present benchmark specification (Zhao et al., 9 Jul 2025).
The future directions are correspondingly clear. The dataset paper points to improved multi-task modeling with better shared-layer architectures and loss functions, algorithmic handling of long-tail imbalance through re-weighting, re-sampling, or long-tail aware losses, dataset expansion toward underrepresented dynasties and calligraphers, and richer annotations including material, tool, layout, region, purpose, higher-resolution images, bounding boxes for components, and stroke-level data (Zhao et al., 9 Jul 2025). This suggests that MCCD is not only a benchmark in its own right, but also a foundation for a broader research program in which Chinese calligraphy is modeled simultaneously as text, image, style, authorship, and historical evolution.