HCSU
收藏资源简介:
HCSU(细粒度历史书法风格理解数据集)是首个专为细粒度历史书法风格理解设计的大规模数据集和评估基准。该数据集旨在解决现有大型视觉语言模型在书法艺术领域‘有知识但无感知’的困境,即缺乏对书法作品微观笔法、墨法和结构的精细感知能力。通过一套开创性的严格数据处理流程,HCSU成功地将真实的墨迹本(帖)与石刻拓本(碑)解耦,彻底解决了长期困扰数字文化遗产领域的‘模态混淆’问题。数据集包含39,307张经过精心标注的高清汉字图像(统一归一化为256×256分辨率),并附有专家级的层次化美学描述。数据覆盖了从三国到近现代的10个主要历史朝代,囊括了49位中国艺术史上具有里程碑意义的书法名家,并全面涵盖了篆、隶、楷、行、草五种主要书体。根据数据来源和处理程度,HCSU划分为三个子域:1) 墨迹本(帖):3,780张图像,高度保留了真实的墨韵动态(如墨色浓淡、飞白纹理);2) 石刻拓本(碑):3,240张图像,提取了纯粹的几何结构和笔画张力;3) 野生数据:32,287张图像,为未经后续严格处理流程的原始书法作品字符图像。HCSU引入了超越传统‘扁平标签’的多维度、可解释的层次化注释框架。每个数据实例提供丰富的元数据:1) 语义与上下文真值:包括字符ID、书法家、朝代、来源媒介和保存质量;2) 细粒度视觉属性分解:具体描述墨法风格(墨色浓淡、枯润、渗化、飞白纹理)、笔法风格(微观笔触特征,如中锋、藏锋、方折)和字法结构(整体空间布局与几何比例);3) 专家自然语言描述:由书法鉴赏专家撰写的专业美学评述(50字以内),作为生成任务的真实参考。数据集适用于图像分类和图像到文本任务,特别是细粒度视觉识别、风格理解、文化遗产数字化分析和可解释人工智能等领域的研究。数据采用CC BY-NC 4.0许可,仅供学术和非商业研究使用。
HCSU (Fine-Grained Historical Calligraphy Style Understanding Dataset) is the first large-scale dataset and evaluation benchmark specifically designed for fine-grained historical calligraphy style understanding. It aims to address the knowledgeable but unperceptive dilemma of existing large vision-language models in the field of calligraphy art, i.e., the lack of fine-grained perception of micro-level brushwork, ink techniques, and structure in calligraphy works. Through a pioneering strict data processing pipeline, HCSU successfully decouples authentic ink manuscripts (Tie) from stone rubbings (Bei), completely solving the long-standing modal confusion problem in digital cultural heritage. The dataset contains 39,307 meticulously annotated high-definition Chinese character images (uniformly normalized to 256×256 resolution), accompanied by expert-level hierarchical aesthetic descriptions. The data spans 10 major historical dynasties from the Three Kingdoms to modern times, includes 49 milestone calligraphy masters in Chinese art history, and comprehensively covers five main script styles: Seal, Clerical, Regular, Running, and Cursive. Based on data source and processing level, HCSU is divided into three subdomains: 1) Ink Manuscripts (Tie): 3,780 images, highly preserving authentic ink dynamics (e.g., ink shades, flying white textures); 2) Stone Rubbings (Bei): 3,240 images, extracting pure geometric structures and stroke tension; 3) Wild Data: 32,287 images, raw calligraphy character images without subsequent strict processing. HCSU introduces a multi-dimensional, interpretable hierarchical annotation framework that goes beyond traditional flat labels. Each data instance provides rich metadata: 1) Semantic and Contextual Ground Truth: including character ID, calligrapher, dynasty, source medium, and preservation quality; 2) Fine-Grained Visual Attribute Decomposition: detailed descriptions of ink style (ink shades, dryness/wetness, diffusion, flying white textures), brushwork style (micro-brushstroke features, such as centered tip, hidden tip, square turns), and character structure (overall spatial layout and geometric proportions); 3) Expert Natural Language Descriptions: professional aesthetic comments (within 50 words) written by calligraphy appraisal experts, serving as ground truth for generation tasks. The dataset is suitable for image classification and image-to-text tasks, particularly in research areas like fine-grained visual recognition, style understanding, digital cultural heritage analysis, and explainable AI. The data is licensed under CC BY-NC 4.0 and intended solely for academic and non-commercial research use.
HCSU:细粒度历史书法风格理解数据集
数据集概览
HCSU(Fine-Grained Historical Calligraphy Style Understanding Dataset)是首个专门用于细粒度历史书法风格理解的大规模数据集和评估基准,由同济大学团队构建,已被ECCV 2026收录。该数据集共包含39,307张经过精心标注的高清汉字图像,并配有专家级分层审美描述。
核心统计
- 总规模:39,307张高分辨率汉字图像(归一化为256×256)
- 时间跨度:覆盖从三国到现代的10个主要历史朝代
- 书法家:涵盖49位中国艺术史里程碑式的书法大师
- 书体分类:完整覆盖5种主要书体——篆书、隶书、楷书、行书、草书
- 数据子域:
- 🖌️ 墨迹本(帖):3,780张图像,真实保留墨韵(墨色、飞白等)
- 🪨 碑刻拓本(碑):3,240张图像,提取纯几何结构与笔画张力
- 📜 原始数据:32,287张图像,为未经管道处理的书法作品原始汉字图像
数据标注体系
HCSU引入多维度、可解释的分层标注框架,每个实例提供以下元数据:
- 语义与上下文标注:
- 字符ID(
char) - 书法家(
author) - 朝代(
dynasty) - 来源介质(
source_type) - 保存质量(
quality)
- 字符ID(
- 细粒度视觉属性分解:
ink_style:墨色密度、水分、渗透与飞白纹理stroke_style:微观笔法特征(如中锋、藏锋、方折)character_structure:整体空间布局与几何比例(如中宫收紧、欹正相生)
- 专家自然语言描述:由书法鉴赏家撰写的专业审美评论(50字以内)
数据处理管道
为解决历史档案中不同物理介质(纸本与石刻)造成的干扰,HCSU设计了多阶段严格处理流程:
阶段一:领域自适应极性归一化与视觉启发式纠正
- 对拓本图像进行全局极性反转(255 - 灰度值),统一为“黑底白字”标准配置
- 通过四角区域像素强度检查自动纠正元数据错误
阶段二:几何无损归一化
- 将原始图像嵌入白色画布中心,保持原始宽高比
- 使用Lanczos插值算法下采样至256×256分辨率,有效抑制锯齿效应
阶段三:形态学去噪与“真实墨韵”重建
- 对二值掩码进行连通域分析,过滤小于50像素的独立噪声点
- 将纯净语义掩码映射回几何对齐的原始高分辨率RGB图像
- 所有背景像素替换为纯白色[255,255,255],在剥离背景干扰的同时100%保留墨色密度、笔画纹理与书写压力变化
评估协议
基于处理后的高质量数据,HCSU提出两项严格评估任务:
- 细粒度风格辨别(8选1候选选择):检验模型能否在1个正确风格与7个强干扰项中正确配对风格
- 可解释审美推理(风格描述生成):给定2-shot提示,模型需使用专业术语生成50字以内的书法评论,采用
BERTScore与LLM-as-a-Judge双轨评估框架
许可协议
- 代码与文档:Apache License 2.0
- 数据集图像、压缩包、标注、样本及相关视觉资产:Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0),仅限学术与非商业研究用途
引用
bibtex @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} }




