VisuCeram_500x500-corrected

Sanitary-ware ceramic surface defect detection — cleaned, split and standardized version of the VisuCeram 500×500 dataset. 卫浴陶瓷表面缺陷检测 —— VisuCeram 500×500 数据集的清理、划分与规整版。

English · 中文

---

English

What is this?

A cleaned, split and standardized object-detection version of the VisuCeram 500×500 dataset — sanitary-ware ceramic surface defect detection across 7 defect categories (body crack, crawling, glaze crack, no-defect, pinhole, sign, spot). Upstream repository: Md-Ali-Azad/VisuCeram.

Reference: Md Ali Azad & Md Mahadi Hasan Nahid, "VisuCeram: A Comprehensive Dataset for Sanitary Ware Ceramic Defect Detection with YOLO Models and Benchmarking", IEEE, 2025.

The original VisuCeram-500x500-labelled archive ships as class-named folders (images/<class>/ + labels/<class>/), YOLO format only, with no split, no classes.txt/data.yaml, and it contains 15 out-of-range boxes (normalised centre coordinates up to 1.65 — boxes placed outside the frame). Because a single image may contain boxes of several classes, the per-class folder layout does not describe its contents.

This release removes the invalid boxes, flattens the folders into a single pool, keeps near-duplicate images on one side of a reproducible split, and ships ready-to-train YOLO TXT + Pascal VOC XML annotations with configuration files.

✅ Upstream license: Apache-2.0. Unlike many of our releases, VisuCeram carries a real open-source licence. See License & attribution.

Corrections vs. the official release

#	Correction	Detail
1	Removed 15 invalid boxes	Fifteen boxes had a normalised centre outside the frame (e.g. cx = 1.65) — clearly mis-placed annotations. They were removed, and the 1 image left with no boxes (f8bc9c1d-1_138) was dropped. No valid coordinate was altered.
2	Flattened the class folders	Upstream stores images/labels under seven class-named folders, but a single image may hold boxes of several classes (e.g. a body_crack image containing spot boxes). The pool is flattened back into images/<split> + labels/<split>.
3	Reproducible train/val/test split	Upstream ships no split at all. This release provides a seeded, class-stratified 70 / 15 / 15 split (seed 42).
4	Near-duplicates kept together	Three perceptually near-identical pairs (8×8-dHash ≤ 2 and 32×32 pixel-MSE ≤ 10) are grouped so each stays on one side of the split.
5	Dual format + packaging	Added Pascal VOC XML from the same coordinates alongside the upstream YOLO TXT, plus classes.txt, data.yaml and per-split CSV tables.

Image pixels are never modified. All 3,264 images are byte-for-byte identical to the upstream files (500×500 RGB).

Scope — structural clean-up only. This is not a re-annotation. We drop invalid boxes, de-folder and re-package the data; we do not re-verify the remaining boxes.

Structural checks performed on the released set:

Check	Method	Result
Filename ↔ image ↔ label pairing	bidirectional	0 mismatches / orphans
Out-of-range boxes	after removal	0 remaining
Degenerate boxes	zero / negative width or height	0
Cross-split duplicate images	byte-identical md5 across splits	0
EXIF orientation anomalies	tag scan	0

Dataset at a glance

Property	Value
Images	3,264 (500 × 500, RGB .jpg)
Classes	7 — body_crack, crawling, glaze_crack, no_defect, pinhole, sign, spot
Bounding boxes	6,042
Formats	Pascal VOC XML · YOLO TXT
Split	train 2,283 / val 491 / test 490 (seeded, class-stratified)

Boxes per class: spot 2,030 · body_crack 997 · glaze_crack 973 · no_defect 548 · crawling 532 · pinhole 509 · sign 453

Boxes per split: train 4,198 · val 931 · test 913

Note on no_defect: in the upstream scheme no_defect is itself one of the seven annotated categories (it is not an empty/background image class). The name is kept verbatim from the source for compatibility.

Structure

VisuCeram_500x500-corrected/
├── images/{train,val,test}/      # 500x500 RGB .jpg
├── labels/{train,val,test}/      # YOLO: cls cx cy w h (normalized)
├── Annotations/                  # Pascal VOC XML (flat, from the same coordinates)
├── JPEGImages/                   # copy of the images (VOC layout)
├── csv/
│   ├── train_labels.csv          # filename,width,height,class,xmin,ymin,xmax,ymax
│   ├── val_labels.csv
│   └── test_labels.csv
├── classes.txt                   # 7 class names, one per line
├── data.yaml                     # Ultralytics YOLO configuration
├── LICENSE                       # Apache-2.0 + notice of modifications
└── README.md

Every image has exactly one YOLO .txt and one VOC .xml; there are no empty label files.

Quick start (Ultralytics YOLO)

yolo detect train data=data.yaml model=yolov8n.pt epochs=100 imgsz=640

Citation

1. The original dataset / paper — please always cite this.

@inproceedings{azad2025visuceram,
  title     = {VisuCeram: A Comprehensive Dataset for Sanitary Ware Ceramic Defect
               Detection with YOLO Models and Benchmarking},
  author    = {Azad, Md Ali and Nahid, Md Mahadi Hasan},
  booktitle = {IEEE},
  year      = {2025},
  note      = {Dataset: https://github.com/Md-Ali-Azad/VisuCeram}
}

2. This corrected release — please cite it as well. It is not identical to the official archive: invalid boxes were removed, the folders were flattened, a reproducible split was added and VOC annotations were generated, so a citation to the original alone does not describe this version.

@misc{visuceram_500x500_corrected,
  author    = {KeenForgeAI},
  title     = {VisuCeram_500x500-corrected: a cleaned release of the VisuCeram sanitary-ware ceramic defect dataset},
  year      = {2026},
  version   = {1.0},
  publisher = {KeenForgeAI},
  url       = {https://huggingface.co/datasets/KeenForgeAI/VisuCeram_500x500-corrected},
  note      = {Curated by Lu Gan and Sam Li. Derived from VisuCeram (Azad & Nahid, IEEE 2025), Apache-2.0.}
}

3. The annotation tool (optional).

@software{keenforge,
  author    = {KeenForgeAI},
  title     = {KeenForge: a local-first, offline image annotation and model-training desktop tool},
  year      = {2026},
  publisher = {KeenForgeAI},
  url       = {https://github.com/KeenForgeAI/KeenForge},
  note      = {MIT licensed. Developed by Lu Gan and Sam Li.}
}

License & attribution

The upstream VisuCeram dataset is licensed under the Apache License 2.0. This corrected release is distributed under the same licence; the LICENSE file contains the full licence text preceded by a notice of the modifications made (as required by Apache-2.0 §4(b)).

Original work: VisuCeram, © Md Ali Azad — https://github.com/Md-Ali-Azad/VisuCeram.

Provenance

• Source: VisuCeram (VisuCeram-500x500-labelled archive), GitHub Md-Ali-Azad/VisuCeram.
• Upstream: 3,265 images / 6,057 boxes / 7 classes; no split; 15 out-of-range boxes; class-named folders.
• Verified: 3,264 images (6,042 boxes, byte-identical); 15 invalid boxes removed + 1 empty image dropped; 0 cross-split leakage, 0 invalid boxes remaining.

---

中文

这是什么？

VisuCeram 500×500 数据集的清理、划分与标准化目标检测版——卫浴陶瓷表面缺陷检测， 覆盖 7 个缺陷类别（body_crack、crawling、glaze_crack、no_defect、pinhole、sign、spot）。 上游仓库：Md-Ali-Azad/VisuCeram。

论文：Md Ali Azad & Md Mahadi Hasan Nahid，《VisuCeram: A Comprehensive Dataset for Sanitary Ware Ceramic Defect Detection with YOLO Models and Benchmarking》，IEEE，2025。

原始 VisuCeram-500x500-labelled 压缩包以按类别命名的文件夹 （images/<class>/ + labels/<class>/）发布，仅 YOLO 格式，无划分、无 classes.txt/data.yaml，且包含 15 个越界框（归一化中心坐标最高 1.65 —— 框被放到 画面之外）。由于单张图可能含有多个类别的框，按类别的文件夹结构并不能描述其内容。

本版本删除无效框、把文件夹重新摊平为统一图池、把近重复图固定到同一划分，并提供开箱即用的 YOLO TXT + Pascal VOC XML 双格式标注及配置文件。

✅ 上游许可证：Apache-2.0。 与我们多数发布不同，VisuCeram 带有真正的开源许可证。 详见 许可证与署名。

相对官方版的修正

#	修正	说明
1	删除 15 个无效框	有 15 个框的归一化中心落在画面之外（如 cx = 1.65），显然是错放的标注。已删除；某图删后无框（f8bc9c1d-1_138）时一并丢弃。未修改任何有效坐标。
2	摊平类别文件夹	上游把图片/标注放在七个按类别命名的文件夹中，但单张图可能同时含有多类框（例如 body_crack 图里含 spot 框）。本版把图池摊平为 images/<split> + labels/<split>。
3	可复现的训练/验证/测试划分	上游完全没有划分。本版提供固定随机种子、按类别分层的 70 / 15 / 15 划分（随机种子 42）。
4	近重复固定同侧	三对感知近似图（8×8 dHash ≤ 2 且 32×32 像素 MSE ≤ 10）归组，确保各自只出现在划分的一侧。
5	双格式 + 规范化打包	由同一坐标生成 Pascal VOC XML（与上游 YOLO TXT 并行），并补充 classes.txt、data.yaml 与按划分的 CSV 表格。

图像像素从不修改——全部 3,264 张图与上游文件逐字节一致（500×500 RGB）。

本版范围——仅结构化清理。 这是结构性清理，不是重新标注。我们只删除无效框、摊平目录、重新打包； 不重新核对剩余标注框。

已执行的结构化检查（针对发布的集合）：

检查	方法	结果
文件名↔图片↔标注 配对	双向	0 错配 / 孤儿
越界框	删除后	0 残留
退化框	宽或高为零/负	0
跨划分重复图	跨划分逐字节 md5	0
EXIF 方向异常	标签扫描	0

数据集概览

属性	值
图片	3,264（500 × 500，RGB .jpg）
类别	7 —— body_crack、crawling、glaze_crack、no_defect、pinhole、sign、spot
标注框	6,042
格式	Pascal VOC XML · YOLO TXT
划分	train 2,283 / val 491 / test 490（固定种子、类别分层）

各类框数： spot 2,030 · body_crack 997 · glaze_crack 973 · no_defect 548 · crawling 532 · pinhole 509 · sign 453

各划分框数： train 4,198 · val 931 · test 913

关于 no_defect： 在上游方案中，no_defect 本身就是七个可标注类别之一（并非空图/背景类）。 为保持兼容，此处沿用原始名称。

目录结构

VisuCeram_500x500-corrected/
├── images/{train,val,test}/      # 500x500 RGB .jpg
├── labels/{train,val,test}/      # YOLO：cls cx cy w h（归一化）
├── Annotations/                  # Pascal VOC XML（扁平，来自同一坐标）
├── JPEGImages/                   # 图片副本（VOC 布局）
├── csv/
│   ├── train_labels.csv          # filename,width,height,class,xmin,ymin,xmax,ymax
│   ├── val_labels.csv
│   └── test_labels.csv
├── classes.txt                   # 7 个类别名称，每行一个
├── data.yaml                     # Ultralytics YOLO 配置文件
├── LICENSE                       # Apache-2.0 + 修改说明
└── README.md

每张图片都有且仅有一个 YOLO .txt 与一个 VOC .xml，不存在空标签文件。

快速开始 (Ultralytics YOLO)

yolo detect train data=data.yaml model=yolov8n.pt epochs=100 imgsz=640

引用

使用本数据集请同时引用原始工作与本修正版：

1. 原始数据集 / 论文（请务必引用） —— 见上方英文部分 azad2025visuceram。

2. 本修正版（请一并引用） —— 本版与官方压缩包并不相同：删除了无效框、摊平了目录、 补充了可复现划分并生成了 VOC 标注，因此只引用原数据集无法描述本版本。 见上方英文部分 visuceram_500x500_corrected。

3. 标注工具（可选） —— 见上方英文部分 keenforge。

许可证与署名

上游 VisuCeram 数据集采用 Apache License 2.0。 本修正版以相同许可证发布； LICENSE 文件包含 Apache-2.0 全文，并在其前附上本版的修改说明（符合 Apache-2.0 §4(b) 要求）。

原始作品：VisuCeram，© Md Ali Azad —— https://github.com/Md-Ali-Azad/VisuCeram。

来源

• 来源：VisuCeram（VisuCeram-500x500-labelled 压缩包），GitHub Md-Ali-Azad/VisuCeram
• 上游：3,265 图 / 6,057 框 / 7 类；无划分；15 个越界框；按类别文件夹存放
• 已校验：3,264 图（6,042 框，逐字节一致）；删除 15 个无效框 + 丢弃 1 张空图；0 跨划分泄漏、0 残留越界框