MagneticTile-corrected

Magnetic tile surface defect detection — cleaned and bounding-box standardized version of the Magnetic Tile Surface Defect Dataset.
磁瓦表面缺陷检测 —— 磁瓦表面缺陷数据集的目标检测清洗与边界框规整版。

English · 中文

---

English

What is this?

A cleaned and standardized object-detection version of the Magnetic Tile Surface Defect Dataset (Huang, Qiu & Yuan — "Saliency of magnetic tile surface defects", arXiv:1805.01168, 2018; upstream repo: abin24/Magnetic-tile-defect-datasets).

The original benchmark provides only continuous probability grayscale saliency masks (PNG) without bounding boxes, contains high-frequency industrial camera burst-shot duplicate frames that cause benchmark evaluation data leakage, has blank defect annotations, and lacks standard dataset partitions.

This release converts the saliency masks into precise YOLO TXT and Pascal VOC XML bounding boxes, strips out redundant burst-shot duplicates, removes corrupt blank masks, provides a stratified 70/10/20 train/val/test split, and supplies ready-to-train YOLO configuration files.

✅ Upstream license: Creative Commons Attribution 4.0 International (CC BY 4.0). This release maintains the same open license.

Corrections vs. the official release

#	Correction	Detail
1	Added Bounding Boxes (YOLO + Pascal VOC)	Upstream released only grayscale saliency masks (0–255) and zero bounding boxes. Adaptive Otsu thresholding and morphological closing were applied to extract tight, high-precision bounding boxes (437 BBoxes) across all 5 defect categories.
2	Removed 22 Burst-Shot Duplicates	Industrial cameras captured identical tiles in rapid burst intervals (e.g. exp2_num_86854 vs exp2_num_85846, MAE = 1.64). 22 near-identical temporal duplicates were isolated into _removed_duplicates/, preventing evaluation data leakage between splits.
3	Removed 4 Zero-Defect Dirty Images	In category MT_Uneven, 4 images were labeled as defective but had 100% black masks (0 foreground defect pixels: exp3_num_45042, exp4_num_124690, exp4_num_352527, exp4_num_45057). Removed to prevent harmful negative-supervision.
4	Added Reproducible Stratified Split	Partitioned images into Train (70%) / Val (10%) / Test (20%) using seeded stratified sampling to guarantee rare classes (such as Fray) are well-represented across all splits.
5	Standard Packaging	Added data.yaml, classes.txt, and synchronized dual-format annotations (Annotations/ + labels/).
6	Defect-free images ship without a label file	The 941 MT_Free negatives keep their image and their (object-free) VOC XML, but have no YOLO .txt. In YOLO a missing label file is behaviorally identical to an empty one — both mean "no objects" — and empty files additionally break dataset indexing on Hugging Face.

Dataset at a glance

Property	Value
Images (total)	1,318 (941 defect-free normal negatives, 377 defective)
Classes	5 — Blowhole, Break, Crack, Fray, Uneven
Bounding boxes	437
Formats	Pascal VOC XML · YOLO TXT
Split	train 920 (70%) / val 129 (10%) / test 269 (20%)

Boxes per class: Break 121 · Blowhole 113 · Uneven 105 · Crack 61 · Fray 37

Structure

MagneticTile-corrected/
├── images/
│   ├── train/                 # 920 images
│   ├── val/                   # 129 images
│   └── test/                  # 269 images
├── labels/
│   ├── train/                 # 262 YOLO TXT files (defective images only)
│   ├── val/                   # 35 YOLO TXT files
│   └── test/                  # 80 YOLO TXT files
├── Annotations/               # Pascal VOC XML annotations (1,318 files)
├── JPEGImages/                # Full image collection (1,318 images)
├── classes.txt                # 5 defect category names
├── data.yaml                  # Ultralytics YOLO configuration
├── _KeenForge_MagneticTile_compare.png  # Visual inspection proof
└── README.md

Note on label files. Only the 377 defective images have a YOLO .txt (377 files total). The 941 defect-free MT_Free negatives have no label file — a missing file and an empty file are equivalent in YOLO (both = "no objects"), and empty files break dataset indexing. Annotations/ still contains a VOC XML for every image (negatives have zero <object> entries).

Quick Start (Ultralytics YOLO)

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

Citation

1. The original dataset — please always cite this.

@article{huang2018saliency,
  title   = {Saliency of magnetic tile surface defects},
  author  = {Huang, Yibin and Qiu, Congying and Yuan, Kui},
  journal = {arXiv preprint arXiv:1805.01168},
  year    = {2018}
}

2. This corrected release — please cite it as well.

@misc{magnetictile_corrected,
  author    = {KeenForgeAI},
  title     = {MagneticTile-corrected: a cleaned and bounding-box standardized release of the Magnetic Tile Surface Defect dataset},
  year      = {2026},
  version   = {1.0},
  publisher = {KeenForgeAI},
  url       = {https://huggingface.co/datasets/KeenForgeAI/MagneticTile-corrected},
  note      = {Curated by Lu Gan and Sam Li. Derived from Huang et al. (2018),
               arXiv:1805.01168. CC BY 4.0 licensed.}
}

3. The annotation and quality inspection tool.

@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

CC BY 4.0 — Creative Commons Attribution 4.0 International, matching the upstream repository's terms. Please also credit the original authors.

---

中文

这是什么？

磁瓦表面缺陷数据集（Magnetic Tile Surface Defect Dataset）（Huang 等，arXiv:1805.01168，2018）的目标检测清洗与标准化重构版。

原始数据集仅提供连续灰度显著性掩码（Mask PNG），缺少矩形目标检测标注框（BBox）；且工业相机连拍导致大量高相似度冗余帧，容易造成评估集数据泄漏；此外还存在纯黑空标脏样本以及目录结构割裂等问题。

本版本利用 KeenForge 质检引擎对原始数据进行去重与清洗，通过自适应分割算法提取紧密 BBox，补充 YOLO TXT 与 Pascal VOC XML 双格式标注，剔除 22 张时序连拍冗余图与 4 张纯黑空标脏图，并提供开箱即用的分层划分与训练配置文件。

✅ 上游许可证：CC BY 4.0。本版沿用同一开源协议。

相对官方版的修正

#	修正项	说明
1	补充检测框（YOLO + Pascal VOC）	官方仅发布连续灰度掩码（0~255），无任何目标检测框。本版采用自适应分割与连通域提取，在 5 类缺陷中生成 437 个高精度 Bounding Box 矩形框。
2	剔除 22 张连拍重复图	工业相机在微秒级间隔抓拍了几乎相同的工件（如 exp2_num_86854 与 exp2_num_85846 平均像素差仅 1.64）。本版将 **22 张时序冗余帧隔离至 _removed_duplicates/**，彻底根除划分集间的数据泄漏。
3	删除 4 张纯黑空标脏图	在 MT_Uneven 类中发现 4 张标记为缺陷但掩码全黑（0 缺陷像素）的异常图（exp3_num_45042、exp4_num_124690、exp4_num_352527、exp4_num_45057），已全部剔除，杜绝负向误学习。
4	补充可复现的分层划分	采用随机种子按 70% 训练 / 10% 验证 / 20% 测试 进行分层抽样，确保稀缺类别（如 Fray 磨损）在验证与测试集中稳定存在。
5	标准化打包	补充 data.yaml、classes.txt 以及标准组织目录，支持主流框架直接训练。
6	无缺陷图不附带标签文件	941 张 MT_Free 正常底板保留图片与（无目标的）VOC XML，但不提供 YOLO .txt。在 YOLO 中"缺少标签文件"与"空标签文件"完全等价（均表示无目标），而空文件还会导致 Hugging Face 数据集索引失败。

数据集概览

属性	值
图像总数	1,318（941 张无缺陷正常底板，377 张缺陷样本）
缺陷类别	5 —— Blowhole（气孔）、Break（断裂）、Crack（裂纹）、Fray（磨损）、Uneven（表面不平）
标注框总数	437
标注格式	Pascal VOC XML · YOLO TXT
划分集	train 920（70%） / val 129（10%） / test 269（20%）

各类框数： Break 121 · Blowhole 113 · Uneven 105 · Crack 61 · Fray 37

目录结构

MagneticTile-corrected/
├── images/
│   ├── train/                 # 920 张图片
│   ├── val/                   # 129 张图片
│   └── test/                  # 269 张图片
├── labels/
│   ├── train/                 # 262 个 YOLO TXT 标注文件（仅缺陷图）
│   ├── val/                   # 35 个 YOLO TXT 标注文件
│   └── test/                  # 80 个 YOLO TXT 标注文件
├── Annotations/               # Pascal VOC XML 标注 (1,318 个文件)
├── JPEGImages/                # 完整图片集合 (1,318 张图片)
├── classes.txt                # 5 个缺陷类别名称
├── data.yaml                  # Ultralytics YOLO 配置文件
├── _KeenForge_MagneticTile_compare.png  # 清洗与框提取效果对比图
└── README.md

关于标签文件。 仅 377 张缺陷图 有 YOLO .txt 标签（合计 377 个文件）。 941 张无缺陷 MT_Free 正常底板不附带标签文件 —— YOLO 中"缺少文件"与"空文件"等价（均表示无目标）， 且空文件会导致数据集索引失败。Annotations/ 仍为每张图保留 VOC XML（正常底板的 <object> 为空）。

快速开始 (Ultralytics YOLO)

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

引用

1. 原始数据集（请务必引用） —— 见上方英文部分 huang2018saliency。

2. 本修正版（请一并引用） —— 见上方英文部分 magnetictile_corrected。

3. 标注与质检工具（可选） —— 见上方英文部分 keenforge。

许可证

CC BY 4.0 —— 与上游 Magnetic Tile 数据集开源协议一致。请同时注明原作者与本整理版本。