NEU-DET-corrected

Hot-rolled steel strip surface defect detection — cleaned version of the NEU Surface Defect Database. 热轧带钢表面缺陷检测 —— NEU 表面缺陷数据库的清理修正版。

English · 中文

---

English

What is this?

A cleaned, complete and consistently packaged version of the NEU Surface Defect Database (NEU-DET) from Northeastern University (Song Kechen & Yan Yunhui).

The upstream release ships as a flat IMAGES/ + ANNOTATIONS/ (Pascal VOC) pair with no official split and a few annotation defects. This release keeps every image pixel unchanged, removes a small number of duplicates, and packages the data in both Pascal VOC and YOLO format with a reproducible train/val/test split.

⚠️ Upstream license: none stated. The original authors have never declared a license for NEU-DET. See License & attribution — verify before commercial use.

Corrections vs. the official release

#	Correction	Detail
1	Removed 1 exact-duplicate image	patches_105.jpg is pixel-identical to patches_101.jpg (the same image annotated twice, with slightly different boxes).
2	Removed 2 near-duplicate images	scratches_247.jpg is a 90°-rotated copy of scratches_232.jpg; pitted_surface_292.jpg is a 2-px-shifted copy of pitted_surface_171.jpg. Both show a sharp, isolated minimum in the alignment profile (random same-class pairs never drop below MAE ≈ 13).
3	Removed 3 duplicate boxes	crazing_120, inclusion_62, patches_198 each contained one bounding box listed twice verbatim.
4	Added a reproducible split	Upstream has no split. This release provides a seeded, class-stratified 80/10/10 split.
5	Dual format	Upstream is Pascal VOC only. This release ships VOC + YOLO from a single verified conversion.
6	Packaging	classes.txt, data.yaml, ImageSets/ added.

Image pixels are never modified. All 1,797 images are byte-for-byte identical to the upstream files.

Scope — structural clean-up only. This is not a re-annotation. We remove duplicates and re-package the data; we do not re-verify individual boxes (missed defects, wrong class labels, box tightness) — that requires domain expertise and is left untouched.

Structural checks performed on the released 1,797-image set:

Check	Method	Result
Exact duplicate images	byte-wise MD5	1 found → removed
Rotation / mirror duplicates	8 dihedral transforms, perceptual hash	1 found → removed
Shifted duplicates	aligned pixel MAE + sharp-minimum test	1 found → removed
Near-duplicate images	32×32/64×64 MAE + full aligned scan	0 remaining
Duplicate boxes within an image	verbatim box comparison	3 found → removed
Empty / invalid labels	0-box files, out-of-range, zero-area boxes	0
Image ↔ label ↔ XML pairing	bidirectional	0 orphans
Class-name consistency	unique label set	6 classes, no anomalies
Split leakage	train/val/test intersection	0

Dataset at a glance

Property	Value
Images	1,797 (200 × 200, grayscale)
Classes	6 — crazing, inclusion, patches, pitted_surface, rolled-in_scale, scratches
Bounding boxes	4,177 (≈ 2.32 per image)
Formats	Pascal VOC XML · YOLO TXT
Split	train 1,437 / val 180 / test 180 (seeded, class-stratified)

Boxes per class: inclusion 1,010 · patches 875 · crazing 688 · rolled-in_scale 628 · scratches 546 · pitted_surface 430

Structure

NEU-DET-corrected/
├── images/{train,val,test}/   # 200x200 grayscale JPG
├── labels/{train,val,test}/   # YOLO: cls cx cy w h (normalized)
├── annotations/               # Pascal VOC XML (flat, cleaned)
├── ImageSets/{train,val,test}.txt
├── classes.txt
├── data.yaml                  # YOLO dataset config
├── LICENSE
└── README.md

annotations/*.xml and labels/**/*.txt are two representations of the same cleaned annotations and are verified to agree.

Citation

1. The original dataset — please always cite this.

@article{he2020neudet,
  title   = {An End-to-end Steel Surface Defect Detection Approach via Fusing Multiple Hierarchical Features},
  author  = {He, Yu and Song, Kechen and Meng, Qinggang and Yan, Yunhui},
  journal = {IEEE Transactions on Instrumentation and Measurement},
  volume  = {69}, number = {4}, pages = {1493--1504}, year = {2020},
  doi     = {10.1109/TIM.2019.2915404}
}

@article{song2013neudet,
  title   = {A noise robust method based on completed local binary patterns for hot-rolled steel strip surface defects},
  author  = {Song, Kechen and Yan, Yunhui},
  journal = {Applied Surface Science},
  volume  = {285}, pages = {858--864}, year = {2013},
  doi     = {10.1016/j.apsusc.2013.08.123}
}

2. This corrected release — please cite it as well. It is not identical to the official release: 3 duplicate images (1 exact, 2 near-duplicates under rotation / shift) and 3 duplicate boxes were removed, a reproducible split was added, and the data was re-packaged, so a citation to the papers alone does not describe this version.

@misc{neudet_corrected,
  author    = {KeenForgeAI},
  title     = {NEU-DET-corrected: a cleaned release of the NEU Surface Defect Database},
  year      = {2026},
  version   = {1.0},
  publisher = {KeenForgeAI},
  url       = {https://huggingface.co/datasets/KeenForgeAI/NEU-DET-corrected},
  note      = {Curated by Lu Gan and Sam Li. Derived from He et al. (2020),
               doi:10.1109/TIM.2019.2915404. Upstream licence unstated.}
}

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 NEU-DET dataset states no license. The original authors provide it for research use via their official page (Google Drive / Baidu Netdisk) and only request a citation. The copyright status of the original images and annotations is therefore undetermined — verify it before any commercial use. See LICENSE in this repository.

Our modifications (de-duplication, re-packaging, split, documentation) are released openly. Image pixels are unmodified and remain subject to the upstream terms.

Provenance

• Source: official NEU-DET release (Google Drive), retrieved 2026-09-21.
• Upstream: 1,800 images + 1,800 Pascal VOC XML, no split.
• Verified: 1,797 images retained, byte-identical to upstream; 4,177 boxes after removing 3 duplicate boxes and 3 duplicate images (1 exact, 2 near-duplicates).

---

中文

这是什么？

NEU 表面缺陷数据库（NEU-DET）（东北大学，宋克臣、颜云辉）的清理、补全、规范化打包版。

上游发布是扁平的 IMAGES/ + ANNOTATIONS/（Pascal VOC），无官方划分，且存在少量标注缺陷。 本版不修改任何图像像素，仅删除少量重复项，并以 Pascal VOC + YOLO 双格式打包， 附带可复现的 train/val/test 划分。

⚠️ 上游未声明许可证。 原作者从未为 NEU-DET 声明许可证。详见 许可证与署名——商用前请自行确认。

相对官方版的修正

#	修正	说明
1	删除 1 张完全重复图	patches_105.jpg 与 patches_101.jpg 像素完全相同（同一张图被标注两次，框略有差异）。
2	删除 2 张近似重复图	scratches_247.jpg 是 scratches_232.jpg 旋转 90° 的副本；pitted_surface_292.jpg 是 pitted_surface_171.jpg 平移 2px 的副本。两者在对齐剖面中都呈尖锐唯一极小值（随机同类对从未低于 MAE ≈ 13）。
3	删除 3 个重复框	crazing_120、inclusion_62、patches_198 各有 1 个边界框逐字重复出现两次。
4	补充可复现划分	上游无划分；本版提供固定随机种子、按类别分层的 80/10/10 划分。
5	双格式	上游仅有 Pascal VOC；本版同时提供 VOC + YOLO（由同一次转换生成并校验一致）。
6	规范化打包	新增 classes.txt、data.yaml、ImageSets/。

图像像素从不修改——1,797 张图与上游文件逐字节一致。

本版范围——仅结构化清理。 这是结构性清理，不是重新标注。我们只去重、重新打包； 不重新核对单个标注框（漏标、类别标错、框松紧）——那需要领域专业知识，一律保持原样。

已执行的结构化检查（针对发布的 1,797 张）：

检查	方法	结果
样本重复 · 精确	逐字节 MD5	1 组 → 已删
样本重复 · 旋转/镜像	8 种二面体变换感知哈希	1 组 → 已删
样本重复 · 平移	对齐像素 MAE + 尖锐极小值判定	1 组 → 已删
样本重复 · 近似	32×32/64×64 MAE + 全库对齐扫描	0 剩余
标签重复 · 图内	逐字比对同一图内的框	3 个 → 已删
空/非法标签	0 框文件、越界框、零面积框	0
图片↔标签↔XML 配对	双向	0 孤儿
类别名一致性	唯一类别名集合	6 类，无杂名
划分泄漏	train/val/test 交集	0

数据集概览

属性	值
图片	1,797（200 × 200 灰度）
类别	6 —— crazing、inclusion、patches、pitted_surface、rolled-in_scale、scratches
标注框	4,177（约 2.32 框/图）
格式	Pascal VOC XML · YOLO TXT
划分	train 1,437 / val 180 / test 180（固定种子、类别分层）

各类框数：inclusion 1,010 · patches 875 · crazing 688 · rolled-in_scale 628 · scratches 546 · pitted_surface 430

目录结构

NEU-DET-corrected/
├── images/{train,val,test}/   # 200x200 灰度 JPG
├── labels/{train,val,test}/   # YOLO：cls cx cy w h（归一化）
├── annotations/               # Pascal VOC XML（扁平，已清理）
├── ImageSets/{train,val,test}.txt
├── classes.txt
├── data.yaml                  # YOLO 配置
├── LICENSE
└── README.md

annotations/*.xml 与 labels/**/*.txt 是同一份清理后标注的两种表示，已校验一致。

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（请务必引用）

@article{he2020neudet,
  title   = {An End-to-end Steel Surface Defect Detection Approach via Fusing Multiple Hierarchical Features},
  author  = {He, Yu and Song, Kechen and Meng, Qinggang and Yan, Yunhui},
  journal = {IEEE Transactions on Instrumentation and Measurement},
  volume  = {69}, number = {4}, pages = {1493--1504}, year = {2020},
  doi     = {10.1109/TIM.2019.2915404}
}

@article{song2013neudet,
  title   = {A noise robust method based on completed local binary patterns for hot-rolled steel strip surface defects},
  author  = {Song, Kechen and Yan, Yunhui},
  journal = {Applied Surface Science},
  volume  = {285}, pages = {858--864}, year = {2013},
  doi     = {10.1016/j.apsusc.2013.08.123}
}

2. 本修正版（请一并引用） —— 本版与官方发布并不相同：删除了 1 张重复图与 3 个重复框， 补充了可复现划分，并重新打包，因此只引用论文无法描述本版本。

@misc{neudet_corrected,
  author    = {KeenForgeAI},
  title     = {NEU-DET-corrected: a cleaned release of the NEU Surface Defect Database},
  year      = {2026},
  version   = {1.0},
  publisher = {KeenForgeAI},
  url       = {https://huggingface.co/datasets/KeenForgeAI/NEU-DET-corrected},
  note      = {Curated by Lu Gan and Sam Li. Derived from He et al. (2020),
               doi:10.1109/TIM.2019.2915404. Upstream licence unstated.}
}

3. 标注工具（可选）

@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.}
}

许可证与署名

上游 NEU-DET 未声明许可证。 原作者通过 官方主页（Google Drive / 百度网盘） 以研究用途提供，仅要求引用。因此原始图像与标注的版权状态未定——商用前请自行确认。 详见本仓库 LICENSE。

我们对数据集的修改部分（去重、重新打包、划分、文档）开源发布。图像像素未修改， 仍受上游条款约束。

来源

• 来源：官方 NEU-DET 发布（Google Drive），获取于 2026-09-21
• 上游：1,800 图 + 1,800 Pascal VOC XML，无划分
• 已校验：保留 1,797 图，与上游逐字节一致；删除 3 个重复框与 3 张重复图（1 张完全重复 + 2 张近似重复）后共 4,177 框