README.md

---
license: cc-by-sa-4.0
task_categories:
- image-classification
language:
- en
- zh
tags:
- ball-screw
- surface-defect
- defect-classification
- industrial-inspection
- quality-control
- manufacturing
- image-classification
- keenforge
size_categories:
- 10K<n<100K
---

# KIT_BallScrew_Defects-corrected

[![DOI](https://img.shields.io/badge/DOI-10.57967%2Fhf%2F11010-blue)](https://doi.org/10.57967/hf/11010)

**Ball-screw surface defect **classification** — cleaned, re-indexed, split and standardized version of the KIT-BS-recls dataset.**
**滚珠丝杠表面缺陷**分类** —— KIT-BS-recls 数据集的清理、重编号、划分与规整版。**

[English](#english) · [中文](#中文)

---

## English

### What is this?

A **cleaned, re-indexed, split and standardized image-classification version** of the
**Tri-classified Dataset for Ball Screw Surface Defects (KIT-BS-recls)** — three surface
conditions: **pitting**, **contaminated** and **normal**. This is a **classification**
dataset (one label per image, **no bounding boxes**).

Reference: Xue, F., Zhang, W., Xie, S., Puchta, A. & Fleischer, J. *"Domain-unique
group- and its subgroups-aware fault diagnostics for machine components: An open-set
domain adaptation approach."* **Neurocomputing**, 2025 — https://doi.org/10.1016/j.neucom.2025.131291.
Upstream: [fenghsueh/Triclass-BSDefects](https://github.com/fenghsueh/Triclass-BSDefects) (also on Kaggle).

The original release ships as three class folders plus a single `labels.txt` that maps
each path to a label, where the labels are the **non-contiguous ids 3 / 6 / 8**, with
**no train/val/test split** and filenames containing parentheses (`P(123).png`).

This release converts the archive into a folder-based classification layout, **remaps the
labels to 0 / 1 / 2**, adds a reproducible stratified split and normalises the filenames.

> ✅ **Upstream license: CC BY-SA 4.0** (Attribution-ShareAlike). This is a real open
> licence, but it is **ShareAlike** — any derivative must be released under the same
> licence. See [License & attribution](#license--attribution).

### Corrections vs. the official release

| # | Correction | Detail |
|---|---|---|
| 1 | **Classification layout** | Converted the archive into `images/<split>/<class>/` (Ultralytics-style folder classification). The redundant `labels.txt` is replaced by `labels.csv` and the folder structure. |
| 2 | **Re-indexed labels** ⭐ | The upstream ids are the **non-contiguous 3 / 6 / 8** (pitting / contaminated / normal). They are remapped to the contiguous **0 / 1 / 2**, so a model's output layer has no wasted units. |
| 3 | **Reproducible split** | Upstream ships **no split**. This release provides a seeded, class-stratified **70 / 15 / 15** split (seed 42): **15,288 / 3,276 / 3,271**. |
| 4 | **Near-duplicates kept together** | Near-identical images (identical 8×8-dHash and 32×32 pixel-MSE ≤ 1) are grouped and assigned entirely to one split (50 such groups). |
| 5 | **Filename hygiene** | Filenames contained parentheses, e.g. `P(123).png` → `P_123.png`; `P(1000)_.png` → `P_1000_.png`. |
| 6 | **Packaging** | Added `classes.txt`, `data.yaml` (Ultralytics classification) and `labels.csv` (filename, split, class, index, original name/dir). |

**Image pixels are never modified.** All 21,835 images are byte-for-byte identical to upstream (150×150 RGB `.png`; 9 images differ slightly in size — see note).

**Scope — structural clean-up only.** This is **not** a re-annotation; each image keeps its original single class label.

**Structural checks performed** on the released set:

| Check | Method | Result |
|---|---|---|
| `labels.txt` ↔ class folder consistency | per line | **0** mismatches |
| Files in `samples/` missing from `labels.txt` | set diff | **0** |
| Exact duplicate images | md5 | **0** |
| Cross-split near-identical leakage | cluster assignment | 0 groups straddle a split |
| Image format / channels | scan | RGB PNG, 150×150 (9 images slightly off-size) |

### Dataset at a glance

| Property | Value |
|---|---|
| Images | **21,835** (150 × 150 RGB `.png`) |
| Task | **Image classification** (single label per image) |
| Classes | **3** — `pitting` (0), `contaminated` (1), `normal` (2) |
| Split | train **15,288** / val **3,276** / test **3,271** (seeded, class-stratified) |

**Images per class:** `pitting` 10,760 · `contaminated` 7,523 · `normal` 3,552

**Per split:** train (`pitting` 7,532 / `contaminated` 5,267 / `normal` 2,489) ·
val (1,614 / 1,130 / 532) · test (1,614 / 1,126 / 531)

> **Note:** 9 of the 21,835 images are not exactly 150×150 (e.g. 146×150, 127×150). Their
> pixels are left unmodified; resize on load if a fixed input size is required.

### Structure

```
KIT_BallScrew_Defects-corrected/
├── images/
│   ├── train/{pitting,contaminated,normal}/   # 150x150 RGB .png
│   ├── val/{pitting,contaminated,normal}/
│   └── test/{pitting,contaminated,normal}/
├── classes.txt     # pitting / contaminated / normal (one per line, index order)
├── data.yaml       # Ultralytics classification configuration
├── labels.csv      # filename,split,class,class_index,original_name,original_dir
├── LICENSE         # CC BY-SA 4.0 (full text + notice of changes)
└── README.md
```

### Quick start (Ultralytics)

```python
from ultralytics import YOLO

model = YOLO("yolo11n-cls.pt")
model.train(data=".", epochs=50, imgsz=150)
```

### Citation

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

```bibtex
@article{xue2025domaingroup,
  title   = {Domain-unique group- and its subgroups-aware fault diagnostics for machine
             components: An open-set domain adaptation approach},
  author  = {Xue, Feng and Zhang, Weimin and Xie, Shulian and Puchta, Alexander and
             Fleischer, J{\"u}rgen},
  journal = {Neurocomputing},
  year    = {2025},
  doi     = {10.1016/j.neucom.2025.131291}
}
```

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

```bibtex
@misc{kit_ballscrew_corrected,
  author    = {KeenForgeAI},
  title     = {KIT_BallScrew_Defects-corrected: a cleaned classification release of the Tri-classified Ball Screw Surface Defects dataset (KIT-BS-recls)},
  year      = {2026},
  version   = {1.0},
  publisher = {KeenForgeAI},
  url       = {https://huggingface.co/datasets/KeenForgeAI/KIT_BallScrew_Defects-corrected},
  note      = {Curated by Lu Gan and Sam Li. Derived from KIT-BS-recls
               (Xue et al., Neurocomputing 2025), licensed CC BY-SA 4.0.}
}
```

**3. The annotation tool (optional).**

```bibtex
@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

**CC BY-SA 4.0 (Attribution-ShareAlike 4.0 International).** The upstream KIT-BS-recls is
licensed under CC BY-SA 4.0. Because ShareAlike is a copyleft condition, **this corrected
release is distributed under the same CC BY-SA 4.0 licence** — you must credit the original
authors, indicate if changes were made, and license any derivative under CC BY-SA 4.0. The
full licence text and a notice of the changes are in `LICENSE`.

Original work: Xue, Zhang, Xie, Puchta & Fleischer — https://github.com/fenghsueh/Triclass-BSDefects.

### Provenance

- Source: Tri-classified Dataset for Ball Screw Surface Defects (KIT-BS-recls), GitHub `fenghsueh/Triclass-BSDefects`.
- Upstream: 21,835 images / 3 classes (ids 3/6/8); no split; filenames with parentheses; 0 duplicate images.
- Verified: 21,835 images (byte-identical); labels remapped 3/6/8→0/1/2; split 15288/3276/3271; 0 label/folder mismatches, 0 cross-split near-identical leakage.

---

## 中文

### 这是什么？

**滚珠丝杠表面缺陷三分类数据集（KIT-BS-recls）**的**清理、重编号、划分与标准化图像分类版**——
三种表面状态：**点蚀（pitting）**、**污染（contaminated）**、**正常（normal）**。这是
**分类**数据集（每图一个标签，**无边界框**）。

论文：Xue, F., Zhang, W., Xie, S., Puchta, A. & Fleischer, J.《Domain-unique group- and its
subgroups-aware fault diagnostics for machine components: An open-set domain adaptation
approach》，**Neurocomputing**，2025 —— https://doi.org/10.1016/j.neucom.2025.131291。
上游：[fenghsueh/Triclass-BSDefects](https://github.com/fenghsueh/Triclass-BSDefects)（Kaggle 亦有）。

原始版本以三个类别文件夹 + 单个 `labels.txt`（每行"路径 + 标签"）发布，标签为**不连续的
3 / 6 / 8**，**无 train/val/test 划分**，且文件名含括号（`P(123).png`）。

本版本将其转换为基于文件夹的分类布局、把标签**重映射为 0 / 1 / 2**、补充可复现的分层划分，
并净化文件名。

> ✅ **上游许可证：CC BY-SA 4.0**（署名—相同方式共享）。这是真正的开源许可，但为 **ShareAlike
> （传染性）**——任何衍生作品必须以同一许可发布。详见 [许可证与署名](#许可证与署名)。

### 相对官方版的修正

| # | 修正 | 说明 |
|---|---|---|
| 1 | **分类布局** | 转换为 `images/<split>/<class>/`（Ultralytics 风格的文件夹分类）。冗余的 `labels.txt` 由 `labels.csv` 与目录结构取代。 |
| 2 | **标签重编号** ⭐ | 上游标签为**不连续的 3 / 6 / 8**（pitting / contaminated / normal），已重映射为连续的 **0 / 1 / 2**，避免模型输出层出现空洞。 |
| 3 | **可复现划分** | 上游**完全没有划分**。本版提供固定种子、按类别分层的 **70 / 15 / 15** 划分（随机种子 42）：**15,288 / 3,276 / 3,271**。 |
| 4 | **近重复固定同侧** | 近重复图（8×8 dHash 相同且 32×32 像素 MSE ≤ 1）归组并整体分配到一个划分（共 50 组）。 |
| 5 | **文件名净化** | 文件名含括号，如 `P(123).png` → `P_123.png`；`P(1000)_.png` → `P_1000_.png`。 |
| 6 | **规范化打包** | 新增 `classes.txt`、`data.yaml`（Ultralytics 分类配置）与 `labels.csv`（文件名、划分、类别、索引、原始名/目录）。 |

**图像像素从不修改**——全部 21,835 张图与上游文件逐字节一致（150×150 RGB `.png`；9 张尺寸略有不同，见说明）。

**本版范围——仅结构化清理**，**不是重新标注**；每张图保留其原始单标签。

**已执行的结构化检查**：`labels.txt`↔类别目录一致（0 错配）/ `samples/` 无遗漏文件 /
0 完全重复 / 0 近重复簇跨越划分 / 全部 RGB PNG。

### 数据集概览

| 属性 | 值 |
|---|---|
| 图片 | **21,835**（150 × 150 RGB `.png`） |
| 任务 | **图像分类**（每图单一标签） |
| 类别 | **3** —— `pitting` (0)、`contaminated` (1)、`normal` (2) |
| 划分 | train **15,288** / val **3,276** / test **3,271**（固定种子、类别分层） |

**各类图像数：** `pitting` 10,760 · `contaminated` 7,523 · `normal` 3,552

**各划分：** train（`pitting` 7,532 / `contaminated` 5,267 / `normal` 2,489）·
val（1,614 / 1,130 / 532）· test（1,614 / 1,126 / 531）

> **说明：** 21,835 张中有 9 张并非精确 150×150（如 146×150、127×150）。其像素保持原样；
> 若需固定输入尺寸，请在加载时缩放。

### 目录结构

```
KIT_BallScrew_Defects-corrected/
├── images/
│   ├── train/{pitting,contaminated,normal}/   # 150x150 RGB .png
│   ├── val/{pitting,contaminated,normal}/
│   └── test/{pitting,contaminated,normal}/
├── classes.txt     # pitting / contaminated / normal（每行一个，按索引顺序）
├── data.yaml       # Ultralytics 分类配置
├── labels.csv      # filename,split,class,class_index,original_name,original_dir
├── LICENSE         # CC BY-SA 4.0（全文 + 修改说明）
└── README.md
```

### 快速开始 (Ultralytics)

```python
from ultralytics import YOLO

model = YOLO("yolo11n-cls.pt")
model.train(data=".", epochs=50, imgsz=150)
```

### 引用

请**同时引用**原始论文与本修正版：**1.** 原始论文见上方英文 `xue2025domaingroup`；
**2.** 本修正版见上方英文 `kit_ballscrew_corrected`；**3.** 标注工具（可选）见 `keenforge`。

### 许可证与署名

**CC BY-SA 4.0（署名—相同方式共享 4.0 国际）。** 上游 KIT-BS-recls 采用 CC BY-SA 4.0。
由于 ShareAlike 是传染性条款，**本修正版以同一 CC BY-SA 4.0 许可发布**——你必须署名原作者、
注明是否修改，并以 CC BY-SA 4.0 发布任何衍生作品。完整许可文本与修改说明见 `LICENSE`。

原始作品：Xue, Zhang, Xie, Puchta & Fleischer —— https://github.com/fenghsueh/Triclass-BSDefects。

### 来源

- 来源：滚珠丝杠表面缺陷三分类数据集（KIT-BS-recls），GitHub `fenghsueh/Triclass-BSDefects`
- 上游：21,835 图 / 3 类（标签 id 3/6/8）；无划分；文件名含括号；0 完全重复
- 已校验：21,835 图（逐字节一致）；标签 3/6/8→0/1/2；划分 15288/3276/3271；0 标签/目录错配、0 跨划分近重复泄漏