Datasets:

KeenForgeAI
/

CPLID_Insulators-corrected

---
license: other
license_name: see-license-file-upstream-unstated
license_link: LICENSE
task_categories:
- object-detection
language:
- en
- zh
tags:
- power-line
- insulator
- insulator-defect
- uav
- aerial-imagery
- electrical-inspection
- cpld
- object-detection
- pascal-voc
- yolo
- keenforge
size_categories:
- 1K<n<10K
---
# CPLID_Insulators-corrected
[![DOI](https://img.shields.io/badge/DOI-10.57967%2Fhf%2F10878-blue)](https://doi.org/10.57967/hf/10878)
**Power-line insulator defect detection — cleaned, split and standardized version of the Chinese Power Line Insulator Dataset (CPLID).**
**电力巡线绝缘子缺陷检测 —— 中国电力线绝缘子数据集（CPLID）的清理、划分与规整版。**
[English](#english) · [中文](#中文)
---
## English
### What is this?
A **cleaned, split and standardized object-detection version** of the
**Chinese Power Line Insulator Dataset (CPLID)** — normal insulators captured by
UAVs plus synthetically defective insulators. Images were provided by the
**State Grid Corporation of China**; the dataset was assembled by **WANG Zi-Hao**.
Reference: Tao, Zhang, Wang, Liu, Zhang & Xu — *"Detection of Power Line Insulator
Defects Using Aerial Images Analyzed With Convolutional Neural Networks"*,
**IEEE Transactions on Systems, Man, and Cybernetics: Systems**, 2018.
Upstream repository: [InsulatorData/InsulatorDataSet](https://github.com/InsulatorData/InsulatorDataSet).
The original release ships as two flat, unsplit folders (`Normal_Insulators` and
`Defective_Insulators`), where each defective image is annotated **twice** (once
for the defect, once for the whole insulator) in two parallel label folders, the
VOC metadata is polluted with hard-coded personal paths, a stray non-image file
sits in the archive, and 3 images are **pixel-identical duplicates**.

This release removes the duplicates, keeps consecutive UAV-capture frames on one
side of the split, merges the paired label folders into a single annotation per
image, cleans the VOC metadata, drops the stray file, provides a reproducible
**train/val/test** split, and ships ready-to-train **YOLO TXT** + **Pascal VOC XML**
annotations with configuration files.

> ⚠️ **Upstream license: none stated.** CPLID ships no LICENSE file and no terms of use.
> See [License & attribution](#license--attribution) — verify before commercial use.

### Corrections vs. the official release

| # | Correction | Detail |
|---|---|---|
| 1 | **Removed 3 duplicate images** | Three pairs in `Normal_Insulators` are **pixel-identical** (`0150` = `2984`, `0852` = `1876`, `1608` = `1938`). The higher-numbered copy of each pair was removed, so identical content can never straddle the new split. |
| 2 | **Consecutive frames kept together** | The 600 normal images are UAV video frames; 112 of them sit next to a numerically consecutive neighbour. Frames are grouped into **486 capture chains** (longest: 5 frames) and each chain is assigned **entirely** to one split, so no near-identical burst is split across train/val/test. |
| 3 | **Merged the paired label folders** | Upstream annotates every defective image twice in parallel folders (`labels/defect/` + `labels/insulator/`). Here the two are merged into **one XML / one YOLO file per image**, containing the insulator box and the defect box together. |
| 4 | **Reproducible train/val/test split** | Upstream ships **no split at all**. This release provides a seeded, class-stratified **70 / 15 / 15** split (seed 42) over the frame-chain groups. |
| 5 | **VOC metadata clean-up** | The upstream XML carried a bogus `<folder>` value (a mix of `images`, `long` and empty) and a hard-coded local `<path>` exposing personal absolute paths with Chinese characters (e.g. `/media/laptop/vbox/硕士毕业设计/小论文/DataSet/...`). Both are normalised, and `<source><database>` is set to `CPLID`. |
| 6 | **Removed a stray file** | `OriginalImageOnWebDriver` is a 48-byte non-image text file containing only a Baidu-pan URL. It was removed. |
| 7 | **Standard packaging** | Added `data.yaml`, `classes.txt`, synchronized dual-format annotations (`Annotations/` + `labels/`) and refreshed per-split CSV tables. |

**Image pixels are never modified.** All 845 image files are byte-for-byte identical to the upstream files.

**Scope — structural clean-up only.** This is **not** a re-annotation. We separate,
de-duplicate and re-package the data; we do **not** re-verify individual boxes.
In particular, the **defective** images are **synthetic** — the original authors
generated them with the TVSeg + affine-augmentation + U-Net pipeline described
upstream and pasted the defects onto different backgrounds.

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

| Check | Method | Result |
|---|---|---|
| Filename ↔ image ↔ XML pairing | bidirectional | **0** mismatches / orphans |
| Image size vs. XML `<size>` | pixel dimensions vs. header | **0** discrepancies (all 1152×864) |
| Out-of-range boxes | xmin/ymin ≥ 0, xmax ≤ W, ymax ≤ H | **0** |
| Degenerate boxes | zero / negative width or height | **0** |
| Cross-split duplicate images | byte-identical md5 across splits | **0** |
| Consecutive-frame leakage | +1 adjacency across splits | **0** |
| Class-name consistency | unique label set | 2 classes, no anomalies |

### Dataset at a glance

| Property | Value |
|---|---|
| Images | **845** (1152 × 864, RGB `.jpg`) |
| Classes | **2** — `insulator`, `defect` |
| Bounding boxes | **1,566** |
| Formats | Pascal VOC XML · YOLO TXT |
| Split | train **590** / val **129** / test **126** (seeded, class-stratified) |

**Boxes per class:** `insulator` 1,318 · `defect` 248

**Boxes per split:** train **1,104** (`insulator` 931 / `defect` 173) ·
val **237** (200 / 37) · test **225** (187 / 38)

### Structure

```
CPLID_Insulators-corrected/
├── images/{train,val,test}/      # 1152x864 RGB .jpg
├── labels/{train,val,test}/      # YOLO: cls cx cy w h (normalized)
├── Annotations/                  # Pascal VOC XML (flat, cleaned)
├── 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                   # 2 class names, one per line
├── data.yaml                     # Ultralytics YOLO configuration
├── LICENSE
└── README.md
```

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

### Quick start (Ultralytics YOLO)

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

### Citation

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

```bibtex
@article{tao2018detection,
  title   = {Detection of Power Line Insulator Defects Using Aerial Images
             Analyzed With Convolutional Neural Networks},
  author  = {Tao, Xian and Zhang, Dapeng and Wang, Zihao and Liu, Xilong
             and Zhang, Hongyan and Xu, De},
  journal = {IEEE Transactions on Systems, Man, and Cybernetics: Systems},
  year    = {2018},
  publisher = {IEEE}
}
```

**2. This corrected release — please cite it as well.** It is not identical to the official
release: duplicates were removed, a reproducible split was added, the paired labels were
merged, and the VOC metadata was cleaned, so a citation to the original alone does not
describe this version.

```bibtex
@misc{cpld_insulators_corrected,
  author    = {KeenForgeAI},
  title     = {CPLID_Insulators-corrected: a cleaned release of the Chinese Power Line Insulator Dataset (CPLID)},
  year      = {2026},
  version   = {1.0},
  publisher = {KeenForgeAI},
  url       = {https://huggingface.co/datasets/KeenForgeAI/CPLID_Insulators-corrected},
  note      = {Curated by Lu Gan and Sam Li. Derived from the Chinese Power Line
               Insulator Dataset (CPLID; images from the State Grid Corporation of
               China, assembled by WANG Zi-Hao). Upstream licence unstated.}
}
```

**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

**The upstream CPLID dataset states no license.** The repository ships no LICENSE file and
no terms of use. The normal images are credited to the State Grid Corporation of China and
the defective images are synthetic; the copyright status of the underlying images and
annotations is therefore **undetermined** — verify it before any commercial use.
See `LICENSE` in this repository.

Our modifications (de-duplication, split, label merging, format conversion, documentation)
are released under **CC BY 4.0**. Image pixels are unmodified and remain subject to the
upstream terms.

### Provenance

- Source: Chinese Power Line Insulator Dataset (CPLID), GitHub `InsulatorData/InsulatorDataSet`.
- Upstream: `Normal_Insulators` 600 img / 1,073 box · `Defective_Insulators` 248 img / 496 box (248 defect + 248 insulator).
- Verified: 845 images (1,566 boxes, byte-identical); 3 duplicates removed, 486 frame chains, 0 cross-split leakage, 0 invalid boxes.

---

## 中文

### 这是什么？

**中国电力线绝缘子数据集（CPLID）**的**清理、划分与标准化目标检测版**——包含无人机拍摄的
正常绝缘子图像，以及合成生成的缺陷绝缘子图像。图像由**国家电网公司**提供，数据集由
**王自浩（WANG Zi-Hao）** 整理。

论文：Tao, Zhang, Wang, Liu, Zhang & Xu，*《利用卷积神经网络分析航拍图像检测电力线绝缘子缺陷》*，
**IEEE Transactions on Systems, Man, and Cybernetics: Systems**，2018。
上游仓库：[InsulatorData/InsulatorDataSet](https://github.com/InsulatorData/InsulatorDataSet)。

原始版本以两个扁平、无划分的目录发布（`Normal_Insulators` 与 `Defective_Insulators`），其中
每张缺陷图被**标注了两次**（一次为缺陷，一次为整个绝缘子）并分别存放在两个平行的标签目录中；
VOC 元数据混入硬编码的个人绝对路径；压缩包内还留有一个垃圾非图片文件；并有 **3 张像素级完全重复**的图像。

本版本删除重复图、将同一段无人机连拍帧固定在同一划分内、把成对的标签目录合并为每图一份标注、
清理 VOC 元数据、剔除垃圾文件，提供可复现的**训练/验证/测试**划分，并附带开箱即用的
**YOLO TXT** 与 **Pascal VOC XML** 双格式标注及训练配置。

> ⚠️ **上游未声明许可证。** CPLID 未附带 LICENSE 文件，也无使用条款。详见
> [许可证与署名](#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 3 张重复图片** | `Normal_Insulators` 中有三对图像**像素级完全相同**（`0150` = `2984`、`0852` = `1876`、`1608` = `1938`）。本版删除每对中编号较大的一张，确保相同内容不会跨越新划分。 |
| 2 | **连拍帧固定同侧** | 600 张正常图均为无人机视频帧，其中 112 张与相邻帧编号连续。本版将其归并为 **486 个连拍链**（最长 5 帧），整条链**完整**分配到一个划分，避免近似重复帧被切散到 train/val/test。 |
| 3 | **合并成对标签目录** | 上游把每张缺陷图在两个平行目录中标注两次（`labels/defect/` + `labels/insulator/`）。本版合并为**每图一份 XML / 一份 YOLO**，同时包含绝缘子框与缺陷框。 |
| 4 | **可复现的训练/验证/测试划分** | 上游**完全没有划分**。本版基于连拍链分组，提供固定随机种子、按类别分层的 **70 / 15 / 15** 划分（随机种子 42）。 |
| 5 | **VOC 元数据清理** | 上游 XML 带有错误的 `<folder>` 值（`images`、`long`、空 混杂）以及硬编码的本地 `<path>`，暴露了含中文的个人绝对路径（如 `/media/laptop/vbox/硕士毕业设计/小论文/DataSet/...`）。均归一化，且 `<source><database>` 统一为 `CPLID`。 |
| 6 | **删除垃圾文件** | `OriginalImageOnWebDriver` 是一个 48 字节的非图片文本文件，内容仅为一个百度网盘链接，已删除。 |
| 7 | **规范化打包** | 新增 `data.yaml`、`classes.txt`、双格式同步标注（`Annotations/` + `labels/`）以及按划分刷新的 CSV 表格。 |

**图像像素从不修改**——全部 845 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**。我们只做分离、去重、重新打包；
**不**重新核对标注框。特别说明：**缺陷图是合成的**——由原作者按上游所述的
TVSeg + 仿射增强 + U-Net 流程生成并贴到不同背景上。

**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↔XML 配对 | 双向 | **0** 错配 / 孤儿 |
| 图片尺寸 vs. XML `<size>` | 像素尺寸 vs. 头信息 | **0** 不一致（全部 1152×864） |
| 越界框 | xmin/ymin ≥ 0，xmax ≤ W，ymax ≤ H | **0** |
| 退化框 | 宽或高为零/负 | **0** |
| 跨划分重复图 | 跨划分逐字节 md5 | **0** |
| 连拍帧泄漏 | 跨划分 +1 邻接 | **0** |
| 类别名一致性 | 唯一类别名集合 | 2 类，无杂名 |

### 数据集概览

| 属性 | 值 |
|---|---|
| 图片 | **845**（1152 × 864，RGB `.jpg`） |
| 类别 | **2** —— `insulator`、`defect` |
| 标注框 | **1,566** |
| 格式 | Pascal VOC XML · YOLO TXT |
| 划分 | train **590** / val **129** / test **126**（固定种子、类别分层） |

**各类框数：** `insulator` 1,318 · `defect` 248

**各划分框数：** train **1,104**（`insulator` 931 / `defect` 173）·
val **237**（200 / 37）· test **225**（187 / 38）

### 目录结构

```
CPLID_Insulators-corrected/
├── images/{train,val,test}/      # 1152x864 RGB .jpg
├── labels/{train,val,test}/      # YOLO：cls cx cy w h（归一化）
├── Annotations/                  # Pascal VOC XML（扁平，已清理）
├── 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                   # 2 个类别名称，每行一个
├── data.yaml                     # Ultralytics YOLO 配置文件
├── LICENSE
└── 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)

```bash
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（请务必引用）** —— 见上方英文部分 `tao2018detection`。

**2. 本修正版（请一并引用）** —— 本版**与官方发布并不相同**：删除了重复图、补充了可复现划分、
合并了成对标签并清理了 VOC 元数据，因此**只引用原数据集无法描述本版本**。见上方英文部分
`cpld_insulators_corrected`。

**3. 标注工具（可选）** —— 见上方英文部分 `keenforge`。

### 许可证与署名

**上游 CPLID 数据集未声明许可证。** 仓库内无 LICENSE 文件、无使用条款。正常图由国家电网公司
提供，缺陷图为合成图像；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、划分、标签合并、格式转换、文档）以 **CC BY 4.0** 发布。
图像像素未修改，仍受上游条款约束。

### 来源

- 来源：中国电力线绝缘子数据集（CPLID），GitHub `InsulatorData/InsulatorDataSet`
- 上游：`Normal_Insulators` 600 图 / 1,073 框 · `Defective_Insulators` 248 图 / 496 框（248 缺陷 + 248 绝缘子）
- 已校验：845 图（1,566 框，逐字节一致）；删除 3 张重复图、486 个连拍链、0 跨划分泄漏、0 非法框

	---
	license: other
	license_name: see-license-file-upstream-unstated
	license_link: LICENSE
	task_categories:
	- object-detection
	language:
	- en
	- zh
	tags:
	- power-line
	- insulator
	- insulator-defect
	- uav
	- aerial-imagery
	- electrical-inspection
	- cpld
	- object-detection
	- pascal-voc
	- yolo
	- keenforge
	size_categories:
	- 1K<n<10K
	---

	# CPLID_Insulators-corrected

	[![DOI](https://img.shields.io/badge/DOI-10.57967%2Fhf%2F10878-blue)](https://doi.org/10.57967/hf/10878)

	Power-line insulator defect detection — cleaned, split and standardized version of the Chinese Power Line Insulator Dataset (CPLID).
	电力巡线绝缘子缺陷检测 —— 中国电力线绝缘子数据集（CPLID）的清理、划分与规整版。

	[English](#english) · [中文](#中文)

	---

	## English

	### What is this?

	A cleaned, split and standardized object-detection version of the
	Chinese Power Line Insulator Dataset (CPLID) — normal insulators captured by
	UAVs plus synthetically defective insulators. Images were provided by the
	State Grid Corporation of China; the dataset was assembled by WANG Zi-Hao.

	Reference: Tao, Zhang, Wang, Liu, Zhang & Xu — *"Detection of Power Line Insulator
	Defects Using Aerial Images Analyzed With Convolutional Neural Networks"*,
	IEEE Transactions on Systems, Man, and Cybernetics: Systems, 2018.
	Upstream repository: [InsulatorData/InsulatorDataSet](https://github.com/InsulatorData/InsulatorDataSet).

	The original release ships as two flat, unsplit folders (`Normal_Insulators` and
	`Defective_Insulators`), where each defective image is annotated twice (once
	for the defect, once for the whole insulator) in two parallel label folders, the
	VOC metadata is polluted with hard-coded personal paths, a stray non-image file
	sits in the archive, and 3 images are pixel-identical duplicates.

	This release removes the duplicates, keeps consecutive UAV-capture frames on one
	side of the split, merges the paired label folders into a single annotation per
	image, cleans the VOC metadata, drops the stray file, provides a reproducible
	train/val/test split, and ships ready-to-train YOLO TXT + Pascal VOC XML
	annotations with configuration files.

	> ⚠️ Upstream license: none stated. CPLID ships no LICENSE file and no terms of use.
	> See [License & attribution](#license--attribution) — verify before commercial use.

	### Corrections vs. the official release

	\| # \| Correction \| Detail \|
	\|---\|---\|---\|
	\| 1 \| Removed 3 duplicate images \| Three pairs in `Normal_Insulators` are pixel-identical (`0150` = `2984`, `0852` = `1876`, `1608` = `1938`). The higher-numbered copy of each pair was removed, so identical content can never straddle the new split. \|
	\| 2 \| Consecutive frames kept together \| The 600 normal images are UAV video frames; 112 of them sit next to a numerically consecutive neighbour. Frames are grouped into 486 capture chains (longest: 5 frames) and each chain is assigned entirely to one split, so no near-identical burst is split across train/val/test. \|
	\| 3 \| Merged the paired label folders \| Upstream annotates every defective image twice in parallel folders (`labels/defect/` + `labels/insulator/`). Here the two are merged into one XML / one YOLO file per image, containing the insulator box and the defect box together. \|
	\| 4 \| Reproducible train/val/test split \| Upstream ships no split at all. This release provides a seeded, class-stratified 70 / 15 / 15 split (seed 42) over the frame-chain groups. \|
	\| 5 \| VOC metadata clean-up \| The upstream XML carried a bogus `<folder>` value (a mix of `images`, `long` and empty) and a hard-coded local `<path>` exposing personal absolute paths with Chinese characters (e.g. `/media/laptop/vbox/硕士毕业设计/小论文/DataSet/...`). Both are normalised, and `<source><database>` is set to `CPLID`. \|
	\| 6 \| Removed a stray file \| `OriginalImageOnWebDriver` is a 48-byte non-image text file containing only a Baidu-pan URL. It was removed. \|
	\| 7 \| Standard packaging \| Added `data.yaml`, `classes.txt`, synchronized dual-format annotations (`Annotations/` + `labels/`) and refreshed per-split CSV tables. \|

	Image pixels are never modified. All 845 image files are byte-for-byte identical to the upstream files.

	Scope — structural clean-up only. This is not a re-annotation. We separate,
	de-duplicate and re-package the data; we do not re-verify individual boxes.
	In particular, the defective images are synthetic — the original authors
	generated them with the TVSeg + affine-augmentation + U-Net pipeline described
	upstream and pasted the defects onto different backgrounds.

	Structural checks performed on the released set:

	\| Check \| Method \| Result \|
	\|---\|---\|---\|
	\| Filename ↔ image ↔ XML pairing \| bidirectional \| 0 mismatches / orphans \|
	\| Image size vs. XML `<size>` \| pixel dimensions vs. header \| 0 discrepancies (all 1152×864) \|
	\| Out-of-range boxes \| xmin/ymin ≥ 0, xmax ≤ W, ymax ≤ H \| 0 \|
	\| Degenerate boxes \| zero / negative width or height \| 0 \|
	\| Cross-split duplicate images \| byte-identical md5 across splits \| 0 \|
	\| Consecutive-frame leakage \| +1 adjacency across splits \| 0 \|
	\| Class-name consistency \| unique label set \| 2 classes, no anomalies \|

	### Dataset at a glance

	\| Property \| Value \|
	\|---\|---\|
	\| Images \| 845 (1152 × 864, RGB `.jpg`) \|
	\| Classes \| 2 — `insulator`, `defect` \|
	\| Bounding boxes \| 1,566 \|
	\| Formats \| Pascal VOC XML · YOLO TXT \|
	\| Split \| train 590 / val 129 / test 126 (seeded, class-stratified) \|

	Boxes per class: `insulator` 1,318 · `defect` 248

	Boxes per split: train 1,104 (`insulator` 931 / `defect` 173) ·
	val 237 (200 / 37) · test 225 (187 / 38)

	### Structure

	```
	CPLID_Insulators-corrected/
	├── images/{train,val,test}/ # 1152x864 RGB .jpg
	├── labels/{train,val,test}/ # YOLO: cls cx cy w h (normalized)
	├── Annotations/ # Pascal VOC XML (flat, cleaned)
	├── 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 # 2 class names, one per line
	├── data.yaml # Ultralytics YOLO configuration
	├── LICENSE
	└── README.md
	```

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

	### Quick start (Ultralytics YOLO)

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

	### Citation

	1. The original dataset — please always cite this.

	```bibtex
	@article{tao2018detection,
	title = {Detection of Power Line Insulator Defects Using Aerial Images
	Analyzed With Convolutional Neural Networks},
	author = {Tao, Xian and Zhang, Dapeng and Wang, Zihao and Liu, Xilong
	and Zhang, Hongyan and Xu, De},
	journal = {IEEE Transactions on Systems, Man, and Cybernetics: Systems},
	year = {2018},
	publisher = {IEEE}
	}
	```

	2. This corrected release — please cite it as well. It is not identical to the official
	release: duplicates were removed, a reproducible split was added, the paired labels were
	merged, and the VOC metadata was cleaned, so a citation to the original alone does not
	describe this version.

	```bibtex
	@misc{cpld_insulators_corrected,
	author = {KeenForgeAI},
	title = {CPLID_Insulators-corrected: a cleaned release of the Chinese Power Line Insulator Dataset (CPLID)},
	year = {2026},
	version = {1.0},
	publisher = {KeenForgeAI},
	url = {https://huggingface.co/datasets/KeenForgeAI/CPLID_Insulators-corrected},
	note = {Curated by Lu Gan and Sam Li. Derived from the Chinese Power Line
	Insulator Dataset (CPLID; images from the State Grid Corporation of
	China, assembled by WANG Zi-Hao). Upstream licence unstated.}
	}
	```

	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

	The upstream CPLID dataset states no license. The repository ships no LICENSE file and
	no terms of use. The normal images are credited to the State Grid Corporation of China and
	the defective images are synthetic; the copyright status of the underlying images and
	annotations is therefore undetermined — verify it before any commercial use.
	See `LICENSE` in this repository.

	Our modifications (de-duplication, split, label merging, format conversion, documentation)
	are released under CC BY 4.0. Image pixels are unmodified and remain subject to the
	upstream terms.

	### Provenance

	- Source: Chinese Power Line Insulator Dataset (CPLID), GitHub `InsulatorData/InsulatorDataSet`.
	- Upstream: `Normal_Insulators` 600 img / 1,073 box · `Defective_Insulators` 248 img / 496 box (248 defect + 248 insulator).
	- Verified: 845 images (1,566 boxes, byte-identical); 3 duplicates removed, 486 frame chains, 0 cross-split leakage, 0 invalid boxes.

	---

	## 中文

	### 这是什么？

	中国电力线绝缘子数据集（CPLID）的清理、划分与标准化目标检测版——包含无人机拍摄的
	正常绝缘子图像，以及合成生成的缺陷绝缘子图像。图像由国家电网公司提供，数据集由
	王自浩（WANG Zi-Hao）整理。

	论文：Tao, Zhang, Wang, Liu, Zhang & Xu，《利用卷积神经网络分析航拍图像检测电力线绝缘子缺陷》，
	IEEE Transactions on Systems, Man, and Cybernetics: Systems，2018。
	上游仓库：[InsulatorData/InsulatorDataSet](https://github.com/InsulatorData/InsulatorDataSet)。

	原始版本以两个扁平、无划分的目录发布（`Normal_Insulators` 与 `Defective_Insulators`），其中
	每张缺陷图被标注了两次（一次为缺陷，一次为整个绝缘子）并分别存放在两个平行的标签目录中；
	VOC 元数据混入硬编码的个人绝对路径；压缩包内还留有一个垃圾非图片文件；并有 3 张像素级完全重复的图像。

	本版本删除重复图、将同一段无人机连拍帧固定在同一划分内、把成对的标签目录合并为每图一份标注、
	清理 VOC 元数据、剔除垃圾文件，提供可复现的训练/验证/测试划分，并附带开箱即用的
	YOLO TXT 与 Pascal VOC XML 双格式标注及训练配置。

	> ⚠️ 上游未声明许可证。 CPLID 未附带 LICENSE 文件，也无使用条款。详见
	> [许可证与署名](#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 3 张重复图片 \| `Normal_Insulators` 中有三对图像像素级完全相同（`0150` = `2984`、`0852` = `1876`、`1608` = `1938`）。本版删除每对中编号较大的一张，确保相同内容不会跨越新划分。 \|
	\| 2 \| 连拍帧固定同侧 \| 600 张正常图均为无人机视频帧，其中 112 张与相邻帧编号连续。本版将其归并为 486 个连拍链（最长 5 帧），整条链完整分配到一个划分，避免近似重复帧被切散到 train/val/test。 \|
	\| 3 \| 合并成对标签目录 \| 上游把每张缺陷图在两个平行目录中标注两次（`labels/defect/` + `labels/insulator/`）。本版合并为每图一份 XML / 一份 YOLO，同时包含绝缘子框与缺陷框。 \|
	\| 4 \| 可复现的训练/验证/测试划分 \| 上游完全没有划分。本版基于连拍链分组，提供固定随机种子、按类别分层的 70 / 15 / 15 划分（随机种子 42）。 \|
	\| 5 \| VOC 元数据清理 \| 上游 XML 带有错误的 `<folder>` 值（`images`、`long`、空混杂）以及硬编码的本地 `<path>`，暴露了含中文的个人绝对路径（如 `/media/laptop/vbox/硕士毕业设计/小论文/DataSet/...`）。均归一化，且 `<source><database>` 统一为 `CPLID`。 \|
	\| 6 \| 删除垃圾文件 \| `OriginalImageOnWebDriver` 是一个 48 字节的非图片文本文件，内容仅为一个百度网盘链接，已删除。 \|
	\| 7 \| 规范化打包 \| 新增 `data.yaml`、`classes.txt`、双格式同步标注（`Annotations/` + `labels/`）以及按划分刷新的 CSV 表格。 \|

	图像像素从不修改——全部 845 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。我们只做分离、去重、重新打包；
	不重新核对标注框。特别说明：缺陷图是合成的——由原作者按上游所述的
	TVSeg + 仿射增强 + U-Net 流程生成并贴到不同背景上。

	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↔XML 配对 \| 双向 \| 0 错配 / 孤儿 \|
	\| 图片尺寸 vs. XML `<size>` \| 像素尺寸 vs. 头信息 \| 0 不一致（全部 1152×864） \|
	\| 越界框 \| xmin/ymin ≥ 0，xmax ≤ W，ymax ≤ H \| 0 \|
	\| 退化框 \| 宽或高为零/负 \| 0 \|
	\| 跨划分重复图 \| 跨划分逐字节 md5 \| 0 \|
	\| 连拍帧泄漏 \| 跨划分 +1 邻接 \| 0 \|
	\| 类别名一致性 \| 唯一类别名集合 \| 2 类，无杂名 \|

	### 数据集概览

	\| 属性 \| 值 \|
	\|---\|---\|
	\| 图片 \| 845（1152 × 864，RGB `.jpg`） \|
	\| 类别 \| 2 —— `insulator`、`defect` \|
	\| 标注框 \| 1,566 \|
	\| 格式 \| Pascal VOC XML · YOLO TXT \|
	\| 划分 \| train 590 / val 129 / test 126（固定种子、类别分层） \|

	各类框数： `insulator` 1,318 · `defect` 248

	各划分框数： train 1,104（`insulator` 931 / `defect` 173）·
	val 237（200 / 37）· test 225（187 / 38）

	### 目录结构

	```
	CPLID_Insulators-corrected/
	├── images/{train,val,test}/ # 1152x864 RGB .jpg
	├── labels/{train,val,test}/ # YOLO：cls cx cy w h（归一化）
	├── Annotations/ # Pascal VOC XML（扁平，已清理）
	├── 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 # 2 个类别名称，每行一个
	├── data.yaml # Ultralytics YOLO 配置文件
	├── LICENSE
	└── 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)

	```bash
	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（请务必引用） —— 见上方英文部分 `tao2018detection`。

	2. 本修正版（请一并引用） —— 本版与官方发布并不相同：删除了重复图、补充了可复现划分、
	合并了成对标签并清理了 VOC 元数据，因此只引用原数据集无法描述本版本。见上方英文部分
	`cpld_insulators_corrected`。

	3. 标注工具（可选） —— 见上方英文部分 `keenforge`。

	### 许可证与署名

	上游 CPLID 数据集未声明许可证。仓库内无 LICENSE 文件、无使用条款。正常图由国家电网公司
	提供，缺陷图为合成图像；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、划分、标签合并、格式转换、文档）以 CC BY 4.0 发布。
	图像像素未修改，仍受上游条款约束。

	### 来源

	- 来源：中国电力线绝缘子数据集（CPLID），GitHub `InsulatorData/InsulatorDataSet`
	- 上游：`Normal_Insulators` 600 图 / 1,073 框 · `Defective_Insulators` 248 图 / 496 框（248 缺陷 + 248 绝缘子）
	- 已校验：845 图（1,566 框，逐字节一致）；删除 3 张重复图、486 个连拍链、0 跨划分泄漏、0 非法框