README.md

---
license: other
license_name: see-license-file-upstream-unstated
license_link: LICENSE
task_categories:
- object-detection
language:
- en
- zh
tags:
- photovoltaic
- pv-panel
- solar-panel
- defect-detection
- industrial-inspection
- renewable-energy
- object-detection
- pascal-voc
- yolo
- keenforge
size_categories:
- 1K<n<10K
---

# PV_Multi_Defect-corrected

[![DOI](https://img.shields.io/badge/DOI-10.57967%2Fhf%2F10882-blue)](https://doi.org/10.57967/hf/10882)

**Photovoltaic (PV) panel multi-defect detection — cleaned, split and standardized version of the PV-Multi-Defect dataset.**
**光伏组件多缺陷检测 —— PV-Multi-Defect 数据集的清理、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
**PV-Multi-Defect** dataset — photovoltaic panel imagery with **5 defect classes**:
broken areas, hot spots (bright spots), black/gray borders, scratches, and
non-electrified (black) areas. Upstream repository:
[houhou34/PV-Multi-Defect](https://github.com/houhou34/PV-Multi-Defect).

Reference: Li, L., Wang, Z. & Zhang, T. (2023). *"GBH-YOLOv5: Ghost Convolution with
BottleneckCSP and Tiny Target Prediction Head Incorporating YOLOv5 for PV Panel Defect
Detection."* **Electronics**, 12(3), 561 — https://doi.org/10.3390/electronics12030561.

The original release ships as a flat, unsplit `JPEGImages/` + `Annotations/` pair
(Pascal VOC, 600×600 grayscale, no split), containing **61 groups of byte-identical
duplicate images**, **2 images with no annotation** and **1 annotation with no objects**,
and VOC metadata polluted with a hard-coded personal Windows path and a bogus
`<folder>` value.

This release removes the duplicates and orphan files, keeps the one near-duplicate
cluster on a single side of the split, cleans the VOC metadata, provides a
reproducible **train/val/test** split, and ships ready-to-train **Pascal VOC XML** +
**YOLO TXT** annotations with configuration files.

> ⚠️ **Upstream license: none stated.** The repository ships no LICENSE file. The
> associated article is open access (MDPI). See
> [License & attribution](#license--attribution) — verify before commercial use.

### Corrections vs. the official release

| # | Correction | Detail |
|---|---|---|
| 1 | **Removed 61 duplicate images** | 61 pairs are **byte-identical** (md5). The redundant copy of each pair was removed so identical content cannot straddle the new split. |
| 2 | **Removed orphan / empty files** | 2 images had **no annotation** (`img3059`, `img9513`) and 1 annotation (`img7863`) contained **no objects**; all 3 were dropped. |
| 3 | **Near-duplicate cluster kept together** | One cluster of **12 perceptually identical images** (8×8-dHash ≤ 2 and 32×32 pixel-MSE ≤ 10) is assigned **entirely to one split**, so it cannot leak across train/val/test. |
| 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 near-duplicate groups. |
| 5 | **VOC metadata clean-up** | The upstream XML carried a bogus `<folder>` value (`new` / `result2`), a hard-coded local `<path>` exposing a personal Windows path (`C:\Users\86151\Desktop\...`), and `<database>Unknown</database>`. All are normalised. |
| 6 | **Dual format + packaging** | Added YOLO TXT from the same coordinates alongside the upstream VOC XML, plus `classes.txt`, `data.yaml` and per-split CSV tables. |

**Image pixels are never modified.** All 1,044 images are byte-for-byte identical to the upstream files (600×600, single-channel grayscale).

**Scope — structural clean-up only.** This is **not** a re-annotation. We
de-duplicate, drop orphans and re-package the data; we do **not** re-verify
individual boxes.

**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 600×600) |
| 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** |
| EXIF orientation anomalies | tag scan | **0** |

### Dataset at a glance

| Property | Value |
|---|---|
| Images | **1,044** (600 × 600, single-channel grayscale `.jpg`) |
| Classes | **5** — `black_border`, `broken`, `hot_spot`, `no_electricity`, `scratch` |
| Bounding boxes | **3,403** |
| Formats | Pascal VOC XML · YOLO TXT |
| Split | train **723** / val **155** / test **166** (seeded, class-stratified) |

**Boxes per class:** `hot_spot` 1,553 · `scratch` 1,349 · `black_border` 251 ·
`no_electricity` 158 · `broken` 92

**Boxes per split:** train **2,358** · val **519** · test **526**

**Defect types (from the upstream README):** broken areas · obvious bright-spot areas ·
black or gray border areas · scratched areas · non-electrified (black) areas.

### Structure

```
PV_Multi_Defect-corrected/
├── images/{train,val,test}/      # 600x600 grayscale .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                   # 5 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 / paper — please always cite this.**

```bibtex
@article{li2023gbh,
  title   = {GBH-YOLOv5: Ghost Convolution with BottleneckCSP and Tiny Target
             Prediction Head Incorporating YOLOv5 for PV Panel Defect Detection},
  author  = {Li, Longlong and Wang, Zhifeng and Zhang, Tingting},
  journal = {Electronics},
  volume  = {12},
  number  = {3},
  pages   = {561},
  year    = {2023},
  publisher = {MDPI},
  doi     = {10.3390/electronics12030561}
}
```

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

```bibtex
@misc{pv_multi_defect_corrected,
  author    = {KeenForgeAI},
  title     = {PV_Multi_Defect-corrected: a cleaned release of the PV-Multi-Defect photovoltaic panel defect dataset},
  year      = {2026},
  version   = {1.0},
  publisher = {KeenForgeAI},
  url       = {https://huggingface.co/datasets/KeenForgeAI/PV_Multi_Defect-corrected},
  note      = {Curated by Lu Gan and Sam Li. Derived from the PV-Multi-Defect
               dataset (Li, Wang & Zhang, Electronics 2023). 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 PV-Multi-Defect dataset states no license.** The GitHub repository ships
no LICENSE file and no terms of use; the associated article is published open access by
MDPI. 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, orphan removal, split, metadata clean-up, format
conversion, documentation) are released under **CC BY 4.0**. Image pixels are unmodified
and remain subject to the upstream terms.

### Provenance

- Source: PV-Multi-Defect dataset (Pascal VOC), GitHub `houhou34/PV-Multi-Defect`.
- Upstream: 1,108 images / 1,106 XML / 3,981 objects; no split; 61 duplicate groups; 2 unannotated images; 1 empty annotation.
- Verified: 1,044 images (3,403 boxes, byte-identical); 61 duplicates removed, 3 orphan/empty files dropped, 0 cross-split leakage, 0 invalid boxes.

---

## 中文

### 这是什么？

**PV-Multi-Defect 数据集**的**清理、划分与标准化目标检测版**——光伏组件图像，覆盖 **5 类缺陷**：
破损区域、亮斑（热斑）、黑/灰边框、划痕、无电（黑区）。上游仓库：
[houhou34/PV-Multi-Defect](https://github.com/houhou34/PV-Multi-Defect)。

论文：Li, L., Wang, Z. & Zhang, T. (2023).《GBH-YOLOv5: Ghost Convolution with
BottleneckCSP and Tiny Target Prediction Head Incorporating YOLOv5 for PV Panel Defect
Detection》，**Electronics** 12(3), 561 —— https://doi.org/10.3390/electronics12030561。

原始版本以扁平的、无划分的 `JPEGImages/` + `Annotations/`（Pascal VOC，600×600 灰度，无划分）发布，
包含 **61 组逐字节完全重复的图片**、**2 张无标注图片**与 **1 个无目标的标注文件**，
且 VOC 元数据混入了硬编码的个人 Windows 路径与错误的 `<folder>` 值。

本版本删除重复与孤儿文件、把唯一的近重复簇固定到同一划分、清理 VOC 元数据、
提供可复现的**训练/验证/测试**划分，并附带开箱即用的 **Pascal VOC XML** + **YOLO TXT**
双格式标注及配置文件。

> ⚠️ **上游未声明许可证。** 仓库无 LICENSE 文件；关联论文由 MDPI 开放获取。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 61 张重复图** | 有 61 对图片**逐字节完全相同**（md5）。本版删除每对中多余的副本，确保相同内容不会跨越新划分。 |
| 2 | **删除孤儿/空文件** | 2 张图片**没有标注**（`img3059`、`img9513`），1 个标注文件（`img7863`）**不含任何目标**；三者均已剔除。 |
| 3 | **近重复簇固定同侧** | 唯一一个由 **12 张感知近似图**组成的簇（8×8 dHash ≤ 2 且 32×32 像素 MSE ≤ 10）**整组分配到一个划分**，不会跨越 train/val/test 泄漏。 |
| 4 | **可复现的训练/验证/测试划分** | 上游**完全没有划分**。本版基于近重复分组，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（`new` / `result2`）、泄露个人 Windows 路径的硬编码 `<path>`（`C:\Users\86151\Desktop\...`）以及 `<database>Unknown</database>`。均归一化。 |
| 6 | **双格式 + 规范化打包** | 由同一坐标生成 YOLO TXT（与上游 VOC XML 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 1,044 张图与上游文件逐字节一致（600×600，单通道灰度）。

**本版范围——仅结构化清理。** 这是**结构性清理，不是重新标注**。我们只做去重、剔除孤儿、重新打包；
**不**重新核对单个标注框。

**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↔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 600×600） |
| 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** |
| EXIF 方向异常 | 标签扫描 | **0** |

### 数据集概览

| 属性 | 值 |
|---|---|
| 图片 | **1,044**（600 × 600，单通道灰度 `.jpg`） |
| 类别 | **5** —— `black_border`、`broken`、`hot_spot`、`no_electricity`、`scratch` |
| 标注框 | **3,403** |
| 格式 | Pascal VOC XML · YOLO TXT |
| 划分 | train **723** / val **155** / test **166**（固定种子、类别分层） |

**各类框数：** `hot_spot` 1,553 · `scratch` 1,349 · `black_border` 251 ·
`no_electricity` 158 · `broken` 92

**各划分框数：** train **2,358** · val **519** · test **526**

**缺陷类型（引自上游 README）：** 破损区域 · 明显亮斑区域 · 黑/灰边框区域 · 划痕区域 · 无电（黑区）。

### 目录结构

```
PV_Multi_Defect-corrected/
├── images/{train,val,test}/      # 600x600 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）
├── 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                   # 5 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 / 论文（请务必引用）** —— 见上方英文部分 `li2023gbh`。

**2. 本修正版（请一并引用）** —— 本版**与官方发布并不相同**：删除了重复与孤儿、补充了可复现划分、
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 `pv_multi_defect_corrected`。

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

### 许可证与署名

**上游 PV-Multi-Defect 数据集未声明许可证。** 其 GitHub 仓库无 LICENSE 文件、无使用条款；
关联论文由 MDPI 开放获取。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、剔除孤儿、划分、元数据清理、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：PV-Multi-Defect 数据集（Pascal VOC），GitHub `houhou34/PV-Multi-Defect`
- 上游：1,108 图 / 1,106 XML / 3,981 框；无划分；61 组重复；2 张未标注图；1 个空标注
- 已校验：1,044 图（3,403 框，逐字节一致）；删除 61 张重复、剔除 3 个孤儿/空文件、0 跨划分泄漏、0 非法框