PKU-Market-PCB-corrected

PCB defect detection — cleaned and consistently packaged version of the PKU-Market-PCB dataset. PCB 缺陷检测 —— PKU-Market-PCB 数据集的清理与规范化打包版。

English · 中文

---

English

What is this?

A cleaned and consistently packaged version of PKU-Market-PCB (Huang, Weibo & Wei, Peng. "A PCB Dataset for Defects Detection and Classification", arXiv:1901.08204, 2019), released by the Open Lab on Human Robot Interaction, Peking University.

The images are synthesized from 10 standard template boards; each image carries 3–5 programmatically generated defects of a single class. The upstream release provides COCO annotations and a split whose two halves share the same boards; this release fixes the packaging and adds a board-disjoint split that matches the split used in the literature.

⚠️ Upstream license: none stated. The original authors have never declared a license for this dataset. See License — verify before commercial use.

Corrections vs. the official release

#	Correction	Detail
1	Added YOLO + VOC formats	Upstream ships COCO JSON only. This release adds Pascal VOC XML and YOLO TXT (all three verified to agree).
2	Added a board-disjoint split	The upstream train / val COCO files put all 10 boards in both halves (100 % board-level leakage — see below). This release also ships train_board.txt / val_board.txt (541 / 152), which matches the split sizes used in the literature.
3	Packaging	classes.txt, data.yaml, ImageSets/ added.

No bounding box was modified and every image is byte-identical to the upstream file.

⚠️ The board-level leakage in the upstream split

Every image is a render of one of 10 base boards with different synthetic defects. The upstream split distributes images randomly, so each of the 10 boards appears in both the train and the val half:

	boards present
upstream train.json (555 images)	01, 04, 05, 06, 07, 08, 09, 10, 11, 12
upstream val.json (138 images)	01, 04, 05, 06, 07, 08, 09, 10, 11, 12

A model evaluated on that val half sees boards it has already memorized, which inflates detection scores. This release therefore also provides a board-disjoint split:

split	boards	images
ImageSets/train_board.txt	01, 04, 05, 06, 07, 08, 09	541
ImageSets/val_board.txt	10, 11, 12	152

data.yaml points at the board-disjoint split by default; the upstream split is preserved in ImageSets/train.txt / val.txt for comparability.

Not a defect — intentional by design

Images from the same board are 99.99 % identical (the defects are a few dozen pixels). A naive duplicate scan reports ~27,000 near-identical pairs — these are the dataset's design (same board + different synthetic defects), not duplicates, and they are left untouched.

Dataset at a glance

Property	Value
Images	693
Classes	6 — missing_hole, mouse_bite, open_circuit, short, spur, spurious_copper
Bounding boxes	2,953 (3–5 per image)
Base boards	10
Image size	2240 × 2016 … 3056 × 2464 (10 distinct sizes, one per board)
Formats	COCO JSON · Pascal VOC XML · YOLO TXT
Split	board-disjoint 541 / 152 · upstream 555 / 138

Boxes per class: spurious_copper 503 · missing_hole 497 · mouse_bite 492 · short 491 · spur 488 · open_circuit 482

The upstream release also contains 693 rotated copies (1,386 images in total) for a registration task. This release covers the 693 upright images only.

Structure

PKU-Market-PCB-corrected/
├── images/               # 693 jpg
├── labels/               # YOLO: cls cx cy w h (normalized)
├── annotations/          # Pascal VOC XML
├── annotations_raw/      # upstream COCO: train.json, val.json
├── ImageSets/
│   ├── train_board.txt   # 541  (boards 01,04,05,06,07,08,09)  ← recommended
│   ├── val_board.txt     # 152  (boards 10,11,12)
│   ├── train.txt         # 555  (upstream split — has board leakage)
│   └── val.txt           # 138  (upstream split — has board leakage)
├── classes.txt
├── data.yaml
├── LICENSE
└── README.md

labels/, annotations/ and annotations_raw/ are three representations of the same annotations and are verified to agree. Class IDs are 0..5 throughout.

Citation

1. The original dataset — please always cite this.

@misc{huang2019pcb,
  title         = {A PCB Dataset for Defects Detection and Classification},
  author        = {Huang, Weibo and Wei, Peng},
  year          = {2019},
  eprint        = {1901.08204},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV}
}

2. This corrected release — please cite it as well. It is not identical to the official release: YOLO and VOC annotations were added, and a board-disjoint split was added.

@misc{pku_market_pcb_corrected,
  author    = {KeenForgeAI},
  title     = {PKU-Market-PCB-corrected: a cleaned release of the PKU-Market-PCB PCB-defect dataset},
  year      = {2026},
  version   = {1.0},
  publisher = {KeenForgeAI},
  url       = {https://huggingface.co/datasets/KeenForgeAI/PKU-Market-PCB-corrected},
  note      = {Curated by Lu Gan and Sam Li. Derived from Huang & Wei (2019),
               arXiv:1901.08204. 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

The upstream PKU-Market-PCB dataset states no license. The dataset is offered for download by the Open Lab on Human Robot Interaction (Peking University) and the associated paper is on arXiv; neither grants an explicit license for the data. The copyright status of the images and annotations is therefore undetermined.

• The modifications in this release (format conversion, split, packaging, documentation) are released under CC BY 4.0.
• The underlying images and annotations remain under whatever terms the original authors intend. Verify the upstream license before any commercial or redistributive use.

See LICENSE.

---

中文

这是什么？

PKU-Market-PCB 数据集（Huang, Weibo & Wei, Peng，*"A PCB Dataset for Defects Detection and Classification"*，arXiv:1901.08204，2019）的清理与规范化 打包版。该数据集由北京大学 人机交互机器人开放实验室发布。

图像由 10 块标准模板板合成，每张含 3–5 个程序生成的同类缺陷。上游提供 COCO 标注，但其 划分的两半共享同一批底板；本版修复打包问题，并补上底板不重叠划分（与文献使用的划分一致）。

⚠️ 上游未声明许可证。 原作者从未为本数据集声明许可证。详见许可证——商用前请自行确认。

相对官方版的修正

#	修正	说明
1	补充 YOLO + VOC	上游仅 COCO JSON；本版新增 Pascal VOC XML 与 YOLO TXT（三种格式已校验一致）。
2	补底板不重叠划分	上游 train/val 把 10 块板全部放进两边（100% 底板级泄漏，见下）。本版另提供 train_board.txt/val_board.txt（541/152），与文献使用的划分规模一致。
3	规范化打包	新增 classes.txt、data.yaml、ImageSets/。

未改动任何标注框；每张图与上游文件逐字节一致。

⚠️ 上游划分的底板级泄漏

每张图都是 10 块底板之一 + 不同合成缺陷。上游按图片随机划分，导致10 块板全部同时出现在 train 和 val 两边：

	包含的底板
上游 train.json（555 图）	01, 04, 05, 06, 07, 08, 09, 10, 11, 12
上游 val.json（138 图）	01, 04, 05, 06, 07, 08, 09, 10, 11, 12

模型在这个 val 上评测时，看到的都是已经背下来的板子，检测分数会被虚高。因此本版另外提供 底板不重叠划分：

划分	底板	图片数
ImageSets/train_board.txt	01, 04, 05, 06, 07, 08, 09	541
ImageSets/val_board.txt	10, 11, 12	152

data.yaml 默认指向底板不重叠划分；上游划分保留在 ImageSets/train.txt/val.txt 以便对照。

不是缺陷 —— 设计使然

同一块板的图像之间99.99% 相同（缺陷只有几十个像素）。天真的去重扫描会报出约 27,000 对近似图 —— 这是数据集的设计（同板 + 不同合成缺陷），不是重复，本版一律保留。

数据集概览

属性	值
图片	693
类别	6 —— missing_hole、mouse_bite、open_circuit、short、spur、spurious_copper
标注框	2,953（每图 3–5 个）
底板	10 块
图像尺寸	2240×2016 … 3056×2464（10 种，每种对应一块板）
格式	COCO JSON · Pascal VOC XML · YOLO TXT
划分	底板不重叠 541 / 152 · 上游 555 / 138

各类框数：spurious_copper 503 · missing_hole 497 · mouse_bite 492 · short 491 · spur 488 · open_circuit 482

上游还包含 693 张旋转副本（共 1,386 张）用于配准任务。本版仅含 693 张正向图。

目录结构

PKU-Market-PCB-corrected/
├── images/               # 693 张 jpg
├── labels/               # YOLO：cls cx cy w h（归一化）
├── annotations/          # Pascal VOC XML
├── annotations_raw/      # 上游 COCO：train.json、val.json
├── ImageSets/
│   ├── train_board.txt   # 541（底板 01,04,05,06,07,08,09）← 推荐
│   ├── val_board.txt     # 152（底板 10,11,12）
│   ├── train.txt         # 555（上游划分 —— 有底板泄漏）
│   └── val.txt           # 138（上游划分 —— 有底板泄漏）
├── classes.txt
├── data.yaml
├── LICENSE
└── README.md

labels/、annotations/、annotations_raw/ 是同一份标注的三种表示，已校验一致。类 ID 统一为 0..5。

引用

1. 原始数据集（请务必引用） —— 见上方英文部分 huang2019pcb。

2. 本修正版（请一并引用） —— 本版与官方发布并不相同：补充了 YOLO 与 VOC 标注、 新增了底板不重叠划分。见上方 pku_market_pcb_corrected。

3. 标注工具（可选） —— 见上方 keenforge。

许可证

上游 PKU-Market-PCB 未声明许可证。 数据集由北京大学人机交互机器人开放实验室提供下载，相关论文 挂在 arXiv；两者都未给出明确的数据许可。因此图像与标注的版权状态未定。

• 本版的修改部分（格式转换、划分、打包、文档）以 CC BY 4.0 发布。
• 底层图像与标注仍归原作者所声明的条款（目前未声明）。商用或再分发前请自行确认上游许可。

详见 LICENSE。