README.md

---
license: mit
task_categories:
- object-detection
language:
- en
tags:
- deeppcb
- pcb
- pcb-defect
- industrial-inspection
- object-detection
- pascal-voc
- yolo
- keenforge
size_categories:
- 1K<n<10K
---

# DeepPCB-corrected

[![DOI](https://img.shields.io/badge/DOI-10.57967%2Fhf%2F10551-blue)](https://doi.org/10.57967/hf/10551)

**PCB defect detection — cleaned version of the DeepPCB dataset.**
**PCB 缺陷检测 —— DeepPCB 数据集的清理修正版。**

[English](#english) · [中文](#中文)

---

## English

### What is this?

A **cleaned and consistently packaged** version of the **DeepPCB** dataset
(Tang, Mao, Wang, Zhu & Ye — *"A Robust PCB Defect Detection Method via Fusing Multiple
Hierarchical Features"*, [arXiv:1902.06197](https://arxiv.org/abs/1902.06197)).

DeepPCB pairs each **defect-free template** image with an **aligned tested** image carrying
bounding-box annotations for six common PCB defects. The upstream release ships a custom
TXT format and a split list that **does not resolve**; this release fixes the packaging,
removes the few broken samples, and adds the two formats the tool ecosystem expects.

> ✅ **Upstream license: MIT** (© 2018 tangsanli5201). This release keeps the same license.

### Corrections vs. the official release

| # | Correction | Detail |
|---|---|---|
| 1 | **Fixed the split lists** | Upstream `trainval.txt` / `test.txt` reference `group…/…/XXXXnnn.jpg`, but the released files are `XXXXnnn_test.jpg` — **all 1,500 image paths fail to resolve**. This release rewrites both lists with paths that exist. |
| 2 | **Removed 1 contaminated image** | `44000020_test.jpg` has **annotation boxes and labels burned into the pixels** (a green overlay, RGB mode). Training on it would teach the model to detect the drawn boxes. Removed together with its template. |
| 3 | **Removed 1 orphan template** | `90100034_temp.jpg` has no matching tested image and appears in no split list. |
| 4 | **Normalised image mode** | 61 images were stored as **RGB although R = G = B** (grayscale content). Re-encoded to single-channel grayscale — **pixel content is identical**. |
| 5 | **Added YOLO + VOC formats** | Upstream ships only the custom `x1 y1 x2 y2 type` TXT. This release adds Pascal VOC XML and YOLO TXT (all three verified to agree). |
| 6 | **Packaging** | `classes.txt`, `data.yaml`, `ImageSets/` added. |

**Boxes and image content are otherwise untouched.** Every retained image is byte-identical
to the upstream file (except the 61 RGB→grayscale re-encodings, which do not change content).

### Not a defect — intentional by design

The dataset **reuses templates** and gives each tested image **synthetic defects**: the README
states *"we manually argument some artificial defects on each tested image"*. Consequently a
tested image is almost identical to its template (and templates recur across images). A naive
duplicate scan will report ~350 near-identical pairs — **these are the dataset's design, not
errors**, and they are left untouched.

### Dataset at a glance

| Property | Value |
|---|---|
| Tested (defective) images | **1,499** (640 × 640, grayscale) |
| Template (defect-free) images | **1,499** |
| Classes | **6** — `open`, `short`, `mousebite`, `spur`, `copper`, `pin-hole` |
| Bounding boxes | **10,004** (≈ 6.7 per image) |
| Formats | custom TXT · Pascal VOC XML · YOLO TXT |
| Split | trainval **1,000** / test **499** (upstream split, paths fixed) |

**Boxes per class:** `mousebite` 1,963 · `open` 1,940 · `spur` 1,624 · `short` 1,504 ·
`pin-hole` 1,500 · `copper` 1,473

> Upstream class labels (per its README) are `open, short, mousebite, spur, copper, pin-hole`.
> Some third-party re-exports rename the last two to `spurious_copper` / `pin_hole`.

### Structure

```
DeepPCB-corrected/
├── images/            # 1499 tested (defective) images: <stem>_test.jpg
├── templates/         # 1499 defect-free templates:    <stem>_temp.jpg
├── labels/            # YOLO: cls cx cy w h (normalized)
├── annotations/       # Pascal VOC XML
├── annotations_raw/   # original upstream format: x1 y1 x2 y2 type
├── ImageSets/
│   ├── trainval.txt   # 1000 stems
│   └── test.txt       #  499 stems
├── classes.txt
├── data.yaml          # YOLO dataset config
├── LICENSE
└── README.md
```

`labels/`, `annotations/` and `annotations_raw/` are three representations of the **same**
annotations and are verified to agree. Class IDs are `1..6` in `annotations_raw`
(upstream convention) and `0..5` in YOLO/VOC.

### Citation

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

```bibtex
@article{tang2019deeppcb,
  title   = {A Robust PCB Defect Detection Method via Fusing Multiple Hierarchical Features},
  author  = {Tang, Sanli and Mao, Fan and Wang, Zhipeng and Zhu, Zhikai and Ye, Shenghua},
  journal = {arXiv preprint arXiv:1902.06197},
  year    = {2019}
}
```

**2. This corrected release — please cite it as well.** It is not identical to the official
release: the split lists were fixed, one contaminated image and one orphan template were
removed, image modes were normalised, and the data was re-packaged.

```bibtex
@misc{deeppcb_corrected,
  author    = {KeenForgeAI},
  title     = {DeepPCB-corrected: a cleaned release of the DeepPCB PCB-defect dataset},
  year      = {2026},
  version   = {1.0},
  publisher = {KeenForgeAI},
  url       = {https://huggingface.co/datasets/KeenForgeAI/DeepPCB-corrected},
  note      = {Curated by Lu Gan and Sam Li. Derived from Tang et al. (2019),
               arXiv:1902.06197. MIT licensed.}
}
```

**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

**MIT** — the same license as the upstream DeepPCB repository
(© 2018 tangsanli5201). See [`LICENSE`](LICENSE). Please also credit the original authors.

---

## 中文

### 这是什么？

**DeepPCB** 数据集（Tang 等，*arXiv:1902.06197*）的**清理与规范化打包版**。

DeepPCB 把每张**无缺陷模板图**与一张**对齐后的缺陷图**配对，缺陷图带有六类常见 PCB 缺陷的
边界框标注。上游发布的是自定义 TXT 格式，且划分列表**无法解析**；本版修复打包问题、删除少量
损坏样本，并补上生态常用的两种格式。

> ✅ **上游许可证：MIT**（© 2018 tangsanli5201）。本版沿用同一许可。

### 相对官方版的修正

| # | 修正 | 说明 |
|---|---|---|
| 1 | **修复划分列表** | 上游 `trainval.txt` / `test.txt` 引用 `group…/…/XXXXnnn.jpg`，但实际文件是 `XXXXnnn_test.jpg` —— **1,500 条图片路径全部无法解析**。本版重写为可解析的路径。 |
| 2 | **删除 1 张污染图** | `44000020_test.jpg` 把**标注框与标签烧进了像素**（绿色叠加层，RGB 模式）。用它训练会让模型学会"检测画上去的框"。已连同其模板一起删除。 |
| 3 | **删除 1 张孤立模板** | `90100034_temp.jpg` 没有对应的缺陷图，也不在任何划分列表里。 |
| 4 | **统一图片模式** | 61 张图存成了 **RGB 但 R=G=B**（灰度内容）。重新编码为单通道灰度 —— **像素内容完全一致**。 |
| 5 | **补充 YOLO + VOC 格式** | 上游仅有自定义 `x1 y1 x2 y2 type` TXT。本版新增 Pascal VOC XML 与 YOLO TXT（三种格式已校验一致）。 |
| 6 | **规范化打包** | 新增 `classes.txt`、`data.yaml`、`ImageSets/`。 |

**除此之外，标注框与图像内容未改动。** 保留的每张图与上游文件逐字节一致
（仅上述 61 张 RGB→灰度 的重编码，内容不变）。

### 不是缺陷 —— 设计使然

数据集**刻意复用模板**并为每张缺陷图**合成缺陷**：上游 README 明说
*"we manually argument some artificial defects on each tested image"*。因此缺陷图与其模板几乎
相同（模板也会跨图复用）。天真的去重扫描会报出约 350 对近似图 —— **这是数据集的设计，不是错误**，
本版一律保留。

### 数据集概览

| 属性 | 值 |
|---|---|
| 缺陷图（tested） | **1,499**（640 × 640 灰度） |
| 模板图（template） | **1,499** |
| 类别 | **6** —— `open`、`short`、`mousebite`、`spur`、`copper`、`pin-hole` |
| 标注框 | **10,004**（约 6.7 框/图） |
| 格式 | 自定义 TXT · Pascal VOC XML · YOLO TXT |
| 划分 | trainval **1,000** / test **499**（沿用官方划分，仅修复路径） |

**各类框数**：`mousebite` 1,963 · `open` 1,940 · `spur` 1,624 · `short` 1,504 ·
`pin-hole` 1,500 · `copper` 1,473

> 官方 README 给出的类别名为 `open, short, mousebite, spur, copper, pin-hole`。
> 部分第三方镜像把后两类改名为 `spurious_copper` / `pin_hole`。

### 目录结构

```
DeepPCB-corrected/
├── images/            # 1499 张缺陷图：<stem>_test.jpg
├── templates/         # 1499 张无缺陷模板：<stem>_temp.jpg
├── labels/            # YOLO：cls cx cy w h（归一化）
├── annotations/       # Pascal VOC XML
├── annotations_raw/   # 上游原格式：x1 y1 x2 y2 type
├── ImageSets/
│   ├── trainval.txt   # 1000 条
│   └── test.txt       #  499 条
├── classes.txt
├── data.yaml          # YOLO 配置
├── 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 在 `annotations_raw` 中是 `1..6`（上游约定），在 YOLO/VOC 中是 `0..5`。

### 引用

**1. 原始数据集（请务必引用）** —— 见上方英文部分 `tang2019deeppcb`。

**2. 本修正版（请一并引用）** —— 本版与官方发布并不相同：修复了划分列表、删除了 1 张污染图与
1 张孤立模板、统一了图片模式，并重新打包。见上方 `deeppcb_corrected`。

**3. 标注工具（可选）** —— 见上方 `keenforge`。

### 许可证

**MIT** —— 与上游 DeepPCB 仓库相同（© 2018 tangsanli5201）。详见 [`LICENSE`](LICENSE)。
请同时注明原作者。