evan6007/Innodisk_PCB_datasets

🚀🔬 Innodisk PCB Image Dataset

The Innodisk PCB Image Dataset is a collaboration between Innodisk Corporation and Tamkang University.
Designed for research on printed‑circuit‑board (PCB) defect inspection and small‑sample learning. 🧑‍🔧💡

📂 Split	🖼️ Images	📁 Source directories
🏋️‍♂️ Train	500	7 train_* dirs
🧪 Test	240	4 test* dirs
📝 Val	30	2 val* dirs

🎯 Resolutions vary; the modal size is 188 × 128 px.

🔖 Filenames embed pseudo‑labels: pass (=acceptable) or ng (=defective).

---

🗂️ Folder layout

Innodisk_PCB_datasets/
├── train_20_20/      # 🏋️‍♂️
├── train_20_20_R/
├── train_40_20/
├── train_40_20_R/
├── train_60_20/
├── train_80_20/
├── train_100_20/
├── test/             # 🧪
├── test_R/
├── test_canny/
├── test_R_canny/
├── val/              # 📝
└── val_R/

---

⚡ Quick start

from datasets import load_dataset

# 🔄 Load images
ds_train = load_dataset("evan6007/Innodisk_PCB_datasets", split="train")

# 🏷️  Derive labels from filenames
def add_label(example):
    fname = example["image"].filename.lower()
    example["label"] = 0 if "ng" in fname else 1  # 0 = defective, 1 = acceptable
    return example

ds_train = ds_train.map(add_label)

---

🏭 Image sources & preprocessing

1. 📷 AOI line cameras – raw production images (train_*, test*, val*).
2. ✂️ ROI versions (*_R) – cropped component regions to reduce background noise.
3. 🖊️ Canny edge maps (*_canny) – generated with cv2.Canny(img, 50, 150).

---

📜 License

Released under Creative Commons Attribution‑NonCommercial 4.0 (CC BY‑NC‑4.0).
Commercial applications require prior approval from Innodisk Corporation.

---

📚 Citation

@dataset{innodisk_pcb_2025,
  author = {{Innodisk Corporation} and {Su, Huan-Che} and {Hsiao, Chao-Hsiang}},
  title  = {Innodisk PCB Image Dataset},
  year   = {2025},
  url    = {https://huggingface.co/datasets/evan6007/Innodisk_PCB_datasets},
  note   = {CC BY-NC 4.0}
}