ParsaVictor/pcb-bare-board-defect-detection

PCB Bare-Board Defect Detection Dataset

A YOLO-format object detection dataset for finding manufacturing defects on bare printed circuit boards (before any components are soldered) — the kind of defects an Automated Optical Inspection (AOI) system looks for on a PCB fabrication line.

This dataset powers ParsaVictor/pcb-bare-board-defect-detection-yolo26, part of the PCB Inspection Suite.

Why this dataset is different

Most public "PCB defect" datasets on the internet are the same ~693 photos (the PKU-Market-PCB release) copy-pasted into dozens of re-uploads at low resolution. This dataset instead merges:

• Full-resolution PKU-Market-PCB images (official originals, not the downscaled re-exports common on other platforms)
• DeepPCB (1,500 paired template/test image pairs)
• A manually captured and annotated set of real boards ("aaa_hr") photographed at high resolution specifically for this project

Images were tiled/cropped intelligently to keep defects visible at training resolution, and every split was checked to ensure test/val images are real, unmodified photographs (no synthetic augmentation leaked into evaluation).

Classes (7)

Class	# instances
open_circuit	8,594
mouse_bite	8,437
spur	8,079
spurious_copper	7,798
short	7,755
missing_hole	2,802
pin_hole	2,633

Splits

Split	Images
train	18,459
val	2,180
test	1,017

Format

Standard YOLO format, packaged as three separate archives (one per split, so you only need to download what you use):

• train_images_and_labels.zip (3.9 GB)
• val_images_and_labels.zip (4.0 GB)
• test_images_and_labels.zip (167 MB)

Each archive extracts to:

final_bare/images/<split>/*.jpg
final_bare/labels/<split>/*.txt   # class_id x_center y_center width height (normalized)

Quick start

from ultralytics import YOLO

model = YOLO("yolo26s.pt")
model.train(data="data.yaml", imgsz=1024, epochs=80)

Or download and unzip just the splits you need:

from huggingface_hub import hf_hub_download
import zipfile

for split in ["train", "val", "test"]:
    zip_path = hf_hub_download(
        repo_id="ParsaVictor/pcb-bare-board-defect-detection",
        repo_type="dataset",
        filename=f"{split}_images_and_labels.zip",
    )
    with zipfile.ZipFile(zip_path) as zf:
        zf.extractall(".")

Related

• Model trained on this data: pcb-bare-board-defect-detection-yolo26
• Sibling dataset: pcb-solder-defect-detection
• Sibling dataset: pcb-component-detection
• Full suite overview: pcb-inspection-suite

License / attribution

This is a mixed-source dataset for research and development use. It combines the public PKU-Market-PCB and DeepPCB releases with originally captured photographs. If you use the PKU-Market-PCB or DeepPCB portions, please also cite their original sources. See LICENSE.md for details. Contact the author before any commercial redistribution of the full merged package.