Solo · 2025 · live demo

Factory Vision: defect detection with YOLOv8

A YOLOv8 detector for scratches and dents, with a Streamlit app for uploads, webcam and a simulated production line. Trained and validated on synthetic images.

Live demo Source on GitHub Full write-up

What it shows

mAP50 of 0.995 and mAP50-95 of 0.948 on 60 synthetic validation images, which shows the pipeline works end to end.
On a few real photos it misses obvious dents, so the next step is training on real defect data such as NEU-DET or MVTec AD.

Python
YOLOv8
OpenCV
Streamlit

Files

README.md
app.py
dataset.yaml
demo/export.py
requirements.txt
run_infer.py
scripts/synth_defects.py
train_model.py

Validation images with predicted scratch and dent boxes — runs/detect/train3/val_batch0_pred.jpg

Confusion matrix for the scratch and dent classes — runs/detect/train3/confusion_matrix.png