License plate segmentation with U-Net
3-class semantic segmentation of license plates using a U-Net in TensorFlow/Keras — background, border ring, and plate interior.
Task
Segment license plate images into three pixel classes: outside background, white border ring, and black plate interior. Pixel-level accuracy: 0.9982.
Architecture
Standard U-Net encoder-decoder with skip connections. Input images paired with masks encoded as either alpha-channel PNG or single-channel 3-class PNG — both formats supported at training time.
Pipeline
# Train
python -m src.train_unet --images data/images --masks data/masks --outdir models --epochs 50
# Predict with overlay
python -m src.predict_unet --images data/test --model models/unet.keras --out results --overlayNotes
- Mask format flexibility was a deliberate design choice: different labelling tools export differently.
- 0.9982 pixel accuracy is achievable here because the classes are visually distinct; the harder problem would be detecting plates in-the-wild before segmentation.