CCT-S for US License Plate OCR
Compact Convolutional Transformer (CCT-S) fine-tuned for US license plate character recognition using fast-plate-ocr. Reads 8-character US plates from cropped 140x70 grayscale images.
Results
| Metric | Value |
|---|---|
| Plate accuracy | 83.6% |
| Character accuracy | 95.2% |
| Plate length accuracy | 95.0% |
| Top-3 accuracy | 97.8% |
Model Details
| Parameter | Value |
|---|---|
| Architecture | CCT-S (6 transformer layers, 2 heads, dim 128) |
| Tokenizer | 4 Conv2D (48→80→96→128), MaxBlurPooling2D, GELU |
| Input | 140x70 grayscale |
| Max plate slots | 8 |
| Alphabet | 0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ_ (37 chars) |
| Normalization | DYT (Dynamic Tanh) |
| Pretrained from | Global (Argentinian plates) |
Training
See Autolane/alpr-finetuning for the full training pipeline.
W&B project: goautolane/alpr-ocr
Hyperparameters
- Optimizer: AdamW with EMA
- Learning rate: 5e-4 (cosine decay, 5% warmup)
- Weight decay: 5e-4
- Batch size: 64
- Max epochs: 150
- Early stopping: patience 50 on val_plate_acc
- Precision: mixed_float16
- Augmentation: Albumentations (Affine, BrightnessContrast, blur, noise, dropout)
Usage
from fast_plate_ocr import ONNXPlateRecognizer
# Using ONNX runtime (recommended for inference)
ocr = ONNXPlateRecognizer("us-plate-model")
plates = ocr.run(["plate_crop.jpg"])
Or with Keras directly:
import keras
model = keras.models.load_model("best_checkpoint.keras")
# Preprocess: resize to 140x70, grayscale, normalize to [0, 1]
predictions = model.predict(preprocessed_images)
Files
best_checkpoint.keras— Best checkpoint (83.6% plate accuracy, epoch 21)model_config.yaml— CCT-S architecture configplate_config.yaml— Plate dimensions, alphabet, preprocessing
Dataset
Trained on ~20k US license plate crops (80/20 train/val split, stratified by US State). See Autolane/us-license-plates.