Babblu2821/alpr-plate-detector

ALPR plate detector (YOLOv8s)

Single-class license plate detector, trained as Phase 2 of fayazhussain2821/ALPR — an end-to-end pipeline that detects plates in video, reads them, validates them against Indian and German plate grammars, and logs one deduplicated row per vehicle to an Excel workbook.

Results

Evaluated on a held-out test split of 465 images the model never saw.

Metric	Test
mAP@50	0.9921
mAP@50-95	0.8377
Precision	0.9816
Recall	0.9917
Inference	4.7 ms/image (T4) · 29 ms/frame (Apple M4, MPS)

Recall is the metric that matters for a plate pipeline. A plate the detector misses can never be read downstream — that error is unrecoverable. A false positive produces a crop, OCR emits noise, and a grammar check rejects it. The error profile is the right way round: 1 missed plate against 23 false positives across the whole test split.

Detection by plate size

Ground-truth width	n	Precision	Recall
tiny (<32 px)	8	1.0000	1.0000
small (32–64 px)	64	0.9000	0.9844
medium (64–128 px)	191	0.9598	1.0000
large (≥128 px)	190	0.9845	1.0000

Small plates were expected to cap accuracy. They do not — every plate under 32 px was found.

The test split was audited for leakage

A perceptual-hash audit found 5.8% of test images had a near-duplicate in train (consecutive video frames whose filenames the grouped-splitting logic did not recognise as frames of one clip). Re-scoring on only the 438 uncontaminated images:

	Full test (465)	Uncontaminated (438)
Recall	0.9979	0.9978
Precision	0.9545	0.9556

The leak was not carrying the score. Grouping was fixed so future splits cluster duplicates.

Training

Architecture	YOLOv8s (11.1M parameters)
Epochs	100
Image size	640
Hardware	Google Colab T4, 1.25 h
Data	3,105 images / 3,273 plates, split 2174/466/465

Augmentation was tuned for plates rather than for COCO: no vertical flip (a plate is never upside down), and rotation and perspective enabled (Ultralytics defaults both to zero, which under-trains the variation this task actually has).

Full arguments are in results/args.yaml. Ultralytics changes augmentation defaults between minor releases, so reproducing this needs the arguments and the version that consumed them.

Usage

from ultralytics import YOLO

model = YOLO("best.pt")
results = model.predict("car.jpg", conf=0.25)

Or through the project's pipeline, which adds tracking, OCR, plate-grammar validation and Excel logging:

alpr run --source 0 --weights best.pt --out plates.xlsx

Training data and attribution

Derived from two Roboflow Universe datasets, both CC BY 4.0:

• European License Plates — 1,455 images
• Indian License Plate (NIVU) — 1,650 images

Limitations

It only finds plates. It does not read them. Recognition is a separate stage in the pipeline.

Boxes are axis-aligned, so a plate photographed obliquely comes out as a rectangle around a slanted plate. Downstream OCR cannot un-skew it without the four corners, which this model does not predict.

The training data is European and Indian. Plates from other regions may work — a detector largely learns "small bright rectangle on a vehicle" — but that is untested here.

Test images averaged 119×40 px plates. Performance on much smaller plates, heavy motion blur or night footage is not characterised by these numbers.

Intended use and responsible use

Built as a portfolio and learning project. Automatic plate recognition is surveillance technology, and a plate is personal data — under the GDPR in the EU, and under comparable regimes elsewhere. Deploying it against public traffic engages legal obligations around lawful basis, retention, and notice that are the deployer's responsibility, not the model's.

Reasonable uses: research, private property you control, and datasets you have the right to process. This model should not be used to track individuals.