faisalthaheem/open-lpr-models

open-lpr models

Model artifacts for open-lpr, a free and open-source licence plate recognition system. The application code is in the open-lpr repository; this repository holds only the weights.

The detector is trained by this project. The recogniser and dictionary are conversions of an upstream model — see Provenance before assuming otherwise.

What is here

artifact	role	ours?
plate_yolox_tiny_640.onnx	plate detector, YOLOX-tiny, 640×640	trained here
plate_ocr_ppocrv5_mobile.onnx	plate text recogniser, CTC	upstream, converted
plate_ocr_dict.json	character dictionary (18,383 chars)	upstream, converted
best_ckpt.pth	detector training checkpoint	trained here

All three runtime artifacts are required. The application refuses to start without them; see Usage.

Verify any artifact against the sha256 in manifest.json before loading it. That file is generated from the artifact bytes themselves, not written by hand.

Demo

Detected and read by the weights in this repository, using the project's own public test image (open-lpr/blackbox/jeep.jpg, 640×360):

plate box	confidence 0.875
text read	349253, confidence 0.99
end-to-end	110 ms (cold, CPU, onnxruntime)

What this demo does not show. The plate also carries the Arabic word قطر, which was not transcribed — the default alphanumeric charset profile emits only alphanumerics, so non-Latin characters are stripped by configuration, not by failure. The numeric part is read correctly. This is a single image and is not evidence of accuracy; see below.

Measured accuracy

Scored on split v4 — 1,275 images, 1,143 plates, human-corrected ground truth. Detection IoU 0.3, confidence 0.3.

	detector (this repo)	superseded model
recall	0.926	0.906
precision	0.897	0.835
false positives / image	0.095	0.161
recall, plates ≥100px tall	0.958	0.899
recall, plates <40px tall	0.832	0.829
latency, mean / p95	20.8 ms / 38.6 ms	21.7 ms / 39.6 ms

Sample sizes matter here, because several plausible-looking differences are not real. A seed-repeat of the identical configuration on 1,275 images gave:

axis	run-to-run spread
recall	0.013
precision	0.013
recall <40px	0.021

Read against that noise floor:

• Real: precision +0.063 and recall ≥100px +0.059 — roughly 5× the spread.
• Marginal: recall +0.019 — about 1.5× the spread.
• Noise: recall <40px +0.003 — no measurable effect.

So the defensible claim is that this model is substantially better at precision and large plates, with recall and small-plate performance unchanged. A previous measurement showing a 2-point regression on small plates was an artefact of incomplete teacher labels and does not exist on corrected ground truth.

The noise floor above comes from one seed pair (n=2). It indicates scale, not a confidence interval.

Text accuracy is not measured. The public corpus annotates plate boxes and carries no transcriptions. The recogniser's accuracy on non-Latin scripts is therefore unknown — see Provenance.

Usage

The application loads these from PIPELINE_MODEL_DIR, default model/plate/, and expects these filenames:

PIPELINE_MODEL_DIR=model/plate/
PIPELINE_DETECTOR_MODEL=plate_yolox_tiny_640.onnx
PIPELINE_OCR_MODEL=plate_ocr_ppocrv5_mobile.onnx
PIPELINE_OCR_DICT=plate_ocr_dict.json

Inference needs onnxruntime only. ROCm is not required to load the checkpoint — only to train. CPU torch reads a .pth fine.

The three runtime artifacts are a set: the recogniser and dictionary must come from the same export. Decoding assumes len(dict) + 2 output classes, so substituting one without the other yields silently wrong text rather than an error.

best_ckpt.pth is a pickle. Loading it executes constructors. Only load checkpoints you trust.

Provenance

Detector — trained here

YOLOX-tiny, Apache-2.0 for both the code and the same-repository release weights. Pretrained from COCO yolox_tiny.pth, then trained on the Simanno annotated plate corpus (frame-grouped split with a deliberate gap) plus community-contributed images pseudo-labelled by a teacher model.

Ultralytics YOLO weights are AGPL-3.0 and are not used anywhere in this project. The detector is YOLOX.

The community images were pseudo-labelled by an 8B Qwen3-VL teacher. Human review found that teacher omitted roughly 7% of plates and misplaced a further batch. Val ground truth here is human-corrected; 251 training images containing unlabelled plates were dropped, because an unlabelled plate is false-negative supervision — the model is penalised for detecting something that is really there. Teacher identity is operator-attested, not verifiable from artifacts.

Recogniser and dictionary — upstream, converted by us

These were not trained by this project, and their exact upstream release was never recorded. It cannot be recovered: the ONNX graph carries no producer, version or doc string, and no PaddlePaddle installation, wheel or shell history survives on the build host.

What can be stated:

• Converted from a PP-OCR recognition model using Paddle 3.x (graph name PaddlePaddle Graph in PIR mode, opset 13).
• PaddleOCR is Apache-2.0. That licence is asserted at project level; the specific release is unverified.

One unresolved question, flagged rather than hidden. The dictionary is 18,383 characters. PaddleOCR's stock ppocr_keys_v1.txt is 6,622 — verified against refs release/2.7, release/2.8, v2.7.0 and v2.9.0. So this is not a stock PaddleOCR dictionary: it is either a different recognition variant or a custom extension. Regenerating it from stock PP-OCR would change which characters the model can emit, and there is no labelled non-Latin data to detect a regression. Do not regenerate it without first labelling non-Latin plates.

Training imagery is not published

The photographs this model was trained on are community-contributed images containing real licence plates, faces and vehicles, uploaded without consent obtained. Publishing them would create a searchable index of real registration numbers, which is a privacy harm the licence terms cannot excuse.

The reusable contribution is the tooling — dataset conversion, evaluation and export in lpr_app/ml/ in the open-lpr repository — and it is already public and needs none of that imagery to be useful.

Licence

Apache-2.0. See LICENSE and NOTICE.

The repository as a whole is Apache-2.0, but that does not mean every artifact here is ours. The recogniser and dictionary are conversions of an upstream Apache-2.0 project and are attributed as such in manifest.json.