ANGPR — Automatic Number-Plate Recognition (Computer-Vision Engine)

Python system that detects, crops, and reads vehicle license plates from images, video, and live IP cameras — tuned for Turkish truck/trailer plates with an international fallback.

 ·  Python · YOLO11-pose · ONNX · OpenCV · fully offline

Portfolio edition. This is a sanitized subset of a system built for a commercial client. It contains the computer-vision engine only. All field data, trained weights, credentials, the operational gate/database layer, and anything client-identifying have been removed. See Scope below.

---

Why it is interesting (engineering highlights)

• 4-corner pose detection, not bounding boxes. An axis-aligned box physically cannot wrap a plate seen at an angle — it grabs background or clips a corner. The detector regresses the plate's four corners (YOLO11-pose, 1 class, kpt_shape=[4,3]), which feed directly into cv2.getPerspectiveTransform for a perspective-correct crop. No contour hunting, no threshold tuning.
• Synthetic composite data to close a measured domain gap. Ablations showed the confidence collapse on Turkish plates was brightness polarity (training plates dark-on-light vs. Turkish light-on-dark), not colour; and the 4.7:1 vs 3.1:1 aspect ratio hurt corner precision, not confidence. Fix: composite synthetic Turkish plates onto real plate quads (label is free and exact). Held-out recall went 15% → 98%.
• Torch-free ONNX deployment. The detector layer loads .pt (dev/GPU) or .onnx (prod/CPU) behind one interface; the rest of the pipeline is agnostic. Production runs on CPU with onnxruntime only (no torch/ultralytics), fully offline.
• Temporal K-frame voting OCR. Per vehicle, the best K frames are read and votes are combined per character position, weighted by character probability — the cure for single-frame I↔J, 0↔O confusions.
• Position-aware Turkish validation. Province code 01–81, no Q/W/X, digit/letter coercion by position (4ZBHE495 → 42 BHE 495), with a guard against turning all-digit fleet/container codes into fake plates.

Pipeline

flowchart LR
    SRC["source<br/>image · video · rtsp"] --> DET["detector<br/>4 corners + conf<br/>.pt or .onnx"]
    DET --> FLT["filters<br/>conf + valid quad"]
    FLT --> TRK["tracker<br/>IoU + center fallback<br/>1 verdict / vehicle"]
    TRK --> RCT["rectify<br/>perspective crop"]
    RCT --> OCR["ocr<br/>fast-plate-ocr<br/>+ K-frame vote"]
    OCR --> VAL["validate_tr<br/>TR format + correction"]
    VAL --> ANN["annotate"]

Loading

• detector — 4 corners + confidence; loads .pt (dev/GPU) or .onnx (prod/CPU), one interface
• filters — is it a plate? (confidence + valid quadrilateral)
• tracker — IoU + center-distance fallback, one verdict per vehicle, motion-direction evidence, best-K frames (bounded memory)
• rectify — 4 corners → getPerspectiveTransform → upright crop
• ocr — fast-plate-ocr (ONNX) + K-frame character voting
• validate_tr — TR format + position-aware correction (international plates pass through)
• annotate — draw result (audit output)

Each module has a single responsibility; thresholds live in one config.yaml (see config.example.yaml), no magic constants in code. Design rationale: DECISIONS.md.

Measured results (detection model)

Metric	Value
Turkish held-out recall@0.25	98.1%
CCPD corner acceptance (error ≤ 2% of plate width)	75.0%
CCPD quad IoU ≥ 0.95	52.6%
Detection recall (CCPD val)	100%

Cropping quality is measured with normalized corner error and quad IoU, not mAP (which is misleading for this task).

Full methodology, the model journey, and the root-cause ablation experiments: RESULTS.md. Real pipeline outputs (rectified crops + OCR): samples/.

Tech stack

Python · YOLO11-pose (Ultralytics) · ONNX Runtime · OpenCV · NumPy · fast-plate-ocr · PyODBC (operational layer, not in this edition).

Scope of this repository

Included: the computer-vision engine (angpr/ — detector, rectify, filters, tracker, ocr, validate_tr, annotate, sources).

Excluded on purpose: trained weights, all datasets (CCPD is referenced, not bundled — MIT, https://github.com/detectRecog/CCPD), field imagery, runs/ outputs, the gate state-machine + database layer, credentials/IP addresses, and the deployment package. This keeps the repository free of client data and secrets.

Data & licensing notes

• CCPD (MIT) is used for pre-training — referenced, not redistributed here.
• Turkish-plate appearance is handled with self-generated synthetic composites (own drawing + CCPD quads), no third-party plate dataset redistributed.
• Runtime is fully local — no external API, no cloud inference.

Author

Hamza Akkaymak — sole author and developer (design, model, and pipeline).