Architecture

One image (POST /api/v1/infer, anpr infer)

  decode ─▶ detect ─▶ crop ─▶ OCR ─▶ format check + look-alike fixes ─▶ JSON

A stream (/ws/stream)

  decode ─▶ every Nth frame ─▶ detect ─▶ track ─▶ crop + resize ─▶ OCR
         ─▶ format check + look-alike fixes ─▶ vote per track ─▶ WebSocket event
                                                     │
                                                     └─ first confirmation ─▶ HMAC ─▶ database

A single image goes straight from detection to the format check. A stream adds frame skipping, tracking and the vote, and is the only path that writes to the database.

Modules (src/anpr/)

Path	Responsibility
detector/	The Detector protocol and its fast-alpr implementation. Returns boxes with confidences.
ocr/	The PlateReader protocol and its fast-plate-ocr implementation. Returns text and per-character confidences.
tracker/iou.py	Greedy IoU tracker. Stable track IDs, expiry after 30 unseen frames.
postprocess/turkish.py	The 81 province codes and the plate pattern. Turns 34ABC1234 into province, letters and digits.
postprocess/confusion.py	When a reading doesn’t fit, tries the three possible letter-block lengths and swaps look-alike characters to fit each.
postprocess/temporal.py	Confidence-weighted, per-character majority vote over a track’s best reads.
postprocess/dewarp.py	On the stream, scales the crop to a fixed height. A placeholder for a real perspective correction.
pipeline.py	infer_image for one image, and the async Pipeline for streams.
imaging.py	Decodes uploads and frames, and rejects them by declared size before decoding.
storage/	The Detection table (SQLModel), HMAC hashing, the repository and the retention task.
api/	The FastAPI app, its routes, the web page and the middleware for request IDs and metrics.
observability.py	Prometheus metrics and optional OpenTelemetry.
config.py	Settings from ANPR_ environment variables, with pydantic-settings.
cli.py	The anpr command, with Typer.

Design choices

Protocols at the model boundary. The pipeline only knows Detector and PlateReader. Another detector or OCR engine can be added by writing one class, and the tests use fakes so that most of them run in a fraction of a second without any model.

Format rules after OCR, not inside it. The OCR model is general. Knowing that a Turkish plate is NN L{1-3} N{2-4} lets the parser fix the characters a general model gets wrong, such as an O in a digit position. Other countries would need their own parser; see the Roadmap.

Hash at the edge of storage. The plate text exists in memory and in the response, and is hashed right where a row is written, so no code path can store it by mistake.

Bounded everything on the stream. A queue of four frames, frames dropped when it is full, and size checks before decoding: a slow model or a hostile client can’t make the service use unbounded memory.

The 2022 version

The first version, built on a YOLOv5 fork, EasyOCR and Flask, is kept unchanged in legacy/ and tagged v0.1.0-legacy. It isn’t maintained.