CV-LPR — license-plate detection + recognition pipeline

A two-stage license-plate recognition pipeline, built and trained end-to-end, not just scaffolding.

• Two-stage detection + recognition. A YOLO model localizes the plate (bounding box), then an LPRNet-based model reads the character sequence off the cropped plate — detection and recognition are separate trained models (lp_detection.pt, lpr_recognition.pt), not one end-to-end network.
• Streaming pipeline, not a notebook. Results flow through a Kafka producer/consumer setup (result_consumer.py, result_cropped_consumer.py) rather than a one-off batch script — built to process a stream of detections, not a fixed test set.
• Open design questions tracked in the repo itself — e.g. whether the character-recognition stage should output direct characters vs. coordinates, and whether plates need de-rotation before character recognition — left visible rather than papered over.

Code is private.