khal94/lpr-system

End-to-end Tunisian license plate recognition: YOLOv8 detection + CRNN/CTC OCR, trained on real road-camera footage. # LPR System — License Plate Detection & Recognition An end-to-end **License Plate Recognition (LPR)** pipeline for Tunisian plates, built from footage captured on real road-side cameras. It has two independently developed stages: ``` camera frame ──▶ [ detection ] ──▶ cropped plate ──▶ [ recognition ] ──▶ "123 تونس 8071" YOLOv8 CRNN + CTC ``` | Stage | Folder | Model | Job | |-------|--------|-------|-----| | Detection | `detection/` | YOLOv8 | Find and crop license plates in an image | | Recognition | `recognition/` | CRNN + CTC | Read the characters off a cropped plate | Each stage has its own README with full details; this page is the overview. ## Pipeline at a glance 1. **Detection** — a YOLOv8 model, trained on manually labeled frames from 4 real cameras, locates plates and crops them with padding. 2. **Recognition** — a CRNN (CNN + BiLSTM, CTC loss) reads the cropped plate as a character sequence over the vocabulary `0-9`, `تونس`, and space. See `detection/README.md` and `recognition/README.md` for setup, training, and usage. ## Example detection A detection from the trained YOLOv8 model (`license_plate`, confidence 0.65). The plate is blurred for privacy; the box and label are the raw model output. ## Datasets The two stages have different data needs: - **Detection** was trained purely on the project's **own camera frames**, labeled by hand (single `license_plate` class). - **Recognition** needed far more character-level data than the real set could provide, so its training data **combines three sources**: 1. Real cropped plates from the project's cameras (limited). 2. An open-source real license-plate dataset. 3. Synthetically generated plates for volume and character coverage. The recognition model is therefore trained in two stages: pre-trained on the synthetic + open-source data, then fine-tuned on the real crops. The open-source dataset used is **Tunisian Licensed Plates** from Dataset Ninja — please refer …