Çekicin plate reader (ONNX, browser-ready) — Turkish licence plates
Two small ONNX models used by the Çekicin Vehicle AI v3 page to find and read licence plates inside the browser (the photo never leaves the device):
| file | what | size | origin (MIT) |
|---|---|---|---|
plate_detector_yolov9t_384.onnx |
plate detector, YOLOv9-t 384 px, NMS inside | 7.4 MB | ankandrew/open-image-models yolo-v9-t-384-license-plate-end2end |
plate_ocr_cct_s_v2.onnx |
plate OCR (10 slots × 37 chars) + country head (66 regions) | 5.0 MB | ankandrew/fast-plate-ocr cct-s-v2-global-model |
All training credit goes to ankandrew (fast-alpr / fast-plate-ocr, MIT licence, weights MIT per issue #99). Çekicin adds a Turkish post-processor (province 01–81 + 1–3 letters + 2–4 digits; fixes O/0, I/1, S/5, B/8 …) and KVKK masking.
Measured on Turkish plates (2026-10-27)
Zenodo record 23369685 Turkish License Plate 250 (250 real photos, CC-BY-4.0), no fine-tuning, CPU (2 vCPU):
| OCR | plate found | read fully correct | characters correct | ms / photo |
|---|---|---|---|---|
| cct-s-v2 (this repo) | 100 % | 92.8 % | 98.4 % | 64 |
| cct-xs-v2 | 100 % | 84.8 % | 97.0 % | 39 |
See plate_config.json for exact pre/post-processing. Plates are personal data under KVKK/GDPR: mask them before storing or showing.