ANPR Training Console

Upload dataset images and YOLO label files, then start detector training from the browser. This is intended for your hosted ANPR panel. Training is CPU-heavy, so on shared cPanel it may be slow.

Open the review console to approve or correct detections before exporting better OCR and detector ground truth.

Checking training status...
No training run active yet.
Training progress updates here automatically every few seconds, including runs started from this page and terminal runs that use the project training script.
Model safety: the current production detector is kept unless the new candidate scores better on the same validation set.
Train Images0
Train Labels0
Val Images0
Val Labels0
Train Plate Texts0
Val Plate Texts0
Training Progress 0%
Epoch0 / 0
Elapsed0s
ETA-
Last Update-
WorkerIdle
Patience-
mAP50-
Precision-
Recall-

Latest Completed Run

No completed run yet.
Epochs-
mAP50-
Precision-
Recall-

Best Run

No best run yet.
Epochs-
mAP50-
Precision-
Recall-

Batch Upload

Upload multiple images and matching `.txt` label files together. File names should match, for example `car1.jpg` and `car1.txt`.

Labels must use YOLO format like: 0 0.512500 0.683333 0.295000 0.116667
Optional OCR CSV format: filename,plate_text or filename,plate_text,split for example car1.jpg,MP04CX3993,train

Backfill Existing OCR Text

Use this when images and YOLO labels are already uploaded on the server and you only want to attach actual plate numbers now.

Recommended CSV format: filename,plate_text,split. Add split when the same filename exists in both train and validation.

Single Upload

Use this when you want to add one image and label pair quickly from the browser.

This saves the real plate text and an OCR crop, which is needed for OCR training later.

Start Training

Trigger detector training from the uploaded dataset. This retrains plate detection and then refreshes the OCR learning model from your saved OCR ground-truth data. After training, the app compares the new candidate with the current production detector and promotes it only when its ANPR validation score is better.

Patience is the number of epochs allowed without improvement before training stops early. Use 30 for the current dataset; 0 disables early stopping.
This uses `dataset/ocr_ground_truth.csv` and `dataset/ocr_crops` to teach the OCR correction layer from your real plate text examples.
If cPanel limits CPU or long-running tasks, training can fail or stop. In that case use a VPS or local machine for training and upload only the new `models/plate_detector.pt` to hosting.

Training Logs

No logs yet.