bobarna / yolov7-number-plates

Transfer learning for recognizing number plates on vehicles. Main repository of the project: https://github.com/bobarna/bme-image-processing

Transfer Learning to Number Plate Recognition

Main repository of the project: https://github.com/bobarna/bme-image-processing

TLDR usage for cutting out all deteted number plates

1. Download some dataset of cars into the number-plates-hun/images folder.
   • Mirror 1
   • Mirror 2
   • or get any other data source of some cars
2. Download yolov7 weights trained for number plate recognition:
   • yolov7-number-plates-trained.pt
3. Run inference with trained yolov7 weights: python detect.py --weights yolov7-number-plates-trained.pt --img-size 448 --source number-plates-hun/images --name number-plates-recognition --save-txt --save-conf --nosave --project inference --exist-ok
4. Move detected *.txt labels into folder inference/labels: mv number-plates-hun/number-plates/recognition/labels number-plates-hun/labels
5. Cut out all detected objects: python cutout.py number-plates-hun
6. Results are in the number-plates-hun/found-classes folder.

Running inference (detecting licence plates)

python detect.py --weights weights-number-plates.pt --img-size 448 --source number-plates-hun/ --name test-number-plates --save-txt --save-conf

• --weights: pretrained weights (result of the transfer learning)
• --img-size: size used for the inference
• --source: folder containing the images
• --name: name for this inference
• --save-txt: also saves the labels as *.txt files
• --save-conf: also saves the confidence in the *.txt files

(detect.py could also take in single images instead of a whole directory.)

Each line of a detection (image_name.txt) takes the following form:

object_id x_min x_max y_min y_max confidence

• object_id: describes which object is detected (in our case, this is always 0 for the number plate class)
• x_min, x_max, y_min, y_max: describe the dimension of the bounding box
• confidence: 0..1 value for the confidence of the given detection.

(We modified detect.py to output detections in image-space, instead of the original relative dimensions.)

Run transfer learning (for reproducibility)

python3 train.py --workers 8 --device 0 --batch-size 8 --data data/number-plates.yaml --img 420 --cfg cfg/training/yolov7-number-plates.yaml --weights yolov7_training.pt --name yolov7-custom --hyp data/hyp.scratch.custom.yaml

Forked from the official YOLOv7 implementation

For more details, see:

• original implementation
• paper