subash1652007/indian-plate-ocr

Indian plate OCR

Non-commercial research and educational use only. See the Licence section.

Two models for reading Indian vehicle number plates:

File	What it is
yolo_best.pt	YOLO plate detector (Ultralytics, AGPL-3.0)
trocr_plate/	microsoft/trocr-small-printed fine-tuned on Indian plate crops

Full pipeline (detection, tracking, format correction, voting across video frames): https://github.com/subash9940/indian-number-plate-recognition

Usage

git clone https://github.com/subash9940/indian-number-plate-recognition.git
cd indian-number-plate-recognition
pip install -r requirements.txt
python anpr.py --video your_video.mp4

The script downloads these weights automatically. To use the reader alone on a cropped plate image:

from huggingface_hub import snapshot_download
from transformers import TrOCRProcessor, VisionEncoderDecoderModel
from PIL import Image

path = snapshot_download("subash1652007/indian-plate-ocr")
proc = TrOCRProcessor.from_pretrained(f"{path}/trocr_plate")
model = VisionEncoderDecoderModel.from_pretrained(f"{path}/trocr_plate").eval()

img = Image.open("plate_crop.jpg").convert("RGB")
ids = model.generate(proc(img, return_tensors="pt").pixel_values, max_length=20, num_beams=3)
print(proc.batch_decode(ids, skip_special_tokens=True)[0])

Training

• Reader: fine-tuned on plate crops from the Kaggle dataset Indian vehicle license plate dataset by Sai Sirisha N and collaborators. About 1,500 training crops with ground-truth plate text (8–11 characters), 15 epochs, lr 4e-5, colour/blur/box-jitter augmentation. Validation split by source video.
• Detector: trained on a Roboflow Universe "Indian License Plate Detection" dataset (YOLO format).

Evaluation

Exact match on 190 held-out plate crops (ground-truth boxes):

Reader	Exact match
EasyOCR (raw)	22%
EasyOCR + format post-processing	31%
This TrOCR model	67%

The 67% is the best of 15 epochs, selected on the same small validation set, so it is optimistic (individual epochs ranged from 44% to 67%).

Limitations

• Indian plates only; other formats will be misread.
• Trained on 8–11 character plates. Two-line, blurred, low-resolution or partly hidden plates are weak spots.
• Outputs can be wrong even when they look like valid plates. Treat results as suggestions, not evidence.
• License plates are personal data in many jurisdictions. Use only on footage you are allowed to process.

Licence

• trocr_plate/: trained on a dataset listed on Kaggle as CC BY-NC-ND (Attribution-NonCommercial-NoDerivatives). These weights are shared for non-commercial research and educational use only. Do not use them commercially.
• yolo_best.pt: AGPL-3.0 (Ultralytics YOLO), trained on a Roboflow dataset with its own terms.
• The training images come from various sources and may belong to third parties. No warranty is given for any output.