ocr_engine.py

from transformers import DonutProcessor, VisionEncoderDecoderModel
from PIL import Image
import re
import torch

# Load processor + model
processor = DonutProcessor.from_pretrained("naver-clova-ix/donut-base-finetuned-cord-v2")
model = VisionEncoderDecoderModel.from_pretrained("naver-clova-ix/donut-base-finetuned-cord-v2")

def extract_weight(image: Image.Image) -> str:
    image = image.convert("RGB")
    pixel_values = processor(image, return_tensors="pt").pixel_values

    # Generate text prediction
    outputs = model.generate(pixel_values, max_length=512)
    decoded = processor.batch_decode(outputs, skip_special_tokens=True)[0]

    # Clean & extract weight
    cleaned = decoded.lower().replace(" ", "")
    match = re.search(r"(\d+(\.\d+)?)", cleaned)
    weight = match.group(1) if match else None

    # Detect unit
    if any(u in cleaned for u in ["kg", "kgs", "kilogram", "kilo"]):
        unit = "kg"
    elif any(u in cleaned for u in ["g", "gram", "grams"]):
        unit = "grams"
    else:
        unit = "kg" if weight and float(weight) >= 5 else "grams"

    return f"{weight} {unit}" if weight else "No valid weight detected"