ocr_engine.py

from transformers import TrOCRProcessor, VisionEncoderDecoderModel
from PIL import Image, ImageEnhance
import re

# Load processor + model
processor = TrOCRProcessor.from_pretrained("microsoft/trocr-base-handwritten")
model = VisionEncoderDecoderModel.from_pretrained("microsoft/trocr-base-handwritten")

def extract_weight(image: Image.Image) -> str:
    # Crop only display region (adjust based on your image format)
    width, height = image.size
    display_area = image.crop((width * 0.35, height * 0.1, width * 0.65, height * 0.25))  # crop display center

    # Enhance contrast & sharpness
    display_area = display_area.convert("L")  # grayscale
    display_area = ImageEnhance.Contrast(display_area).enhance(2.0)
    display_area = ImageEnhance.Sharpness(display_area).enhance(2.5)
    display_area = display_area.convert("RGB")

    # OCR
    pixel_values = processor(images=display_area, return_tensors="pt").pixel_values
    generated_ids = model.generate(pixel_values, max_length=32)
    full_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]

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

    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) >= 20 else "grams"

    return f"{weight} {unit}" if weight else ""