bloombase/plants/services/card_ocr.py

"""
Extract structured plant data from card photos using Tesseract OCR.

Designed for hellogarden.nl / Elburg & Smit style cards but the regex patterns
are generic enough to handle most Dutch/international plant labels.
"""
import re
from PIL import Image, ImageEnhance, ImageFilter

from plants.services.vpc import _parse_months


# Tesseract config: treat as a single block of text, best for card layouts
_TESS_CONFIG = '--psm 6'
_TESS_LANGS = 'nld+eng'


def extract_from_card(image) -> dict:
    """
    Run OCR on a card image (path or PIL Image) and return extracted fields.

    Returned dict keys: scientific_name, bloom_months, max_height_cm,
    frost_hardiness_c, planting_density_m2, bee_attractant.
    Values are None / [] / False when not found.
    """
    from PIL import ImageOps
    if not isinstance(image, Image.Image):
        image = Image.open(image)
    image = ImageOps.exif_transpose(image)

    text = _ocr_image(image)
    return parse_card_text(text)


def extract_plant_photo(image) -> Image.Image | None:
    """
    Crop just the plant photo from the front of a card.
    Scans rows from the bottom upward to find where the dark banner ends,
    then returns everything above that boundary.
    Returns None if detection fails.
    """
    from PIL import ImageOps
    if not isinstance(image, Image.Image):
        image = Image.open(image)
    image = ImageOps.exif_transpose(image)

    w, h = image.size
    gray = image.convert('L')

    # Scan upward from the bottom half to find the first bright row
    # (transition from dark banner to colourful plant photo)
    boundary = None
    for y in range(h - 1, h // 3, -1):
        samples = [
            gray.getpixel((x, y))
            for x in range(w // 4, 3 * w // 4, max(1, w // 20))
        ]
        if samples and sum(samples) / len(samples) > 100:
            boundary = y
            break

    if boundary is None or boundary < h // 4:
        return None

    return image.crop((0, 0, w, min(boundary + 20, h)))


_OCR_MAX_WIDTH = 1200  # cap before upscaling — phone photos are 4K+


def _ocr_image(img: Image.Image) -> str:
    """Downscale, auto-rotate, upscale 2×, enhance contrast, run OCR."""
    import pytesseract

    max_side = max(img.width, img.height)
    if max_side > _OCR_MAX_WIDTH:
        scale = _OCR_MAX_WIDTH / max_side
        img = img.resize((int(img.width * scale), int(img.height * scale)), Image.LANCZOS)

    # Auto-detect orientation and rotate to upright
    try:
        osd = pytesseract.image_to_osd(img, config='--psm 0')
        angle = int(re.search(r'Rotate: (\d+)', osd).group(1))
        if angle:
            img = img.rotate(angle, expand=True)
    except Exception:
        pass

    img = img.resize((img.width * 2, img.height * 2), Image.LANCZOS)

    def _run(image, cfg):
        try:
            return pytesseract.image_to_string(image, lang=_TESS_LANGS, config=cfg)
        except Exception:
            return pytesseract.image_to_string(image, config=cfg)

    # Pass 1: color image, block mode — best for scientific name / structured text
    color = ImageEnhance.Contrast(img).enhance(1.5)

    # Pass 2: OpenCV CLAHE + adaptive threshold, sparse mode — best for data-strip values
    import cv2
    import numpy as np
    arr = np.array(img.convert('L'))
    clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8))
    arr = clahe.apply(arr)
    arr = cv2.adaptiveThreshold(arr, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C,
                                cv2.THRESH_BINARY, 31, 10)
    binarized = Image.fromarray(arr)

    return _run(color, '--psm 6') + '\n' + _run(binarized, '--psm 11')


def parse_card_text(text: str) -> dict:
    """
    Parse raw OCR text into structured fields.
    Exposed as a standalone function so it can be unit-tested without Tesseract.
    """
    result = {
        'scientific_name': None,
        'bloom_months': [],
        'max_height_cm': None,
        'frost_hardiness_c': None,
        'planting_density_m2': None,
        'bee_attractant': False,
    }

    # Scientific name strategy:
    # 1. Look for genus in the plant passport section ("A Lupinus", "B NL-..." pattern).
    #    This is the most reliable signal on Elburg & Smit cards.
    # 2. Look for a cultivar in quotes anywhere in the text.
    # 3. Fall back to scanning for a genus+species pair in sequence.

    passport_match = re.search(r'(?:^|\s)A\s+([A-Z][a-z]{3,})(?:\s|$)', text, re.MULTILINE)
    cultivar_match = re.search(r"['‘’ʼ]([A-Z][^'‘’ʼ\n]{2,30})['‘’ʼ]", text)

    if passport_match:
        genus = passport_match.group(1)
        if cultivar_match:
            result['scientific_name'] = f"{genus} '{cultivar_match.group(1).strip()}'"
        else:
            # Try to find species epithet on the line after the genus in the text body
            genus_line_match = re.search(rf'^{genus}\s*$', text, re.MULTILINE)
            if genus_line_match:
                after = text[genus_line_match.end():].lstrip('\n')
                epithet_line = after.splitlines()[0].strip() if after.splitlines() else ''
                epithet_clean = re.sub(r'[^\w\s\'\-]', '', epithet_line).strip()
                first_epithet_word = epithet_clean.split()[0] if epithet_clean.split() else ''
                if re.match(r'^[a-z]{3,}$', first_epithet_word):
                    result['scientific_name'] = f"{genus} {epithet_clean}"
                else:
                    result['scientific_name'] = genus
            else:
                result['scientific_name'] = genus
    else:
        # Fallback: scan lines for a strictly title-case genus word
        name_lines = []
        for line in text.splitlines():
            line = re.sub(r'[^\w\s\'\-]', '', line).strip()
            if not line:
                continue
            if not name_lines:
                first_word = line.split()[0] if line.split() else ''
                if re.match(r'^[A-Z][a-z]{4,}$', first_word) and not re.search(r'\d', line):
                    name_lines.append(line)
            else:
                if re.match(r'^[A-Za-z]', line) and not re.search(r'\d', line) and len(line) < 40:
                    name_lines.append(line)
                break
        if name_lines:
            result['scientific_name'] = ' '.join(name_lines)

    # Bloom months: Roman numeral range — OCR misreads 'I' as 'E'/'e'
    roman = re.search(r'\b([IVXEe]{1,4}\s*[-–]\s*[IVXEe]{1,4})\b', text)
    if roman:
        normalized_roman = roman.group(1).replace(' ', '').upper().replace('E', 'I')
        result['bloom_months'] = _parse_months(normalized_roman)

    # Height: "120 cm" — OCR misreads: '0'→'o', '5'→'s'/'S'
    height = re.search(r'([0-9soSO]{2,3})\s*cm', text, re.I)
    if height:
        h_str = height.group(1).lower().replace('s', '5').replace('o', '0')
        if h_str.isdigit():
            result['max_height_cm'] = int(h_str)

    # Frost hardiness: "-20 °C" — OCR misreads: '-'→'„'/'=', '1'→'i'/'l'/'|', 'C'→'G'
    frost = re.search(r'[-−„=]\s*([0-9il|]{1,3})\s*[o°]?\s*[CcGg]', text)
    if frost:
        f_str = frost.group(1).replace('i', '1').replace('l', '1').replace('|', '1')
        if f_str.isdigit():
            result['frost_hardiness_c'] = -int(f_str)

    # Planting density: "7/m²" or "7/m2" or "7/m" (OCR often reads ² as ? or 2 or drops it)
    density = re.search(r'(\d+)\s*/\s*m[²2?]?(?!\w)', text)
    if density:
        result['planting_density_m2'] = int(density.group(1))

    # Bee attractant: Dutch, English, German keywords
    if re.search(r'bijenlokker|attracts bees|bienen', text, re.I):
        result['bee_attractant'] = True

    return result