readme.md

Hello!

I present You a simple script, that can help you to scan Game Cards, or Photos, multiple at once. I've made it as a project for my university, it was meant to preserve old images and cards I had.

It basically finds the individual cards/photos, extracts them and rotates them to straight orientation.

I share it with You under Creative Commons Licence. Please respect that.

Few assumptions:

• Place game cards/Photos in ( more or less ) straight orientation, avoid 45° rotation (0° = best quality, 45° = worst quality)
• Scan to multi-page PDF, place the file name in line 93
• DO NOT overlap images, about 3-5 mm margin is safe, otherwise script may detect two images as one!
• Remember - usually scanners are not scanning anything that's placed to about 3 mm from edges
• If you'd like to scan cards with white borders - do not close the lid, change white_cards to True
• For Photos you may need to change the cv.threshold(...) on line 17 to something like cv.threshold(gray, 240, 255, cv.THRESH_BINARY_INV)

Needed packages:

• pip install opencv-python PyMuPDF numpy

Showdown ( Those are Munchkin 3 cards, blurred due to copyright. )

Before ( original scan )

After ( extracted images )

CardScanExtractor.py

import cv2 as cv
import numpy as np
import fitz
import os
from datetime import datetime
from pathlib import Path

white_cards = True  # DO NOT CLOSE THE LID WHEN SCANNING WHITE CARDS!


def create_mask(img):
    gray = cv.cvtColor(img, cv.COLOR_BGR2GRAY)

    if white_cards:
        gray = cv.bitwise_not(gray)

    _, threshold = cv.threshold(gray, 250, 255, cv.THRESH_BINARY_INV + cv.THRESH_OTSU)

    contours, _ = cv.findContours(threshold, cv.RETR_CCOMP, cv.CHAIN_APPROX_SIMPLE)

    for cnt in contours:
        cv.drawContours(threshold, [cnt], 0, 255, -1)

    threshold = cv.bitwise_not(threshold)
    kernel = cv.getStructuringElement(cv.MORPH_RECT, (5, 5))

    mask = cv.bitwise_not(cv.morphologyEx(threshold, cv.MORPH_CLOSE, kernel, iterations=3))

    return mask


def trim_border(array):
    return np.mean(array) > 127.5 if not white_cards else np.mean(array) < 127.5


def detect_and_cut(mask, img):
    contours, _ = cv.findContours(mask, 1, 2)
    img_return = []  # array with processed images.

    for cnt in contours:

        # get boundigRect ( w/o rotation )
        x, y, w, h = cv.boundingRect(cnt)

        # get minAreaRect ( w/ rotation )
        rect = cv.minAreaRect(cnt)
        angle = rect[2]
        if angle > 45:
            angle -= 90

        image = img[y:y + h, x:x + w]  # crop image by boudingRect

        image_center = tuple(np.array(image.shape[1::-1]) / 2)  # get center point of img
        rot_mat = cv.getRotationMatrix2D(image_center, angle, 1.0)  # get rotationMatrix from angle
        result = cv.warpAffine(image, rot_mat, image.shape[1::-1], flags=cv.INTER_LINEAR)  # warp by rotationMatrix

        # fill conrers after rotation
        height, width, _ = result.shape
        for point in [(0, 0), (width - 1, 0), (0, height - 1), (width - 1, height - 1)]:
            for _ in range(2):
                cv.floodFill(result, None, seedPoint=point, newVal=(0,) * 3 if white_cards else (255,) * 3, loDiff=(0,) * 4, upDiff=(10,) * 4)

        # final croping
        mask2 = create_mask(result)
        contours2, _ = cv.findContours(mask2, 1, 2)

        for cnt2 in contours2:
            x2, y2, w2, h2 = cv.boundingRect(cnt2)  # get final boundingRect

            result = result[y2:y2 + h2, x2:x2 + w2]

            untrimmed = True
            while untrimmed:
                untrimmed = False
                if trim_border(result[:1]):
                    result = result[1:, :]
                    untrimmed = True
                if trim_border(result[-1:]):
                    result = result[:-1]
                    untrimmed = True
                if trim_border(result[:, :1]):
                    result = result[:, 1:]
                    untrimmed = True
                if trim_border(result[:, -1:]):
                    result = result[:, :-1]
                    untrimmed = True

        img_return.append(result)

    return img_return


file = "filename.pdf"

# open the file
pdf_file = fitz.Document(file)
image_count = 0

folder_name = datetime.now().strftime("extracted %m-%d-%y %H-%M-%S")

for page_index in range(len(pdf_file)):

    # get the page itself
    page = pdf_file[page_index]
    image_list = page.get_images()

    for img in page.get_images():
        # get the XREF of the image
        xref = img[0]

        # extract the image bytes
        base_image = pdf_file.extract_image(xref)
        image_bytes = base_image["image"]
        nparr = np.frombuffer(image_bytes, np.uint8)
        img = cv.imdecode(nparr, cv.IMREAD_COLOR)

        mask = create_mask(img)  # create a mask
        images = detect_and_cut(mask, img)  # frame images
        for img in images:
            Path(folder_name).mkdir(exist_ok=True)
            cv.imwrite(os.getcwd() + "\\" + folder_name + "\\" + str(image_count) + ".jpg", img)
            image_count += 1

Seems that the chat don't support PDFs

You can probably create a secret gist here, and share a link with me, or contact me via discord:)