card_separation.py

import math
import cv2
import numpy as np

CARD_RATIO = 4.3/3.3 # dimensions from Amazon
CARD_IMG_WIDTH = 1000
CARD_IMG_HEIGHT = int(CARD_IMG_WIDTH/CARD_RATIO)

WHITE_MIN = np.array([0, 0, 150],np.uint8)
WHITE_MAX = np.array([180, 100, 255],np.uint8)

def card_segment(img):
    # Extract white cards    
    img_hsv = cv2.cvtColor(img, cv2.COLOR_BGR2HSV)
    img_thresh = cv2.inRange(img_hsv, WHITE_MIN, WHITE_MAX)

    # Find Contours
    contours, hierarchy = cv2.findContours(img_thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
    img_copy1 = img.copy()
    cv2.drawContours(img_copy1, contours, -1, (0, 255, 0), 2, cv2.LINE_AA)
    #mmcv2.imshow("image", img_copy1)

    # Find Cards
    card_contours = []
    for contour in contours:
        # Get rid of all the contours that are not four-sided
        perimeter = cv2.arcLength(contour, True)
        vertices = cv2.approxPolyDP(contour, 0.03 * perimeter, True)
        if len(vertices) != 4:
            continue
        card_contours.append(vertices)

    # We know there are exactly 12 cards, so find the 12 largest contours
    card_contours.sort(key=cv2.contourArea, reverse=True)
    card_contours = card_contours[:12]

    # Sort
    card_contours.sort(key=lambda k:k[0][0][1])
    card_contours_sorted = []
    for i in range(0,12,3):
        card_contours_sorted += sorted(card_contours[i:i+3], key=lambda k:k[0][0][0])
    card_contours = card_contours_sorted

    img_copy2 = img.copy()
    cv2.drawContours(img_copy2, card_contours, -1, (0, 255, 0), 2, cv2.LINE_AA)

    # Perspective-skew each card
    vertices = np.float32([[0,CARD_IMG_HEIGHT], [0,0], [CARD_IMG_WIDTH,0], [CARD_IMG_WIDTH,CARD_IMG_HEIGHT]])
    cards = []
    for card in card_contours:
        # Reformat card contour because findContour gives strange results
        card = card.reshape(4,2)
        card = np.array(card, np.float32)

        # Orient points the right way
        lengths = []
        for i in range(3):
            p1, p2 = card[i], card[i-1]
            dist = math.sqrt((p1[0]-p2[0])**2 + (p1[1]-p2[1])**2)
            lengths.append(dist)
        if (lengths[0] < lengths[1]):
            card = np.roll(card, 2)

        # Dewarp
        matrix = cv2.getPerspectiveTransform(card, vertices)
        card_dewarped = cv2.warpPerspective(img, matrix, (CARD_IMG_WIDTH, CARD_IMG_HEIGHT))
        card_dewarped = cv2.flip(card_dewarped, 1) # Temporary fix until I figure out what's flipping it
        cards.append(card_dewarped)
    
    return cards

if __name__ == "__main__":
    # Read image
    img = cv2.imread("test/set.jpg")
    # Output
    cards = card_segment(img)
    for i in range(len(cards)):
        cv2.imwrite(f"cards/card{i}.png", cards[i])
    cv2.waitKey()