[Python] Understanding some kind of grid structure of a rubick's cube using OpenCV

So, I am trying to extract all the colors out of a rubicks cube, Right now, I have tried the following:

import cv2
import PIL
import numpy as np
import math
import matplotlib.pyplot as plt

def detect_colored_cubes(image_path):
img = cv2.imread(image_path)
hsv = cv2.cvtColor(img, cv2.COLOR_BGR2HSV)

h, w, c = img.shape
area_min = h*w//(30*30)
area_max = h*w//5

print(area_min, area_max)

color_ranges = {
'red': ([0, 50, 50], [6, 255, 255]),
'orange': ([7., 50, 50], [16.9, 255, 255]),
'yellow': ([17.0, 50, 50], [34., 255, 255]),
'green': ([34., 50, 50], [70., 255, 255]),
'blue': ([71, 50, 50], [135, 255, 255]),
'pink': ([136, 50, 50], [180, 255, 255]),
'white': ([0, 0, 100], [180, 30, 255])
}

detected_cubes = []

for color_name, (lower, upper) in color_ranges.items():
# Create mask for color
mask = cv2.inRange(hsv, np.array(lower), np.array(upper))
masked_img = cv2.bitwise_and(img, img, mask=mask)
bw_masked_img = cv2.cvtColor(masked_img, cv2.COLOR_BGR2GRAY)
# _, bw_masked_img = cv2.threshold(bw_masked_img, 100, 255, cv2.THRESH_BINARY)
bw_masked_img_rgb = cv2.cvtColor(bw_masked_img, cv2.COLOR_GRAY2RGB)

cv2_imshow(bw_masked_img)

small_kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (3, 3))
medium_kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5, 5))
large_kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (7, 7))

cleaned = cv2.morphologyEx(bw_masked_img, cv2.MORPH_OPEN, small_kernel, iterations=2)

cleaned = cv2.morphologyEx(cleaned, cv2.MORPH_CLOSE, medium_kernel, iterations=1)


dist = cv2.distanceTransform(cleaned, cv2.DIST_L2, 5)

thres = cv2.threshold(dist, 0.15*dist.max(), 255, 0)[1].astype(np.uint8)

#cv2_imshow(dist)
#cv2_imshow(thres)

contours, _ = cv2.findContours(thres, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
print(color_name)
# cv2_imshow(cleaned)

markers = np.zeros_like(img)
for i, contour in enumerate(contours):
x, y, w, h = cv2.boundingRect(contour)
area = w*h
# print(area)
# print("area", area)
if area > area_min and area < area_max:
detected_cubes.append({
'color': color_name,
'position': cv2.boundingRect(contour),
'area': area
})


return detected_cubes

img_file = 'cube4.jpg'
cubes = detect_colored_cubes(img_file)


Basically, I am looking at all colors seperately and trying to get components from them by using distance transform. However, I am having difficulty in removing outliers and getting only the useful data from them.

Sometimes, it detects components which span multiple cubes since same colored subes are adjacent. So, I want to ask, how do I exactly get the set of 9 colors from one face as well as leave out the outliers?

img_face = cv2.imread(img_file)
cubes = sorted(cubes, key=lambda x: x['area'], reverse=True)
new = np.zeros((img_face.shape[0], img_face.shape[1], 3), dtype=np.uint8)
for cube in cubes:
new[cube['position'][1]:cube['position'][1] + cube['position'][3], cube['position'][0]:cube['position'][0] + cube['position'][2]] = color_codes[cube['color']]
cv2.putText(new, cube['color'], (cube['position'][0], cube['position'][1] - 10), cv2.FONT_HERSHEY_SIMPLEX, 0.5, color_codes[cube['color']], 2)
cv2_imshow(new)


gives me the following which is majorly correct but has outliers:



Real image: cube4.jpg

Continue reading...