White Balance a photo from a known point - python-3.x

Question

White Balancing is a rather well-covered topic, but most of the answers I have seen cover automatic white balancing techniques for an entire image that does not have a known point for what is white, gray, and black. I cannot seem to find many that cover white balancing from a known point. I have the script (below) that takes an image of a color card (Spyder Checkr 48) and returns the white, 20% Gray, and Black color card blocks:
Color L A B sR sG sB aR aG aB
Card White 96.04 2.16 2.6 249 242 238 247 242 237
20% Gray 80.44 1.17 2.05 202 198 195 199 196 193
Card Black 16.91 1.43 -0.81 43 41 43 46 46 47
Question: Since I know the ground truth LAB, sRGB and AdobeRGB values for specific parts of the image, what would be the best way to white balance the image?
Here is a link to the images I am working with. This is the code for extracting the color card blocks (I currently am running this on Windows, Python 3.7):
from __future__ import print_function
import cv2
import imutils
import numpy as np
from matplotlib import pyplot as plt
import os
import sys
image = cv2.imread("PATH_TO_IMAGE")
template = cv2.imread("PATH_TO_TEMPLATE")
rtemplate = cv2.imread("PATH_TO_RIGHT_TEMPLATE")
def sift(image):
sift = cv2.xfeatures2d.SIFT_create()
kp, des = sift.detectAndCompute(image, None)
return kp, des
def sift_match(im1, im2, vis=False, save=False):
MIN_MATCH_COUNT = 10
FLANN_INDEX_KDTREE = 0
kp1, des1 = sift(im1)
kp2, des2 = sift(im2)
index_params = dict(algorithm=FLANN_INDEX_KDTREE, trees=7)
search_params = dict(checks=100)
flann = cv2.FlannBasedMatcher(index_params, search_params)
matches = flann.knnMatch(des1, des2, k=2)
# Need to draw only good matches, so create a mask
matchesMask = [[0, 0] for i in range(len(matches))]
if vis is True:
draw_params = dict(matchColor=(0, 255, 0),
singlePointColor=(255, 0, 0),
matchesMask=matchesMask,
flags=0)
im3 = cv2.drawMatchesKnn(im1, kp1, im2, kp2, matches, None, **draw_params)
if save:
cv2.imwrite("tempSIFT_Match.png", im3)
plt.imshow(im3), plt.show()
good = []
for m, n in matches:
if m.distance < 0.75 * n.distance:
good.append(m)
return kp1, des1, kp2, des2, good
def smartextractor(im1, im2, vis=False):
# Detect features and compute descriptors.
kp1, d1, kp2, d2, matches = sift_match(im1, im2, vis)
kp1 = np.asarray(kp1)
kp2 = np.asarray(kp2)
# Extract location of good matches
points1 = np.zeros((len(matches), 2), dtype=np.float32)
points2 = np.zeros((len(matches), 2), dtype=np.float32)
for i, match in enumerate(matches):
points1[i, :] = kp1[match.queryIdx].pt
points2[i, :] = kp2[match.trainIdx].pt
# Find homography
h, mask = cv2.findHomography(points1, points2, cv2.RANSAC)
if h is None:
print("could not find homography")
return None, None
# Use homography
height, width, channels = im2.shape
im1Reg = cv2.warpPerspective(im1, h, (width, height))
return im1Reg, h
def show_images(images, cols=1, titles=None):
"""
Display a list of images in a single figure with matplotlib.
"""
assert ((titles is None) or (len(images) == len(titles)))
n_images = len(images)
if titles is None: titles = ['Image (%d)' % i for i in range(1, n_images + 1)]
fig = plt.figure()
for n, (image, title) in enumerate(zip(images, titles)):
a = fig.add_subplot(cols, np.ceil(n_images / float(cols)), n + 1)
if image.ndim == 2:
plt.gray()
plt.imshow(image)
a.set_title(title)
fig.set_size_inches(np.array(fig.get_size_inches()) * n_images)
plt.show()
def Sobel(img, bilateralFilter=True):
# timestart = time.clock()
try:
img = cv2.imread(img, 0)
except TypeError:
None
try:
rheight, rwidth, rdepth = img.shape
img1 = cv2.cvtColor(img, cv2.COLOR_BGR2HSV)
except ValueError:
raise TypeError
# cv2.imwrite('temp.png',img)
_, s, v = cv2.split(img1)
b, g, r = cv2.split(img)
if bilateralFilter is True:
s = cv2.bilateralFilter(s, 11, 17, 17)
v = cv2.bilateralFilter(v, 11, 17, 17)
b = cv2.bilateralFilter(b, 11, 17, 17)
g = cv2.bilateralFilter(g, 11, 17, 17)
r = cv2.bilateralFilter(r, 11, 17, 17)
# calculate sobel in x,y,diagonal directions with the following kernels
sobelx = np.array([[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]], dtype=np.float32)
sobely = np.array([[-1, -2, -1], [0, 0, 0], [1, 2, 1]], dtype=np.float32)
sobeldl = np.array([[0, 1, 2], [-1, 0, 1], [-2, -1, 0]], dtype=np.float32)
sobeldr = np.array([[2, 1, 0], [1, 0, -1], [0, -1, -2]], dtype=np.float32)
# calculate the sobel on value of hsv
gx = cv2.filter2D(v, -1, sobelx)
gy = cv2.filter2D(v, -1, sobely)
gdl = cv2.filter2D(v, -1, sobeldl)
gdr = cv2.filter2D(v, -1, sobeldr)
# combine sobel on value of hsv
xylrv = 0.25 * gx + 0.25 * gy + 0.25 * gdl + 0.25 * gdr
# calculate the sobel on saturation of hsv
sx = cv2.filter2D(s, -1, sobelx)
sy = cv2.filter2D(s, -1, sobely)
sdl = cv2.filter2D(s, -1, sobeldl)
sdr = cv2.filter2D(s, -1, sobeldr)
# combine sobel on value of hsv
xylrs = 0.25 * sx + 0.25 * sy + 0.25 * sdl + 0.25 * sdr
# combine value sobel and saturation sobel
xylrc = 0.5 * xylrv + 0.5 * xylrs
xylrc[xylrc < 6] = 0
# calculate the sobel on value on green
grx = cv2.filter2D(g, -1, sobelx)
gry = cv2.filter2D(g, -1, sobely)
grdl = cv2.filter2D(g, -1, sobeldl)
grdr = cv2.filter2D(g, -1, sobeldr)
# combine sobel on value on green
xylrgr = 0.25 * grx + 0.25 * gry + 0.25 * grdl + 0.25 * grdr
# calculate the sobel on blue
bx = cv2.filter2D(b, -1, sobelx)
by = cv2.filter2D(b, -1, sobely)
bdl = cv2.filter2D(b, -1, sobeldl)
bdr = cv2.filter2D(b, -1, sobeldr)
# combine sobel on value on blue
xylrb = 0.25 * bx + 0.25 * by + 0.25 * bdl + 0.25 * bdr
# calculate the sobel on red
rx = cv2.filter2D(r, -1, sobelx)
ry = cv2.filter2D(r, -1, sobely)
rdl = cv2.filter2D(r, -1, sobeldl)
rdr = cv2.filter2D(r, -1, sobeldr)
# combine sobel on value on red
xylrr = 0.25 * rx + 0.25 * ry + 0.25 * rdl + 0.25 * rdr
# combine value sobel and saturation sobel
xylrrgb = 0.33 * xylrgr + 0.33 * xylrb + 0.33 * xylrr
xylrrgb[xylrrgb < 6] = 0
# combine HSV and RGB sobel outputs
xylrc = 0.5 * xylrc + 0.5 * xylrrgb
xylrc[xylrc < 6] = 0
xylrc[xylrc > 25] = 255
return xylrc
print("extracting image")
extractedImage, _ = smartextractor(image, template)
print("extracting right image")
rextractedImage, _ = smartextractor(extractedImage, rtemplate, vis=False)
grextractedImage = cv2.cvtColor(rextractedImage, cv2.COLOR_BGR2GRAY)
bfsobelImg = Sobel(rextractedImage)
sobelImg = Sobel(rextractedImage, bilateralFilter=False)
csobelImg = cv2.add(bfsobelImg, sobelImg)
csobelImg[csobelImg < 6] = 0
csobelImg[csobelImg > 18] = 255
csobelImg = csobelImg.astype(np.uint8)
img2 = csobelImg.copy()
ret, thresh = cv2.threshold(img2, 18, 255, 0)
contours = cv2.findContours(thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
contours = imutils.grab_contours(contours)
contours = sorted(contours, key=cv2.contourArea, reverse=True)
count = 0
trigger = False
for c in contours:
# approximate the contour
peri = cv2.arcLength(c, True)
contours[count] = cv2.approxPolyDP(c, 0.05 * peri, True)
if len(contours[count]) == 4:
if trigger is False:
screenCnt = contours[count]
trigger = True
count += 1
tl = screenCnt[0]
tr = screenCnt[1]
bl = screenCnt[3]
br = screenCnt[2]
tLy, tLx = tl[0]
tRy, tRx = tr[0]
bLy, bLx = bl[0]
bRy, bRx = br[0]
ratio = .15
realSpace = (3/16)
boxwidth = int(((tRx - tLx) + (bRx - bLx))*.5 - (tLx + bLx)*.5)
boxheight = int(((bRy - tRy) + (bLy - tLy))*.5 - (tRy + tLy)*.5)
spaceWidth = int((boxwidth + boxheight)*.5*realSpace)
boxcenter = [int(((bRy - tRy)*.5 + (bLy - tLy)*.5)*.5), int(((tRx - tLx)*.5 + (bRx - bLx)*.5)*.5)]
roitl = [boxcenter[0] - int(ratio*boxheight), boxcenter[1] - int(ratio*boxwidth)]
roitr = [boxcenter[0] - int(ratio*boxheight), boxcenter[1] + int(ratio*boxwidth)]
roibl = [boxcenter[0] + int(ratio*boxheight), boxcenter[1] - int(ratio*boxwidth)]
roibr = [boxcenter[0] + int(ratio*boxheight), boxcenter[1] + int(ratio*boxwidth)]
spacing = int((boxwidth + boxheight)*.5)+spaceWidth
roiWhite = np.array((roitl, roitr, roibr, roibl))
roiGray = np.array(([roitl[1], roitl[0]+spacing*1], [roitr[1], roitr[0]+spacing*1],
[roibr[1], roibr[0]+spacing*1], [roibl[1], roibl[0]+spacing*1]))
roiBlack = np.array(([roitl[1], roitl[0]+spacing*6], [roitr[1], roitr[0]+spacing*6],
[roibr[1], roibr[0]+spacing*6], [roibl[1], roibl[0]+spacing*6]))
whiteAvgb, whiteAvgg, whiteAvgr, _ = cv2.mean(rextractedImage[(roitl[0]+spacing*0):(roibr[0]+spacing*0),
roitl[1]:roibr[1]])
grayAvgb, grayAvgg, grayAvgr, _ = cv2.mean(rextractedImage[(roitl[0]+spacing*1):(roibr[0]+spacing*1),
roitl[1]:roibr[1]])
blackAvgb, blackAvgg, blackAvgr, _ = cv2.mean(rextractedImage[(roitl[0]+spacing*6):(roibr[0]+spacing*6),
roitl[1]:roibr[1]])
whiteROI = rextractedImage[(roitl[0]+spacing*0):(roibr[0]+spacing*0), roitl[1]:roibr[1]]
grayROI = rextractedImage[(roitl[0]+spacing*1):(roibr[0]+spacing*1), roitl[1]:roibr[1]]
blackROI = rextractedImage[(roitl[0]+spacing*6):(roibr[0]+spacing*6), roitl[1]:roibr[1]]
imageList = [whiteROI, grayROI, blackROI]
show_images(imageList, cols=1)
correctedImage = rextractedImage.copy()
whiteROI[:, :, 0] = whiteAvgb
whiteROI[:, :, 1] = whiteAvgg
whiteROI[:, :, 2] = whiteAvgr
grayROI[:, :, 0] = grayAvgb
grayROI[:, :, 1] = grayAvgg
grayROI[:, :, 2] = grayAvgr
blackROI[:, :, 0] = blackAvgb
blackROI[:, :, 1] = blackAvgg
blackROI[:, :, 2] = blackAvgr
imageList = [whiteROI, grayROI, blackROI]
show_images(imageList, cols=1)
# SPYDER COLOR CHECKR Values: http://www.bartneck.de/2017/10/24/patch-color-definitions-for-datacolor-spydercheckr-48/
blank = np.zeros_like(csobelImg)
maskedImg = blank.copy()
maskedImg = cv2.fillConvexPoly(maskedImg, roiWhite, 255)
maskedImg = cv2.fillConvexPoly(maskedImg, roiGray, 255)
maskedImg = cv2.fillConvexPoly(maskedImg, roiBlack, 255)
res = cv2.bitwise_and(rextractedImage, rextractedImage, mask=maskedImg)
# maskedImg = cv2.fillConvexPoly(maskedImg, roi2Black, 255)
cv2.drawContours(blank, contours, -1, 255, 3)
outputSquare = np.zeros_like(csobelImg)
cv2.drawContours(outputSquare, [screenCnt], -1, 255, 3)
imageList = [rextractedImage, grextractedImage, bfsobelImg, sobelImg, csobelImg, blank, outputSquare, maskedImg, res]
show_images(imageList, cols=3)
sys.exit()

Answer 1

Given the RGB value of a white patch, the image can be corrected for white balance by dividing by that value. That is, applying a linear transformation that makes the white patch have the same level in the three channels:
lum = (whiteR + whiteG + whiteB)/3
imgR = imgR * lum / whiteR
imgG = imgG * lum / whiteG
imgB = imgB * lum / whiteB
Multiplying by lum makes it so that the average intensity doesn’t change.
(The computation of lum will be better if using proper weights: 0.2126, 0.7152, 0.0722, but I wanted to keep it simple. This would only make a big difference if the input white is way off the mark, in which case you'll have other issues too.)
Note that this transformation is best applied in linear RGB space. Both the image and the RGB values for white should first be converted to linear RGB if the image is stored in sRGB or similar (a raw image from the camera would be linear RGB, a JPEG would be sRGB). See here for the relevant equations.
For better precision, you can apply the above using also the RGB values of the grey patch. Take the average multiplication factor (whiteR/lum) derived from the white and grey patches, for each channel, and apply those to the image.
The black level could be subtracted from the image, prior to determining the white RGB values and correcting for white balance. This will improve contrast and color perception, but not part of white balancing.
A full color correction is way more complex, I will not go into that.