使用 python opencv 进行照明标准化

Question

I am doing a project in image classification with deep learning, and want to implement the illumination normalization which provided in section 4.4 of paper https://arxiv.org/pdf/1907.09449.pdf ， but I didn't get the expected results.

In section 4.4 of paper, it using gaussian kerenl filter to get the background image, then subtract it from Y channel in YCrCb color space, I try to reproduce the result but is not the same as paper do in Figure 4

below is my code

image = cv2.imread(file)
cv2.imshow("origin", image)

# illumination normalize
ycrcb = cv2.cvtColor(image, cv2.COLOR_BGR2YCrCb)
y, cr, cb = cv2.split(ycrcb)

# get background which paper says (gaussian method using standard deviation 5 pixel)
gaussian = cv2.GaussianBlur(y, (0, 0), 5, 5)

# subtract background from Y channel which paper says
y = y - gaussian
ycrcb = cv2.merge([y, cr, cb])

output = cv2.cvtColor(ycrcb, cv2.COLOR_YCrCb2BGR)
cv2.imshow("output", output)

but the output result image (see below) is not the same as paper says (Figure 4)

original image -> illumination normalized image

does anyone knows how to do that? thank you for your greate help:-)

Answer 1

You have two issues. First, you need to adjust the sigma for the Gaussian blur in proportion to the image size vs the 299 pixel size used in the reference. Second you need to bias the difference near mid-gray. Adjust the bias as desired for brightness.

So here is how I achieved it in Python/OpenCV.

Input:

import cv2
import numpy as np

# read input
image = cv2.imread('retina2.jpg')
hh, ww = image.shape[:2]
print(hh, ww)
max = max(hh, ww)

# illumination normalize
ycrcb = cv2.cvtColor(image, cv2.COLOR_BGR2YCrCb)

# separate channels
y, cr, cb = cv2.split(ycrcb)

# get background which paper says (gaussian blur using standard deviation 5 pixel for 300x300 size image)
# account for size of input vs 300
sigma = int(5 * max / 300)
print('sigma: ',sigma)
gaussian = cv2.GaussianBlur(y, (0, 0), sigma, sigma)

# subtract background from Y channel
y = (y - gaussian + 100)

# merge channels back
ycrcb = cv2.merge([y, cr, cb])

#convert to BGR
output = cv2.cvtColor(ycrcb, cv2.COLOR_YCrCb2BGR)

# save results
cv2.imwrite('retina2_proc.jpg', output)

# show results
cv2.imshow("output", output)
cv2.waitKey(0)

Results: