Kuwahara filter in Python (numpy + OpenCV).
The Kuwahara filter is a non-linear smoothing filter used in image processing for adaptive noise reduction. It is able to apply smoothing on the image while preserving the edges. Source: Wikipedia
This implementation provide two variants of the filter:
- The classic one, using a uniform kernel to compute the window mean.
- A gaussian based filter, by computing the window gaussian mean. This is inspired by the ImageMagick approach.
pip install pykuwahara
import cv2
from pykuwahara import kuwahara
image = cv2.imread('lena_std.jpg')
filt1 = kuwahara(image, method='mean', radius=3)
filt2 = kuwahara(image, method='gaussian', radius=3) # default sigma: computed by OpenCV
cv2.imwrite('lena-kfilt-mean.jpg', filt1)
cv2.imwrite('lena-kfilt-gaus.jpg', filt2)
Kuwahara filter can be used to apply a painting effet on pictures.
import cv2
from pykuwahara import kuwahara
image = cv2.imread('photo.jpg')
# Set radius according to the image dimensions and the desired effect
filt1 = kuwahara(image, method='mean', radius=4)
# NOTE: with sigma >= radius, this is equivalent to using 'mean' method
# NOTE: with sigma << radius, the radius has no effect
filt2 = kuwahara(image, method='gaussian', radius=4, sigma=1.5)
cv2.imwrite('photo-kfilt-mean.jpg', filt1)
cv2.imwrite('photo-kfilt-gaus.jpg', filt2)
Original image (source: wikipedia)
Color image are supported by grayscaling the source image and using the gray channel to calculate the variance.
The user can provide another channel at his convenience, and alternatively give the right color conversion code (default is COLOR_BGR2GARY
).
import cv2
from pykuwahara import kuwahara
image = cv2.imread('selfie.jpg')
image = (image / 255).astype('float32') # pykuwahara supports float32 as well
lab_image = cv2.cvtColor(image, cv2.COLOR_BGR2Lab)
l, a, b = cv2.split(lab_image)
hsv_image = cv2.cvtColor(image, cv2.COLOR_BGR2HSV)
h, s, v = cv2.split(hsv_image)
filt1 = kuwahara(image, method='gaussian', radius=5, sigma=2., image_2d=l)
filt2 = kuwahara(image, method='gaussian', radius=5, sigma=2., image_2d=v)
cv2.imwrite('selfie-kfilt-gaus1.jpg', filt1 * 255)
cv2.imwrite('selfie-kfilt-gaus2.jpg', filt2 * 255)