I came across an image filter with a clearly Japanese name called Kuwahara filter, and when I tried it, it was amazing, so I will introduce it.

Introduction

** What is Kuwahara filter ** The Kuwahara filter is a kind of smoothing filter devised by a university professor named Michiyoshi Kuwahara (Wikipedia says). If you want to see, take a look at the bottom of the article) Kuwahara filter -Wikipedia Data processing for SPECT (original paper?)

Contents of Kuwahara filter

https://upload.wikimedia.org/wikipedia/commons/4/49/Kuwahara.jpg

To briefly explain the Kuwahara filter, This filter uses the color of each pixel as the average color of the area with the smallest sum of variances in the upper left, upper right, lower left, and lower right square areas of any width around it.

In the image above, to determine the color of the center pixel,

1. Calculate the variance of the ** a ** area, the ** b ** area, the ** c ** area, and the ** d ** area for each RGB color, and for each area. Add all RGB to.
2. Compare the variances obtained in 1 and calculate the average of the colors in the smallest area for each RGB.
3. Set the average color obtained in 2 as the color of the center pixel.

Go through the steps above or the process to get equivalent results. Also, in the image, one side of the square area is 3 pixels, but it can be any width.

Implementation

** Definition **

import numpy as np
import cv2

 def kuwahara (pic, r = 5, resize = False, rate = 0.5): #Original image, square area width -1, Resize or resize ratio
    h,w,_=pic.shape
    if resize:pic=cv2.resize(pic,(int(w*rate),int(h*rate)));h,w,_=pic.shape
    filtered_pic=np.empty_like(pic)
    pic=np.pad(pic,((r,r),(r,r),(0,0)),"edge")
    ave,var=cv2.integral2(pic)
 ave = (ave [: -r-1,: -r-1] + ave [r + 1 :, r + 1:]-ave [r + 1 :,: -r-1] -ave [: -r -1, r + 1 :]) / (r + 1) ** 2 #Batch derivation of average color of area
 var = ((var [: -r-1,: -r-1] + var [r + 1 :, r + 1:]-var [r + 1 :,: -r-1]-var [:- r-1, r + 1:]) / (r + 1) ** 2-ave ** 2) .sum (axis = 2) #Batch derivation of regional variance

    for i in range(h):
        for j in range(w):
 filtered_pic [i, j] = np.array ([ave [i, j], ave [i + r, j], ave [i, j + r], ave [i + r, j + r]]) [ np.array ([var [i, j], var [i + r, j], var [i, j + r], var [i + r, j + r]]). argmin ()] #color Decision

    return filtered_pic

** Run **

import matplotlib.pyplot as plt

 pic = np.array (plt.imread ("input_picture.png ")) Change # input_picture.png to the path of the image you want to filter
filtered_pic=kuwahara(pic,7,True,0.2)
plt.imshow(filtered_pic)

result

Here are some of the photos I took when I went on a trip to France last year.

The original image

** After applying the filter ** ** A smaller square area ver. **

People, people, people, people ＞ ** Serious painting ** < ￣Y^Y^Y^Y^Y^Y^￣

Summary

I introduced it because I was impressed with the texture like drawing on canvas. It's easy so give it a try. I couldn't think of a method that doesn't use the for statement this time, so if the resizing setting is off for a large image, it will take some time. Don't forget to set the third argument to True.