2nd variation on Perlin Noise (with Python code)

[Michael Lee]

One of the disadvantage of the first variation I tried was that the resizing is naturally tied to the grid of pixels of the image.

Therefore, I figured out a similar idea without this artifact. Although the pictures (with pseudorandom noise) are not quite as cool, it does have a very "fractally" nature to it.

 

 

#---------------------------------
# Original author: Michael S. Lee
# Date: 9/18/2021
#---------------------------------

import numpy as np
import imageio
import cv2
from scipy.signal import convolve,get_window
import matplotlib.pyplot as plt

def shrink(x, gamma = 30):
    beta = gamma / (x.std())
    y = 1.0 / (1.0 + np.exp(-beta*(x-x.mean())))
    return(y)
    
N = 1024 # Image size

noise = np.random.rand(N,N,3) # White noise in RGB channels
noise2 = np.copy(noise) # setup output array
noise2 = cv2.resize(noise2,(1024,1024))
noise2 = noise
image = noise2

blur = np.copy(image)
sum1 = np.zeros_like(image)
thresh = 0.375

i = 1024

num_levels = 19
factor = 3/4

for j in range(num_levels):
   
   blur2 = blur
   ii = ((i+1) // 2) * 2
   blur = cv2.GaussianBlur(noise2, (ii-1,ii-1), 0) #convolve(noise2,kernel3,'same')
   
   if (j > 0):
     diff = blur - blur2
     diff = shrink(diff, gamma = 25)
     sum1 = sum1 + (i**0.25)*diff 
   #sum1 = sum1 + diff
   print (i)
   i = int(i * factor)
   
sum1 = (sum1 - sum1.min())/ (sum1.max()-sum1.min())
sum1 = sum1 ** 3

# Save on disk
index = np.random.randint(0,262144) # Random filename
imageio.imsave('temp.'+str(index)+'.png',sum1)