Perlin noise image augmentation

perlin_forward_warp.py

import sys
import numpy as np
import cv2
from perlin_noise import PerlinNoise
import glob

def vector_field(x, y, noise, scale=0.01, w=100):
  noise_x = w * noise([scale*x, scale*y])
  noise_y = w * noise([100+scale*x, scale*y])
  return noise_x, noise_y


if __name__ == '__main__':
    w = 400
    scale = 0.008
    print (len(glob.glob('*.png')),  'images') ## change to refect path
    for img_path in glob.glob('*.png'):
        for n in range(1, 5):
            for w in [300, 400, 500, 600]:
                noise = PerlinNoise(n)
                print ('reading', img_path, 'w:', w, 'n:', n)
                x = cv2.imread(img_path)
                x = x / 255.
                result2d = np.empty_like(x)
                for i in range(x.shape[0]):
                    for j in range(x.shape[1]):
                        new_i, new_j = vector_field(i, j, noise, scale, w)
                        out_x = int(np.clip(np.floor(i+new_i), 0, x.shape[0]-1))
                        out_y = int(np.clip(np.floor(j+new_j), 0, x.shape[1]-1))
                        result2d[out_x, out_y] = x[i, j]
                print (result2d.min(), result2d.max())
                result2d[result2d == 0.] = 1.0
                result2d = (result2d*255.).astype(np.uint8)
                cv2.imwrite(f'{img_path.split(".")[0]}n{n}w{w}.png', result2d)