-
Notifications
You must be signed in to change notification settings - Fork 0
/
generate_corruption.py
462 lines (361 loc) · 16.2 KB
/
generate_corruption.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
# Based on https://github.com/hendrycks/robustness
# --------------------------------------------------
import random
random.seed(123)
import numpy as np
np.random.seed(123)
import torch
torch.manual_seed(123)
import skimage as sk
from skimage.filters import gaussian
from io import BytesIO
from wand.image import Image as WandImage
from wand.api import library as wandlibrary
import ctypes
from PIL import Image as PILImage
import cv2
from scipy.ndimage import zoom as scizoom
from scipy.ndimage.interpolation import map_coordinates
import warnings, math
warnings.simplefilter("ignore", UserWarning)
def disk(radius, alias_blur=0.1, dtype=np.float32):
if radius <= 8:
L = np.arange(-8, 8 + 1)
ksize = (3, 3)
else:
L = np.arange(-radius, radius + 1)
ksize = (5, 5)
X, Y = np.meshgrid(L, L)
aliased_disk = np.array((X ** 2 + Y ** 2) <= radius ** 2, dtype=dtype)
aliased_disk /= np.sum(aliased_disk)
# supersample disk to antialias
return cv2.GaussianBlur(aliased_disk, ksize=ksize, sigmaX=alias_blur)
# Tell Python about the C method
wandlibrary.MagickMotionBlurImage.argtypes = (ctypes.c_void_p, # wand
ctypes.c_double, # radius
ctypes.c_double, # sigma
ctypes.c_double) # angle
# Extend wand.image.Image class to include method signature
class MotionImage(WandImage):
def motion_blur(self, radius=0.0, sigma=0.0, angle=0.0):
wandlibrary.MagickMotionBlurImage(self.wand, radius, sigma, angle)
# modification of https://github.com/FLHerne/mapgen/blob/master/diamondsquare.py
def plasma_fractal(mapsize=256, wibbledecay=3):
"""
Generate a heightmap using diamond-square algorithm.
Return square 2d array, side length 'mapsize', of floats in range 0-255.
'mapsize' must be a power of two.
"""
assert (mapsize & (mapsize - 1) == 0)
maparray = np.empty((mapsize, mapsize), dtype=np.float_)
maparray[0, 0] = 0
stepsize = mapsize
wibble = 100
def wibbledmean(array):
return array / 4 + wibble * np.random.uniform(-wibble, wibble, array.shape)
def fillsquares():
"""For each square of points stepsize apart,
calculate middle value as mean of points + wibble"""
cornerref = maparray[0:mapsize:stepsize, 0:mapsize:stepsize]
squareaccum = cornerref + np.roll(cornerref, shift=-1, axis=0)
squareaccum += np.roll(squareaccum, shift=-1, axis=1)
maparray[stepsize // 2:mapsize:stepsize,
stepsize // 2:mapsize:stepsize] = wibbledmean(squareaccum)
def filldiamonds():
"""For each diamond of points stepsize apart,
calculate middle value as mean of points + wibble"""
mapsize = maparray.shape[0]
drgrid = maparray[stepsize // 2:mapsize:stepsize, stepsize // 2:mapsize:stepsize]
ulgrid = maparray[0:mapsize:stepsize, 0:mapsize:stepsize]
ldrsum = drgrid + np.roll(drgrid, 1, axis=0)
lulsum = ulgrid + np.roll(ulgrid, -1, axis=1)
ltsum = ldrsum + lulsum
maparray[0:mapsize:stepsize, stepsize // 2:mapsize:stepsize] = wibbledmean(ltsum)
tdrsum = drgrid + np.roll(drgrid, 1, axis=1)
tulsum = ulgrid + np.roll(ulgrid, -1, axis=0)
ttsum = tdrsum + tulsum
maparray[stepsize // 2:mapsize:stepsize, 0:mapsize:stepsize] = wibbledmean(ttsum)
while stepsize >= 2:
fillsquares()
filldiamonds()
stepsize //= 2
wibble /= wibbledecay
maparray -= maparray.min()
return maparray / maparray.max()
def clipped_zoom(img, zoom_factor):
h = img.shape[0]
w = img.shape[1]
ch = int(np.ceil(h / zoom_factor))
cw = int(np.ceil(w / zoom_factor))
top = (h - ch) // 2
img_zoomed = scizoom(img[top:top + ch, top:top + cw], (zoom_factor, zoom_factor, 1), order=1)
trim_top = (img_zoomed.shape[0] - h) // 2
trim_top_w = (img.shape[1] - w) // 2
return img_zoomed[trim_top:trim_top + h, trim_top_w:trim_top_w + w]
def gaussian_noise(x, severity=1):
c = [.08, .12, 0.18, 0.26, 0.38][severity - 1]
x = np.array(x) / 255.
return np.clip(x + np.random.normal(size=x.shape, scale=c), 0, 1) * 255
def shot_noise(x, severity=1):
c = [60, 25, 12, 5, 3][severity - 1]
x = np.array(x) / 255.
return np.clip(np.random.poisson(x * c) / c, 0, 1) * 255
def impulse_noise(x, severity=1):
c = [.03, .06, .09, 0.17, 0.27][severity - 1]
x = sk.util.random_noise(np.array(x) / 255., mode='s&p', amount=c, seed=123)
return np.clip(x, 0, 1) * 255
def speckle_noise(x, severity=1):
c = [.15, .2, 0.35, 0.45, 0.6][severity - 1]
x = np.array(x) / 255.
return np.clip(x + x * np.random.normal(size=x.shape, scale=c), 0, 1) * 255
def gaussian_blur(x, severity=1):
c = [1, 2, 3, 4, 6][severity - 1]
x = gaussian(np.array(x) / 255., sigma=c, multichannel=True)
return np.clip(x, 0, 1) * 255
def glass_blur(x, severity=1):
c = [(0.7, 1, 2), (0.9, 2, 1), (1, 2, 3), (1.1, 3, 2), (1.5, 4, 2)][severity - 1]
orig_size = x.size
x = np.uint8(gaussian(np.array(x) / 255., sigma=c[0], multichannel=True) * 255)
for i in range(c[2]):
for h in range(orig_size[1] - c[1], c[1], -1):
for w in range(orig_size[0] - c[1], c[1], -1):
dx, dy = np.random.randint(-c[1], c[1], size=(2,))
h_prime, w_prime = h + dy, w + dx
# swap
x[h, w], x[h_prime, w_prime] = x[h_prime, w_prime], x[h, w]
return np.clip(gaussian(x / 255., sigma=c[0], multichannel=True), 0, 1) * 255
def defocus_blur(x, severity=1):
c = [(3, 0.1), (4, 0.5), (6, 0.5), (8, 0.5), (10, 0.5)][severity - 1]
x = np.array(x) / 255.
kernel = disk(radius=c[0], alias_blur=c[1])
channels = []
for d in range(3):
channels.append(cv2.filter2D(x[:, :, d], -1, kernel))
channels = np.array(channels).transpose((1, 2, 0)) # 3x224x224 -> 224x224x3
return np.clip(channels, 0, 1) * 255
def motion_blur(x, severity=1):
c = [(10, 3), (15, 5), (15, 8), (15, 12), (20, 15)][severity - 1]
output = BytesIO()
x.save(output, format='PNG')
x = MotionImage(blob=output.getvalue())
x.motion_blur(radius=c[0], sigma=c[1], angle=np.random.uniform(-45, 45))
x = cv2.imdecode(np.fromstring(x.make_blob(), np.uint8),
cv2.IMREAD_UNCHANGED)
if len(x.shape) == 3:
return np.clip(x[..., [2, 1, 0]], 0, 255) # BGR to RGB
else:
return np.clip(np.array([x, x, x]).transpose((1, 2, 0)), 0, 255)
def zoom_blur(x, severity=1):
c = [np.arange(1, 1.11, 0.01),
np.arange(1, 1.16, 0.01),
np.arange(1, 1.21, 0.02),
np.arange(1, 1.26, 0.02),
np.arange(1, 1.31, 0.03)][severity - 1]
x = (np.array(x) / 255.).astype(np.float32)
out = np.zeros_like(x)
for zoom_factor in c:
out += clipped_zoom(x, zoom_factor)
x = (x + out) / (len(c) + 1)
return np.clip(x, 0, 1) * 255
def fractal_map_size_mapper(largest_dim):
map_size = 0.
#For AWA2, largest dim after resizing is 913 and smallest dim is 256. Therefore, pick 256,512 or 1024.
if largest_dim == 256:
map_size = 256
elif largest_dim > 256 and largest_dim <= 512:
map_size = 512
else:
map_size = 1024
return map_size
def fog(x, severity=1):
cx = [(1.5, 2), (2, 2), (2.5, 1.7), (2.5, 1.5), (3, 1.4)][severity - 1]
x = np.array(x) / 255.
#Get largest dimension, get a map of size equal to nearest power of two, crop to image size for processing
h = x.shape[0]
w = x.shape[1]
if h >= w:
largest_dim = h
else:
largest_dim = w
map_size = fractal_map_size_mapper(largest_dim)
max_val = x.max()
x += cx[0] * plasma_fractal(mapsize=map_size, wibbledecay=cx[1])[:h, :w][..., np.newaxis]
return np.clip(x * max_val / (max_val + cx[0]), 0, 1) * 255
def round_up(n, decimals=0):
multiplier = 10 ** decimals
return math.ceil(n * multiplier) / multiplier
def frost(x, severity=1):
c = [(1, 0.4),
(0.8, 0.6),
(0.7, 0.7),
(0.65, 0.7),
(0.6, 0.75)][severity - 1]
idx = np.random.randint(5)
filename = ['frost_images/frost1.png', 'frost_images/frost2.png', 'frost_images/frost3.png', 'frost_images/frost4.jpg',
'frost_images/frost5.jpg', 'frost_images/frost6.jpg'][idx]
frost = cv2.imread(filename)
x = np.array(x)
frost_im_size = frost.shape[0:2]
orig_im_size = x.shape[0:2]
# IF our image is, in any dimension than, larger the frost image, upscale the frost image.
scale_ratio = 1.0
if orig_im_size[0] > frost_im_size[0] or orig_im_size[1] > frost_im_size[1]:
if orig_im_size[0] > frost_im_size[0] and orig_im_size[1] <= frost_im_size[1]:
scale_ratio = orig_im_size[0] / frost_im_size[0]
if orig_im_size[1] > frost_im_size[1] and orig_im_size[0] <= frost_im_size[0]:
scale_ratio = orig_im_size[1] / frost_im_size[1]
scale_ratio = round_up(scale_ratio, 1)
frost = cv2.resize(frost, None, fx=scale_ratio, fy=scale_ratio)
x_start, y_start = np.random.randint(0, frost.shape[0] +1 - orig_im_size[0]), np.random.randint(0, frost.shape[1] +1 - orig_im_size[1])
frost = frost[x_start:x_start + orig_im_size[0], y_start:y_start + orig_im_size[1]][..., [2, 1, 0]]
return np.clip(c[0] * x + c[1] * frost, 0, 255)
def snow(x, severity=1):
c = [(0.1, 0.3, 3, 0.5, 10, 4, 0.8),
(0.2, 0.3, 2, 0.5, 12, 4, 0.7),
(0.55, 0.3, 4, 0.9, 12, 8, 0.7),
(0.55, 0.3, 4.5, 0.85, 12, 8, 0.65),
(0.55, 0.3, 2.5, 0.85, 12, 12, 0.55)][severity - 1]
x = np.array(x, dtype=np.float32) / 255.
h = x.shape[0]
w = x.shape[1]
snow_layer = np.random.normal(size=x.shape[:2], loc=c[0], scale=c[1]) # [:2] for monochrome
snow_layer = clipped_zoom(snow_layer[..., np.newaxis], c[2])
snow_layer[snow_layer < c[3]] = 0
snow_layer = PILImage.fromarray((np.clip(snow_layer.squeeze(), 0, 1) * 255).astype(np.uint8), mode='L')
output = BytesIO()
snow_layer.save(output, format='PNG')
snow_layer = MotionImage(blob=output.getvalue())
snow_layer.motion_blur(radius=c[4], sigma=c[5], angle=np.random.uniform(-135, -45))
snow_layer = cv2.imdecode(np.fromstring(snow_layer.make_blob(), np.uint8),
cv2.IMREAD_UNCHANGED) / 255.
snow_layer = snow_layer[..., np.newaxis]
x = c[6] * x + (1 - c[6]) * np.maximum(x, cv2.cvtColor(x, cv2.COLOR_RGB2GRAY).reshape(h, w, 1) * 1.5 + 0.5)
return np.clip(x[0:snow_layer.shape[0], 0:snow_layer.shape[1]] + snow_layer + np.rot90(snow_layer, k=2), 0, 1) * 255
def spatter(x, severity=1):
c = [(0.65, 0.3, 4, 0.69, 0.6, 0),
(0.65, 0.3, 3, 0.68, 0.6, 0),
(0.65, 0.3, 2, 0.68, 0.5, 0),
(0.65, 0.3, 1, 0.65, 1.5, 1),
(0.67, 0.4, 1, 0.65, 1.5, 1)][severity - 1]
x = np.array(x, dtype=np.float32) / 255.
liquid_layer = np.random.normal(size=x.shape[:2], loc=c[0], scale=c[1])
liquid_layer = gaussian(liquid_layer, sigma=c[2])
liquid_layer[liquid_layer < c[3]] = 0
if c[5] == 0:
liquid_layer = (liquid_layer * 255).astype(np.uint8)
dist = 255 - cv2.Canny(liquid_layer, 50, 150)
dist = cv2.distanceTransform(dist, cv2.DIST_L2, 5)
_, dist = cv2.threshold(dist, 20, 20, cv2.THRESH_TRUNC)
dist = cv2.blur(dist, (3, 3)).astype(np.uint8)
dist = cv2.equalizeHist(dist)
ker = np.array([[-2, -1, 0], [-1, 1, 1], [0, 1, 2]])
dist = cv2.filter2D(dist, cv2.CV_8U, ker)
dist = cv2.blur(dist, (3, 3)).astype(np.float32)
m = cv2.cvtColor(liquid_layer * dist, cv2.COLOR_GRAY2BGRA)
m /= np.max(m, axis=(0, 1))
m *= c[4]
color = np.concatenate((175 / 255. * np.ones_like(m[..., :1]),
238 / 255. * np.ones_like(m[..., :1]),
238 / 255. * np.ones_like(m[..., :1])), axis=2)
color = cv2.cvtColor(color, cv2.COLOR_BGR2BGRA)
x = cv2.cvtColor(x, cv2.COLOR_BGR2BGRA)
return cv2.cvtColor(np.clip(x + m * color, 0, 1), cv2.COLOR_BGRA2BGR) * 255
else:
m = np.where(liquid_layer > c[3], 1, 0)
m = gaussian(m.astype(np.float32), sigma=c[4])
m[m < 0.8] = 0
# m = np.abs(m) ** (1/c[4])
# mud brown
color = np.concatenate((63 / 255. * np.ones_like(x[..., :1]),
42 / 255. * np.ones_like(x[..., :1]),
20 / 255. * np.ones_like(x[..., :1])), axis=2)
color *= m[..., np.newaxis]
x *= (1 - m[..., np.newaxis])
return np.clip(x + color, 0, 1) * 255
def contrast(x, severity=1):
c = [0.4, .3, .2, .1, .05][severity - 1]
x = np.array(x) / 255.
means = np.mean(x, axis=(0, 1), keepdims=True)
return np.clip((x - means) * c + means, 0, 1) * 255
def brightness(x, severity=1):
c = [.1, .2, .3, .4, .5][severity - 1]
x = np.array(x) / 255.
x = sk.color.rgb2hsv(x)
x[:, :, 2] = np.clip(x[:, :, 2] + c, 0, 1)
x = sk.color.hsv2rgb(x)
return np.clip(x, 0, 1) * 255
def saturate(x, severity=1):
c = [(0.3, 0), (0.1, 0), (2, 0), (5, 0.1), (20, 0.2)][severity - 1]
x = np.array(x) / 255.
x = sk.color.rgb2hsv(x)
x[:, :, 1] = np.clip(x[:, :, 1] * c[0] + c[1], 0, 1)
x = sk.color.hsv2rgb(x)
return np.clip(x, 0, 1) * 255
def jpeg_compression(x, severity=1):
c = [25, 18, 15, 10, 7][severity - 1]
output = BytesIO()
x.save(output, 'JPEG', quality=c)
x = PILImage.open(output)
return x
def pixelate(x, severity=1):
c = [0.6, 0.5, 0.4, 0.3, 0.25][severity - 1]
orig_size = (x.size[0], x.size[1])
x = x.resize((int(orig_size[0] * c), int(orig_size[1]* c)), PILImage.BOX)
x = x.resize((orig_size[0], orig_size[1]), PILImage.BOX)
return x
# mod of https://gist.github.com/erniejunior/601cdf56d2b424757de5
def elastic_transform(image, severity=1):
c = [(244 * 2, 244 * 0.7, 244 * 0.1),
(244 * 2, 244 * 0.08, 244 * 0.2),
(244 * 0.05, 244 * 0.01, 244 * 0.02),
(244 * 0.07, 244 * 0.01, 244 * 0.02),
(244 * 0.12, 244 * 0.01, 244 * 0.02)][severity - 1]
image = np.array(image, dtype=np.float32) / 255.
shape = image.shape
shape_size = shape[:2]
# random affine
center_square = np.float32(shape_size) // 2
square_size = min(shape_size) // 3
pts1 = np.float32([center_square + square_size,
[center_square[0] + square_size, center_square[1] - square_size],
center_square - square_size])
pts2 = pts1 + np.random.uniform(-c[2], c[2], size=pts1.shape).astype(np.float32)
M = cv2.getAffineTransform(pts1, pts2)
image = cv2.warpAffine(image, M, shape_size[::-1], borderMode=cv2.BORDER_REFLECT_101)
dx = (gaussian(np.random.uniform(-1, 1, size=shape[:2]),
c[1], mode='reflect', truncate=3) * c[0]).astype(np.float32)
dy = (gaussian(np.random.uniform(-1, 1, size=shape[:2]),
c[1], mode='reflect', truncate=3) * c[0]).astype(np.float32)
dx, dy = dx[..., np.newaxis], dy[..., np.newaxis]
x, y, z = np.meshgrid(np.arange(shape[1]), np.arange(shape[0]), np.arange(shape[2]))
indices = np.reshape(y + dy, (-1, 1)), np.reshape(x + dx, (-1, 1)), np.reshape(z, (-1, 1))
return np.clip(map_coordinates(image, indices, order=1, mode='reflect').reshape(shape), 0, 1) * 255
def get_corruptions():
import collections
d = collections.OrderedDict()
#Test corruptions
#Noise category
d['gaussian_noise'] = gaussian_noise
d['shot_noise'] = shot_noise
d['impulse_noise'] = impulse_noise
#Blur category
d['zoom_blur'] = zoom_blur
d['glass_blur'] = glass_blur
d['defocus_blur'] = defocus_blur
d['motion_blur'] = motion_blur
#Weather category
d['brightness'] = brightness
d['snow'] = snow
d['fog'] = fog
d['frost'] = frost
#Digital category
d['contrast'] = contrast
d['jpeg_compression'] = jpeg_compression
d['pixelate'] = pixelate
d['elastic_transform'] = elastic_transform
# Validation corruptions, one for each category
d['speckle_noise'] = speckle_noise
d['gaussian_blur'] = gaussian_blur
d['spatter'] = spatter
d['saturate'] = saturate
return d