forked from yzhouas/PD-Denoising-pytorch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
utils.py
529 lines (480 loc) · 20 KB
/
utils.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
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
import math
import torch
import torch.nn as nn
import numpy as np
from skimage.measure.simple_metrics import compare_psnr
from torch.autograd import Variable
import cv2
import scipy.ndimage
import scipy.io as sio
import matplotlib as mpl
mpl.use('Agg')
import matplotlib.pyplot as plt
def weights_init_kaiming(m):
classname = m.__class__.__name__
if classname.find('Conv') != -1:
nn.init.kaiming_normal(m.weight.data, a=0, mode='fan_in')
elif classname.find('Linear') != -1:
nn.init.kaiming_normal(m.weight.data, a=0, mode='fan_in')
elif classname.find('BatchNorm') != -1:
# nn.init.uniform(m.weight.data, 1.0, 0.02)
m.weight.data.normal_(mean=0, std=math.sqrt(2./9./64.)).clamp_(-0.025,0.025)
nn.init.constant(m.bias.data, 0.0)
def batch_PSNR(img, imclean, data_range):
Img = img.data.cpu().numpy().astype(np.float32)
Iclean = imclean.data.cpu().numpy().astype(np.float32)
PSNR = 0
for i in range(Img.shape[0]):
PSNR += compare_psnr(Iclean[i,:,:,:], Img[i,:,:,:], data_range=data_range)
return (PSNR/Img.shape[0])
def data_augmentation(image, mode):
out = np.transpose(image, (1,2,0))
if mode == 0:
# original
out = out
elif mode == 1:
# flip up and down
out = np.flipud(out)
elif mode == 2:
# rotate counterwise 90 degree
out = np.rot90(out)
elif mode == 3:
# rotate 90 degree and flip up and down
out = np.rot90(out)
out = np.flipud(out)
elif mode == 4:
# rotate 180 degree
out = np.rot90(out, k=2)
elif mode == 5:
# rotate 180 degree and flip
out = np.rot90(out, k=2)
out = np.flipud(out)
elif mode == 6:
# rotate 270 degree
out = np.rot90(out, k=3)
elif mode == 7:
# rotate 270 degree and flip
out = np.rot90(out, k=3)
out = np.flipud(out)
return np.transpose(out, (2,0,1))
def visual_va2np(Out, mode=1, ps=0, pss=1, scal=1, rescale=0, w=10, h=10, c=3, refill=0, refill_img=0, refill_ind=[0, 0]):
if mode == 0 or mode == 1 or mode==3:
out_numpy = Out.data.squeeze(0).cpu().numpy()
elif mode == 2:
out_numpy = Out.data.squeeze(1).cpu().numpy()
if out_numpy.shape[0] == 1:
out_numpy = np.tile(out_numpy, (3, 1, 1))
if mode == 0 or mode == 1:
out_numpy = (np.transpose(out_numpy, (1, 2, 0))) * 255.0 * scal
else:
out_numpy = (np.transpose(out_numpy, (1, 2, 0)))
if ps == 1:
out_numpy = reverse_pixelshuffle(out_numpy, pss, refill, refill_img, refill_ind)
if rescale == 1:
out_numpy = cv2.resize(out_numpy, (h, w))
print(out_numpy.shape)
return out_numpy
def temp_ps_4comb(Out, In):
pass
def np2ts(x, mode=0): #now assume the input only has one channel which is ignored
w, h, c= x.shape
x_ts = x.transpose(2, 0, 1)
x_ts = torch.from_numpy(x_ts).type(torch.FloatTensor)
if mode == 0 or mode == 1:
x_ts = x_ts.unsqueeze(0)
elif mode == 2:
x_ts = x_ts.unsqueeze(1)
return x_ts
def np2ts_4d(x):
x_ts = x.transpose(0, 3, 1, 2)
x_ts = torch.from_numpy(x_ts).type(torch.FloatTensor)
return x_ts
def get_salient_noise_in_maps(lm, thre = 0., chn=3):
'''
Description: To find out the most frequent estimated noise level in the images
----------
[Input]
a multi-channel tensor of noise map
[Output]
A list of noise level value
'''
lm_numpy = lm.data.cpu().numpy()
lm_numpy = (np.transpose(lm_numpy, (0, 2, 3, 1)))
nl_list = np.zeros((lm_numpy.shape[0], chn,1))
for n in range(lm_numpy.shape[0]):
for c in range(chn):
selected_lm = np.reshape(lm_numpy[n,:,:,c], (lm_numpy.shape[1]*lm_numpy.shape[2], 1))
selected_lm = selected_lm[selected_lm>thre]
if selected_lm.shape[0] == 0:
nl_list[n, c] = 0
else:
hist = np.histogram(selected_lm, density=True)
nl_ind = np.argmax(hist[0])
#print(nl_ind)
#print(hist[0])
#print(hist[1])
nl = ( hist[1][nl_ind] + hist[1][nl_ind+1] ) / 2.
nl_list[n, c] = nl
return nl_list
def get_cdf_noise_in_maps(lm, thre=0.8, chn=3):
'''
Description: To find out the most frequent estimated noise level in the images
----------
[Input]
a multi-channel tensor of noise map
[Output]
A list of noise level value
'''
lm_numpy = lm.data.cpu().numpy()
lm_numpy = (np.transpose(lm_numpy, (0, 2, 3, 1)))
nl_list = np.zeros((lm_numpy.shape[0], chn,1))
for n in range(lm_numpy.shape[0]):
for c in range(chn):
selected_lm = np.reshape(lm_numpy[n,:,:,c], (lm_numpy.shape[1]*lm_numpy.shape[2], 1))
H, x = np.histogram(selected_lm, normed=True)
dx = x[1]-x[0]
F = np.cumsum(H)*dx
F_ind = np.where(F>0.9)[0][0]
nl_list[n, c] = x[F_ind]
print(nl_list[n,c])
return nl_list
def get_pdf_in_maps(lm, mark, chn=1):
'''
Description: get the noise estimation cdf of each channel
----------
[Input]
a multi-channel tensor of noise map and channel dimension
chn: the channel number for gaussian
[Output]
CDF function of each sample and each channel
'''
lm_numpy = lm.data.cpu().numpy()
lm_numpy = (np.transpose(lm_numpy, (0, 2, 3, 1)))
pdf_list = np.zeros((lm_numpy.shape[0], chn, 10))
for n in range(lm_numpy.shape[0]):
for c in range(chn):
selected_lm = np.reshape(lm_numpy[n,:,:,c], (lm_numpy.shape[1]*lm_numpy.shape[2], 1))
H, x = np.histogram(selected_lm, range=(0.,1.), bins=10, normed=True)
dx = x[1]-x[0]
F = H * dx
pdf_list[n, c, :] = F
#sio.savemat(mark + str(c) + '.mat',{'F':F})
plt.bar(range(10), F)
#plt.savefig(mark + str(c) + '.png')
plt.close()
return pdf_list
def get_pdf_matching_score(F1, F2):
'''
Description: Given two sets of CDF, get the overall matching score for each channel
-----------
[Input] F1, F2
[Output] score for each channel
'''
return np.mean((F1-F2)**2)
def decide_scale_factor(noisy_image, estimation_model, color=1, thre = 0, plot_flag = 1, stopping = 4, mark=''):
'''
Description: Given a noisy image and the noise estimation model, keep multiscaling the image\\
using pixel-shuffle methods, and estimate the pdf and cdf of AWGN channel
Compare the changes of the density function and decide the optimal scaling factor
------------
[Input] noisy_image, estimation_model, plot_flag, stopping
[Output] plot the middle vector
score_seq: the matching score sequence between the two subsequent pdf
opt_scale: the optimal scaling factor
'''
if color == 1:
c = 3
elif color == 0:
c = 1
score_seq = []
Pre_CDF = None
flag = 0
for pss in range(1, stopping+1): #scaling factor from 1 to the limit
noisy_image = pixelshuffle(noisy_image, pss)
INoisy = np2ts(noisy_image, color)
INoisy = Variable(INoisy.cuda(), volatile=True)
EMap = torch.clamp(estimation_model(INoisy), 0., 1.)
EPDF = get_pdf_in_maps(EMap, mark + str(pss), c)[0]
if flag != 0:
score = get_pdf_matching_score(EPDF, Pre_PDF) #TODO: How to match these two
print(score)
score_seq.append(score)
if score <= thre:
print('optimal scale is %d:' % (pss-1))
return (pss-1, score_seq)
Pre_PDF = EPDF
flag = 1
return (stopping, score_seq)
def get_max_noise_in_maps(lm, chn=3):
'''
Description: To find out the maximum level of noise level in the images
----------
[Input]
a multi-channel tensor of noise map
[Output]
A list of noise level value
'''
lm_numpy = lm.data.cpu().numpy()
lm_numpy = (np.transpose(lm_numpy, (0, 2, 3, 1)))
nl_list = np.zeros((lm_numpy.shape[0], chn, 1))
for n in range(lm_numpy.shape[0]):
for c in range(chn):
nl = np.amax(lm_numpy[n, :, :, c])
nl_list[n, c] = nl
return nl_list
def get_smooth_maps(lm, dilk = 50, gsd = 10):
'''
Description: To return the refined maps after dilation and gaussian blur
[Input] a multi-channel tensor of noise map
[Output] a multi-channel tensor of refined noise map
'''
kernel = np.ones((dilk, dilk))
lm_numpy = lm.data.squeeze(0).cpu().numpy()
lm_numpy = (np.transpose(lm_numpy, (1, 2, 0)))
ref_lm_numpy = lm_numpy.copy() #a refined map
for c in range(lm_numpy.shape[2]):
nmap = lm_numpy[:, :, c]
nmap_dilation = cv2.dilate(nmap, kernel, iterations=1)
ref_lm_numpy[:, :, c] = nmap_dilation
#ref_lm_numpy[:, :, c] = scipy.ndimage.filters.gaussian_filter(nmap_dilation, gsd)
RF_tensor = np2ts(ref_lm_numpy)
RF_tensor = Variable(RF_tensor.cuda(),volatile=True)
def zeroing_out_maps(lm, keep=0):
'''
Only Keep one channel and zero out other channels
[Input] a multi-channel tensor of noise map
[Output] a multi-channel tensor of noise map after zeroing out items
'''
lm_numpy = lm.data.squeeze(0).cpu().numpy()
lm_numpy = (np.transpose(lm_numpy, (1, 2, 0)))
ref_lm_numpy = lm_numpy.copy() #a refined map
for c in range(lm_numpy.shape[2]):
if np.isin(c,keep)==0:
ref_lm_numpy[:, :, c] = 0.
print(ref_lm_numpy)
RF_tensor = np2ts(ref_lm_numpy)
RF_tensor = Variable(RF_tensor.cuda(),volatile=True)
return RF_tensor
def level_refine(NM_tensor, ref_mode, chn=3):
'''
Description: To refine the estimated noise level maps
[Input] the noise map tensor, and a refinement mode
Mode:
[0] Get the most salient (the most frequent estimated noise level)
[1] Get the maximum value of noise level
[2] Gaussian smooth the noise level map to make the regional estimation more smooth
[3] Get the average maximum value of the noise level
[5] Get the CDF thresholded value
[Output] a refined map tensor with four channels
'''
#RF_tensor = NM_tensor.clone() #get a clone version of NM tensor without changing the original one
if ref_mode == 0 or ref_mode == 1 or ref_mode == 4 or ref_mode==5: #if we use a single value for the map
if ref_mode == 0 or ref_mode == 4:
nl_list = get_salient_noise_in_maps(NM_tensor, 0., chn)
if ref_mode == 4: #half the estimation
nl_list = nl_list - nl_list
print(nl_list)
elif ref_mode == 1:
nl_list = get_max_noise_in_maps(NM_tensor, chn)
elif ref_mode == 5:
nl_list = get_cdf_noise_in_maps(NM_tensor, 0.999, chn)
noise_map = np.zeros((NM_tensor.shape[0], chn, NM_tensor.size()[2], NM_tensor.size()[3])) #initialize the noise map before concatenating
for n in range(NM_tensor.shape[0]):
noise_map[n,:,:,:] = np.reshape(np.tile(nl_list[n], NM_tensor.size()[2] * NM_tensor.size()[3]),
(chn, NM_tensor.size()[2], NM_tensor.size()[3]))
RF_tensor = torch.from_numpy(noise_map).type(torch.FloatTensor)
RF_tensor = Variable(RF_tensor.cuda(),volatile=True)
elif ref_mode == 2:
RF_tensor = get_smooth_maps(NM_tensor, 10, 5)
elif ref_mode == 3:
lb = get_salient_noise_in_maps(NM_tensor)
up = get_max_noise_in_maps(NM_tensor)
nl_list = ( lb + up ) * 0.5
noise_map = np.zeros((1, chn, NM_tensor.size()[2], NM_tensor.size()[3])) #initialize the noise map before concatenating
noise_map[0, :, :, :] = np.reshape(np.tile(nl_list, NM_tensor.size()[2] * NM_tensor.size()[3]),
(chn, NM_tensor.size()[2], NM_tensor.size()[3]))
RF_tensor = torch.from_numpy(noise_map).type(torch.FloatTensor)
RF_tensor = Variable(RF_tensor.cuda(),volatile=True)
return (RF_tensor, nl_list)
def normalize(a, len_v, min_v, max_v):
'''
normalize the sequence of factors
'''
norm_a = np.reshape(a, (len_v,1))
norm_a = (norm_a - float(min_v)) / float(max_v - min_v)
return norm_a
def generate_training_noisy_image(current_image, s_or_m, limit_set, c, val=0):
noise_level_list = np.zeros((c, 1))
if s_or_m == 0: #single noise type
if val == 0:
for chn in range(c):
noise_level_list[chn] = np.random.uniform(limit_set[0][0], limit_set[0][1])
elif val == 1:
for chn in range(c):
noise_level_list[chn] = 35
noisy_img = generate_noisy(current_image, 0, noise_level_list /255.)
return (noisy_img, noise_level_list)
def generate_ground_truth_noise_map(noise_map, n, noise_level_list, limit_set, c, pn, pw, ph):
for chn in range(c):
noise_level_list[chn] = normalize(noise_level_list[chn], 1, limit_set[0][0], limit_set[0][1]) #normalize the level value
noise_map[n, :, :, :] = np.reshape(np.tile(noise_level_list, pw * ph), (c, pw, ph)) #total number of channels
return noise_map
#Add noise to the original images
def generate_noisy(image, noise_type, noise_level_list=0, sigma_s=20, sigma_c=40):
'''
Description: To generate noisy images of different types
----------
[Input]
image : ndarray of float type: [0,1] just one image, current support gray or color image input (w,h,c)
noise_type: 0,1,2,3
noise_level_list: pre-defined noise level for each channel, without normalization: only information of 3 channels
[0]'AWGN' Multi-channel Gaussian-distributed additive noise
[1]'RVIN' Replaces random pixels with 0 or 1. noise_level: ratio of the occupation of the changed pixels
[2]'Gaussian-Poisson' GP noise approximator, the combinatin of signal-dependent and signal independent noise
[Output]
A noisy image
'''
w, h, c = image.shape
#Some unused noise type: Poisson and Uniform
#if noise_type == *:
#vals = len(np.unique(image))
#vals = 2 ** np.ceil(np.log2(vals))
#noisy = np.random.poisson(image * vals) / float(vals)
#if noise_type == *:
#uni = np.random.uniform(-factor,factor,(w, h, c))
#uni = uni.reshape(w, h, c)
#noisy = image + uni
noisy = image.copy()
if noise_type == 0: #MC-AWGN model
gauss = np.zeros((w, h, c))
for chn in range(c):
gauss[:,:,chn] = np.random.normal(0, noise_level_list[chn], (w, h))
noisy = image + gauss
elif noise_type == 1: #MC-RVIN model
for chn in range(c): #process each channel separately
prob_map = np.random.uniform(0.0, 1.0, (w, h))
noise_map = np.random.uniform(0.0, 1.0, (w, h))
noisy_chn = noisy[: , :, chn]
noisy_chn[ prob_map < noise_level_list[chn] ] = noise_map[ prob_map < noise_level_list[chn] ]
elif noise_type == 2:
#sigma_s = np.random.uniform(0.0, 0.16, (3,))
#sigma_c = np.random.uniform(0.0, 0.06, (3,))
sigma_c = [sigma_c]*3
sigma_s = [sigma_s]*3
sigma_s = np.reshape(sigma_s, (1, 1, c)) #reshape the sigma factor to [1,1,c] to multiply with the image
noise_s_map = np.multiply(sigma_s, image) #according to x or temp_x?? (according to clean image or irradience)
#print(noise_s_map) # different from the official code, here we use the original clean image x to compute the variance
noise_s = np.random.randn(w, h, c) * noise_s_map #use the new variance to shift the normal distribution
noisy = image + noise_s
#add signal_independent noise to L
noise_c = np.zeros((w, h, c))
for chn in range(3):
noise_c [:, :, chn] = np.random.normal(0, sigma_c[chn], (w, h))
noisy = noisy + noise_c
return noisy
#generate AWGN-RVIN noise together
def generate_comp_noisy(image, noise_level_list):
'''
Description: To generate mixed AWGN and RVIN noise together
----------
[Input]
image: a float image between [0,1]
noise_level_list: AWGN and RVIN noise level
[Output]
A noisy image
'''
w, h, c = image.shape
noisy = image.copy()
for chn in range(c):
mix_thre = noise_level_list[c+chn] #get the mix ratio of AWGN and RVIN
gau_std = noise_level_list[chn] #get the gaussian std
prob_map = np.random.uniform( 0, 1, (w, h) ) #the prob map
noise_map = np.random.uniform( 0, 1, (w, h) ) #the noisy map
noisy_chn = noisy[: ,: ,chn]
noisy_chn[prob_map < mix_thre ] = noise_map[prob_map < mix_thre ]
gauss = np.random.normal(0, gau_std, (w, h))
noisy_chn[prob_map >= mix_thre ] = noisy_chn[prob_map >= mix_thre ] + gauss[prob_map >= mix_thre]
return noisy
def generate_denoise(image, model, noise_level_list):
'''
Description: Generate Denoised Blur Images
----------
[Input]
image:
model:
noise_level_list:
[Output]
A blur image patch
'''
#input images
ISource = np2ts(image)
ISource = torch.clamp(ISource, 0., 1.)
ISource = Variable(ISource.cuda(),volatile=True)
#input denoise conditions
noise_map = np.zeros((1, 6, image.shape[0], image.shape[1])) #initialize the noise map before concatenating
noise_map[0, :, :, :] = np.reshape(np.tile(noise_level_list, image.shape[0] * image.shape[1]), (6, image.shape[0], image.shape[1]))
NM_tensor = torch.from_numpy(noise_map).type(torch.FloatTensor)
NM_tensor = Variable(NM_tensor.cuda(),volatile=True)
#generate blur images
Res = model(ISource, NM_tensor)
Out = torch.clamp(ISource-Res, 0., 1.)
out_numpy = Out.data.squeeze(0).cpu().numpy()
out_numpy = np.transpose(out_numpy, (1, 2, 0))
return out_numpy
#TODO: two pixel shuffle functions to process the images
def pixelshuffle(image, scale):
'''
Discription: Given an image, return a reversible sub-sampling
[Input]: Image ndarray float
[Return]: A mosic image of shuffled pixels
'''
if scale == 1:
return image
w, h ,c = image.shape
mosaic = np.array([])
for ws in range(scale):
band = np.array([])
for hs in range(scale):
temp = image[ws::scale, hs::scale, :] #get the sub-sampled image
band = np.concatenate((band, temp), axis = 1) if band.size else temp
mosaic = np.concatenate((mosaic, band), axis = 0) if mosaic.size else band
return mosaic
def reverse_pixelshuffle(image, scale, fill=0, fill_image=0, ind=[0,0]):
'''
Discription: Given a mosaic image of subsampling, recombine it to a full image
[Input]: Image
[Return]: Recombine it using different portions of pixels
'''
w, h, c = image.shape
real = np.zeros((w, h, c)) #real image
wf = 0
hf = 0
for ws in range(scale):
hf = 0
for hs in range(scale):
temp = real[ws::scale, hs::scale, :]
wc, hc, cc = temp.shape #get the shpae of the current images
if fill==1 and ws==ind[0] and hs==ind[1]:
real[ws::scale, hs::scale, :] = fill_image[wf:wf+wc, hf:hf+hc, :]
else:
real[ws::scale, hs::scale, :] = image[wf:wf+wc, hf:hf+hc, :]
hf = hf + hc
wf = wf + wc
return real
def scal2map(level, h, w, min_v=0., max_v=255.):
'''
Change a single normalized noise level value to a map
[Input]: level: a scaler noise level(0-1), h, w
[Return]: a pytorch tensor of the cacatenated noise level map
'''
#get a tensor from the input level
level_tensor = torch.from_numpy(np.reshape(level, (1,1))).type(torch.FloatTensor)
#make the noise level to a map
level_tensor = level_tensor.view(stdN_tensor.size(0), stdN_tensor.size(1), 1, 1)
level_tensor = level_tensor.repeat(1, 1, h, w)
return level_tensor
def scal2map_spatial(level1, level2, h, w):
stdN_t1 = scal2map(level1, int(h/2), w)
stdN_t2 = scal2map(level2, h-int(h/2), w)
stdN_tensor = torch.cat([stdN_t1, stdN_t2], dim=2)
return stdN_tensor