/
transparency_separation.py
411 lines (361 loc) · 18.8 KB
/
transparency_separation.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
from net import skip
from net.losses import ExclusionLoss, plot_image_grid, StdLoss
from net.noise import get_noise
from utils.image_io import *
from skimage.measure import compare_psnr
import numpy as np
import torch
import torch.nn as nn
from collections import namedtuple
TwoImagesSeparationResult = namedtuple("TwoImagesSeparationResult",
["reflection", "transmission", "psnr", "alpha1", "alpha2"])
class TwoImagesSeparation(object):
def __init__(self, image1_name, image2_name, image1, image2, plot_during_training=True, show_every=500, num_iter=4000,
original_reflection=None, original_transmission=None):
# we assume the reflection is static
self.image1 = image1
self.image2 = image2
self.plot_during_training = plot_during_training
self.psnrs = []
self.show_every = show_every
self.image1_name = image1_name
self.image2_name = image2_name
self.num_iter = num_iter
self.loss_function = None
self.parameters = None
self.learning_rate = 0.001
self.input_depth = 2
self.reflection_net_input = None
self.transmission_net_input = None
self.original_transmission = original_transmission
self.original_reflection = original_reflection
self.reflection_net = None
self.transmission_net = None
self.total_loss = None
self.multiscale_loss = None
self.multiscale = None
self.reflection_out = None
self.transmission_out = None
self.current_result = None
self.best_result = None
self._init_all()
def _init_all(self):
self._init_images()
self._init_nets()
self._init_inputs()
self._init_parameters()
self._init_losses()
def _init_images(self):
self.image1_torch = np_to_torch(self.image1).type(torch.cuda.FloatTensor)
self.image2_torch = np_to_torch(self.image2).type(torch.cuda.FloatTensor)
def _init_inputs(self):
input_type = 'noise'
# input_type = 'meshgrid'
data_type = torch.cuda.FloatTensor
self.reflection_net_input = get_noise(self.input_depth, input_type,
(self.image1.shape[1], self.image1.shape[2])).type(data_type).detach()
self.alpha_net1_input = get_noise(self.input_depth, input_type,
(self.image1.shape[1], self.image1.shape[2])).type(data_type).detach()
self.alpha_net2_input = get_noise(self.input_depth, input_type,
(self.image1.shape[1], self.image1.shape[2])).type(data_type).detach()
self.transmission_net_input = get_noise(self.input_depth, input_type,
(self.image1.shape[1], self.image1.shape[2])).type(data_type).detach()
def _init_parameters(self):
self.parameters = [p for p in self.reflection_net.parameters()] + \
[p for p in self.transmission_net.parameters()]
self.parameters += [p for p in self.alpha1.parameters()]
self.parameters += [p for p in self.alpha2.parameters()]
def _init_nets(self):
data_type = torch.cuda.FloatTensor
pad = 'reflection'
reflection_net = skip(
self.input_depth, 3,
num_channels_down=[8, 16, 32, 64, 128],
num_channels_up=[8, 16, 32, 64, 128],
num_channels_skip=[0, 0, 0, 4, 4],
upsample_mode='bilinear',
filter_size_down=5,
filter_size_up=5,
need_sigmoid=True, need_bias=True, pad=pad, act_fun='LeakyReLU')
self.reflection_net = reflection_net.type(data_type)
transmission_net = skip(
self.input_depth, 3,
num_channels_down=[8, 16, 32, 64, 128],
num_channels_up=[8, 16, 32, 64, 128],
num_channels_skip=[0, 0, 0, 4, 4],
upsample_mode='bilinear',
filter_size_down=5,
filter_size_up=5,
need_sigmoid=True, need_bias=True, pad=pad, act_fun='LeakyReLU')
self.transmission_net = transmission_net.type(data_type)
alpha_net1 = skip(
self.input_depth, 1,
num_channels_down=[8, 16, 32, 64, 128],
num_channels_up=[8, 16, 32, 64, 128],
num_channels_skip=[0, 0, 0, 4, 4],
upsample_mode='bilinear',
filter_size_down=5,
filter_size_up=5,
need_sigmoid=True, need_bias=True, pad=pad, act_fun='LeakyReLU')
self.alpha1 = alpha_net1.type(data_type)
alpha_net2 = skip(
self.input_depth, 1,
num_channels_down=[8, 16, 32, 64, 128],
num_channels_up=[8, 16, 32, 64, 128],
num_channels_skip=[0, 0, 0, 4, 4],
upsample_mode='bilinear',
filter_size_down=5,
filter_size_up=5,
need_sigmoid=True, need_bias=True, pad=pad, act_fun='LeakyReLU')
self.alpha2 = alpha_net2.type(data_type)
def _init_losses(self):
data_type = torch.cuda.FloatTensor
self.mse_loss = torch.nn.MSELoss().type(data_type)
self.exclusion_loss = ExclusionLoss().type(data_type)
self.blur_loss = StdLoss().type(data_type)
def optimize(self):
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = True
optimizer = torch.optim.Adam(self.parameters, lr=self.learning_rate)
for j in range(self.num_iter):
optimizer.zero_grad()
self._optimization_closure(j)
self._obtain_current_result(j)
if self.plot_during_training:
self._plot_closure(j)
optimizer.step()
def _optimization_closure(self, step):
reg_noise_std = 0
reflection_net_input = self.reflection_net_input + (self.reflection_net_input.clone().normal_() * reg_noise_std)
transmission_net_input = self.transmission_net_input + \
(self.transmission_net_input.clone().normal_() * reg_noise_std)
self.reflection_out = self.reflection_net(reflection_net_input)
self.transmission_out = self.transmission_net(transmission_net_input)
alpha_net_input = self.alpha_net1_input + (self.alpha_net1_input.clone().normal_() * reg_noise_std)
self.current_alpha1 = self.alpha1(alpha_net_input)[:, :,
self.image1_torch.shape[2] // 2:self.image1_torch.shape[2] // 2 + 1,
self.image1_torch.shape[3] // 2:self.image1_torch.shape[3] // 2 + 1] * 0.9 + 0.05
alpha_net_input = self.alpha_net2_input + (self.alpha_net2_input.clone().normal_() * reg_noise_std)
self.current_alpha2 = self.alpha2(alpha_net_input)[:, :,
self.image1_torch.shape[2] // 2:self.image1_torch.shape[2] // 2 + 1,
self.image1_torch.shape[3] // 2:self.image1_torch.shape[3] // 2 + 1]* 0.9 + 0.05
self.total_loss = self.mse_loss(self.current_alpha1 * self.reflection_out +
(1 - self.current_alpha1) * self.transmission_out,
self.image1_torch)
self.total_loss += self.mse_loss(self.current_alpha2 * self.reflection_out +
(1 - self.current_alpha2) * self.transmission_out,
self.image2_torch)
self.exclusion = self.exclusion_loss(self.reflection_out, self.transmission_out)
self.total_loss += 0.1 * self.exclusion
# self.total_loss += self.blur_loss(self.current_alpha2) + self.blur_loss(self.current_alpha1)
if step < 1000:
self.total_loss += 0.5 * self.mse_loss(self.current_alpha1,
torch.tensor([[[[0.5]]]]).type(torch.cuda.FloatTensor))
self.total_loss += 0.5 * self.mse_loss(self.current_alpha2,
torch.tensor([[[[0.5]]]]).type(torch.cuda.FloatTensor))
self.total_loss.backward()
def _obtain_current_result(self, j):
"""
puts in self.current result the current result.
also updates the best result
:return:
"""
reflection_out_np = np.clip(torch_to_np(self.reflection_out), 0, 1)
transmission_out_np = np.clip(torch_to_np(self.transmission_out), 0, 1)
# print(reflection_out_np.shape)
alpha1 = np.clip(torch_to_np(self.current_alpha1), 0, 1)
alpha2 = np.clip(torch_to_np(self.current_alpha2), 0, 1)
v = alpha1 * reflection_out_np + (1 - alpha1) * transmission_out_np
# print(v.shape, self.image2.shape)
psnr1 = compare_psnr(self.image1, v)
psnr2 = compare_psnr(self.image2, alpha2 * reflection_out_np + (1 - alpha2) * transmission_out_np)
self.psnrs.append(psnr1+psnr2)
self.current_result = TwoImagesSeparationResult(reflection=reflection_out_np, transmission=transmission_out_np,
psnr=psnr1, alpha1=alpha1, alpha2=alpha2)
if self.best_result is None or self.best_result.psnr < self.current_result.psnr:
self.best_result = self.current_result
def _plot_closure(self, step):
print('Iteration {:5d} Loss {:5f} Exclusion {:5f} PSRN_gt: {:f}'.format(step,
self.total_loss.item(),
self.exclusion.item(),
self.current_result.psnr),
'\r', end='')
if self.plot_during_training and step % self.show_every == self.show_every - 1:
plot_image_grid("reflection_transmission_{}".format(step),
[self.current_result.reflection, self.current_result.transmission])
# plot_image_grid("learned_mask_{}".format(step),
# [self.current_result.alpha1, self.current_result.alpha2])
save_image("sum1_{}".format(step), self.current_result.alpha1 * self.current_result.reflection +
(1-self.current_result.alpha1)* self.current_result.transmission)
save_image("sum2_{}".format(step), self.current_result.alpha2 * self.current_result.reflection +
(1 - self.current_result.alpha2) * self.current_result.transmission)
def finalize(self):
save_graph(self.image1_name + "_psnr", self.psnrs)
save_image(self.image1_name + "_reflection", self.best_result.reflection)
save_image(self.image1_name + "_transmission", self.best_result.transmission)
save_image(self.image1_name + "_original", self.image1)
save_image(self.image2_name + "_original", self.image2)
class Separation(object):
def __init__(self, image_name, image, plot_during_training=True, show_every=500, num_iter=8000,
original_reflection=None, original_transmission=None):
self.image = image
self.plot_during_training = plot_during_training
# self.ratio = ratio
self.psnrs = []
self.show_every = show_every
self.image_name = image_name
self.num_iter = num_iter
self.loss_function = None
# self.ratio_net = None
self.parameters = None
self.learning_rate = 0.0005
self.input_depth = 3
self.reflection_net_inputs = None
self.transmission_net_inputs = None
self.original_transmission = original_transmission
self.original_reflection = original_reflection
self.reflection_net = None
self.transmission_net = None
self.total_loss = None
self.reflection_out = None
self.transmission_out = None
self.current_result = None
self.best_result = None
self._init_all()
def _init_all(self):
self._init_images()
self._init_nets()
self._init_inputs()
self._init_parameters()
self._init_losses()
def _init_images(self):
self.images = create_augmentations(self.image)
self.images_torch = [np_to_torch(image).type(torch.cuda.FloatTensor) for image in self.images]
def _init_inputs(self):
input_type = 'noise'
# input_type = 'meshgrid'
data_type = torch.cuda.FloatTensor
origin_noise = torch_to_np(get_noise(self.input_depth,
input_type,
(self.images_torch[0].shape[2],
self.images_torch[0].shape[3])).type(data_type).detach())
self.reflection_net_inputs = [np_to_torch(aug).type(data_type).detach() for aug in create_augmentations(origin_noise)]
origin_noise = torch_to_np(get_noise(self.input_depth,
input_type,
(self.images_torch[0].shape[2],
self.images_torch[0].shape[3])).type(data_type).detach())
self.transmission_net_inputs = [np_to_torch(aug).type(data_type).detach() for aug in
create_augmentations(origin_noise)]
def _init_parameters(self):
self.parameters = [p for p in self.reflection_net.parameters()] + \
[p for p in self.transmission_net.parameters()]
def _init_nets(self):
data_type = torch.cuda.FloatTensor
pad = 'reflection'
reflection_net = skip(
self.input_depth, self.images[0].shape[0],
num_channels_down=[8, 16, 32, 64, 128],
num_channels_up=[8, 16, 32, 64, 128],
num_channels_skip=[0, 0, 0, 4, 4],
upsample_mode='bilinear',
filter_size_down=5,
filter_size_up=5,
need_sigmoid=True, need_bias=True, pad=pad, act_fun='LeakyReLU')
self.reflection_net = reflection_net.type(data_type)
transmission_net = skip(
self.input_depth, self.images[0].shape[0],
num_channels_down=[8, 16, 32, 64, 128],
num_channels_up=[8, 16, 32, 64, 128],
num_channels_skip=[0, 0, 0, 4, 4],
upsample_mode='bilinear',
filter_size_down=5,
filter_size_up=5,
need_sigmoid=True, need_bias=True, pad=pad, act_fun='LeakyReLU')
self.transmission_net = transmission_net.type(data_type)
def _init_losses(self):
data_type = torch.cuda.FloatTensor
self.l1_loss = nn.L1Loss().type(data_type)
self.exclusion_loss = ExclusionLoss().type(data_type)
def optimize(self):
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = True
optimizer = torch.optim.Adam(self.parameters, lr=self.learning_rate)
for j in range(self.num_iter):
optimizer.zero_grad()
self._optimization_closure(j)
self._obtain_current_result(j)
if self.plot_during_training:
self._plot_closure(j)
optimizer.step()
def _get_augmentation(self, iteration):
if iteration % 2 == 1:
return 0
# return 0
iteration //= 2
return iteration % 8
def _optimization_closure(self, step):
if step == self.num_iter - 1:
reg_noise_std = 0
elif step < 1000:
reg_noise_std = (1 / 1000.) * (step // 100)
else:
reg_noise_std = 1 / 1000.
aug = self._get_augmentation(step)
if step == self.num_iter - 1:
aug = 0
reflection_net_input = self.reflection_net_inputs[aug] + (self.reflection_net_inputs[aug].clone().normal_() * reg_noise_std)
transmission_net_input = self.transmission_net_inputs[aug] + (self.transmission_net_inputs[aug].clone().normal_() * reg_noise_std)
self.reflection_out = self.reflection_net(reflection_net_input)
self.transmission_out = self.transmission_net(transmission_net_input)
self.total_loss = self.l1_loss(self.reflection_out + self.transmission_out, self.images_torch[aug])
self.total_loss += 0.01 * self.exclusion_loss(self.reflection_out, self.transmission_out)
self.total_loss.backward()
def _obtain_current_result(self, step):
"""
puts in self.current result the current result.
also updates the best result
:return:
"""
if step == self.num_iter - 1 or step % 8 == 0:
reflection_out_np = np.clip(torch_to_np(self.reflection_out), 0, 1)
transmission_out_np = np.clip(torch_to_np(self.transmission_out), 0, 1)
psnr = compare_psnr(self.images[0], reflection_out_np + transmission_out_np)
self.psnrs.append(psnr)
self.current_result = SeparationResult(reflection=reflection_out_np, transmission=transmission_out_np,
psnr=psnr)
if self.best_result is None or self.best_result.psnr < self.current_result.psnr:
self.best_result = self.current_result
def _plot_closure(self, step):
print('Iteration {:5d} Loss {:5f} PSRN_gt: {:f}'.format(step,
self.total_loss.item(),
self.current_result.psnr),
'\r', end='')
if step % self.show_every == self.show_every - 1:
plot_image_grid("left_right_{}".format(step), [self.current_result.reflection, self.current_result.transmission])
def _plot_distance_map(self):
calculated_left = self.best_result.reflection
calculated_right = self.best_result.transmission
# this is to left for reason
# print(distance_to_left.shape)
pass
def finalize(self):
save_graph(self.image_name + "_psnr", self.psnrs)
save_image(self.image_name + "_reflection", self.best_result.reflection)
save_image(self.image_name + "_transmission", self.best_result.transmission)
save_image(self.image_name + "_reflection2", 2 * self.best_result.reflection)
save_image(self.image_name + "_transmission2", 2 * self.best_result.transmission)
save_image(self.image_name + "_original", self.images[0])
SeparationResult = namedtuple("SeparationResult", ['reflection', 'transmission', 'psnr'])
if __name__ == "__main__":
# Separation from two images
input1 = prepare_image('images/input1.jpg')
input2 = prepare_image('images/input2.jpg')
t = TwoImagesSeparation('input1', 'input2', input1, input2, num_iter=7000)
t.optimize()
t.finalize()
# Separation of textures
t1 = prepare_image('images/texture12.jpg')
t2 = prepare_image('images/texture16.jpg')
s = Separation('textures', (t1+t2)/2)
s.optimize()
s.finalize()