/
model_blind_rgb.py
357 lines (273 loc) · 16.7 KB
/
model_blind_rgb.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
from __future__ import print_function
from __future__ import division
import time
from utils import *
def dncnn(input, is_training=True, output_channels=3):
with tf.variable_scope('block1'):
output = tf.layers.conv2d(input, 64, 3, padding='same', activation=tf.nn.relu)
for layers in xrange(2, 19 + 1):
with tf.variable_scope('block%d' % layers):
output = tf.layers.conv2d(output, 64, 3, padding='same', name='conv%d' % layers, use_bias=False)
output = tf.nn.relu(tf.layers.batch_normalization(output, training=is_training, name='bn%d' % layers))
with tf.variable_scope('block20'):
output = tf.layers.conv2d(output, output_channels, 3, padding='same')
return input-output
class denoiser(object):
def __init__(self, sess, sigma, cost_str, ckpt_dir, sample_dir, log_dir):
self.sess = sess
self.sigma = sigma
self.ckpt_dir = ckpt_dir
self.sample_dir = sample_dir
self.log_dir = log_dir
if not os.path.exists(self.sample_dir):
os.makedirs(self.sample_dir)
# build model
#placeholders for clean and noisy image batches
self.GT = tf.placeholder(tf.float32, [None, None, None, 3], name='gt_true_image')
self.X = tf.placeholder(tf.float32, [None, None, None, 3], name='gt_fake_image')
self.Y = tf.placeholder(tf.float32, [None, None, None, 3], name='noisy_image')
self.Z = tf.placeholder(tf.float32, [None, None, None, 3], name='less_noisy')
self.is_training = tf.placeholder(tf.bool, name='is_training') #for batchnorm
#vector containing sigma values for each image
self.sigma_tf = tf.placeholder(tf.float32, [None], name='sigma_vector')
self.sigma_tf_less = tf.placeholder(tf.float32, [None], name='sigma_vector')
#tuned epsilon values
self.eps_tf = 1.6*self.sigma_tf*0.0001 # for gt range [0-1]
#self.eps_tf = tf.maximum(.001 * tf.reduce_max(tf.reshape(self.Y, shape=[tf.shape(self.Y)[0], -1]), axis=-1), .00001) # for gt range [0-255] and remove normalization for sigma
self.eps_tf = self.eps_tf[:, tf.newaxis, tf.newaxis, tf.newaxis] #for vectorized calculation of the sure
#forward propagation
with tf.variable_scope('DnCNN'):
self.Y_ = dncnn(self.Y, is_training=self.is_training)
#forward propagation of the perturbed input
self.b_prime = tf.random_normal(shape=tf.shape(self.Y), stddev=1.0)
self.Yptb = self.Y + tf.multiply(self.b_prime, self.eps_tf)
with tf.variable_scope('DnCNN', reuse=True):
self.Yptb_ = dncnn(self.Yptb, is_training=self.is_training)
self.Ht = tf.to_float(tf.shape(self.Y)[1]) #height of the image
self.Wt = tf.to_float(tf.shape(self.Y)[2]) #width of the image
batch = tf.to_float(tf.shape(self.Y)[0]) #size of the minibatch
self.var_Y = tf.square(self.sigma_tf/255.0, name = 'var_vector')
self.var_Y = self.var_Y[:, tf.newaxis, tf.newaxis, tf.newaxis] #for vectorized calculation of the sure
self.var_Z = tf.square(self.sigma_tf_less/255.0, name = 'var_vector_less') #tensor of variance values
self.var_Z = self.var_Z[:, tf.newaxis, tf.newaxis, tf.newaxis] #for vectorized calculation of the sure
self.divergence = tf.multiply((1.0/self.eps_tf), tf.multiply(self.b_prime, (self.Yptb_-self.Y_)))
self.divergence_sum_Y = tf.reduce_sum(tf.multiply(self.var_Y, self.divergence))
self.divergence_sum_Z = tf.reduce_sum(tf.multiply(self.var_Z, self.divergence))
############################################-COSTS-############################################
# MSE
self.mse = (1.0 / batch) * tf.nn.l2_loss(self.Y_ - self.GT)
# N2N
self.n2n = (1.0 / batch) * tf.nn.l2_loss(self.Y_ - self.X)
# SURE
self.var_sum_Y = self.Ht * self.Wt * 3 * tf.reduce_sum(self.var_Y)/2.0
self.sure = (1.0 / batch)*(tf.nn.l2_loss(self.Y - self.Y_) - self.var_sum_Y + self.divergence_sum_Y)
# NEW SURE
self.var_sum_Z = self.Ht * self.Wt * 3 * tf.reduce_sum(self.var_Z)/2.0
self.new_sure = (1.0 / batch)*(tf.nn.l2_loss(self.Z - self.Y_) - self.var_sum_Z + self.divergence_sum_Z)
#which cost function to use for training
if cost_str=='SURE':
self.cost = self.sure
print('[*] Optimizing SURE!')
elif cost_str=='N2N':
self.cost = self.n2n
print('[*] Optimizing Noise2Noise!')
elif cost_str=='MSE':
self.cost = self.mse
print('[*] Optimizing MSE!')
elif cost_str == 'e-SURE':
self.cost = self.new_sure
print('[*] Optimizing e-SURE!')
else:
print("UNKNOWN COST")
self.lr = tf.placeholder(tf.float32, name='learning_rate')
self.eva_psnr = tf.placeholder(tf.float32, name='eva_psnr')
#optimizer
optimizer = tf.train.AdamOptimizer(self.lr, name='AdamOptimizer')
#for batchnorm
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
self.train_op = optimizer.minimize(self.cost) ##CHANGED
#checkpoint saver
self.saver = tf.train.Saver(max_to_keep=15)
init = tf.global_variables_initializer()
self.sess.run(init)
print("[*] Initialize model successfully...")
#function to evaluate the performance after each epoch
def evaluate(self, test_files, iter_num, summary_merged, summary_writer):
mse_sum_ev = 0
sigma_img = np.reshape(np.asarray(np.squeeze(self.sigma), dtype=np.float32), [1, ])
print('Sigma image shape: ', np.shape(sigma_img))
psnr_sum = 0
print("[*] " + 'noise level: ' + str(self.sigma) + " start testing...")
for idx in xrange(len(test_files)):
clean_image = load_images_RGB(test_files[idx]).astype(np.float32) / 255.0
noisy_image = clean_image + np.random.normal(0, self.sigma/255.0, np.shape(clean_image)).astype('float32')
# not important, just needed to denoise an image
noisy_image2 = clean_image + np.random.normal(0, self.sigma/255.0, np.shape(clean_image)).astype('float32')
less_image = noisy_image2
output_clean_image, mse_ev = self.sess.run([self.Y_, self.mse], feed_dict={self.GT: clean_image,
self.X: clean_image,
self.Y: noisy_image,
self.Z: less_image,
self.sigma_tf: sigma_img,
self.is_training: False})
groundtruth = np.clip(255 * clean_image, 0, 255).astype('uint8')
noisyimage = np.clip(255 * noisy_image, 0, 255).astype('uint8')
outputimage = np.clip(255 * output_clean_image, 0, 255).astype('uint8')
# calculate PSNR
psnr = cal_psnr(groundtruth, outputimage)
if idx<10:
print("img%d PSNR: %.2f" % (idx+1, psnr))
psnr_sum += psnr
if idx < 5:
save_images_RGB(os.path.join(self.sample_dir, 'eval%d_%d.png' % (idx + 1, iter_num)), groundtruth, noisyimage, outputimage)
# statistics
mse_sum_ev += mse_ev
avg_psnr = psnr_sum / len(test_files)
mse_ave_ev = mse_sum_ev / len(test_files)
# statistics
print("Test set length: ", len(test_files))
print("Iter: %d MSE: %.5f \n" % (iter_num, mse_ave_ev))
print("--- Evaluation on BSD68 Dataset --- Average PSNR %.2f ---" % avg_psnr)
psnr_summary = self.sess.run(summary_merged, feed_dict={self.eva_psnr:avg_psnr})
summary_writer.add_summary(psnr_summary, iter_num)
def train(self, data_path, eval_files, batch_size, epoch, lr, gt_type):
# CLEAN ground-truth trainset
gt_data = np.load(os.path.join(data_path, 'gt_rgb_clean_patches.npy'), mmap_mode='r')
# IMPERFECT ground-truth trainset
gt_name = 'gt_rgb_noisy_' + str(gt_type) + '_patches.npy'
gt_noisy = np.load(os.path.join(data_path, gt_name), mmap_mode='r')
# Noise std. of imperfect ground-truth images
sigmapath_imp = 'sigma_rgb_vector_imperfect_' + str(gt_type) + '.npy'
sigma_imperfect = np.load(os.path.join(data_path, sigmapath_imp), mmap_mode='r') #.astype('float32')/255.
# NOISY patches trainset
noisy_name ='rgb_noisy_' + str(gt_type) + '_patches.npy'
noisy = np.load(os.path.join(data_path, noisy_name), mmap_mode='r')
# Noise std. of noisy trainset images
sigmapath_ny = 'sigma_rgb_vector_blind_' + str(gt_type) + '_' + str(int(self.sigma)) + '.npy'
sigma_vector = np.load(os.path.join(data_path, sigmapath_ny), mmap_mode='r') #.astype('float32')/255.
print('[*] Dataset path: %s Sigma path: %s' % (datapath, sigmapath))
print('[*] GT clean range: [%.2f %.2f]' % (np.amin(gt_data), np.amax(gt_data)))
print('[*] GT noisy range: [%.2f %.2f]' % (np.amin(gt_noisy), np.amax(gt_noisy)))
print('[*] Noisy img range: [%.2f %.2f]' % (np.amin(noisy), np.amax(noisy)))
print('[*] Sigma vec range: [%.2f %.2f]' % (np.amin(sigma_vector), np.amax(sigma_vector)))
print('[*] Imperfect GT sigma vec range: [%.2f %.2f]' % (np.amin(sigma_imperfect), np.amax(sigma_imperfect)))
print('[!] BUT network is trained and tested on normalized images between [0-1]')
numData = np.shape(gt_data)[0]
numBatch = int(numData / batch_size)
# load pretrained model
load_model_status, global_step = self.load(self.ckpt_dir)
if load_model_status:
iter_num = global_step
start_epoch = (global_step) // numBatch
start_step = (global_step) % numBatch
print("[*] Model restore success!")
else:
iter_num = 0
start_epoch = 0
start_step = 0
print("[*] Not find pretrained model!")
# make summary
tf.summary.scalar('MSE', self.mse)
tf.summary.scalar('N2N', self.n2n)
tf.summary.scalar('SURE', self.sure)
tf.summary.scalar('e-SURE', self.new_sure)
tf.summary.scalar('lr', self.lr)
writer = tf.summary.FileWriter(self.log_dir+"/", self.sess.graph)
merged = tf.summary.merge_all()
summary_psnr = tf.summary.scalar('eva_psnr', self.eva_psnr)
print("[*] Start training, with start epoch %d start iter %d : " % (start_epoch, iter_num))
start_time = time.time()
self.evaluate(eval_files, iter_num, summary_merged=summary_psnr, summary_writer=writer)
tf.get_default_graph().finalize() # making sure that the graph is fixed at this point
#training loop
for epoch in xrange(start_epoch, epoch):
print("Model: %s" % (self.ckpt_dir))
print("Learning rate: {}".format(lr[epoch]))
rand_inds=np.random.choice(numData, numData,replace=False)
for batch_id in xrange(0, numBatch):
# No RAM required
batch_rand_inds = rand_inds[batch_id * batch_size:(batch_id + 1) * batch_size]
batch_images = np.array(gt_data[batch_rand_inds]).astype(np.float32) / 255.0 # clean gt
batch_images_noise = np.array(gt_noisy[batch_rand_inds]).astype(np.float32) / 255.0 # imperfect gt
batch_sigma_less = np.array(sigma_imperfect[batch_rand_inds]).astype(np.float32) # imperfect gt Gaussian std.
batch_images_corrupt = np.array(noisy[batch_rand_inds]).astype(np.float32) / 255.0 # noisy
batch_sigma = np.array(sigma_vector[batch_rand_inds]).astype(np.float32) # noisy Gaussian std.
feed_dict = {self.GT: batch_images,
self.X: batch_images_noise,
self.Y: batch_images_corrupt,
self.Z: batch_images_noise,
self.sigma_tf: batch_sigma,
self.sigma_tf_less: batch_sigma_less,
self.lr: lr[epoch],
self.is_training: True}
self.sess.run(self.train_op, feed_dict=feed_dict)
if (iter_num)%100==0:
feed_dict2 = {self.GT: batch_images,
self.X: batch_images_noise,
self.Y: batch_images_corrupt,
self.Z: batch_images_noise,
self.sigma_tf: batch_sigma,
self.sigma_tf_less: batch_sigma_less,
self.lr: lr[epoch],
self.is_training: False}
mse, n2n, sure, new_sure, summary = self.sess.run([self.mse, self.n2n, self.sure, self.new_sure, merged],feed_dict=feed_dict2)
print("Epoch: [%2d] [%4d/%4d] Time: %4.4f MSE: %.6f, n2n: %.6f, sure: %.6f, new_sure: %.6f"
% (epoch + 1, batch_id + 1, numBatch, time.time() - start_time, mse, n2n, sure, new_sure))
writer.add_summary(summary, iter_num)
if (iter_num+1)%973==0:
self.evaluate(eval_files, iter_num+1, summary_merged=summary_psnr, summary_writer=writer)
iter_num += 1
self.save(iter_num, self.ckpt_dir)
print('\n')
print("[*] Finish training.")
def save(self, iter_num, ckpt_dir, model_name='CDnCNN-tensorflow'):
checkpoint_dir = ckpt_dir
if not os.path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir)
print("[*] Saving model...")
self.saver.save(self.sess,
os.path.join(checkpoint_dir, model_name),
global_step=iter_num)
def load(self, checkpoint_dir):
print("[*] Reading checkpoint...")
ckpt = tf.train.get_checkpoint_state(checkpoint_dir)
if ckpt and ckpt.model_checkpoint_path:
full_path = tf.train.latest_checkpoint(checkpoint_dir)
global_step = int(full_path.split('/')[-1].split('-')[-1])
self.saver.restore(self.sess, full_path)
return True, global_step
else:
return False, 0
def test(self, test_files, noisy_files, save_dir):
"""Test CDnCNN"""
# init variables
tf.initialize_all_variables().run()
if not os.path.exists(save_dir):
os.makedirs(save_dir)
load_model_status, global_step = self.load(self.ckpt_dir)
assert load_model_status == True, '[!] Load weights FAILED...'
print("[*] Load weights SUCCESS...")
numData = len(test_files)
psnr_sum = 0
print("[*] " + 'noise level: ' + str(self.sigma) + " start testing...")
for idx in xrange(numData):
clean_image = load_images_RGB(test_files[idx]).astype(np.float32) / 255.0
noisy_image = np.load(noisy_files[idx]).astype(np.float32)
print('cl shape: {}'.format(np.shape(clean_image)))
print(np.amax(clean_image))
print('ny shape: {}'.format(np.shape(noisy_image)))
print(np.amin(noisy_image))
output_clean_image = self.sess.run(self.Y_, feed_dict={self.X: clean_image, self.Y: noisy_image, self.is_training: False})
groundtruth = np.clip(255 * clean_image, 0, 255).astype('uint8')
noisyimage = np.clip(255 * noisy_image, 0, 255).astype('uint8')
outputimage = np.clip(255 * output_clean_image, 0, 255).astype('uint8')
# calculate PSNR
psnr = cal_psnr(groundtruth, outputimage)
psnr_sum += psnr
img_name, img_ext = os.path.splitext(os.path.basename(test_files[idx]))
print("%s PSNR: %.2f" % (os.path.basename(test_files[idx]), psnr))
save_images_RGB(os.path.join(save_dir, 'noisy%s.png' % img_name), noisyimage)
save_images_RGB(os.path.join(save_dir, 'denoised%s.png' % img_name), outputimage)
avg_psnr = psnr_sum / len(test_files)
print("--- Test ---- Average PSNR %.2f ---" % avg_psnr)