forked from 491506870/PRIDNet
-
Notifications
You must be signed in to change notification settings - Fork 0
/
network.py
176 lines (133 loc) · 8.32 KB
/
network.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
import tensorflow as tf
import tensorflow.contrib.slim as slim
from tflearn.layers.conv import global_avg_pool
def lrelu(x):
return tf.maximum(x * 0.2, x)
def upsample_and_concat(x1, x2, output_channels, in_channels):
pool_size = 2
deconv_filter = tf.Variable(tf.truncated_normal([pool_size, pool_size, output_channels, in_channels], stddev=0.02))
deconv = tf.nn.conv2d_transpose(x1, deconv_filter, tf.shape(x2), strides=[1, pool_size, pool_size, 1])
deconv_output = tf.concat([deconv, x2], 3)
deconv_output.set_shape([None, None, None, output_channels * 2])
return deconv_output
def unet(input):
conv1 = slim.conv2d(input, 32, [3, 3], rate=1, activation_fn=lrelu)
conv1 = slim.conv2d(conv1, 32, [3, 3], rate=1, activation_fn=lrelu)
conv1 = slim.conv2d(conv1, 32, [3, 3], rate=1, activation_fn=lrelu)
conv1 = slim.conv2d(conv1, 32, [3, 3], rate=1, activation_fn=lrelu)
pool1 = slim.max_pool2d(conv1, [2, 2], padding='SAME')
conv2 = slim.conv2d(pool1, 64, [3, 3], rate=1, activation_fn=lrelu)
conv2 = slim.conv2d(conv2, 64, [3, 3], rate=1, activation_fn=lrelu)
conv2 = slim.conv2d(conv2, 64, [3, 3], rate=1, activation_fn=lrelu)
conv2 = slim.conv2d(conv2, 64, [3, 3], rate=1, activation_fn=lrelu)
pool2 = slim.max_pool2d(conv2, [2, 2], padding='SAME')
conv3 = slim.conv2d(pool2, 128, [3, 3], rate=1, activation_fn=lrelu)
conv3 = slim.conv2d(conv3, 128, [3, 3], rate=1, activation_fn=lrelu)
conv3 = slim.conv2d(conv3, 128, [3, 3], rate=1, activation_fn=lrelu)
conv3 = slim.conv2d(conv3, 128, [3, 3], rate=1, activation_fn=lrelu)
pool3 = slim.max_pool2d(conv3, [2, 2], padding='SAME')
conv4 = slim.conv2d(pool3, 256, [3, 3], rate=1, activation_fn=lrelu)
conv4 = slim.conv2d(conv4, 256, [3, 3], rate=1, activation_fn=lrelu)
conv4 = slim.conv2d(conv4, 256, [3, 3], rate=1, activation_fn=lrelu)
conv4 = slim.conv2d(conv4, 256, [3, 3], rate=1, activation_fn=lrelu)
pool4 = slim.max_pool2d(conv4, [2, 2], padding='SAME')
conv5 = slim.conv2d(pool4, 512, [3, 3], rate=1, activation_fn=lrelu)
conv5 = slim.conv2d(conv5, 512, [3, 3], rate=1, activation_fn=lrelu)
conv5 = slim.conv2d(conv5, 512, [3, 3], rate=1, activation_fn=lrelu)
conv5 = slim.conv2d(conv5, 512, [3, 3], rate=1, activation_fn=lrelu)
up6 = upsample_and_concat(conv5, conv4, 256, 512)
conv6 = slim.conv2d(up6, 256, [3, 3], rate=1, activation_fn=lrelu)
conv6 = slim.conv2d(conv6, 256, [3, 3], rate=1, activation_fn=lrelu)
conv6 = slim.conv2d(conv6, 256, [3, 3], rate=1, activation_fn=lrelu)
up7 = upsample_and_concat(conv6, conv3, 128, 256)
conv7 = slim.conv2d(up7, 128, [3, 3], rate=1, activation_fn=lrelu)
conv7 = slim.conv2d(conv7, 128, [3, 3], rate=1, activation_fn=lrelu)
conv7 = slim.conv2d(conv7, 128, [3, 3], rate=1, activation_fn=lrelu)
up8 = upsample_and_concat(conv7, conv2, 64, 128)
conv8 = slim.conv2d(up8, 64, [3, 3], rate=1, activation_fn=lrelu)
conv8 = slim.conv2d(conv8, 64, [3, 3], rate=1, activation_fn=lrelu)
conv8 = slim.conv2d(conv8, 64, [3, 3], rate=1, activation_fn=lrelu)
up9 = upsample_and_concat(conv8, conv1, 32, 64)
conv9 = slim.conv2d(up9, 32, [3, 3], rate=1, activation_fn=lrelu)
conv9 = slim.conv2d(conv9, 32, [3, 3], rate=1, activation_fn=lrelu)
conv9 = slim.conv2d(conv9, 32, [3, 3], rate=1, activation_fn=lrelu)
conv10 = slim.conv2d(conv9, 1, [1, 1], rate=1, activation_fn=None)
#out = tf.depth_to_space(conv10, 2)
return conv10
def feature_encoding(input):
conv1 = slim.conv2d(input, 32, [3, 3], rate=1, activation_fn=lrelu, scope='fe_conv1')
conv2 = slim.conv2d(conv1, 32, [3, 3], rate=1, activation_fn=lrelu, scope='fe_conv2')
conv3 = slim.conv2d(conv2, 32, [3, 3], rate=1, activation_fn=lrelu, scope='fe_conv3')
conv4 = slim.conv2d(conv3, 32, [3, 3], rate=1, activation_fn=lrelu, scope='fe_conv4')
conv4 = squeeze_excitation_layer(conv4, 32, 2)
output = slim.conv2d(conv4, 1, [3, 3], rate=1, activation_fn=lrelu, scope='fe_conv5')
return output
def avg_pool(feature_map):
ksize = [[1, 1, 1, 1], [1, 2, 2, 1], [1, 4, 4, 1], [1, 8, 8, 1], [1, 16, 16, 1]]
pool1 = tf.nn.avg_pool(feature_map, ksize=ksize[0], strides=ksize[0], padding='VALID')
pool2 = tf.nn.avg_pool(feature_map, ksize=ksize[1], strides=ksize[1], padding='VALID')
pool3 = tf.nn.avg_pool(feature_map, ksize=ksize[2], strides=ksize[2], padding='VALID')
pool4 = tf.nn.avg_pool(feature_map, ksize=ksize[3], strides=ksize[3], padding='VALID')
pool5 = tf.nn.avg_pool(feature_map, ksize=ksize[4], strides=ksize[4], padding='VALID')
return pool1, pool2, pool3, pool4, pool5
def all_unet(pool1, pool2, pool3, pool4, pool5):
unet1 = unet(pool1)
unet2 = unet(pool2)
unet3 = unet(pool3)
unet4 = unet(pool4)
unet5 = unet(pool5)
return unet1, unet2, unet3, unet4, unet5
def resize_all_image(unet1, unet2, unet3, unet4, unet5):
resize1 = tf.image.resize_images(images=unet1, size=[tf.shape(unet1, out_type=tf.int32)[1],tf.shape(unet1, out_type=tf.int32)[2]], method=tf.image.ResizeMethod.BILINEAR)
resize2 = tf.image.resize_images(images=unet2, size=[tf.shape(unet1, out_type=tf.int32)[1],tf.shape(unet1, out_type=tf.int32)[2]], method=tf.image.ResizeMethod.BILINEAR)
resize3 = tf.image.resize_images(images=unet3, size=[tf.shape(unet1, out_type=tf.int32)[1],tf.shape(unet1, out_type=tf.int32)[2]], method=tf.image.ResizeMethod.BILINEAR)
resize4 = tf.image.resize_images(images=unet4, size=[tf.shape(unet1, out_type=tf.int32)[1],tf.shape(unet1, out_type=tf.int32)[2]], method=tf.image.ResizeMethod.BILINEAR)
resize5 = tf.image.resize_images(images=unet5, size=[tf.shape(unet1, out_type=tf.int32)[1],tf.shape(unet1, out_type=tf.int32)[2]], method=tf.image.ResizeMethod.BILINEAR)
return resize1, resize2, resize3, resize4, resize5
def to_clean_image(feature_map, resize1, resize2, resize3, resize4, resize5):
concat = tf.concat([feature_map, resize1, resize2, resize3, resize4, resize5], 3)
sk_conv1 = slim.conv2d(concat, 7, [3, 3], rate=1, activation_fn=lrelu)
sk_conv2 = slim.conv2d(concat, 7, [5, 5], rate=1, activation_fn=lrelu)
sk_conv3 = slim.conv2d(concat, 7, [7, 7], rate=1, activation_fn=lrelu)
sk_out = selective_kernel_layer(sk_conv1, sk_conv2, sk_conv3, 4, 7)
output = slim.conv2d(sk_out, 1, [3, 3], rate=1, activation_fn=None)
return output
def squeeze_excitation_layer(input_x, out_dim, middle):
squeeze = global_avg_pool(input_x)
excitation = tf.layers.dense(squeeze, use_bias=True, units=middle)
excitation = tf.nn.relu(excitation)
excitation = tf.layers.dense(excitation, use_bias=True, units=out_dim)
excitation = tf.nn.sigmoid(excitation)
excitation = tf.reshape(excitation, [-1, 1, 1, out_dim])
scale = input_x * excitation
return scale
def selective_kernel_layer(sk_conv1, sk_conv2, sk_conv3, middle, out_dim):
sum_u = sk_conv1 + sk_conv2 + sk_conv3
squeeze = global_avg_pool(sum_u)
squeeze = tf.reshape(squeeze, [-1, 1, 1, out_dim])
z = tf.layers.dense(squeeze, use_bias=True, units=middle)
z = tf.nn.relu(z)
a1 = tf.layers.dense(z, use_bias=True, units=out_dim)
a2 = tf.layers.dense(z, use_bias=True, units=out_dim)
a3 = tf.layers.dense(z, use_bias=True, units=out_dim)
before_softmax = tf.concat([a1, a2, a3], 1)
after_softmax = tf.nn.softmax(before_softmax, dim=1)
a1 = after_softmax[:, 0, :, :]
a1 = tf.reshape(a1, [-1, 1, 1, out_dim])
a2 = after_softmax[:, 1, :, :]
a2 = tf.reshape(a2, [-1, 1, 1, out_dim])
a3 = after_softmax[:, 2, :, :]
a3 = tf.reshape(a3, [-1, 1, 1, out_dim])
select_1 = sk_conv1 * a1
select_2 = sk_conv2 * a2
select_3 = sk_conv3 * a3
out = select_1 + select_2 + select_3
return out
def network(in_image):
feature_map = feature_encoding(in_image)
feature_map_2 = tf.concat([in_image, feature_map], 3)
pool1, pool2, pool3, pool4, pool5 = avg_pool(feature_map_2)
unet1, unet2, unet3, unet4, unet5 = all_unet(pool1, pool2, pool3, pool4, pool5)
resize1, resize2, resize3, resize4, resize5 = resize_all_image(unet1, unet2, unet3, unet4, unet5)
out_image = to_clean_image(feature_map_2, resize1, resize2, resize3, resize4, resize5)
return out_image