forked from YonghaoXu/SSFCN
-
Notifications
You must be signed in to change notification settings - Fork 0
/
SSFCN.py
262 lines (192 loc) · 8.8 KB
/
SSFCN.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
# -*- coding: utf-8 -*-
"""
@author: Sonic
"""
import pydensecrf.densecrf as dcrf
from pydensecrf.utils import compute_unary, create_pairwise_bilateral,create_pairwise_gaussian, softmax_to_unary
import time
import sys
import os
from HyperFunctions import *
import tensorflow as tf
resultpath = './SSFCN/'
if not os.path.isdir(resultpath):
os.makedirs(resultpath)
def weight_variable(shape,name=None):
initial = tf.random_uniform(shape,minval=0.01, maxval=0.02)
return tf.Variable(initial,name)
def weight(name=None):
initial = tf.random_uniform([1],minval=1, maxval=2)
return tf.Variable(initial,name)
def bias_variable(shape,name=None):
initial = tf.constant(0., shape=shape)
return tf.Variable(initial,name=None)
def conv2d(x, W, p=0):
if p==1:
return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')
else:
return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='VALID')
def atrousconv2d(x,W,rate):
return tf.nn.atrous_conv2d(x, W,rate,padding='SAME')
def max_pool(x,W,p=0):
if p==1:
return tf.nn.max_pool(x, ksize=[1, W, W, 1],
strides=[1, W, W, 1], padding='SAME')
else:
return tf.nn.max_pool(x, ksize=[1, W, W, 1],
strides=[1, W, W, 1], padding='VALID')
def max_pool_padding(x,W):
return tf.nn.max_pool(x, ksize=[1, W, W, 1],
strides=[1, 1, 1, 1], padding='SAME')
def avg_pool_padding(x,W):
return tf.nn.avg_pool(x, ksize=[1, W, W, 1],
strides=[1, 1, 1, 1], padding='SAME')
def one_hot3d(Y):
num_class = np.max(Y)
row,col = Y.shape
y = np.zeros((row,col,num_class),'uint8')
for i in range(1,num_class+1):
index = np.where(Y==i)
y[index[0],index[1],i-1] = 1
return y
x = tf.placeholder(tf.float32, shape=[1,610,340,103])
y_ = tf.placeholder(tf.float32, shape=[1,610,340,9])
idx1 = tf.placeholder(tf.int32, shape=None)
idx2 = tf.placeholder(tf.int32, shape=None)
idx3 = tf.placeholder(tf.int32, shape=None)
W_spectral_conv1 = weight_variable([1,1, 103, 64],'W_spectral_conv1')
b_spectral_conv1 = bias_variable([64],'b_spectral_conv1')
h_spectral_conv1 = tf.nn.relu(conv2d(x, W_spectral_conv1,1) + b_spectral_conv1)
W_spectral_conv2 = weight_variable([1,1, 64, 64],'W_spectral_conv2')
b_spectral_conv2 = bias_variable([64],'b_spectral_conv2')
h_spectral_conv2 = tf.nn.relu(conv2d(h_spectral_conv1, W_spectral_conv2,1) + b_spectral_conv2)
W_spectral_conv3 = weight_variable([1,1,64,64],'W_spectral_conv3')
b_spectral_conv3 = bias_variable([64],'b_spectral_conv3')
h_spectral_conv3 = tf.nn.relu(conv2d(h_spectral_conv2, W_spectral_conv3,1) + b_spectral_conv3)
h_spectral = h_spectral_conv1 + h_spectral_conv2 + h_spectral_conv3
rate=2
W_dr_conv1 = weight_variable([1,1, 103, 64],'W_dr_conv1')
b_dr_conv1 = bias_variable([64],'b_dr_conv1')
h_dr_conv1 = tf.nn.relu(conv2d(x, W_dr_conv1,1) + b_dr_conv1)
W_spatial_conv1 = weight_variable([3,3,64,64],'W_spatial_conv1')
b_spatial_conv1 = bias_variable([64],'b_spatial_conv1')
h_spatial_conv1 = tf.nn.relu(atrousconv2d(h_dr_conv1, W_spatial_conv1,2) + b_spatial_conv1)
h_spatial_pool1 = avg_pool_padding(h_spatial_conv1,2)
W_spatial_conv2 = weight_variable([3,3, 64, 64],'W_spatial_conv2')
b_spatial_conv2 = bias_variable([64],'b_spatial_conv2')
h_spatial_conv2 = tf.nn.relu(atrousconv2d(h_spatial_pool1, W_spatial_conv2,2) + b_spatial_conv2)
h_spatial_pool2 = avg_pool_padding(h_spatial_conv2,2)
h_spatial = h_dr_conv1 + h_spatial_pool1 + h_spatial_pool2
W_spatial = weight('W_spatial')
W_spectral = weight('W_spectral')
h_SS = W_spatial*h_spatial + W_spectral*h_spectral
W_conv5 = weight_variable([1,1, 64, 9],'W_conv5')
b_conv5 = bias_variable([9],'b_conv5')
y_conv = tf.nn.relu(conv2d(h_SS, W_conv5,1) + b_conv5)
y_prob = tf.nn.softmax(y_conv)
y_label = tf.argmax(y_prob,-1)
indices = tf.stack([idx3,idx1,idx2], axis=1)
y_conv_mask = tf.gather_nd(y_conv, indices)
y_mask = tf.gather_nd(y_, indices)
cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=y_conv_mask, labels=y_mask))
train_step = tf.train.AdamOptimizer(5e-4).minimize(cross_entropy)
correct_prediction = tf.equal(tf.argmax(y_conv_mask,-1), tf.argmax(y_mask,-1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
X = sio.loadmat('./DataSets/PaviaU.mat')['paviaU'].astype('float32')
Y = sio.loadmat('./DataSets/PaviaU_gt.mat')['paviaU_gt']
row,col,n_band = X.shape
X = np.reshape(featureNormalize(np.reshape(X,-1),2),(row,col,n_band))
num_class = np.max(Y)
Y_train = np.zeros(Y.shape).astype('int')
n_sample_train = 0
n_sample_test = 0
FCN_joint = np.zeros((num_class+3))
FCN_crf = np.zeros((num_class+3))
gpu_options = tf.GPUOptions(allow_growth=True)
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
init = tf.global_variables_initializer()
sess.run(init)
for i in range(1,num_class+1):
index = np.where(Y==i)
n_sample = len(index[0])
array = np.random.permutation(n_sample)
n_per = 150
if i==1:
array1_train = index[0][array[:n_per]]
array2_train = index[1][array[:n_per]]
array1_test = index[0][array[n_per:]]
array2_test = index[1][array[n_per:]]
else:
array1_train = np.concatenate((array1_train,index[0][array[:n_per]]))
array2_train = np.concatenate((array2_train,index[1][array[:n_per]]))
array1_test = np.concatenate((array1_test,index[0][array[n_per:]]))
array2_test = np.concatenate((array2_test,index[1][array[n_per:]]))
Y_train[index[0][array[:n_per]],index[1][array[:n_per]]] = i
n_sample_train += n_per
n_sample_test += n_sample-n_per
array3 = np.zeros(array1_train.shape)
y_train = one_hot3d(Y_train)
y_train = np.reshape(y_train,(1,row,col,num_class))
X_train = np.reshape(X,(1,row,col,n_band))
mask_train = np.zeros(y_train.shape)
for i in range(num_class):
mask_train[:,:,:,i] = np.sum(y_train,-1)
time1 = time.time()
num_epoch = 4000
histloss = np.zeros((num_epoch,2))
for i in range(num_epoch):
train_accuracy = 0
train_loss = 0
sess.run(train_step,feed_dict={x:X_train,y_:y_train,idx1:array1_train,idx2:array2_train,idx3:array3})
if (i+1)%100==0:
train_accuracy,train_loss = sess.run([accuracy,cross_entropy],feed_dict={x:X_train,y_:y_train,idx1:array1_train,idx2:array2_train,idx3:array3})
histloss[i,:] = train_accuracy,train_loss
print("epoch %d, train_accuracy %g, train_loss %g"%(i+1, histloss[i,0],histloss[i,1]))
label = np.squeeze(sess.run(y_label,feed_dict={x:X_train,y_: y_train,idx1:array1_train,idx2:array2_train,idx3:array3}))
prob = np.squeeze(sess.run(y_prob,feed_dict={x:X_train,y_: y_train,idx1:array1_train,idx2:array2_train,idx3:array3}))
y_test = Y[array1_test,array2_test]-1
y_pred = label[array1_test,array2_test]
OA,kappa,ProducerA = CalAccuracy(y_pred,y_test)
FCN_joint[:num_class] = ProducerA
FCN_joint[-3] = OA
FCN_joint[-2] = kappa
time2 = time.time()
FCN_joint[-1] = time2 - time1
print("Running time: %g"%(FCN_joint[-1]))
img = DrawResult(np.reshape(label+1,-1),1)
plt.imsave(resultpath+'FCNJoint'+'_'+repr(int(OA*10000))+'.png',img)
time1 = time.time()
CNNMap = np.squeeze(sess.run(h_SS,feed_dict={x:X_train,y_: y_train,idx1:array1_train,idx2:array2_train,idx3:array3}))
n_feature = CNNMap.shape[-1]
PC_num = 5
image = featureNormalize(PCANorm(CNNMap.reshape(row*col, n_feature),PC_num),1).reshape(row,col,PC_num)
softmax = prob.transpose((2, 0, 1))
unary = softmax_to_unary(softmax)
unary = np.ascontiguousarray(unary)
d = dcrf.DenseCRF(image.shape[0] * image.shape[1], 9)
d.setUnaryEnergy(unary)
feats = create_pairwise_gaussian(sdims=(10, 10), shape=image.shape[:2])
d.addPairwiseEnergy(feats, compat=1,
kernel=dcrf.DIAG_KERNEL,
normalization=dcrf.NORMALIZE_SYMMETRIC)
feats = create_pairwise_bilateral(sdims=(30, 30), schan=[5],
img=image, chdim=2)
d.addPairwiseEnergy(feats, compat=10,
kernel=dcrf.DIAG_KERNEL,
normalization=dcrf.NORMALIZE_SYMMETRIC)
Q = d.inference(5)
res = np.argmax(Q, axis=0).reshape((image.shape[0], image.shape[1]))
y_test = Y[array1_test,array2_test]-1
y_pred = res[array1_test,array2_test]
OA,kappa,ProducerA = CalAccuracy(y_pred,y_test)
FCN_crf[:num_class] = ProducerA
FCN_crf[-3] = OA
FCN_crf[-2] = kappa
time2 = time.time()
FCN_crf[-1] = FCN_joint[-1] + time2 - time1
crf_img = DrawResult(np.reshape(res+1,-1),1)
plt.imsave(resultpath+'FCNcrf'+'_'+repr(int(OA*10000))+'.png',crf_img)
FCN_joint[:-1] = FCN_joint[:-1]*100
FCN_crf[:-1] = FCN_crf[:-1]*100
sio.savemat(resultpath+'FCN_joint.mat', {'FCN_joint': FCN_joint})
sio.savemat(resultpath+'FCN_crf.mat', {'FCN_crf': FCN_crf})