-
Notifications
You must be signed in to change notification settings - Fork 4
/
Fusion_losses.py
208 lines (152 loc) · 6.08 KB
/
Fusion_losses.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
import torch
import torch.nn.functional as F
import torch.nn as nn
from math import exp
import numpy as np
def gaussian(window_size, sigma):
gauss = torch.Tensor([exp(-(x - window_size//2)**2/float(2*sigma**2)) for x in range(window_size)])
return gauss/gauss.sum()
def create_window(window_size, channel=1):
_1D_window = gaussian(window_size, 1.5).unsqueeze(1) # sigma = 1.5 shape: [11, 1]
_2D_window = _1D_window.mm(_1D_window.t()).float().unsqueeze(0).unsqueeze(0) # unsqueeze()函数,增加维度 .t() 进行了转置 shape: [1, 1, 11, 11]
window = _2D_window.expand(channel, 1, window_size, window_size).contiguous() # window shape: [1,1, 11, 11]
return window
# 计算 ssim 损失函数
def mssim(img1, img2, window_size=11):
# Value range can be different from 255. Other common ranges are 1 (sigmoid) and 2 (tanh).
max_val = 255
min_val = 0
L = max_val - min_val
padd = window_size // 2
(_, channel, height, width) = img1.size()
# 滤波器窗口
window = create_window(window_size, channel=channel).to(img1.device)
mu1 = F.conv2d(img1, window, padding=padd, groups=channel)
mu2 = F.conv2d(img2, window, padding=padd, groups=channel)
mu1_sq = mu1.pow(2)
mu2_sq = mu2.pow(2)
mu1_mu2 = mu1 * mu2
sigma1_sq = F.conv2d(img1 * img1, window, padding=padd, groups=channel) - mu1_sq
sigma2_sq = F.conv2d(img2 * img2, window, padding=padd, groups=channel) - mu2_sq
sigma12 = F.conv2d(img1 * img2, window, padding=padd, groups=channel) - mu1_mu2
C1 = (0.01 * L) ** 2
C2 = (0.03 * L) ** 2
v1 = 2.0 * sigma12 + C2
v2 = sigma1_sq + sigma2_sq + C2
cs = torch.mean(v1 / v2) # contrast sensitivity
ssim_map = ((2 * mu1_mu2 + C1) * v1) / ((mu1_sq + mu2_sq + C1) * v2)
ret = ssim_map
return ret
def mse(img1, img2, window_size=9):
max_val = 255
min_val = 0
L = max_val - min_val
padd = window_size // 2
(_, channel, height, width) = img1.size()
img1_f = F.unfold(img1, (window_size, window_size), padding=padd)
img2_f = F.unfold(img2, (window_size, window_size), padding=padd)
res = (img1_f - img2_f) ** 2
res = torch.sum(res, dim=1, keepdim=True) / (window_size ** 2)
res = F.fold(res, output_size=(height, width), kernel_size=(1, 1))
return res
# 方差计算
def std(img, window_size=9):
padd = window_size // 2
(_, channel, height, width) = img.size()
window = create_window(window_size, channel=channel).to(img.device)
mu = F.conv2d(img, window, padding=padd, groups=channel)
mu_sq = mu.pow(2)
sigma1 = F.conv2d(img * img, window, padding=padd, groups=channel) - mu_sq
return sigma1
# def sum(img, window_size=9):
# padd = window_size // 2
# (_, channel, height, width) = img.size()
# window = create_window(window_size, channel=channel).to(img.device)
# win1 = torch.ones_like(window)
# res = F.conv2d(img, win1, padding=padd, groups=channel)
# return res
def final_ssim(img_ir, img_vis, img_fuse, mask=None):
ssim_ir = mssim(img_ir, img_fuse)
ssim_vi = mssim(img_vis, img_fuse)
# std_ir = std(img_ir)
# std_vi = std(img_vis)
std_ir = std(img_ir)
std_vi = std(img_vis)
zero = torch.zeros_like(std_ir)
one = torch.ones_like(std_vi)
# m = torch.mean(img_ir)
# w_ir = torch.where(img_ir > m, one, zero)
map1 = torch.where((std_ir - std_vi) > 0, one, zero)
map1 = map1
map2 = torch.where((std_ir - std_vi) >= 0, zero, one)
map2 = map2
ssim = map1 * ssim_ir + map2 * ssim_vi
# ssim = ssim * w_ir
return ssim.mean()
def final_mse(img_ir, img_vis, img_fuse):
mse_ir = mse(img_ir, img_fuse)
mse_vi = mse(img_vis, img_fuse)
std_ir = std(img_ir)
std_vi = std(img_vis)
# std_ir = sum(img_ir)
# std_vi = sum(img_vis)
zero = torch.zeros_like(std_ir)
one = torch.ones_like(std_vi)
m = torch.mean(img_ir)
w_vi = torch.where(img_ir <= m, one, zero)
map1 = torch.where((std_ir - std_vi) > 0, one, zero)
map2 = torch.where((std_ir - std_vi) >= 0, zero, one)
res = map1 * mse_ir + map2 * mse_vi
res = res * w_vi
return res.mean()
def final_mse1(img_ir, img_vis, img_fuse, mask=None):
mse_ir = mse(img_ir, img_fuse)
mse_vi = mse(img_vis, img_fuse)
std_ir = std(img_ir)
std_vi = std(img_vis)
# std_ir = sum(img_ir)
# std_vi = sum(img_vis)
zero = torch.zeros_like(std_ir)
one = torch.ones_like(std_vi)
m = torch.mean(img_ir)
map1 = torch.where((std_ir - std_vi) > 0, one, zero)
# map2 = torch.where((std_ir - std_vi) >= 0, zero, one)
map_ir=torch.where(map1+mask>0, one, zero)
map_vi= 1 - map_ir
res = map_ir * mse_ir + map_vi * mse_vi
# res = res * w_vi
return res.mean()
def corr_loss(image_ir, img_vis, img_fusion, eps=1e-6):
reg = REG()
corr = reg(image_ir, img_vis, img_fusion)
corr_loss = 1./(corr + eps)
return corr_loss
class REG(nn.Module):
"""
global normalized cross correlation (sqrt)
"""
def __init__(self):
super(REG, self).__init__()
def corr2(self, img1, img2):
img1 = img1 - img1.mean()
img2 = img2 - img2.mean()
r = torch.sum(img1*img2)/torch.sqrt(torch.sum(img1*img1)*torch.sum(img2*img2))
return r
def forward(self, a, b, c):
return self.corr2(a, c) + self.corr2(b, c)
if __name__ == '__main__':
criterion = mssim
input = torch.rand([1, 1, 64, 64])
output = torch.rand([1, 1, 64, 64])
img_fuse = torch.rand([1, 1, 64, 64])
mask = torch.zeros_like(img_fuse)
uw = torch.Tensor(np.ones((11, 11), dtype=float)) / 11
uw = uw.float().unsqueeze(0).unsqueeze(0)
# print(uw)
device = torch.device('cuda:{}'.format(2))
input = input.to(device)
output = output.to(device)
img_fuse = img_fuse.to(device)
mask = mask.to(device)
ssim = final_mse(input, output, img_fuse, mask)
print(ssim)