-
Notifications
You must be signed in to change notification settings - Fork 10
/
noise.py
255 lines (203 loc) · 9 KB
/
noise.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
import numpy as np
import scipy.stats as stats
from os.path import join
from scipy.stats import tukeylambda
class RawPacker:
def __init__(self, cfa='bayer'):
self.cfa = cfa
def pack_raw_bayer(self, cfa_img):
# pack Bayer image to 4 channels
img_shape = cfa_img.shape
H = img_shape[0]
W = img_shape[1]
out = np.stack((cfa_img[0:H:2, 0:W:2], # RGBG
cfa_img[0:H:2, 1:W:2],
cfa_img[1:H:2, 1:W:2],
cfa_img[1:H:2, 0:W:2]), axis=0).astype(np.float32)
return out
def pack_raw_xtrans(self, cfa_img):
# pack X-Trans image to 9 channels
img_shape = cfa_img.shape
H = (img_shape[0] // 6) * 6
W = (img_shape[1] // 6) * 6
out = np.zeros((9, H // 3, W // 3), dtype=np.float32)
# 0 R
out[0, 0::2, 0::2] = cfa_img[0:H:6, 0:W:6]
out[0, 0::2, 1::2] = cfa_img[0:H:6, 4:W:6]
out[0, 1::2, 0::2] = cfa_img[3:H:6, 1:W:6]
out[0, 1::2, 1::2] = cfa_img[3:H:6, 3:W:6]
# 1 G
out[1, 0::2, 0::2] = cfa_img[0:H:6, 2:W:6]
out[1, 0::2, 1::2] = cfa_img[0:H:6, 5:W:6]
out[1, 1::2, 0::2] = cfa_img[3:H:6, 2:W:6]
out[1, 1::2, 1::2] = cfa_img[3:H:6, 5:W:6]
# 1 B
out[2, 0::2, 0::2] = cfa_img[0:H:6, 1:W:6]
out[2, 0::2, 1::2] = cfa_img[0:H:6, 3:W:6]
out[2, 1::2, 0::2] = cfa_img[3:H:6, 0:W:6]
out[2, 1::2, 1::2] = cfa_img[3:H:6, 4:W:6]
# 4 R
out[3, 0::2, 0::2] = cfa_img[1:H:6, 2:W:6]
out[3, 0::2, 1::2] = cfa_img[2:H:6, 5:W:6]
out[3, 1::2, 0::2] = cfa_img[5:H:6, 2:W:6]
out[3, 1::2, 1::2] = cfa_img[4:H:6, 5:W:6]
# 5 B
out[4, 0::2, 0::2] = cfa_img[2:H:6, 2:W:6]
out[4, 0::2, 1::2] = cfa_img[1:H:6, 5:W:6]
out[4, 1::2, 0::2] = cfa_img[4:H:6, 2:W:6]
out[4, 1::2, 1::2] = cfa_img[5:H:6, 5:W:6]
out[5, :, :] = cfa_img[1:H:3, 0:W:3]
out[6, :, :] = cfa_img[1:H:3, 1:W:3]
out[7, :, :] = cfa_img[2:H:3, 0:W:3]
out[8, :, :] = cfa_img[2:H:3, 1:W:3]
return out
def unpack_raw_bayer(self, img):
# unpack 4 channels to Bayer image
img4c = img
_, h, w = img.shape
H = int(h * 2)
W = int(w * 2)
cfa_img = np.zeros((H, W), dtype=np.float32)
cfa_img[0:H:2, 0:W:2] = img4c[0, :,:]
cfa_img[0:H:2, 1:W:2] = img4c[1, :,:]
cfa_img[1:H:2, 1:W:2] = img4c[2, :,:]
cfa_img[1:H:2, 0:W:2] = img4c[3, :,:]
return cfa_img
def unpack_raw_xtrans(self, img):
img9c = img
_, h, w = img.shape
H = int(h * 3)
W = int(w * 3)
cfa_img = np.zeros((H, W), dtype=np.float32)
# 0 R
cfa_img[0:H:6, 0:W:6] = img9c[0, 0::2, 0::2]
cfa_img[0:H:6, 4:W:6] = img9c[0, 0::2, 1::2]
cfa_img[3:H:6, 1:W:6] = img9c[0, 1::2, 0::2]
cfa_img[3:H:6, 3:W:6] = img9c[0, 1::2, 1::2]
# 1 G
cfa_img[0:H:6, 2:W:6] = img9c[1, 0::2, 0::2]
cfa_img[0:H:6, 5:W:6] = img9c[1, 0::2, 1::2]
cfa_img[3:H:6, 2:W:6] = img9c[1, 1::2, 0::2]
cfa_img[3:H:6, 5:W:6] = img9c[1, 1::2, 1::2]
# 1 B
cfa_img[0:H:6, 1:W:6] = img9c[2, 0::2, 0::2]
cfa_img[0:H:6, 3:W:6] = img9c[2, 0::2, 1::2]
cfa_img[3:H:6, 0:W:6] = img9c[2, 1::2, 0::2]
cfa_img[3:H:6, 4:W:6] = img9c[2, 1::2, 1::2]
# 4 R
cfa_img[1:H:6, 2:W:6] = img9c[3, 0::2, 0::2]
cfa_img[2:H:6, 5:W:6] = img9c[3, 0::2, 1::2]
cfa_img[5:H:6, 2:W:6] = img9c[3, 1::2, 0::2]
cfa_img[4:H:6, 5:W:6] = img9c[3, 1::2, 1::2]
# 5 B
cfa_img[2:H:6, 2:W:6] = img9c[4, 0::2, 0::2]
cfa_img[1:H:6, 5:W:6] = img9c[4, 0::2, 1::2]
cfa_img[4:H:6, 2:W:6] = img9c[4, 1::2, 0::2]
cfa_img[5:H:6, 5:W:6] = img9c[4, 1::2, 1::2]
cfa_img[1:H:3, 0:W:3] = img9c[5, :, :]
cfa_img[1:H:3, 1:W:3] = img9c[6, :, :]
cfa_img[2:H:3, 0:W:3] = img9c[7, :, :]
cfa_img[2:H:3, 1:W:3] = img9c[8, :, :]
return cfa_img
def pack_raw(self, cfa_img):
if self.cfa == 'bayer':
out = self.pack_raw_bayer(cfa_img)
elif self.cfa == 'xtrans':
out = self.pack_raw_xtrans(cfa_img)
else:
raise NotImplementedError
return out
def unpack_raw(self, img):
if self.cfa == 'bayer':
out = self.unpack_raw_bayer(img)
elif self.cfa == 'xtrans':
out = self.unpack_raw_xtrans(img)
else:
raise NotImplementedError
return out
class NoiseModelBase: # base class
def __call__(self, y, params=None, continuous=False):
if params is None:
K, g_scale, saturation_level, ratio, sigma_r, G_scale, G_shape = self._sample_params(continuous)
else:
K, g_scale, saturation_level, ratio, sigma_r, G_scale, G_shape = params
y = y * saturation_level
y = y / ratio
if "u" in self.model: # quantization noise
y = y + (np.random.uniform(0, 1, y.shape) - 0.5)
y = y.clip(0)
if 'P' in self.model:
z = np.random.poisson(y / K).astype(np.float32) * K
elif 'p' in self.model:
z = y + np.random.randn(*y.shape).astype(np.float32) * np.sqrt(np.maximum(K * y, 1e-10))
else:
z = y
if 'r' in self.model: # row noise
z = self.raw_packer.unpack_raw(z)
z = z + np.tile(np.random.randn(z.shape[0])[:,np.newaxis], (1, z.shape[1])).astype(np.float32) * np.maximum(sigma_r, 1e-10)
z = self.raw_packer.pack_raw(z)
if 'g' in self.model:
z = z + np.random.randn(*y.shape).astype(np.float32) * np.maximum(g_scale, 1e-10) # Gaussian noise
elif 'G' in self.model:
lam = np.random.choice(G_shape)
z = z + tukeylambda.rvs(lam, size=z.shape).astype("float32") * np.maximum(G_scale, 1e-10)
else:
z = z
z = z * ratio
z = z / saturation_level
return z, {"K": K, "ratio": ratio, "g_scale": g_scale, "saturation_level": saturation_level}
# Only support baseline noise models: G / G+P / G+P*
class NoiseModel(NoiseModelBase):
def __init__(self, model='g', cameras=None, include=None, exclude=None, cfa='bayer'):
super().__init__()
assert cfa in ['bayer', 'xtrans']
assert include is None or exclude is None
self.cameras = cameras or ['CanonEOS5D4', 'CanonEOS70D', 'CanonEOS700D', 'NikonD850', 'SonyA7S2']
if include is not None:
self.cameras = [self.cameras[include]]
if exclude is not None:
exclude_camera = set([self.cameras[exclude]])
self.cameras = list(set(self.cameras) - exclude_camera)
self.param_dir = join('camera_params', 'release')
print('[i] NoiseModel with {}'.format(self.param_dir))
print('[i] cameras: {}'.format(self.cameras))
print('[i] using noise model {}'.format(model))
self.camera_params = {}
for camera in self.cameras:
self.camera_params[camera] = np.load(join(self.param_dir, camera+'_params.npy'), allow_pickle=True).item()
self.model = model
self.raw_packer = RawPacker(cfa)
def ISO_to_K(self, ISO):
camera_params = self.camera_params[self.cameras[0]]
Kmin = camera_params['Kmin']
Kmax = camera_params['Kmax']
k = (ISO - 100)/(6400-100) *(Kmax-Kmin)+Kmin
return k
def _sample_params(self, continuous):
camera = np.random.choice(self.cameras)
# print(camera)
saturation_level = 16383 - 800
profiles = ['Profile-1']
camera_params = self.camera_params[camera]
G_shape = camera_params["G_shape"]
Kmin = camera_params['Kmin']
Kmax = camera_params['Kmax']
profile = np.random.choice(profiles)
camera_params = camera_params[profile]
# log_K = np.random.uniform(low=np.log(Kmin), high=np.log(Kmax))
log_K = np.random.uniform(low=np.log(1e-1), high=np.log(30))
log_g_scale = np.random.standard_normal() * camera_params['g_scale']['sigma'] * 1 +\
camera_params['g_scale']['slope'] * log_K + camera_params['g_scale']['bias']
K = np.exp(log_K)
g_scale = np.exp(log_g_scale)
if continuous:
ratio = np.random.uniform(low=20, high=300)
else:
ratio = np.random.uniform(low=100, high=300)
log_r = np.random.standard_normal() * camera_params['R_scale']['sigma'] * 1 +\
camera_params['R_scale']['slope'] * log_K + camera_params['R_scale']['bias']
log_G_scale = np.random.standard_normal() * camera_params['G_scale']['sigma'] * 1 +\
camera_params['G_scale']['slope'] * log_K + camera_params['G_scale']['bias']
sigma_r = np.exp(log_r)
G_scale = np.exp(log_G_scale)
return (K, g_scale, saturation_level, ratio, sigma_r, G_scale, G_shape)