-
Notifications
You must be signed in to change notification settings - Fork 41
/
kpn_data_provider.py
266 lines (223 loc) · 11 KB
/
kpn_data_provider.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
import torch
import torch.nn as nn
import torchvision.transforms as transforms
import torch.nn.functional as F
from torch.utils.data import Dataset, DataLoader
import os
from PIL import Image
import numpy as np
from skimage.color import rgb2xyz
import inspect
from utils.training_util import read_config
from data_generation.data_utils import *
import torch.nn.functional as F
def sRGBGamma(tensor):
threshold = 0.0031308
a = 0.055
mult = 12.92
gamma = 2.4
res = torch.zeros_like(tensor)
mask = tensor > threshold
# image_lo = tensor * mult
# 0.001 is to avoid funny thing at 0.
# image_hi = (1 + a) * torch.pow(tensor + 0.001, 1.0 / gamma) - a
res[mask] = (1 + a) * torch.pow(tensor[mask] + 0.001, 1.0 / gamma) - a
res[1-mask] = tensor[1-mask] * mult
# return mask * image_hi + (1 - mask) * image_lo
return res
def UndosRGBGamma(tensor):
threshold = 0.0031308
a = 0.055
mult = 12.92
gamma = 2.4
res = torch.zeros_like(tensor)
mask = tensor > threshold
# image_lo = tensor / mult
# image_hi = torch.pow(tensor + a, gamma) / (1 + a)
res[1-mask] = tensor[1-mask] / mult
res[mask] = torch.pow(tensor[mask] + a, gamma) / (1 + a)
# return mask * image_hi + (1 - mask) * image_lo
return res
class Random_Horizontal_Flip(object):
def __init__(self, p=0.5):
self.p = p
def __call__(self, tensor):
if np.random.rand() < self.p:
return torch.flip(tensor, dims=[-1])
return tensor
class Random_Vertical_Flip(object):
def __init__(self, p=0.5):
self.p = p
def __call__(self, tensor):
if np.random.rand() < self.p:
return torch.flip(tensor, dims=[-2])
return tensor
class TrainDataSet(torch.utils.data.Dataset):
def __init__(self, config_file,
config_spec=None, img_format='.bmp', degamma=True, color=True, blind=False, train=True):
super(TrainDataSet, self).__init__()
if config_spec is None:
config_spec = self._configspec_path()
config = read_config(config_file, config_spec)
self.dataset_config = config['dataset_configs']
self.dataset_dir = self.dataset_config['dataset_dir']
self.images = list(filter(lambda x: True if img_format in x else False, os.listdir(self.dataset_dir)))
self.burst_size = self.dataset_config['burst_length']
self.patch_size = self.dataset_config['patch_size']
self.upscale = self.dataset_config['down_sample']
self.big_jitter = self.dataset_config['big_jitter']
self.small_jitter = self.dataset_config['small_jitter']
# 对应下采样之前图像的最大偏移量
self.jitter_upscale = self.big_jitter * self.upscale
# 对应下采样之前的图像的patch尺寸
self.size_upscale = self.patch_size * self.upscale + 2 * self.jitter_upscale
# 产生大jitter和小jitter之间的delta 在下采样之前的尺度上
self.delta_upscale = (self.big_jitter - self.small_jitter) * self.upscale
# 对应到原图的patch的尺寸
self.patch_size_upscale = self.patch_size * self.upscale
# 去伽马效应
self.degamma = degamma
# 是否用彩色图像进行处理
self.color = color
# 是否盲估计 盲估计即估计的噪声方差不会作为网络的输入
self.blind = blind
self.train = train
self.vertical_flip = Random_Vertical_Flip(p=0.5)
self.horizontal_flip = Random_Horizontal_Flip(p=0.5)
@staticmethod
def _configspec_path():
current_dir = os.path.dirname(
os.path.abspath(inspect.getfile(inspect.currentframe()))
)
return os.path.join(current_dir,
'dataset_specs/data_configspec.conf')
@staticmethod
def crop_random(tensor, patch_size):
return random_crop(tensor, 1, patch_size)[0]
# get一个item 根据index检索
def __getitem__(self, index):
# print(index)
image = Image.open(os.path.join(self.dataset_dir, self.images[index])).convert('RGB')
# 先转换为Tensor进行degamma
image = transforms.ToTensor()(image)
# if self.degamma:
# image = UndosRGBGamma(tensor=image)
image_crop = self.crop_random(image, self.size_upscale)
# 3*H*W 对应于较小jitter下
image_crop_small = image_crop[:, self.delta_upscale:-self.delta_upscale,
self.delta_upscale:-self.delta_upscale]
# 进一步进行random_crop所需的transform
# burst中的第一个不做偏移 后期作为target
# output shape: N*3*H*W
img_burst = []
for i in range(self.burst_size):
if i == 0:
img_burst.append(
image_crop[:, self.jitter_upscale:-self.jitter_upscale, self.jitter_upscale:-self.jitter_upscale]
)
else:
if np.random.binomial(1, min(1.0, np.random.poisson(lam=1.5) / self.burst_size)) == 0:
img_burst.append(
self.crop_random(
image_crop_small, self.patch_size_upscale
)
)
else: #big
img_burst.append(
self.crop_random(image_crop, self.patch_size_upscale)
)
image_burst = torch.stack(img_burst, dim=0)
image_burst = F.adaptive_avg_pool2d(image_burst, (self.patch_size, self.patch_size))
# label为patch中burst的第一个
if not self.color:
image_burst = 0.2989*image_burst[:, 0, ...] + 0.5870 * image_burst[:, 1, ...] + 0.1140*image_burst[:, 2, ...]
image_burst = torch.clamp(image_burst, 0.0, 1.0)
if self.degamma:
UndosRGBGamma(image_burst)
if self.train:
# data augment
image_burst = self.horizontal_flip(image_burst)
image_burst = self.vertical_flip(image_burst)
gt = image_burst[0, ...]
# 以上得到的patch size为burst*（3）*size*size
"""
数据加噪声等一系列处理 全部基于rgb图像做
"""
# 要产生[log10(0.1), log10(1.0)]之间的均匀分布随机数 也就是[0,1加负号即可]
# 产生pred之后 再除以white_level恢复原来的亮度
# batch中的每一个burst 产生一个white_level
white_level = torch.from_numpy(np.power(10, -np.random.rand(1, 1, 1))).type_as(image_burst)
# 论文中对图像亮度赋值进行线性缩放[0.1, 1]
image_burst = white_level * image_burst
# gray image
if not self.color:
# 生成随机的read和shot噪声方差
sigma_read = torch.from_numpy(
np.power(10, np.random.uniform(-3.0, -1.5, (1, 1, 1)))).type_as(image_burst)
sigma_shot = torch.from_numpy(
np.power(10, np.random.uniform(-4.0, -2.0, (1, 1, 1)))).type_as(image_burst)
# sigma_read = torch.from_numpy(2*np.power(10, np.array([[[-2.0]]]))).type_as(image_burst)
# sigma_shot = torch.from_numpy(6.4 * np.power(10, np.array([[[-3.0]]]))).type_as(image_burst)
# 产生噪声 依据论文中公式产生
sigma_read_com = sigma_read.expand_as(image_burst)
sigma_shot_com = sigma_shot.expand_as(image_burst)
# generate noise
burst_noise = torch.normal(image_burst, torch.sqrt(sigma_read_com**2 + image_burst * sigma_shot_com)).type_as(image_burst)
# burst_noise 恢复到[0,1] 截去外面的值
burst_noise = torch.clamp(burst_noise, 0.0, 1.0)
# 非盲估计 就要估计噪声的方差
if not self.blind:
# 接下来就是根据两个sigma 将估计的噪声标准差也作为输入 用burst中的第一个进行估计
# estimation shape: H*W
sigma_read_est = sigma_read.view(1, 1).expand_as(gt)
sigma_shot_est = sigma_shot.view(1, 1).expand_as(gt)
sigma_estimate = torch.sqrt(sigma_read_est ** 2 + sigma_shot_est.mul(
torch.max(torch.stack([burst_noise[0, ...], torch.zeros_like(burst_noise[0, ...])], dim=0), dim=0)[0]))
# 把噪声的估计和burst图像连接在一起
burst_noise = torch.cat([burst_noise, sigma_estimate.unsqueeze(0)], dim=0)
# 按照文章中的 ref Image作为target进行了训练 输出结果和ref很相似 没能起到太大的去噪作用
# return patches_with_noise, patches_with_noise[:, 0, ...], white_level
# 不含噪声的ref作为target进行测试
return burst_noise, gt, white_level
# color image
else:
# 生成随机的read和shot噪声方差
sigma_read = torch.from_numpy(
np.power(10, np.random.uniform(-3.0, -1.5, (1, 1, 1, 1)))).type_as(image_burst)
sigma_shot = torch.from_numpy(
np.power(10, np.random.uniform(-4.0, -2.0, (1, 1, 1, 1)))).type_as(image_burst)
# 产生噪声 依据论文中公式产生
sigma_read_com = sigma_read.expand_as(image_burst)
sigma_shot_com = sigma_shot.expand_as(image_burst)
# generate noise
burst_noise = torch.normal(image_burst,
torch.sqrt(sigma_read_com ** 2 + image_burst * sigma_shot_com)).type_as(image_burst)
# burst_noise 恢复到[0,1] 截去外面的值
burst_noise = torch.clamp(burst_noise, 0.0, 1.0)
# 非盲估计 就要估计噪声的方差
if not self.blind:
# 接下来就是根据两个sigma 将估计的噪声标准差也作为输入 用burst中的第一个进行估计
# estimation shape: H*W
sigma_read_est = sigma_read.view(1, 1, 1).expand_as(gt)
sigma_shot_est = sigma_shot.view(1, 1, 1).expand_as(gt)
sigma_estimate = torch.sqrt(sigma_read_est ** 2 + sigma_shot_est.mul(
torch.max(torch.stack([burst_noise[0, ...], torch.zeros_like(burst_noise[0, ...])], dim=0), dim=0)[0]))
# 把噪声的估计和burst图像连接在一起
burst_noise = torch.cat([burst_noise, sigma_estimate.unsqueeze(0)], dim=0)
white_level = white_level.unsqueeze(0)
return burst_noise, gt, white_level
def __len__(self):
return len(self.images)
if __name__ == '__main__':
# path = 'F:/BinZhang/Codes/deep-burst-denoising/data/train'
# dataset = TrainDataSet(path, '.jpg', 8, 128, 4, 16, 2, color=False)
# dataloader = DataLoader(dataset,
# batch_size=4,
# shuffle=True,
# num_workers=4)
# dataloader = iter(dataloader)
# a, b, c = next(dataloader)
# print(a.size(), b.size(), c.size())
hf = Random_Horizontal_Flip(0.5)
a = torch.randint(0, 10, (2, 2))
print(a, hf(a))