-
Notifications
You must be signed in to change notification settings - Fork 3
/
test_denoise_LRD.py
192 lines (158 loc) · 9.08 KB
/
test_denoise_LRD.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
import torch
import rawpy
import numpy as np
import os
from tqdm import tqdm
import argparse
from torch.nn import functional as F
from raw_process.SID_post_process import tensor2im, quality_assess
import torchvision.utils as vutils
import glob
from raw_process import LRD_post_process as process
from models.denoise_model import UNetSeeInDark
from utils import get_config
from datasets.Denoise_dataset import pack_raw_LRD, metainfo
def forward_patches_denoise(model, input_img, patch_size=256, pad=32):
shift = patch_size - pad * 2
input_img = F.pad(input_img, (pad, pad, pad, pad), mode='reflect')
denoised = torch.zeros_like(input_img)
_, _, H, W = input_img.shape
for i in np.arange(0, H, shift):
for j in np.arange(0, W, shift):
h_end, w_end = min(i + patch_size, H), min(j + patch_size, W)
h_start, w_start = h_end - patch_size, w_end - patch_size
input_var = input_img[..., h_start: h_end, w_start: w_end]
with torch.no_grad():
out_var = model(input_var)
denoised[..., h_start + pad: h_end - pad, w_start + pad: w_end - pad] = \
out_var[..., pad:-pad, pad:-pad]
denoised = denoised[..., pad:-pad, pad:-pad]
denoised = denoised.clip(0, 1)
return denoised
def denoise_test_unpair(denoise_model, data_dir):
img_list_noisy = sorted(glob.glob(os.path.join(data_dir, 'test_set', '*.DNG')))
total_iter = 0
for idx, noisy_image in tqdm(enumerate(img_list_noisy), total=len(img_list_noisy), desc='Train epoch={}, iter={}'.format(0, total_iter), ncols=80, leave=False):
total_iter += 1
raw_noisy = rawpy.imread(img_list_noisy[idx])
noisy_image = pack_raw_LRD(raw_noisy)
ISO, EXPO, EV, cshot, cread = metainfo(img_list_noisy[idx])
ratio_EV = 1 / 2 ** EV
noisy_image = noisy_image * ratio_EV
noisy_image = np.maximum(np.minimum(noisy_image, 1.0), 0)
noisy_image = np.ascontiguousarray(noisy_image)
noisy_tensor = torch.from_numpy(noisy_image).unsqueeze(0).cuda()
denoise_model.eval()
with torch.no_grad():
denoise_tensor = forward_patches_denoise(denoise_model, noisy_tensor)
with rawpy.imread(img_list_noisy[idx]) as noisy_raw:
camera_whitebalance = torch.Tensor(noisy_raw.camera_whitebalance)
red_gain = camera_whitebalance[0].cuda()
blue_gain = camera_whitebalance[2].cuda()
ccm = [[1.359, -0.357, -0.001], [-0.209, 1.258, -0.049], [-0.049, -0.499, 1.548]]
cam2rgb = torch.FloatTensor(ccm).unsqueeze(0).cuda()
denoise_rbg_image = process.process(denoise_tensor, red_gains=red_gain, blue_gains=blue_gain, cam2rgbs=cam2rgb)
file_name_denoise = './test_unpair_denoise_dir/noisy_denoised/%s.bmp' % total_iter
vutils.save_image(denoise_rbg_image.data, file_name_denoise, nrow=1)
noisy_rbg_image = process.process(noisy_tensor, red_gains=red_gain, blue_gains=blue_gain, cam2rgbs=cam2rgb)
file_name_noisy = './test_unpair_denoise_dir/noisy_input/%s.bmp' % total_iter
vutils.save_image(noisy_rbg_image.data, file_name_noisy, nrow=1)
print("Total test raw image number: %s" % total_iter)
def denoise_test_pair(denoise_model, data_list, data_dir, save_path, set_ev):
img_info = []
with open(data_list, 'r') as f:
for i, img_pair in enumerate(f):
img_pair = img_pair.strip() # ./Sony/short/10003_00_0.04s.ARW ./Sony/long/10003_00_10s.ARW ISO200 F9
clean_file, noisy_file, iso, ev, _ = img_pair.split(' ')
ISO = int(''.join([x for x in iso if x.isdigit()]))
EV = -float(''.join([x for x in ev if x.isdigit()]))
clean_exposure = float(os.path.split(clean_file)[-1][9:-5]) # 0.04
noisy_exposure = float(os.path.split(noisy_file)[-1][9:-5]) # 10
ratio = (100 * clean_exposure) / (ISO * noisy_exposure)
if set_ev - 0.5 < EV < set_ev + 0.5:
img_info.append({
'clean': clean_file,
'noisy': noisy_file,
'clean_exposure': clean_exposure,
'noisy_exposure': noisy_exposure,
'ratio': ratio,
'iso': ISO,
'ev': EV,
})
total_iter = 0
psnr = 0
ssim = 0
count = 0
KLD = []
for idx, info in tqdm(enumerate(img_info), total=len(img_info),
desc='Train epoch={}, iter={}'.format(0, total_iter), ncols=80, leave=False):
clean_fn, noisy_fn, iso, ratio, ev = info['clean'], info['noisy'], info['iso'], info['ratio'], info['ev']
total_iter += 1
count += 1
clean_path = os.path.join(data_dir, clean_fn)
noisy_path = os.path.join(data_dir, noisy_fn)
clean_raw = rawpy.imread(clean_path)
clean_image = pack_raw_LRD(clean_raw)
noise_raw = rawpy.imread(noisy_path)
noisy_image = pack_raw_LRD(noise_raw) * ratio
noisy_image = np.maximum(np.minimum(noisy_image, 1.0), 0)
noisy_image = np.ascontiguousarray(noisy_image)
clean_image = np.ascontiguousarray(clean_image)
clean_tensor = torch.from_numpy(clean_image).unsqueeze(0).cuda()
noisy_tensor = torch.from_numpy(noisy_image).unsqueeze(0).cuda()
denoise_model.eval()
with torch.no_grad():
denoised_tensor = forward_patches_denoise(denoise_model, noisy_tensor)
clean_numpy = tensor2im(clean_tensor)
denoised_numpy = tensor2im(denoised_tensor)
res = quality_assess(denoised_numpy, clean_numpy, data_range=255)
psnr += res['PSNR']
ssim += res['SSIM']
with rawpy.imread(clean_path) as clean_raw:
camera_whitebalance = torch.Tensor(clean_raw.camera_whitebalance)
red_gain = camera_whitebalance[0].cuda()
blue_gain = camera_whitebalance[2].cuda()
ccm = [[1.359, -0.357, -0.001], [-0.209, 1.258, -0.049], [-0.049, -0.499, 1.548]]
cam2rgb = torch.FloatTensor(ccm).unsqueeze(0).cuda()
if not os.path.exists('%s/denoised/' % save_path):
print("Creating directory: {}".format('%s/denoised/' % save_path))
os.makedirs('%s/denoised/' % save_path)
if not os.path.exists('%s/noisy/' % save_path):
print("Creating directory: {}".format('%s/noisy/' % save_path))
os.makedirs('%s/noisy/' % save_path)
if not os.path.exists('%s/clean/' % save_path):
print("Creating directory: {}".format('%s/clean/' % save_path))
os.makedirs('%s/clean/' % save_path)
denoise_rbg_image = process.process(denoised_tensor, red_gains=red_gain, blue_gains=blue_gain, cam2rgbs=cam2rgb)
vutils.save_image(denoise_rbg_image.data, '%s/denoised/%s_ISO%s_%sEV.bmp' % (save_path, total_iter, iso, ev), nrow=1)
noisy_rbg_image = process.process(noisy_tensor, red_gains=red_gain, blue_gains=blue_gain, cam2rgbs=cam2rgb)
vutils.save_image(noisy_rbg_image.data, '%s/noisy/%s_ISO%s_%sEV.bmp' % (save_path, total_iter, iso, ev), nrow=1)
clean_rbg_image = process.process(clean_tensor, red_gains=red_gain, blue_gains=blue_gain, cam2rgbs=cam2rgb)
vutils.save_image(clean_rbg_image.data, '%s/clean/%s_ISO%s_%sEV.bmp' % (save_path, total_iter, iso, ev), nrow=1)
print("===> PSNR: {}, SSIM:{}".format(psnr / count, ssim / count))
print("KL Distance: %s" % (np.sum(KLD) / total_iter))
print("Total test raw image number: %s" % total_iter)
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Denoise args setting')
parser.add_argument('--config', type=str, default='./configs/Noise_Synthesis_LRD.yaml', help='Path to the config file.')
parser.add_argument('--checkpoints_denoise', type=str, default='./checkpoints_denoise/IMX586_Ours_best_model.pth', help='Path to the config file.')
parser.add_argument("--resume", action="store_true")
parser.add_argument('--seed', type=int, default=3407, help="random seed")
parser.add_argument('--EV', type=int, default=-3, help="set up the target exposure bias")
opts = parser.parse_args()
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
torch.manual_seed(opts.seed)
torch.cuda.manual_seed(opts.seed)
config = get_config(opts.config)
raw_dir = config['LRD_root']
test_img_list = config['test_img_list']
unpair_data_dir = '/data/data_ssd/DJI/'
save_dir = './LRD_denoise_results_%sEV/' % opts.EV
if not os.path.exists(save_dir):
print("Creating directory: {}".format(save_dir))
os.makedirs(save_dir)
denoise_model = UNetSeeInDark(in_channels=4, out_channels=4).cuda()
state_dict_denoise = torch.load(opts.checkpoints_denoise, map_location='cpu')
denoise_model.load_state_dict(state_dict_denoise)
denoise_test_pair(denoise_model, test_img_list, raw_dir, save_dir, opts.EV)
# denoise_test_unpair(denoise_model, unpair_data_dir)