-
Notifications
You must be signed in to change notification settings - Fork 0
/
denoise.py
124 lines (108 loc) · 4.32 KB
/
denoise.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
import imageio
import matplotlib.pyplot as plt
import numpy as np
import torch as th
import data
import util
colormap = plt.get_cmap('inferno')
th.set_grad_enabled(False)
color = False
images = data.Set68(color=False).data()[:15]
M, N = images.shape[2:]
# Ordering: GMM7, GMM15, GSM7, wavelet db2, wavelet db4, shearlet
rs = util.get_models(320)
kernel_size = 7
div7 = util.patch2image(
util.image2patch(th.ones((M, N)).cuda(), (kernel_size, kernel_size)),
(M, N), (kernel_size, kernel_size)
)
kernel_size = 15
div15 = util.patch2image(
util.image2patch(th.ones((M, N)).cuda(), (kernel_size, kernel_size)),
(M, N), (kernel_size, kernel_size)
)
ssim = util.SSIM().cuda()
diff_mult = 3
# Parameters for stochastic denoising
sigma_L = 0.01
L = 250
T = 3
def stochastic_image_denoiser(y, score, sigma):
'''https://arxiv.org/pdf/2101.09552.pdf'''
x = y.clone()
for i_s, sigma in enumerate(sigmas):
alpha = epsilon * sigma**2 / sigmas[-1]**2
for t in range(T):
z = th.randn_like(y)
delta = score(x, sigma) + (y - x) / (sigma_0**2 - sigma**2)
x = x + alpha * delta + np.sqrt(2 * alpha) * z
return x
def tweedie(y, score, sigma):
return y + score(y, sigma) * sigma**2
for restore_name, restore_fn in zip(
['tweedie', 'stoch'],
[tweedie, stochastic_image_denoiser],
):
for eval_method in [util.psnr, ssim]:
for i_s, sigma in enumerate([0.025, 0.05, 0.1, 0.2]):
epsilon = 5e-6
sigma_0 = sigma
gamma = (sigma_L / sigma_0)**(1 / L)
sigmas = [sigma_0 * gamma**ll for ll in range(1, L + 1)]
noisy = images + sigma * th.randn_like(images)
print(f'{sigma:.3f}', end=' & ')
print(f'{eval_method(images, noisy):.2f}', end=' & ')
for i_n, (n, gt) in enumerate(zip(noisy, images)):
diff = (th.abs(n - gt) * diff_mult).cpu().numpy()
diff_heat = colormap(diff)
imageio.imsave(
f'./out/denoising/{sigma:.3f}/{restore_name}/noisy/{i_n:03d}_d.png',
(diff_heat.clip(0, 1).squeeze() * 255.).astype(np.uint8)
)
imageio.imsave(
f'./out/denoising/{sigma:.3f}/{restore_name}/noisy/{i_n:03d}.png',
(n.cpu().numpy().clip(0, 1).squeeze() *
255.).astype(np.uint8)
)
for i_n, gt in enumerate(images):
imageio.imsave(
f'./out/denoising/{sigma:.3f}/{restore_name}/gt/{i_n:03d}.png',
(gt.cpu().numpy().clip(0, 1).squeeze() *
255.).astype(np.uint8)
)
for i_r, (R, div, name) in enumerate(
zip(
rs, [div7] + [div15] + [div7] + 3 * [th.ones_like(div7)], [
'gmm7', 'gmm15', 'gsm7', 'wavelet-db2', 'wavelet-db4',
'shearlet'
]
)
):
R.set_sigma(sigma)
def score(x, sigma):
R.set_sigma(sigma)
return -R.grad(x)[1] / div
denoised = restore_fn(noisy, score, sigma)
end = ' & ' if i_r < 5 else ' \\\\'
print(f'{eval_method(images, denoised):.2f}', end=end)
for i_n, den in enumerate(denoised):
imageio.imsave(
f'./out/denoising/{sigma:.3f}/{restore_name}/{name}/{i_n:03d}.png',
(den.cpu().numpy().clip(0, 1).squeeze() *
255.).astype(np.uint8)
)
for i_n, (den, gt) in enumerate(zip(denoised, images)):
diff = (th.abs(den - gt) * diff_mult).cpu().numpy()
diff_heat = colormap(diff)
imageio.imsave(
f'./out/denoising/{sigma:.3f}/{restore_name}/{name}/{i_n:03d}_d.png',
(diff_heat.clip(0, 1).squeeze() *
255.).astype(np.uint8)
)
print()
print('\\midrule')
print('{\\#Params} & & ', end='')
for i_r, R in enumerate(rs):
end = ' & ' if i_r < 5 else ''
print(f'\\num{{{sum(p.numel() for p in R.parameters())}}}', end=end)
print('\\\\\\bottomrule')