Create utils/color_fix.py
Browse files- utils/color_fix.py +115 -0
utils/color_fix.py
ADDED
|
@@ -0,0 +1,115 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
from PIL import Image
|
| 3 |
+
from torch import Tensor
|
| 4 |
+
from torch.nn import functional as F
|
| 5 |
+
|
| 6 |
+
from torchvision.transforms import ToTensor, ToPILImage
|
| 7 |
+
|
| 8 |
+
def adain_color_fix(target: Image, source: Image):
|
| 9 |
+
# Convert images to tensors
|
| 10 |
+
to_tensor = ToTensor()
|
| 11 |
+
target_tensor = to_tensor(target).unsqueeze(0)
|
| 12 |
+
source_tensor = to_tensor(source).unsqueeze(0)
|
| 13 |
+
|
| 14 |
+
# Apply adaptive instance normalization
|
| 15 |
+
result_tensor = adaptive_instance_normalization(target_tensor, source_tensor)
|
| 16 |
+
|
| 17 |
+
# Convert tensor back to image
|
| 18 |
+
to_image = ToPILImage()
|
| 19 |
+
result_image = to_image(result_tensor.squeeze(0).clamp_(0.0, 1.0))
|
| 20 |
+
|
| 21 |
+
return result_image
|
| 22 |
+
|
| 23 |
+
def wavelet_color_fix(target: Image, source: Image):
|
| 24 |
+
if target.size() != source.size():
|
| 25 |
+
source = source.resize((target.size()[-2], target.size()[-1]), Image.LANCZOS)
|
| 26 |
+
# Convert images to tensors
|
| 27 |
+
to_tensor = ToTensor()
|
| 28 |
+
target_tensor = to_tensor(target).unsqueeze(0)
|
| 29 |
+
source_tensor = to_tensor(source).unsqueeze(0)
|
| 30 |
+
|
| 31 |
+
# Apply wavelet reconstruction
|
| 32 |
+
result_tensor = wavelet_reconstruction(target_tensor, source_tensor)
|
| 33 |
+
|
| 34 |
+
# Convert tensor back to image
|
| 35 |
+
to_image = ToPILImage()
|
| 36 |
+
result_image = to_image(result_tensor.squeeze(0).clamp_(0.0, 1.0))
|
| 37 |
+
|
| 38 |
+
return result_image
|
| 39 |
+
|
| 40 |
+
def calc_mean_std(feat: Tensor, eps=1e-5):
|
| 41 |
+
"""Calculate mean and std for adaptive_instance_normalization.
|
| 42 |
+
Args:
|
| 43 |
+
feat (Tensor): 4D tensor.
|
| 44 |
+
eps (float): A small value added to the variance to avoid
|
| 45 |
+
divide-by-zero. Default: 1e-5.
|
| 46 |
+
"""
|
| 47 |
+
size = feat.size()
|
| 48 |
+
assert len(size) == 4, 'The input feature should be 4D tensor.'
|
| 49 |
+
b, c = size[:2]
|
| 50 |
+
feat_var = feat.view(b, c, -1).var(dim=2) + eps
|
| 51 |
+
feat_std = feat_var.sqrt().view(b, c, 1, 1)
|
| 52 |
+
feat_mean = feat.view(b, c, -1).mean(dim=2).view(b, c, 1, 1)
|
| 53 |
+
return feat_mean, feat_std
|
| 54 |
+
|
| 55 |
+
def adaptive_instance_normalization(content_feat:Tensor, style_feat:Tensor):
|
| 56 |
+
"""Adaptive instance normalization.
|
| 57 |
+
Adjust the reference features to have the similar color and illuminations
|
| 58 |
+
as those in the degradate features.
|
| 59 |
+
Args:
|
| 60 |
+
content_feat (Tensor): The reference feature.
|
| 61 |
+
style_feat (Tensor): The degradate features.
|
| 62 |
+
"""
|
| 63 |
+
size = content_feat.size()
|
| 64 |
+
style_mean, style_std = calc_mean_std(style_feat)
|
| 65 |
+
content_mean, content_std = calc_mean_std(content_feat)
|
| 66 |
+
normalized_feat = (content_feat - content_mean.expand(size)) / content_std.expand(size)
|
| 67 |
+
return normalized_feat * style_std.expand(size) + style_mean.expand(size)
|
| 68 |
+
|
| 69 |
+
def wavelet_blur(image: Tensor, radius: int):
|
| 70 |
+
"""
|
| 71 |
+
Apply wavelet blur to the input tensor.
|
| 72 |
+
"""
|
| 73 |
+
# input shape: (1, 3, H, W)
|
| 74 |
+
# convolution kernel
|
| 75 |
+
kernel_vals = [
|
| 76 |
+
[0.0625, 0.125, 0.0625],
|
| 77 |
+
[0.125, 0.25, 0.125],
|
| 78 |
+
[0.0625, 0.125, 0.0625],
|
| 79 |
+
]
|
| 80 |
+
kernel = torch.tensor(kernel_vals, dtype=image.dtype, device=image.device)
|
| 81 |
+
# add channel dimensions to the kernel to make it a 4D tensor
|
| 82 |
+
kernel = kernel[None, None]
|
| 83 |
+
# repeat the kernel across all input channels
|
| 84 |
+
kernel = kernel.repeat(3, 1, 1, 1)
|
| 85 |
+
image = F.pad(image, (radius, radius, radius, radius), mode='replicate')
|
| 86 |
+
# apply convolution
|
| 87 |
+
output = F.conv2d(image, kernel, groups=3, dilation=radius)
|
| 88 |
+
return output
|
| 89 |
+
|
| 90 |
+
def wavelet_decomposition(image: Tensor, levels=5):
|
| 91 |
+
"""
|
| 92 |
+
Apply wavelet decomposition to the input tensor.
|
| 93 |
+
This function only returns the low frequency & the high frequency.
|
| 94 |
+
"""
|
| 95 |
+
high_freq = torch.zeros_like(image)
|
| 96 |
+
for i in range(levels):
|
| 97 |
+
radius = 2 ** i
|
| 98 |
+
low_freq = wavelet_blur(image, radius)
|
| 99 |
+
high_freq += (image - low_freq)
|
| 100 |
+
image = low_freq
|
| 101 |
+
|
| 102 |
+
return high_freq, low_freq
|
| 103 |
+
|
| 104 |
+
def wavelet_reconstruction(content_feat:Tensor, style_feat:Tensor):
|
| 105 |
+
"""
|
| 106 |
+
Apply wavelet decomposition, so that the content will have the same color as the style.
|
| 107 |
+
"""
|
| 108 |
+
# calculate the wavelet decomposition of the content feature
|
| 109 |
+
content_high_freq, content_low_freq = wavelet_decomposition(content_feat)
|
| 110 |
+
del content_low_freq
|
| 111 |
+
# calculate the wavelet decomposition of the style feature
|
| 112 |
+
style_high_freq, style_low_freq = wavelet_decomposition(style_feat)
|
| 113 |
+
del style_high_freq
|
| 114 |
+
# reconstruct the content feature with the style's high frequency
|
| 115 |
+
return content_high_freq + style_low_freq
|