# parameters
n = 256
m = 8
image_size = (n, n)
patch_size = (m, m)
step = 4
patches_recto = []
patches_verso = []
# Extract reference patches from the images
print('Extracting reference patches...')
for i in range(10):
image1, image2 = load_images('../images/set' + str(i + 1) + '_pic1.png',
'../images/set' + str(i + 1) + '_pic2.png',
256,
show_images=False)
patches1 = patchify(image1, patch_size, step)
patches2 = patchify(image2, patch_size, step)
initial_patch_size = patches1.shape
patches1 = patches1.reshape(-1, patch_size[0] * patch_size[1])
patches2 = patches2.reshape(-1, patch_size[0] * patch_size[1])
patches_recto.append(patches1)
patches_recto.append(patches2)
patches_recto = np.reshape(patches_recto, (-1, m * m))
patches_recto -= np.mean(patches_recto, axis=0) # remove the mean
patches_recto /= np.std(patches_recto, axis=0) # normalize each patch
t = np.linspace(1, max_it, int(max_it / 5))
d = 25
method = 'single_dictionary_learning_filtered'
#f.write(f'\nsigma = {sigma}, degrees = {d}, iterations = {max_it}\n method = {method}')
#f.write(f'\n\nLearned on 15 images from images, tested on images hard set 3')
patches_recto = []
patches_verso = []
# Extract reference patches from the first images
print('Extracting reference patches...')
for i in range(15):
image1, image2 = load_images(f'./train_images_bm3d_v/set{i+1}_pic1.png',
f'./train_images_bm3d_v/set{i+1}_pic2.png',
256,
show_images=False)
patches1 = patchify(image1, patch_size, step)
initial_patch_size = patches1.shape
patches1 = patches1.reshape(-1, patch_size[0] * patch_size[1])
patches_recto.append(patches1)
patches2 = patchify(image2, patch_size, step)
patches2 = patches2.reshape(-1, patch_size[0] * patch_size[1])
patches_recto.append(patches2)
patches_recto = np.reshape(patches_recto, (-1, m * m))
patches_recto -= np.mean(patches_recto, axis=0) # remove the mean
patches_recto /= np.std(patches_recto, axis=0) # normalize each patch
print('Learning the recto dictionary...')
This function does the whitening projection with PCA
"""
R = np.dot(S, S.T)
W = la.sqrtm(np.linalg.inv(R))
return W, np.dot(W, S)
n = 256
image_size = (n, n)
max_it = 100
d = 45
i = 4
source1, source2 = load_images(f'../images_hard/set{i+1}_pic1.png',
f'../images_hard/set{i+1}_pic2.png',
256,
show_images=False)
theta = np.radians(d)
mixing_matrix = [[np.cos(theta), -np.sin(theta)],
[np.sin(theta), np.cos(theta)]]
sources, mixtures = linear_mixture(source1,
source2,
mixing_matrix=mixing_matrix,
show_images=False)
W_whiten, mixtures = whiten_projection(mixtures)
mix1, mix2 = unflatten(mixtures, image_size)
plt.figure()
plt.suptitle(f'Mixtures, theta = {d}')
from mixing_models import load_images, linear_mixture
from functions import TV_proj, flatten, unflatten, whiten_projection, learn_dictionary, dictionary_projection
from skimage.metrics import structural_similarity as ssim
from skimage.util import random_noise
from skimage.restoration import denoise_tv_chambolle
n = 256
m = 8
image_size = (n, n)
patch_size = (m, m)
step = 4
n_coef = 2
# 2 images
source1, source2 = load_images('../../images_hard/set2_pic1.png',
'../../images_hard/set2_pic2.png',
n=256,
show_images=False)
print('SSIM of the clean image with itself =', ssim(source1, source1))
# add gaussian noise
sigma = 0.1
noisy_image1 = random_noise(source1, var=sigma**2) #var = 0.01 by default
noisy_image2 = random_noise(source2, var=sigma**2) #var = 0.01 by default
print(
'\nSSIM of source 1 with its noisy one =',
ssim(source1,
noisy_image1,
data_range=noisy_image1.max() - noisy_image1.min()))
print(
'SSIM of source 2 with its noisy one =',