def SSIM(original_stacked, estimated_stacked, image_size, permutation,
X_normalized, B):
C = np.dot(estimated_stacked, X_normalized.T)
D = np.dot(C, B)
(sdr, sir, sar,
perm) = mmetrics.bss_eval_sources(np.asarray(original_stacked),
estimated_stacked)
if np.array_equal(perm, [[1], [0]]):
permutation = True
source1, source2 = unflatten(original_stacked, image_size)
estimated_stacked_normalized = normalizing_for_ssim(
estimated_stacked, D, permutation)
estimated1, estimated2 = unflatten(estimated_stacked_normalized,
image_size)
if permutation:
ssim1 = ssim(source1,
estimated2,
data_range=estimated2.max() - estimated2.min())
ssim2 = ssim(source2,
estimated1,
data_range=estimated1.max() - estimated1.min())
else:
ssim1 = ssim(source1,
estimated1,
data_range=estimated1.max() - estimated1.min())
ssim2 = ssim(source2,
estimated2,
data_range=estimated2.max() - estimated2.min())
return ssim1, ssim2
def sigmoid_proj(S):
S1 = S[0,:]
#S2 = S[1,:]
S[0,:] = 1 / (1 + np.exp(-S1))
return S
# Here begins the algorithm
# whitening processing. It's important
R = np.dot(X, X.T)
W = la.sqrtm(np.linalg.inv(R))
X = np.dot(W, X)
(sdr_ref, sir_ref, sar, perm) = mmetrics.bss_eval_sources(np.asarray(source), np.asarray(X))
# mix = [[0.6992, 0.7275], [0.4784, 0.5548]] #or use the matrix from the paper
print("Reference SDR is: ", sdr_ref)
print("Reference SIR is: ", sir_ref)
lambda_max = 0.02
#lambda_final = lambda_max
lambda_final = 0.00002
max_it = 50
lambda_v = np.logspace(np.log10(lambda_max),np.log10(lambda_final),max_it)
#Se = np.random.randn(2, n*n)
Se = X
cost_it = np.zeros((1,max_it))
SDR_it = np.zeros((2, max_it))
SIR_it = np.zeros((2, max_it))
source2 = np.matrix(img2_gray)
source2 = source2.flatten('F') #column wise
source1 = source1 - np.mean(source1)
source2 = source2 - np.mean(source2)
#source1 = source1/np.linalg.norm(source1)
#source2 = source2/np.linalg.norm(source2)
source = np.stack((source1, source2))
#mixing_matrix = np.random.rand(2,2)
mixing_matrix = np.array([[1, 0.5], [0.5, 1]])
X = np.matmul(source.T, mixing_matrix)
(sdr_ref, sir, sar, perm) = mmetrics.bss_eval_sources(np.asarray(source),
np.asarray(X.T))
print("Reference SDR is: ")
print(sdr_ref)
# DCT of the observation
X1 = np.reshape(X[:, 0], (n, n))
X2 = np.reshape(X[:, 1], (n, n))
X_dct1 = dct(X1)
X_dct2 = dct(X2)
X_dct = np.stack((X_dct1.flatten(), X_dct2.flatten()))
# try the sparsity separation
lambda_max = 0.02
print(i + 1)
# load and mix images
source1, source2 = load_images(
'../images_hard/set' + str(i + 1) + '_pic1.png',
'../images_hard/set' + str(i + 1) + '_pic2.png',
256,
show_images=False)
sources, mixed_sources = linear_mixture(
source1, source2, mixing_matrix=mixing_matrix,
show_images=False) #mixing_matrix = mixing_matrix,
# Here begins the algorithm
# whitening processing. It's important
X = whiten_projection(mixed_sources)
(sdr_ref, sir_ref, sar,
perm) = mmetrics.bss_eval_sources(np.asarray(sources), np.asarray(X))
print('The mean value of the reference SDR is: ', np.mean(sdr_ref))
reference_sdr.append(np.mean(sdr_ref))
Se = np.copy(X)
Se_old = np.copy(Se)
for it in np.arange(max_it):
# 1. denoising
Se = Dic_proj_double(Se_old, num_coeff, sigma)
# 2. get demixing matrix
WW = get_demix(X, Se)
# 3. whiten the demix matrix
WW = whiten_projection(WW)
# 4. get the new source
Se = np.dot(WW, X)
X3, W3 = whiten_projection(mixture_channel_3)
B3 = np.dot(W3, mixing_matrix)"""
stacked_channel_h_mixtures = flatten(mixture1_hvs[:,:,0], mixture2_hvs[:,:,0])
stacked_channel_s_mixtures = flatten(mixture1_hvs[:,:,1], mixture2_hvs[:,:,1])
stacked_channel_v_mixtures = flatten(mixture1_hvs[:,:,2], mixture2_hvs[:,:,2])
stacked_channel_h_source = flatten(source1_hsv[:,:,0], source2_hsv[:,:,0])
stacked_channel_s_source = flatten(source1_hsv[:,:,1], source2_hsv[:,:,1])
stacked_channel_v_source = flatten(source1_hsv[:,:,2], source2_hsv[:,:,2])
mixture_channel_list_hsv = [stacked_channel_h_mixtures, stacked_channel_s_mixtures, stacked_channel_v_mixtures]
source_channel_list_hsv = [stacked_channel_h_source, stacked_channel_s_source, stacked_channel_v_source]
for i in np.arange(3):
X_i, W_i = whiten_projection(mixture_channel_list_hsv[i])
B_i = np.dot(W_i, mixing_matrix)
(sdr_ref, sir_ref, sar, perm) = mmetrics.bss_eval_sources(source_channel_list_hsv[i], X_i)
print('The mean value of the reference SDR is: ', np.mean(sdr_ref), perm)
if np.array_equal(perm, [[1],[0]]):
permutation = True
X_normalized = normalizing_for_ssim(X_i, B_i, permutation) #used for evaluation of SSIM
Se = np.copy(X_i)
Se_old = np.copy(Se)
ssim_estimated_source1 = []
ssim_estimated_source2 = []
estimated_matrix = []
mean_ssim = []
for it in np.arange(max_it):
ref_patches1 = patchify(np.reshape(source1, image_size).T, patch_size, step)
initial_size1 = ref_patches1.shape
ref_patches1 = ref_patches1.reshape((-1, m, m))
ref_patches2 = patchify(np.reshape(source2, image_size).T, patch_size, step)
initial_size2 = ref_patches2.shape
ref_patches2 = ref_patches2.reshape((-1, m, m))
source = np.stack((source1, source2))
mixing_matrix = [[1, 0.5],[0.5, 1]]
# X = source * mixing_matrix - The mixed images
X = np.matmul(mixing_matrix, source)
(sdr_ref, sir, sar, perm) = mmetrics.bss_eval_sources(np.asarray(source), np.asarray(X))
print("***")
print(sdr_ref)
print("***")
# reconstructing the mixed images
X1 = X[0,:]
X1 = np.reshape(X1, (n,n))
X2 = X[1,:]
X2 = np.reshape(X2, (n,n))
#extracting patches from the mixed images
mix_patches1 = patchify(X1.T, patch_size, step)
mix_patches1 = mix_patches1.reshape((-1, m, m))
ref_p = np.stack((refp1.flatten(), refp2.flatten()))
ica = FastICA(n_components=2, fun='cube') #, whiten=False)#, fun='cube')
source_estimated = ica.fit_transform(mix_p.T)
mixing_estimated = ica.mixing_
# Prepare for the permutation here
#source_estimated = oracle_perm(ref_p, source_estimated.T)
source_estimated = source_estimated.T
# remove the original mean
ref_p[0, :] = ref_p[0, :] - np.mean(ref_p[0, :])
ref_p[1, :] = ref_p[1, :] - np.mean(ref_p[1, :])
(sdr_ref, sir, sar, perm) = mmetrics.bss_eval_sources(ref_p, mix_p)
(sdr, sir, sar, perm) = mmetrics.bss_eval_sources(ref_p, source_estimated)
print(sdr_ref)
print(sdr)
print(perm)
# permute here
source_estimated = source_estimated[perm, :]
mixing_estimated = mixing_estimated[:, perm]
# change the sign here
if mixing_estimated[0, 0] < 0:
source_estimated[0, :] = -source_estimated[0, :]
if mixing_estimated[1, 1] < 0:
source_estimated[1, :] = -source_estimated[1, :]
source_estimated[0, :] = n1 * source_estimated[0, :] + m1
for i in np.arange(3):
print("Channel ", i + 1)
source_channel_i, mixture_channel_i = linear_mixture(
image1[:, :, i], image2[:, :, i], mixing_matrix, False)
if i == 0:
mixture_channel_1 = mixture_channel_i
if i == 1:
mixture_channel_2 = mixture_channel_i
if i == 2:
mixture_channel_3 = mixture_channel_i
X_i, W_i = whiten_projection(mixture_channel_i)
B_i = np.dot(W_i, mixing_matrix)
(sdr_ref, sir_ref, sar,
perm) = mmetrics.bss_eval_sources(source_channel_i, X_i)
print('The mean value of the reference SDR is: ', np.mean(sdr_ref), perm)
if np.array_equal(perm, [[1], [0]]):
permutation = True
X_normalized = normalizing_for_ssim(
X_i, B_i, permutation) #used for evaluation of SSIM
Se = np.copy(X_i)
Se_old = np.copy(Se)
ssim_estimated_source1 = []
ssim_estimated_source2 = []
estimated_matrix = []
mean_ssim = []
for it in np.arange(max_it):
#X[:,0] = X[:,0] - mx1
#X[:,1] = X[:,1] - mx2
ica = FastICA(n_components=2, fun='cube')
source_estimated = ica.fit_transform(X)
mixing_estimated = ica.mixing_
ms = np.dot(np.linalg.inv(mixing_estimated), mx)
#source_estimated[:,0] = source_estimated[:,0]+ms[0]
#source_estimated[:,1] = source_estimated[:,1]+ms[1]
#print("mixing matrix estimated = ", mixing_estimated)
(sdr, sir, sar,
perm) = mmetrics.bss_eval_sources(np.asarray(source),
np.asarray(source_estimated.T))
(sdr_ref, sir, sar,
perm) = mmetrics.bss_eval_sources(np.asarray(source), np.asarray(X.T))
SDR_improv[0, it] = np.mean(sdr) - np.mean(sdr_ref)
rdr_it[0, it] = rdr_this
plt.figure
plt.subplot(211)
plt.plot(size_v, SDR_improv[0, :], '-*')
plt.xlabel('Image size')
plt.ylabel('SDR improvement (dB)')
plt.grid()
plt.show
plt.subplot(212)
all_sdr_improved = []
for i in range(15):
print("Set = ", i + 1)
source1, source2 = load_images('./images/set' + str(i + 1) + '_pic1.png',
'./images/set' + str(i + 1) + '_pic2.png',
n,
show_images=False)
S, X = linear_mixture(source1,
source2,
mixing_matrix=mixing_matrix,
show_images=False)
#Reference SDR
(sdr_ref, sir_ref, sar,
perm) = mmetrics.bss_eval_sources(np.asarray(S), np.asarray(X))
print('The mean value of the reference SDR is: ', np.mean(sdr_ref))
#RDC of sources
rdc_i = rdc(S[0, :], S[1, :])
all_rdc.append(rdc_i)
print("rdc of sources = ", rdc_i)
#Kurtosis of sources
kurtosis_i = mean_kurtosis(S)
all_kurtosis_sources.append(kurtosis_i)
print("mean kurtosis of sources = ", kurtosis_i)
print("kurtosis of sources = ", kurtosis(S[0, :]), kurtosis(S[1, :]))
#Skewness of sources
skewness_i = mean_skewness(S)
print("rdc = ", rdc(source1.T, source2.T))
source = np.stack((source1, source2))
print('Covariance matrix is: ')
print(np.matmul(source, source.T))
# randomly generated mixing matrix
np.random.seed(0)
#mixing_matrix = np.random.rand(2,2)
mixing_matrix = np.array([[1, 0.5], [0.5, 1]])
# X = source * mixing_matrix - The mixed images
X = np.matmul(source.T, mixing_matrix)
(sdr_ref, sir, sar, perm) = mmetrics.bss_eval_sources(np.asarray(source),
np.asarray(X.T))
# mix = [[0.6992, 0.7275], [0.4784, 0.5548]] #or use the matrix from the paper
# X = np.matmul(source.T, mix)
X1 = X[:, 0]
mx1 = np.mean(X1)
X1 = np.reshape(X1, (n, n))
#print(X1.min(), X1.max(), X1.mean())
X2 = X[:, 1]
mx2 = np.mean(X2)
X2 = np.reshape(X2, (n, n))
mx = np.array([mx1, mx2])
#print(X2.min(), X2.max(), X2.mean())