#it = 500
start = time.time()
recon, firstRecon = iterative.sirt_frt_complex(
it, N, drtSpace, mValues, t, 2, 1, oversampleFilter=oversampleFilter)
end = time.time()
elapsed = end - start
print("SIRT Reconstruction took " + str(elapsed) + " secs or " +
str(elapsed / 60) + " mins")
testdrtSpace = finite.frt_complex(image, N, mValues=mValues)
recon = np.abs(recon)
diff = image - recon
sample = dftSpace * fractalMine
mse = imageio.immse(image, np.abs(recon))
ssim = imageio.imssim(image.astype(float), recon.astype(float))
psnr = imageio.impsnr(image, np.abs(recon))
print("RMSE:", math.sqrt(mse))
print("SSIM:", ssim)
print("PSNR:", psnr)
plt.figure(1)
plt.imshow(np.abs(drtSpace))
plt.figure(2)
plt.imshow(np.abs(recon))
plt.figure(3)
plt.imshow(np.abs(image))
plt.figure(4)
plt.imshow(np.abs(firstRecon))
plt.figure(5)
for i, g_j in enumerate(drtSpace):
recon[i, :, :], firstRecon[i, :, :] = iterative.abmlem_frt_complex(
it, N, g_j, mValues, BU, BL, 3, 10, oversampleFilter)
print("Image number: ", i)
if i >= 0:
break
end = time.time()
elapsed = end - start
print("ABfMLEM Reconstruction took " + str(elapsed) + " secs or " +
str(elapsed / 60) + " mins")
recon = np.abs(recon)
images = np.abs(images)
diff = np.abs(image - recon)
mse = imageio.immse(np.abs(images[0, :, :]), np.abs(recon[0, :, :]))
ssim = imageio.imssim(
np.abs(images[0, :, :]).astype(float),
np.abs(recon[0, :, :]).astype(float))
psnr = imageio.impsnr(np.abs(images[0, :, :]), np.abs(recon[0, :, :]))
print("RMSE:", math.sqrt(mse))
print("SSIM:", ssim)
print("PSNR:", psnr)
plt.figure(2)
plt.imshow(np.abs(recon[0, :, :]))
plt.figure(3)
plt.imshow(np.abs(images[0, :, :]))
plt.figure(4)
plt.imshow(np.abs(firstRecon[0, :, :]))
plt.figure(9)
sirtPSNRS = sirtReconDict['psnrs']
sirtSSIMS = sirtReconDict['ssims']
print("SIRT Recon Size:", sirtRecon.shape)
print("SIRT Recon Min/Max:", sirtRecon.min(), sirtRecon.max())
#radialReconDict = np.load('result_radial.npz')
radialReconDict = np.load('result_phantom_radial.npz')
#radialReconDict = np.load('result_camera_radial.npz')
radialRecon = radialReconDict['recon']
radialPSNRS = radialReconDict['psnrs'] #1 iteration
radialSSIMS = radialReconDict['ssims']
print("Radial Recon Size:", radialRecon.shape)
print("Radial Recon Min/Max:", radialRecon.min(), radialRecon.max())
#compute metrics
mse = imageio.immse(imageio.immask(image, mask, N, N),
imageio.immask(mlemRecon, mask, N, N))
ssim = imageio.imssim(
imageio.immask(image, mask, N, N).astype(float),
imageio.immask(mlemRecon, mask, N, N).astype(float))
psnr = imageio.impsnr(imageio.immask(image, mask, N, N),
imageio.immask(mlemRecon, mask, N, N))
print("MLEM Results ------")
print("RMSE:", math.sqrt(mse))
print("SSIM:", ssim)
print("PSNR:", psnr)
mse = imageio.immse(imageio.immask(image, mask, N, N),
imageio.immask(sirtRecon, mask, N, N))
ssim = imageio.imssim(
imageio.immask(image, mask, N, N).astype(float),
imageio.immask(sirtRecon, mask, N, N).astype(float))
maxValue = np.max(image)
print("Max Pixel Value:", maxValue)
#csReconMat = loadmat('result_lego_2.mat')
csReconMat = loadmat('result_lego_4.mat')
#csReconMat = loadmat('result_lego_8.mat')
csRecon = np.abs(csReconMat['im_res'])*maxValue
diff2 = image2-csRecon
data2 = np.load('result_radial.npz')
radialRecon = data2['recon']
diff3 = image2-radialRecon
#-------------------------------
#compute metrics
mse = imageio.immse(image, np.abs(recon))
ssim = imageio.imssim(image.astype(float), np.abs(recon).astype(float))
psnr = imageio.impsnr(image, np.abs(recon))
print("ChaoS RMSE:", math.sqrt(mse))
print("ChaoS SSIM:", ssim)
print("ChaoS PSNR:", psnr)
mse = imageio.immse(image2, np.abs(csRecon))
ssim = imageio.imssim(image2.astype(float), np.abs(csRecon).astype(float))
psnr = imageio.impsnr(image2, np.abs(csRecon))
print("CS RMSE:", math.sqrt(mse))
print("CS SSIM:", ssim)
print("CS PSNR:", psnr)
mse = imageio.immse(image2, np.abs(radialRecon))
ssim = imageio.imssim(image2.astype(float), np.abs(radialRecon).astype(float))
psnr = imageio.impsnr(image2, np.abs(radialRecon))
print("Radial RMSE:", math.sqrt(mse))
#-------------------------------
#reconstruct test using MLEM
start = time.time() #time generation
recon = abmlem_expand_complex(iterations, p, drtSpace, subsetsMValues,
finite.frt_complex, finite.ifrt_complex)
recon = fftpack.ifftshift(recon)
recon = np.abs(recon)
print("Done")
end = time.time()
elapsed = end - start
print("OSEM Reconstruction took " + str(elapsed) + " secs or " +
str(elapsed / 60) + " mins in total")
mse = imageio.immse(image, recon)
ssim = imageio.imssim(image.astype(float), recon.astype(float))
psnr = imageio.impsnr(image, recon)
print("RMSE:", math.sqrt(mse))
print("SSIM:", ssim)
print("PSNR:", psnr)
diff = image - recon
#save data
np.savez('result_abmlem.npz',
K=K,
s=s,
p=p,
iterations=iterations,
recon=recon,
diff=diff,
def ossirt_expand_complex(iterations,
t,
p,
g_j,
os_mValues,
projector,
backprojector,
image,
mask,
epsilon=1e3,
dtype=np.int32):
'''
# Gary's implementation
# From Lalush and Wernick;
# f^\hat <- (f^\hat / |\sum h|) * \sum h * (g_j / g) ... (*)
# where g = \sum (h f^\hat) ... (**)
#
# self.f is the current estimate f^\hat
# The following g from (**) is equivalent to g = \sum (h f^\hat)
'''
norm = False
center = False
fdtype = floatType
f = np.zeros((p, p), fdtype)
mses = []
psnrs = []
ssims = []
for i in range(0, iterations):
print("Iteration:", i)
for j, mValues in enumerate(os_mValues):
# print("Subset:", j)
muFinite = len(mValues)
g = projector(f, p, fdtype, mValues)
# form parenthesised term (g_j / g) from (*)
r = np.zeros_like(g)
for m in mValues:
# r[m,:] = g_j[m,:] - g[m,:]
for y in range(p):
r[m, y] = (g[m, y] - g_j[m, y]) / (muFinite * muFinite)
# backproject to form \sum h * (g_j / g)
g_r = backprojector(r, p, norm, center, 1, 0, mValues) / muFinite
# Renormalise backprojected term / \sum h)
# Normalise the individual pixels in the reconstruction
f -= t * g_r
if smoothReconMode > 0 and i % smoothIncrement == 0 and i > 0: #smooth to stem growth of noise
if smoothReconMode == 1:
print("Smooth TV")
f = denoise_tv_chambolle(f, 0.02, multichannel=False)
elif smoothReconMode == 2:
h = parameters[4] #6, phantom; 4, camera
if i > smoothMaxIteration:
h /= 2.0
if i > smoothMaxIteration2:
h /= 4.0
print("Smooth NL h:", h)
fReal = denoise_nl_means(np.real(f),
patch_size=3,
patch_distance=7,
h=h,
multichannel=False,
fast_mode=True).astype(fdtype)
fImag = denoise_nl_means(np.imag(f),
patch_size=3,
patch_distance=7,
h=h,
multichannel=False,
fast_mode=True).astype(fdtype)
f = fReal + 1j * fImag
elif smoothReconMode == 3:
print("Smooth Median")
f = ndimage.median_filter(f, 3)
if i % plotIncrement == 0:
img = imageio.immask(image, mask, N, N)
recon = imageio.immask(f, mask, N, N)
recon = np.abs(recon)
mse = imageio.immse(img, recon)
psnr = imageio.impsnr(img, recon)
ssim = imageio.imssim(img.astype(float), recon.astype(float))
print("RMSE:", math.sqrt(mse), "PSNR:", psnr, "SSIM:", ssim)
mses.append(mse)
psnrs.append(psnr)
ssims.append(ssim)
return f, mses, psnrs, ssims
fftLenaShifted = fftpack.fftshift(fftLena)
#power spectrum
powSpectLena = np.abs(fftLenaShifted)
#add noise to kSpace
noise = finite.noise(fftLenaShifted, SNR)
if addNoise:
fftLenaShifted += noise
#Recover full image with noise
print("Actual noisy image")
reconLena = fftpack.ifft2(fftLenaShifted) #the '2' is important
reconLena = np.abs(reconLena)
reconNoise = lena - reconLena
mse = imageio.immse(lena, np.abs(reconLena))
ssim = imageio.imssim(lena.astype(float), np.abs(reconLena).astype(float))
psnr = imageio.impsnr(lena, np.abs(reconLena))
print("Acutal RMSE:", math.sqrt(mse))
print("Acutal SSIM:", ssim)
print("Acutal PSNR:", psnr)
#compute lines
centered = True
subsetsLines = []
subsetsMValues = []
mu = 0
for angles in subsetsAngles:
lines = []
mValues = []
for angle in angles: