def inspect_pupil(cfg):
from mpl_toolkits.axes_grid1 import make_axes_locatable
fig, axes = plt.subplots(1, 2)
center = 15
image_size = fpmm.get_image_size(cfg)
pupil_radius = fpmm.get_pupil_radius(cfg)
aberrations = [1E-6, 0]
ps = fpmm.get_pixel_size(cfg)
wlen = fpmm.get_wavelength(cfg)
pupil = fpmm.aberrated_pupil(image_size=cfg.patch_size,
pupil_radius=pupil_radius,
aberrations=aberrations,
pixel_size=ps,
wavelength=wlen)
CTF = fpmm.generate_CTF(fx=0,
fy=0,
image_size=cfg.patch_size,
pupil_radius=pupil_radius)
def put_colorbar(ax, im):
divider = make_axes_locatable(ax)
cax = divider.append_axes('right', size='5%', pad=0.05)
fig.colorbar(im, cax=cax, orientation='vertical')
im = axes[0].imshow(np.abs(pupil * CTF), cmap='Blues')
put_colorbar(axes[0], im)
im = axes[1].imshow(np.imag(pupil * CTF), cmap='RdBu')
put_colorbar(axes[1], im)
plt.show()
return ct.set_iterator(cfg)
def fpm_reconstruct_classic(samples=None, it=None, cfg=None, debug=False):
""" FPM reconstructon using the alternating projections algorithm. Here
the complete samples and (optional) background images are loaded and Then
cropped according to the patch size set in the configuration tuple (cfg).
Args:
-----
samples: the acquired samples as a dictionary with angles as keys.
backgrounds: the acquired background as a dictionary with angles as
keys. They must be acquired right after or before taking
the samples.
it: iterator with additional sampling information for each sample.
init_point: [xoff, yoff] center of the patch to be reconstructed.
cfg: configuration (named tuple)
debug: set it to 'True' if you want to see the reconstruction proccess
(it slows down the reconstruction).
Returns:
--------
(ndarray) The reconstructed modulus and phase of the sampled image.
"""
# Getting the maximum angle by the given configuration
# Step 1: initial estimation
# Getting the maximum angle by the given configuration
# Step 1: initial estimation
lrsize = fpmm.get_image_size(cfg)
pupil_radius = fpmm.get_pupil_radius(cfg)
ps = fpmm.get_pixel_size(cfg)
wlen = fpmm.get_wavelength(cfg)
hrshape = fpmm.get_reconstructed_shape(cfg)
# k relative to absolute k factor
npx = fpmm.get_image_size(cfg)
ps_req = fpmm.get_pixel_size_required(cfg)
wlen = fpmm.get_wavelength(cfg)
led_gap = float(cfg.led_gap)
height = float(cfg.sample_height)
objectRecover_mag = (samples[(0, -1)]) / np.amax(samples[(0, -1)])
objectRecover_phase = (samples[(0, -1)] - samples[(-1, 0)]) / (samples[
(0, -1)] + samples[(-1, 0)])
objectRecover_phase = np.pi * (
objectRecover_phase) / np.amax(objectRecover_phase)
objectRecover = objectRecover_mag * np.exp(1j * objectRecover_phase)
objectRecover = np.ones(hrshape)
# plt.imshow(objectRecover_phase)
# plt.show()
center = fpmm.image_center(hrshape)
CTF = fpmm.generate_CTF(image_size=[lrsize, lrsize],
pupil_radius=pupil_radius) * .1
# pupil = fpmm.aberrated_pupil(image_size=[lrsize, lrsize], pupil_radius=pupil_radius,
# aberrations=[0,], pixel_size=ps, wavelength=wlen)
objectRecoverFT = fftshift(fft2(objectRecover)) # shifted transform
# Steps 2-5
factor = (lrsize / hrshape[0])**2
# For the convergence index
# Steps 2-5
for iteration in range(10):
iterator = ct.set_iterator(cfg)
# print('Iteration n. %d' % iteration)
for it in iterator:
# indexes, kx_rel, ky_rel = ct.n_to_krels(it=it, cfg=cfg, xoff=0, yoff=0)
indexes = it['indexes']
lr_sample = np.copy(samples[it['indexes']])
led_number = np.array([it['ny'], it['nx']])
# Convert indexes to pupil center coordinates kx, ky
mc = np.array(cfg.mat_center) # Matrix center
k_rel = np.sin(np.arctan((led_number - mc) * led_gap / height))
k_abs = k_rel * ps_req * npx / wlen
pup_center = k_abs + center
# Pupil borders
lp = np.round(pup_center - (lrsize + 1) / 2).astype(
int) # Extreme left low point
hp = lp + lrsize # Extreme right high point
pupil_square = [slice(lp[0], hp[0]), slice(lp[1], hp[1])]
lowResFT = factor * objectRecoverFT[pupil_square] * CTF
im_lowRes = ifft2(ifftshift(lowResFT))
im_lowRes = (1 / factor) * lr_sample * np.exp(
1j * np.angle(im_lowRes))
lowResFT = fftshift(fft2(im_lowRes)) * CTF
objectRecoverFT[pupil_square] = (
1 - CTF) * objectRecoverFT[pupil_square] + lowResFT
im_out = ifft2(fftshift(objectRecoverFT))
return im_out
def fpm_reconstruct_wrappable(samples=None, it=None, cfg=None, debug=False):
""" FPM reconstructon using the alternating projections algorithm. Here
the complete samples and (optional) background images are loaded and Then
cropped according to the patch size set in the configuration tuple (cfg).
Args:
-----
samples: the acquired samples as a dictionary with angles as keys.
backgrounds: the acquired background as a dictionary with angles as
keys. They must be acquired right after or before taking
the samples.
it: iterator with additional sampling information for each sample.
init_point: [xoff, yoff] center of the patch to be reconstructed.
cfg: configuration (named tuple)
debug: set it to 'True' if you want to see the reconstruction proccess
(it slows down the reconstruction).
Returns:
--------
(ndarray) The reconstructed modulus and phase of the sampled image.
"""
# Getting the maximum angle by the given configuration
# Step 1: initial estimation
lrsize = fpmm.get_image_size(cfg)
pupil_radius = fpmm.get_pupil_radius(cfg)
ps = fpmm.get_pixel_size(cfg)
wlen = fpmm.get_wavelength(cfg)
hrshape = fpmm.get_reconstructed_shape(cfg)
kdsc = fpmm.get_k_discrete(cfg)
objectRecover = np.ones(hrshape)
xc, yc = fpmm.image_center(hrshape)
CTF = fpmm.generate_CTF(0, 0, [lrsize, lrsize], pupil_radius)
objectRecoverFT = fftshift(fft2(objectRecover)) # shifted transform
# Steps 2-5
factor = (lrsize / hrshape[0])**2
# For the convergence index
N = len(samples)
for iteration in range(1):
iterator = ct.set_iterator(cfg)
# Patching for testing
for it in iterator:
# Coordinates manipulation
acqpars = it['acqpars']
indexes, kx_rel, ky_rel = ct.n_to_krels(it=it,
cfg=cfg,
xoff=0,
yoff=0)
lr_sample = np.copy(samples[it['indexes']])
[kx, ky] = kdsc * kx_rel, kdsc * ky_rel
# Iteration step
kyl = int(np.round(yc + ky - (lrsize + 1) / 2))
kyh = kyl + lrsize
kxl = int(np.round(xc + kx - (lrsize + 1) / 2))
kxh = kxl + lrsize
lowResFT = factor * objectRecoverFT[kyl:kyh, kxl:kxh] * CTF
# Step 2: lr of the estimated image using the known pupil
im_lowRes = ifft2(
ifftshift(lowResFT)) # space pupil * fourier image
im_lowRes = 1 / factor * lr_sample * np.exp(
1j * np.angle(im_lowRes))
## pupil correction update
lowResFT2 = fftshift(fft2(im_lowRes)) * CTF * 1. / pupil
ORFT = objectRecoverFT[kyl:kyh, kxl:kxh].ravel()
objectRecoverFT[kyl:kyh,
kxl:kxh] += (lowResFT2 - lowResFT) * np.conjugate(
pupil) / np.max(np.abs(pupil.ravel())**2)
pupil += (lowResFT2 - lowResFT) * np.conjugate(
objectRecoverFT[kyl:kyh, kxl:kxh]) / np.max(np.abs(ORFT)**2)
####################################################################
im_out = ifft2(ifftshift(objectRecoverFT))
return im_out
def fpm_reconstruct_epry(samples=None, it=None, cfg=None, debug=False):
""" FPM reconstructon using the alternating projections algorithm. Here
the complete samples and (optional) background images are loaded and Then
cropped according to the patch size set in the configuration tuple (cfg).
Args:
-----
samples: the acquired samples as a dictionary with angles as keys.
backgrounds: the acquired background as a dictionary with angles as
keys. They must be acquired right after or before taking
the samples.
it: iterator with additional sampling information for each sample.
init_point: [xoff, yoff] center of the patch to be reconstructed.
cfg: configuration (named tuple)
debug: set it to 'True' if you want to see the reconstruction proccess
(it slows down the reconstruction).
Returns:
--------
(ndarray) The reconstructed modulus and phase of the sampled image.
"""
# Getting the maximum angle by the given configuration
# Step 1: initial estimation
lrsize = fpmm.get_image_size(cfg)
pupil_radius = fpmm.get_pupil_radius(cfg)
ps = fpmm.get_pixel_size(cfg)
wlen = fpmm.get_wavelength(cfg)
hrshape = fpmm.get_reconstructed_shape(cfg)
kdsc = fpmm.get_k_discrete(cfg)
objectRecover = np.ones(hrshape)
xc, yc = fpmm.image_center(hrshape)
CTF = fpmm.generate_CTF(0, 0, [lrsize, lrsize], 1. * pupil_radius)
# print(lrsize,hrshape)
# plt.imshow(CTF)
# plt.show()
# pupil = np.ones_like(CTF, dtype='complex')
pupil = 1
# pupil = fpmm.aberrated_pupil(image_size=[lrsize, lrsize], pupil_radius=pupil_radius, aberrations=[-.1E-6,], pixel_size=ps, wavelength=wlen)*CTF
# pupil = (rand(lrsize, lrsize)*(1+0*1j)+1)*0.5
# gfk = exp(1j*rand(hrshape[0], hrshape[1])*pi)*fpmm.generate_CTF(0, 0, [hrshape[0], hrshape[1]], pupil_radius*.1)
# gfk = np.zeros(hrshape, dtype='complex')
gfk = fftshift(fft2(objectRecover))
# gfk += rand(hrshape[0], hrshape[1])+1
# gfk[yc-lrsize//2:yc+lrsize//2, xc-lrsize//2:xc+lrsize//2] = 1+rand(lrsize, lrsize)
# gfk[yc, xc] = 1
# Steps 2-5
factor = (float(lrsize) / hrshape[0])**2
# For convergence index
e_gk = 0
gk_prev = np.zeros_like(gfk)
gk = np.ones_like(gfk) * (1 + 1j)
gpk_prev = np.zeros_like(gfk)
gpk = np.ones_like(gfk)
ek = 1
gm = 0
a = 1
b = 1
sum_lr = 0
for it in ct.set_iterator(cfg):
indexes, kx_rel, ky_rel = ct.n_to_krels(it=it, cfg=cfg)
sum_lr += np.sum(samples[it['indexes']])
for iteration in range(15):
iterator = ct.set_iterator(cfg)
print('Iteration n. %d' % iteration)
# Patching for testing
e_gk = np.sum((abs(gk) - abs(gk_prev))**2) / (np.sum(abs(gk)**2))
e_gpk = np.sum(
(angle(gk) - angle(gk_prev))**2) / (np.sum(angle(gk)**2))
gk_prev = gk
print('Iteration %d | gk error %.2e | gpk error %.4f | ek %.2e |' %
(iteration, e_gk, e_gpk, ek))
# if (ek < 0.00001):
# break
ek = 0
gm = gm * e_gk
for it in iterator:
acqpars = it['acqpars']
indexes, kx_rel, ky_rel = ct.n_to_krels(it=it,
cfg=cfg,
xoff=0.,
yoff=0.)
lr_sample = np.copy(
np.sqrt(samples[it['indexes']][:lrsize, :lrsize]))
# From generate_il
# Calculating coordinates
[kx, ky] = kdsc * kx_rel, kdsc * ky_rel # k_rel is sin(phi)
# print(kdsc)
kyl = int(np.round(yc + ky - (lrsize + 1) / 2))
kyh = kyl + lrsize
kxl = int(np.round(xc + kx - (lrsize + 1) / 2))
kxh = kxl + lrsize
delta_gfk1 = gfk[kyl:kyh, kxl:kxh] * pupil * CTF
# print(kyl, kyh, kxl, kxh, yc+kx, yc+ky)
# Step 2: lr of the estimated image using the known pupil
delta_gk = ifft2(ifftshift(delta_gfk1))
gk_prime = 1 / factor * lr_sample * delta_gk / abs(delta_gk)
delta_gfk2 = fftshift(fft2(gk_prime))
# update of the pupil and the fourier transform of the sample.
delta_phi = delta_gfk2 - delta_gfk1
sampled_gfk = gfk[kyl:kyh, kxl:kxh]
sampled_gfk += a * delta_phi * conj(CTF * pupil) / amax(
abs(CTF * pupil))**2
if iteration > -1:
pupil += b * delta_phi * conj(sampled_gfk) / amax(
abs(sampled_gfk))**2
gfk[kyl:kyh, kxl:kxh] = sampled_gfk
# test_img = delta_phi
# plt.plot(abs(delta_phi[25,:]))
# plt.show()
ek += np.sum(abs(delta_gfk2 - delta_gfk1)**2) / sum_lr**2
gk = ifft2(fftshift(gfk))
# plt.imshow(np.log(abs(gfk)))
# plt.show()
return gk, pupil
def fpm_reconstruct(samples=None, it=None, cfg=None, debug=False):
""" FPM reconstructon using the alternating projections algorithm. Here
the complete samples and (optional) background images are loaded and Then
cropped according to the patch size set in the configuration tuple (cfg).
Args:
-----
samples: the acquired samples as a dictionary with angles as keys.
backgrounds: the acquired background as a dictionary with angles as
keys. They must be acquired right after or before taking
the samples.
it: iterator with additional sampling information for each sample.
init_point: [xoff, yoff] center of the patch to be reconstructed.
cfg: configuration (named tuple)
debug: set it to 'True' if you want to see the reconstruction proccess
(it slows down the reconstruction).
Returns:
--------
(ndarray) The reconstructed modulus and phase of the sampled image.
"""
# Getting the maximum angle by the given configuration
# Step 1: initial estimation
# objectRecover = initialize(hrshape, cfg, 'zero')
lrsize = fpmm.get_image_size(cfg)
pupil_radius = fpmm.get_pupil_radius(cfg)
ps = fpmm.get_pixel_size(cfg)
wlen = fpmm.get_wavelength(cfg)
hrshape = fpmm.get_reconstructed_shape(cfg)
kdsc = fpmm.get_k_discrete(cfg)
# Initialization
xc, yc = fpmm.image_center(hrshape)
CTF = fpmm.generate_CTF(0, 0, [lrsize, lrsize], pupil_radius)
# gk_prime = np.ones(hrshape)*exp(1j*rand(hrshape)*pi)
# gfk = np.ones(hrshape)*exp(1j*rand(hrshape[0], hrshape[1])*pi)
objectRecover = np.ones(hrshape)
gfk = fftshift(fft2(objectRecover)) # shifted transform
factor = (lrsize / hrshape[0])**2
# For the convergence index
N = len(samples)
beta = 0.
gm = 0
# lrarray = np.zeros((lrsize, lrsize, N))
e_gk = 0
gk_prev = np.zeros_like(gfk)
gk = np.ones_like(gfk)
gpk_prev = np.zeros_like(gfk)
gpk = np.ones_like(gfk)
for iteration in range(5):
iterator = ct.set_iterator(cfg)
# Patching for testing
e_gk = np.sum((abs(gk) - abs(gk_prev))**2) / (np.sum(abs(gk)**2))
e_gpk = np.sum(
(angle(gk) - angle(gk_prev))**2) / (np.sum(angle(gk)**2))
gk_prev = gk
print('Iteration %d | gk error %.4f' % (iteration, e_gk))
print(e_gpk)
if (e_gk < 5E-5):
break
for it in iterator:
gm = gm * e_gk
CTF = fpmm.generate_CTF(0, 0, [lrsize, lrsize], pupil_radius)
acqpars = it['acqpars']
# indexes, theta, phi = it['indexes'], it['theta'], it['phi']
indexes, kx_rel, ky_rel = ct.n_to_krels(it, cfg=cfg)
lr_sample = np.copy(samples[it['indexes']][:lrsize, :lrsize])
# Calculating coordinates
[ky, kx] = kdsc * ky_rel, kdsc * kx_rel
kyl = int(np.round(yc + ky - (lrsize + 1) / 2))
kyh = kyl + lrsize
kxl = int(np.round(xc + kx - (lrsize + 1) / 2))
kxh = kxl + lrsize
# Fourier-space update
delta_gfk = factor * gfk[kyl:kyh, kxl:kxh] * CTF
# Step 2: lr of the estimated image using the known pupil
delta_gk = ifft2(ifftshift(delta_gfk)) + gm * rand(lrsize, lrsize)
# phase = arctan(real(delta_gk)/(imag(delta_gk)+0.))
phase = angle(delta_gk)
gk_prime = lr_sample * exp(1j * phase) / factor
delta_gfk = fftshift(fft2(gk_prime)) * CTF
gfk[kyl:kyh,
kxl:kxh] = (1 - CTF) * gfk[kyl:kyh, kxl:kxh] + delta_gfk
gk = ifft2(fftshift(gfk))
return gk, gk