def poop():
global sources
meas_focus = get_measurements(sources=sources,
mode='circular',
psfs=psfs_focus,
real=True)
meas_mid = get_measurements(sources=sources,
mode='circular',
psfs=psfs_mid,
real=True)
meas_outer = get_measurements(sources=sources,
mode='circular',
psfs=psfs_outer,
real=True)
meas_csbs = get_measurements(sources=sources,
mode='circular',
psfs=psfs_csbs,
real=True)
meas_focus = add_noise(meas_focus, model='gaussian', dbsnr=15)
meas_mid = add_noise(meas_mid, model='gaussian', dbsnr=15)
meas_outer = add_noise(meas_outer, model='gaussian', dbsnr=15)
meas_csbs = add_noise(meas_csbs, model='gaussian', dbsnr=15)
# meas_focus = add_noise(meas_focus, model='poisson', max_count=20)
# meas_mid = add_noise(meas_mid, model='poisson', max_count=20)
# meas_outer = add_noise(meas_outer, model='poisson', max_count=20)
# meas_csbs = add_noise(meas_csbs, model='poisson', max_count=20)
recon_focus = tikhonov(psfs=psfs_focus,
measurements=meas_focus,
tikhonov_lam=1e-3)
recon_mid = tikhonov(psfs=psfs_mid,
measurements=meas_mid,
tikhonov_lam=1e-3)
recon_outer = tikhonov(psfs=psfs_outer,
measurements=meas_outer,
tikhonov_lam=1e-3)
recon_csbs = tikhonov(psfs=psfs_csbs,
measurements=meas_csbs,
tikhonov_lam=1e-3)
recon_focus -= np.mean(recon_focus, axis=(1, 2))[:, np.newaxis, np.newaxis]
recon_mid -= np.mean(recon_mid, axis=(1, 2))[:, np.newaxis, np.newaxis]
recon_outer -= np.mean(recon_outer, axis=(1, 2))[:, np.newaxis, np.newaxis]
recon_csbs -= np.mean(recon_csbs, axis=(1, 2))[:, np.newaxis, np.newaxis]
sources_comp = sources - np.mean(sources, axis=(1, 2))[:, np.newaxis,
np.newaxis]
result_focus = np.sum(compare_ssim(sources_comp, recon_focus))
result_mid = np.sum(compare_ssim(sources_comp, recon_mid))
result_outer = np.sum(compare_ssim(sources_comp, recon_outer))
result_csbs = np.sum(compare_ssim(sources_comp, recon_csbs))
print('focused SSIM sum:', result_focus)
print('middle SSIM sum:', result_mid)
print('outer SSIM sum:', result_outer)
print('csbs SSIM sum:', result_csbs)
return [result_focus, result_mid, result_outer, result_csbs]
def measure_reconstruct():
noise = 15 # dB SNR
meas_focus = get_measurements(sources=sources,
mode='circular',
psfs=psfs_focus,
real=True)
meas_csbs = get_measurements(sources=sources,
mode='circular',
psfs=psfs_csbs,
real=True)
meas_focus = add_noise(meas_focus, model='gaussian', dbsnr=noise)
meas_csbs = add_noise(meas_csbs, model='gaussian', dbsnr=noise)
recon_focus = tikhonov(psfs=psfs_focus,
measurements=meas_focus,
tikhonov_lam=1e-3)
recon_csbs = tikhonov(psfs=psfs_csbs,
measurements=meas_csbs,
tikhonov_lam=1e-3)
result_focus = np.sum(compare_ssim(sources, recon_focus))
result_csbs = np.sum(compare_ssim(sources, recon_csbs))
ratio = result_csbs / result_focus
return dict(**locals())
def test_tikhonov(self):
from mas.deconvolution import tikhonov
recon = tikhonov(measurements=np.fft.fftshift(self.measured_noisy,
axes=(1, 2)),
psfs=self.psfs,
tikhonov_lam=5e-2,
tikhonov_order=1)
def cost(tikhonov_lam_exp, csbs_lam_exp):
# csbs_grid
psfs.copies = np.repeat(12, 12)
# csbs_focus
# psfs.copies = np.repeat(12, 2)
csbs(psfs, sse_cost, 12, lam=10**csbs_lam_exp, order=1)
measured = get_measurements(psfs=psfs, sources=truth, real=True)
measured_noisy = add_noise(measured, model='poisson', max_count=500)
recon = tikhonov(
sources=measured,
psfs=psfs,
measurements=measured_noisy,
tikhonov_lam=10**tikhonov_lam_exp
)
plt.imshow(recon[0])
plt.show()
plt.pause(.05)
return compare_ssim(truth[0], recon[0]) + compare_ssim(truth[1], recon[1])
def motion_deblur(*, sv, coadded, drift):
"""
"""
pixel_size_um = sv.pixel_size * 1e6
# width of the final motion blur kernel with CCD pixel size
kernel_size = 11
(x, y) = (pixel_size_um * drift[0], pixel_size_um * drift[1])
N = int(np.ceil(np.max((abs(x), abs(y)))))
# set the shape of the initial kernel with 1 um pixels based on the estimated drift
kernel_um = np.zeros((2 * N + 1, 2 * N + 1))
# calculate the line representing the motion blur
rr, cc, val = line_aa(
N + np.round((y / 2)).astype(int),
N - np.round((x / 2)).astype(int),
N - np.round((y / 2)).astype(int),
N + np.round((x / 2)).astype(int),
)
# update the kernel with the calculated line
kernel_um[rr, cc] = val
# resize the initial 1 um kernel to the given pixel size
kernel = resize(size_equalizer(kernel_um,
[int(pixel_size_um) * kernel_size] * 2),
[kernel_size] * 2,
anti_aliasing=True)
# compute the analytical photon sieve PSF with the given pixel size
psfs = copy.deepcopy(sv.psfs)
# convolve the photon sieve PSF with the motion blur kernel to find the "effective blurring kernel"
psfs.psfs[0, 0] = convolve2d(psfs.psfs[0, 0], kernel, mode='same')
# normalize the kernel
psfs.psfs[0, 0] /= psfs.psfs[0, 0].sum()
# normalize the coadded image (doesn't change anything but helps choosing regularization parameter consistently)
coadded /= coadded.max()
# do a tikhonov regularized deblurring on the coadded image to remove the
# in-frame blur with the calculated "effective blurring kernel"
recon_tik = tikhonov(measurements=coadded[np.newaxis, :, :],
psfs=psfs,
tikhonov_lam=1e1,
tikhonov_order=1)
plt.figure()
plt.imshow(recon_tik[0], cmap='gist_heat')
plt.title('Deblurred Tikhonov')
plt.show()
# do a Plug and Play with BM3D reconstruction with tikhonov initialization
recon = admm(measurements=coadded[np.newaxis, :, :],
psfs=psfs,
regularizer=partial(bm3d_pnp),
recon_init=recon_tik,
plot=False,
iternum=5,
periter=1,
nu=10**-0.0,
lam=[10**-0.5])
plt.figure()
plt.imshow(recon[0], cmap='gist_heat')
plt.title('Deblurred')
plt.show()
return recon
# measured = get_measurements(
# sources=sources,
# mode='circular',
# meas_size=meas_size,
# psfs=psfs,
# blur_sigma=1,
# drift_amount=5
# )
measured_noisy = add_noise(measured, max_count=100, model='Poisson')
# %% recon ---------------------------------
recon = tikhonov(
measurements=measured_noisy,
psfs=psfs,
tikhonov_lam=2e-2,
tikhonov_order=1
)
recon = admm(
sources=sources_ref,
measurements=measured_noisy,
psfs=psfs,
regularizer=partial(
# patch_based
# TV
bm3d_pnp
# dncnn_pnp, model=model
),
recon_init=recon,
iternum=20,
periter=5,
noise = 15 # dB SNR
meas_focus = get_measurements(sources=sources,
mode='circular',
psfs=psfs_focus,
real=True)
meas_csbs = get_measurements(sources=sources,
mode='circular',
psfs=psfs_csbs,
real=True)
meas_focus = add_noise(meas_focus, model='gaussian', dbsnr=noise)
meas_csbs = add_noise(meas_csbs, model='gaussian', dbsnr=noise)
recon_focus = tikhonov(psfs=psfs_focus,
measurements=meas_focus,
tikhonov_lam=1e-3)
recon_csbs = tikhonov(psfs=psfs_csbs,
measurements=meas_csbs,
tikhonov_lam=1e-3)
result_focus = np.sum(compare_ssim(sources, recon_focus))
result_csbs = np.sum(compare_ssim(sources, recon_csbs))
print(result_csbs / result_focus, result_focus)
# print('focused SSIM:', result_focus)
# print('csbs SSIM:', result_csbs)
# recon_focus_mean = recon_focus - np.mean(recon_focus, axis=(1, 2))[:, np.newaxis, np.newaxis]
# recon_csbs_mean = recon_csbs - np.mean(recon_csbs, axis=(1, 2))[:, np.newaxis, np.newaxis]
# sources_mean = sources - np.mean(sources, axis=(1, 2))[:, np.newaxis, np.newaxis]
source_wavelengths=source_wavelengths,
psf_generator=circ_incoherent_psf,
# image_width=psf_width,
num_copies=1)
measured = get_measurements(sources=sources,
mode='valid',
meas_size=meas_size,
psfs=psfs)
measured_noisy = add_noise(measured, max_count=10, model='Poisson')
# %% recon ------------------------------------
recon = tikhonov(sources=sources_ref,
measurements=measured_noisy,
psfs=psfs,
tikhonov_lam=2e-1,
tikhonov_order=1)
if deconvolver == admm:
recon = admm(
sources=sources_ref,
measurements=measured_noisy,
psfs=psfs,
regularizer=partial(
# patch_based
# TV
bm3d_pnp
# dncnn_pnp, model=model
),
recon_init=recon,
iternum=20,
num_copies=6)
psfs_csbs = csbs(psfs_csbs, sse_cost, 12, lam=csbs_lam, order=order)
# %% measure
psnr_focus, psnr_csbs, ssim_focus, ssim_csbs = [], [], [], []
measured_focus = get_measurements(sources=sources, psfs=psfs_focus)
measured_csbs = get_measurements(sources=sources, psfs=psfs_csbs)
for i in range(50):
measured_noisy_focus = add_noise(measured_focus,
dbsnr=15,
model='Gaussian')
measured_noisy_csbs = add_noise(measured_csbs, dbsnr=15, model='Gaussian')
recon_focus = tikhonov(psfs=psfs_focus,
measurements=measured_noisy_focus,
tikhonov_lam=1e-3,
tikhonov_order=order)
recon_csbs = tikhonov(psfs=psfs_csbs,
measurements=measured_noisy_csbs,
tikhonov_lam=1e-3,
tikhonov_order=order)
psnr_focus.append(compare_psnr(sources, recon_focus))
psnr_csbs.append(compare_psnr(sources, recon_csbs))
ssim_focus.append(np.sum(compare_ssim(sources, recon_focus)))
ssim_csbs.append(np.sum(compare_ssim(sources, recon_csbs)))
print('psnr_focus:{} \npsnr_csbs:{} \nssim_focus:{} \nssim_csbs:{}'.format(
np.array(psnr_focus).mean(),
np.array(psnr_csbs).mean(),
np.array(ssim_focus).mean(),