def deconv_wiener(ys, hs, gamma=1.0e-6, n_threads=6, y_is_fft=False, h_is_fft=False, fft_is_unitary=True):
""" wiener deconv
y is/are the recorded image(s)
h is/are the (already fftshifted) kernel(s)
both can be equal sized tuples, in which case a joint deconvolution is performed
if the data or the kernels are already fourier transformed then
set y_is_fft/h_is_fft to True
"""
if not isinstance(ys, (list, tuple)):
ys = [ys]
if not isinstance(hs, (list, tuple)):
hs = [hs]
if len(ys) != len(hs):
raise ValueError("len(y) != len(h) %d != %d" % (len(ys), len(hs)))
if not np.all([_y.shape == _h.shape for _y, _h in zip(ys, hs)]):
raise ValueError("y and h have non compatible shapes...")
if not y_is_fft:
dshape = ys[0].shape
else:
dshape = ys[0].shape[:-1] + (2 * (ys[0].shape[-1] - 1),)
FFTW = MyFFTW(dshape, unitary=fft_is_unitary, n_threads=n_threads)
if not h_is_fft:
Hs = [FFTW.rfftn(h) for h in hs]
else:
Hs = hs
if not y_is_fft:
Ys = [FFTW.rfftn(y) for y in ys]
else:
Ys = ys
U = reduce(np.add, [_H.conjugate() * _Y for _H, _Y in zip(Hs, Ys)])
U /= 1.0 * gamma + reduce(np.add, [np.abs(_H) ** 2 for _H in Hs])
return FFTW.irfftn(U)
def deconv_rl(ys, hs,
Niter=10,
gamma = 1.e-5,
n_threads = 6,
h_is_fft = False,
log_iter = False,
mult_mode = "root",
fft_is_unitary=False):
""" richardson lucy deconvolution
y is/are the recorded image(s)
h is/are the kernel(s)
both can be equal sized tuples, in which case a joint deconvolution is performed
if the data or the kernels are already fourier transformed then
set y_is_fft/h_is_fft to True
increase gamma if something explodes....
mult_mode = "root", "max", "min"
definies how each lr channel is multiplied
"root" and "min" should be preferred
"""
if not isinstance(ys,(list,tuple)):
ys = [ys]
if not isinstance(hs,(list,tuple)):
hs = [hs]
if len(ys) != len(hs):
raise ValueError("len(y) != len(h) %d != %d"%(len(ys),len(hs)))
if not h_is_fft and not np.all([_y.shape == _h.shape for _y, _h in zip(ys,hs)]):
raise ValueError("y and h have non compatible shapes...")
dshape = ys[0].shape
if log_iter:
print "creating FFTW object"
FFTW = MyFFTW(dshape,n_threads = n_threads, unitary=fft_is_unitary)
if not h_is_fft:
if ys[0].ndim==1:
hs_flip = [h[::-1] for h in hs]
elif ys[0].ndim==2:
hs_flip = [h[::-1,::-1] for h in hs]
elif ys[0].ndim==3:
hs_flip = [h[::-1,::-1,::-1] for h in hs]
else:
raise NotImplementedError("data dimension %s not supported"%ys[0].ndim)
else:
if ys[0].ndim==1:
hs_flip = h
elif ys[0].ndim==2:
hs_flip = [h[::-1,:] for h in hs]
elif ys[0].ndim==3:
hs_flip = [h[::-1,::-1,:] for h in hs]
else:
raise NotImplementedError("data dimension %s not supported"%ys[0].ndim)
if log_iter:
print "setting up ffts"
if not h_is_fft:
Hs = [FFTW.rfftn(h) for h in hs]
Hs_flip = [FFTW.rfftn(h) for h in hs_flip]
else:
Hs = hs
Hs_flip = hs_flip
def _single_lucy_multiplier(d,u_f,h_f,h_flip_f):
tmp = np.abs(FFTW.irfftn(u_f*h_f))
tmp2 = FFTW.rfftn(d/(tmp+gamma))
tmp3 = np.abs(FFTW.irfftn(tmp2*h_flip_f))
return tmp3
u = np.mean([np.mean(_y) for _y in ys])*np.ones_like(ys[0])
um = []
for i in range(Niter):
if log_iter:
print "deconv_rl step %s/%s"%(i+1,Niter)
U = FFTW.rfftn(u)
us = np.stack([_single_lucy_multiplier(y,U,H,H_flip) for y,H,H_flip in zip(ys,Hs,Hs_flip)])
if mult_mode=="root":
u *= reduce(np.multiply,us**(1./len(ys)))
elif mult_mode =="max":
u *= np.amax(us,axis=0)
elif mult_mode =="min":
u *= np.amin(us,axis=0)
else:
raise KeyError(multmode)
um.append(us)
#return u,us[0]
return u
def deconv_wiener2(ys, hs, gamma=1.0e-6, n_threads=6, y_is_fft=False, h_is_fft=False):
""" wiener deconv
y is/are the recorded image(s)
h is/are the (already fftshifted) kernel(s)
both can be equal sized tuples, in which case a joint deconvolution is performed
if the data or the kernels are already fourier transformed then
set y_is_fft/h_is_fft to True
noise level is estimated
gamma is the thikoniv regularizer
"""
if not isinstance(ys, (list, tuple)):
ys = [ys]
if not isinstance(hs, (list, tuple)):
hs = [hs]
if len(ys) != len(hs):
raise ValueError("len(y) != len(h) %d != %d" % (len(ys), len(hs)))
if not np.all([_y.shape == _h.shape for _y, _h in zip(ys, hs)]):
raise ValueError("y and h have non compatible shapes...")
if not y_is_fft:
dshape = ys[0].shape
else:
dshape = ys[0].shape[:-1] + (2 * (ys[0].shape[-1] - 1),)
FFTW = MyFFTW(dshape, unitary=True, n_threads=n_threads)
if not h_is_fft:
Hs = [FFTW.rfftn(h) for h in hs]
else:
Hs = hs
if not y_is_fft:
Ys = [FFTW.rfftn(y) for y in ys]
else:
Ys = ys
# estimate noise power spectrum
hs_cut = [np.abs(h_f[0, 0]) * 1.0e-6 for h_f in Hs]
# inds_noise = [2./(1.+np.exp((np.abs(h_f)-h_cut)/h_cut)) for h_cut, h_f in zip(hs_cut,Hs)]
#
inds_noise = [1.0 * (np.abs(h_f) < h_cut) for h_cut, h_f in zip(hs_cut, Hs)]
inds_signal = [1.0 - ind_noise for ind_noise in inds_noise]
# the different power spectra
ps_y = [np.abs(y_f) ** 2 for y_f in Ys]
ps_signal = [ind * p for ind, p in zip(inds_signal, ps_y)]
mean_ps_noise = [np.mean(ind * p) for ind, p in zip(inds_noise, ps_y)]
ps_noise = [
ind_n * p + ind_s * mean_p for ind_n, ind_s, p, mean_p in zip(inds_noise, inds_signal, ps_y, mean_ps_noise)
]
filters = [p_signal / (1.0 * p_signal + p_noise) for p_signal, p_noise in zip(ps_signal, ps_noise)]
U = reduce(np.add, [_H.conjugate() * _Y * _F for _H, _Y, _F in zip(Hs, Ys, filters)])
U /= 1.0 * gamma + reduce(np.add, [np.abs(_H) ** 2 for _H in Hs])
return FFTW.irfftn(U)
def deconv_tv_al(ys, hs,
mu = 1000.,
rho = 2.,
Niter = 10,
n_threads = 6,
tv_norm = "isotropic",
y_is_fft = False,
h_is_fft = False):
""" total variation deconv
y is/are the recorded image(s)
h is/are the kernel(s)
both can be equal sized tuples, in which case a joint deconvolution is performed
if the data or the kernels are already fourier transformed then
set y_is_fft/h_is_fft to True
tv_norm = "isotropic","anisotropic"
"""
if not isinstance(ys,(list,tuple)):
ys = [ys]
if not isinstance(hs,(list,tuple)):
hs = [hs]
if len(ys) != len(hs):
raise ValueError("len(y) != len(h) %d != %d"%(len(ys),len(hs)))
if not np.all([_y.shape == _h.shape for _y, _h in zip(ys,hs)]):
raise ValueError("y and h have non compatible shapes...")
if not y_is_fft:
dshape = ys[0].shape
else:
dshape = ys[0].shape[:-1]+(2*(ys[0].shape[-1]-1),)
FFTW = MyFFTW(dshape,n_threads = n_threads)
if not h_is_fft:
Hs = [FFTW.rfftn(h) for h in hs]
else:
Hs = hs
if not y_is_fft:
Ys = [FFTW.rfftn(y) for y in ys]
else:
Ys = ys
lap = -dft_lap(dshape, use_rfft = True)
if not y_is_fft:
f = 1.*ys[0]
else:
f = np.real(FFTW.irfftn(ys[0]))
#u = np.array(np.gradient(f))
u = grad(f)
y = np.zeros((ys[0].ndim,)+dshape)
rnorm = 0.
for i in range(Niter):
# print i, _fun(f,mu), rho
# f subproblem
f_f = 1.*mu/rho*reduce(np.add,[_H.conjugate()*_Y for _H,_Y in zip(Hs,Ys)])
f_f -= FFTW.rfftn(divergence(u-1./rho*y))
f_f /= 1.*mu/rho*reduce(np.add,[np.abs(_H)**2 for _H in Hs]) + lap
f = np.real(FFTW.irfftn(f_f))
# u subproblem
#df = np.array(np.gradient(f))
df = grad(f)
if tv_norm == "isotropic":
# isotropic case
v = df + 1./rho*y
mv = np.sqrt(np.sum(v**2,axis=0))
mv[mv==0] = 1.
tv = np.maximum(0,mv-1./rho)/mv
u = tv*v
else:
# anisotropic case
u = soft_thresh(df + 1./rho*y,1./rho)
y -= rho*(u-df)
print rho
rnorm_new = np.sqrt(np.mean((u-df)**2))
if rnorm_new > .7*rnorm:
rho *= 2.
rnorm = rnorm_new
return f
def deconv_hyperlap(y, h,
lam = 1000.,
outeriter= 6,
inneriter = 1,
reg0=0.,
alpha = None,
n_threads = 6,
logged = False):
""" hyper laplacian regularized deconvolution
Krishnan et al
see http://people.ee.duke.edu/~lcarin/NIPS2009_0341.pdf
y is/are the recorded image(s)
h is/are the (already fftshifted) kernel(s) of same shape
both can be equal sized tuples, in which case a joint deconvolution is performed
if the data or the kernels are already fourier transformed then
set y_is_fft/h_is_fft to True
"""
# see the paper for definition of the following hyperparameter
beta = 1.
beta_inc = 2.8
if alpha is not None:
look = get_lookup(alpha)
else:
alpha = 1.
#make everything scale invariant
y_mean = np.mean(y)
y = y/y_mean
dshape = y.shape
FFTW = MyFFTW(dshape,unitary = False, n_threads = n_threads)
H = FFTW.rfftn(h)
Y = FFTW.rfftn(y)
fs = finite_deriv_dft_forward(dshape, use_rfft=True)
x = 1.*y
den1 = reduce(np.add,[f.conjugate()*f for f in fs])
den2 = H.conjugate()*H
nom1 = H.conjugate()*Y
res = []
def objective(x):
dxs = np.gradient(x)
X = FFTW.rfftn(x)
x_b = FFTW.irfftn(X*H)
ener = .5*lam*np.sum((x_b-y)**2)
grad = np.sum([abs(dx)**alpha for dx in dxs])
return ener +grad
for i in xrange(outeriter):
print "iteration: %d / %d"%(i+1,outeriter)
for _ in xrange(inneriter):
# w subproblem
#dx1, dx2 = np.gradient(x)
#dxs = [.5*(np.roll(x,-1,i)-np.roll(x,1,i)) for i in range(y.ndim)]
dxs = np.gradient(x)
if alpha == 1.:
ws = [soft_thresh(dx,1./beta) for dx in dxs]
else:
ws = [look(dx,beta) for dx in dxs]
res.append(x)
# x subproblem
w_fs = [FFTW.rfftn(w) for w in ws]
nom2 = reduce(np.add,[f.conjugate()*w_f for f, w_f in zip(fs,w_fs)])
x = FFTW.irfftn((nom1+1.*lam/beta*nom2)/(reg0*lam/beta+den1+1.*lam/beta*den2))
# nom = f1.conjugate()*w1_f + f2.conjugate()*w2_f + 1.*lam/beta*H.conjugate()*Y
# den = reg + f1.conjugate()*f1 + f2.conjugate()*f2 + 1.*lam/beta*H.conjugate()*H
# x = FFTW.irfftn(nom/den)
if logged:
print "objective f = %s"%objective(x)
# return H.conjugate()*Y
beta *= beta_inc
x *= y_mean
return x, res
def deconv_hyperlap_breg(y, h,
lam = 1000.,
gamma = 5.,
niter = 10,
n_threads = 6,
logged = False):
""" hyper laplacian regularized deconvolution via split bregman
Krishnan et al
see http://people.ee.duke.edu/~lcarin/NIPS2009_0341.pdf
y is/are the recorded image(s)
h is/are the (already fftshifted) kernel(s) of same shape
both can be equal sized tuples, in which case a joint deconvolution is performed
if the data or the kernels are already fourier transformed then
set y_is_fft/h_is_fft to True
params:
lam - the higher, the bigger the data term
gamma - the smaller, the more TV enforced
"""
alpha = 1.
#make everything scale invariant
y_mean = np.mean(y)
y = y/y_mean
dshape = y.shape
FFTW = MyFFTW(dshape,unitary = False, n_threads = n_threads)
H = FFTW.rfftn(h)
Y = FFTW.rfftn(y)
dft_stencil = dft_lap(dshape, use_rfft=True)
x = 1.*y
def objective(x):
dxs = np.gradient(x)
X = FFTW.rfftn(x)
x_b = FFTW.irfftn(X*H)
ener = .5*lam*np.sum((x_b-y)**2)
grad = np.sum([abs(dx)**alpha for dx in dxs])
return ener + grad
d = np.zeros((x.ndim,)+x.shape)
b = np.zeros((x.ndim,)+x.shape)
grad_x = np.stack(np.gradient(x))
for i in xrange(niter):
print "iteration: %d / %d"%(i+1,niter)
# d subproblem
d = soft_thresh(grad_x+b,1./gamma)
# x subproblem
nom1 = H.conjugate()*Y
nom2 = FFTW.rfftn(divergence(d-b))
den1 = H.conjugate()*H
den2 = dft_stencil
x = FFTW.irfftn((1.*lam/gamma*nom1 - nom2)/(1.*lam/gamma*den1 - den2))
grad_x = np.stack(np.gradient(x))
# enforcing constraint
b += grad_x - d
x *= y_mean
return x
