def just_denoise(arr,
k=3,
level=5,
noise_std=None,
coefs=None,
supp=None,
min_nscales=2):
if np.iterable(k):
level = len(k)
if coefs is None:
coefs = atrous.decompose(arr, level)
if noise_std is None:
if arr.ndim > 2:
noise_std = atrous.estimate_sigma_mad(coefs[0], True)
else:
noise_std = atrous.estimate_sigma(arr, coefs)
## calculate support taking only positive coefficients (light sources)
if supp is None:
supp = atrous.get_support(coefs,
np.array(k, _dtype_) * noise_std,
modulus=False)
structures = get_structures(coefs, supp)
g = connectivity_graph(structures, min_nscales)
#labels = reduce(lambda a,b:a+b, (n.labels for n in lib.flatten(g)))
new_supp = supp_from_connectivity(g, level)
return atrous.rec_with_support(coefs, new_supp)
def conj_grad_rec(obj, coefs, thresh=0.1, verbose=True):
adjoint = _atilda1
N = len(coefs) - 1
supp = [x for x in supp_from_obj(obj)]
while len(supp) < len(coefs):
supp.append(np.zeros(coefs[0].shape))
WW = _project(coefs, supp)
F0 = atrous.rec_atrous(WW)
x = atrous.decompose(F0, N)
AF0 = _project(x, supp)
Wr = [c1 - c2 for c1, c2 in zip(WW, AF0)]
Fr0 = adjoint(Wr)
AF, Fp, Fr, niter = AF0, F0, Fr0, 0
conv = []
meas_prev = -1e-5
while niter < 1000:
Fr_norm = np.sum(Fr**2)
AF_norm = np.sum([x**2 for x in AF[:-1]])
alpha = Fr_norm / AF_norm
_Fn = Fp + alpha * Fr
Fnext = _Fn * (_Fn >= 0)
AF = _project(atrous.decompose(Fnext, N), supp)
Wrn = [c1 - c2 for c1, c2 in zip(WW, AF)]
meas_next = np.sum([c**2 for c in Wrn[:-1]])
if niter > 0:
meas = 1 - meas_next / meas_prev
else:
meas = 1
conv.append(meas)
if verbose:
print meas, meas_next
if meas < thresh or meas_next < thresh:
return Fnext, conv
meas_prev = meas_next
Frnext = adjoint(Wrn)
beta = np.sum(Frnext**2) / Fr_norm
Frnext = Frnext + beta * Fr
Fr, Wr, Fp = Frnext, Wrn, Fnext
niter += 1
return Fnext
