def PGPD_Denoising(par, model):
im_out = par['nim']
h, w = im_out.shape
# Fill in more parameters
par['nSig0'] = par['nSig']
par['maxr'] = h - par['ps'] + 1
par['maxc'] = w - par['ps'] + 1
par['maxrc'] = par['maxr'] * par['maxc']
par['h'] = h
par['w'] = w
r = range(0, par['maxr'], par['step'])
par['r'] = r + range(r[-1] + 1, par['maxr'])
# Unsure if I translated this line correctly
#par['r'] = [r r()+1:par['maxr']]
c = range(0, par['maxc'], par['step'])
par['c'] = c + range(c[-1] + 1, par['maxc'])
par['lenr'] = len(par['r'])
par['lenc'] = len(par['c'])
par['lenrc'] = par['lenr'] * par['lenc']
par['ps2'] = par['ps'] * par['ps']
for ite in range(par['IteNum']):
# iterative regularization
im_out = im_out + par['delta'] * (par['nim'] - im_out)
# estimation of noise variance
if ite == 0:
par['nSig'] = par['nSig0']
else:
dif = numpy.mean(numpy.mean(numpy.square(par['nim'] - im_out)))
par['nSig'] = numpy.sqrt(
abs(par['nSig0'] * par['nSig0'] - dif)) * par['eta']
# search non-local patch groups
[nDCnlX, blk_arr, DC, par] = CalNonLocal(im_out, par)
# Gaussian dictionary selection by MAP
if (ite) % 2 == 0:
# CHECKPOINT
PYZ = numpy.zeros((model['nmodels'], (DC.shape)[1]))
sigma2I = (par['nSig']**2) * numpy.eye(par['ps2'])
for i in range(model['nmodels']):
sigma = model['covs'].item()[:, :, i] + sigma2I
R = numpy.linalg.cholesky(sigma)
Q = numpy.transpose(
numpy.linalg.lstsq(numpy.transpose(R), nDCnlX)[0])
#print numpy.diag(R).shape, numpy.log(numpy.diag(R)).shape
print numpy.sum(numpy.log(numpy.diag(R))).shape
print(numpy.dot(numpy.transpose(Q), Q) / 2.).shape
#print Q[0].shape, Q[1].shape
#print type(numpy.dot(Q,Q)/2.), numpy.sum(numpy.log(numpy.diag(R)))
TempPYZ = -numpy.sum(numpy.log(numpy.diag(R))) - numpy.dot(
numpy.transpose(Q), Q) / 2.
print DC.shape, TempPYZ.shape, par['nlsp']
TempPYZ = TempPYZ.reshape((par['nlsp'], DC.shape[1]))
#print TempPYZ.shape
#print numpy.sum(TempPYZ)
PYZ[i, :] = numpy.sum(TempPYZ)
# find the most likely component for each patch group
tmp, dicidx = max(PYZ)
dicidx = numpy.transpose(dicidx)
idx, s_idx = numpy.sort(dicidx)
idx2 = idx[:-1] - idx[1:]
seq = numpy.nonzero(idx2)
seg = numpy.concatenate([0, seq, len(dicidx)])
seg = seg.reshape(len(seg), 1)
# Weighted Sparse Coding
X_hat = numpy.zeros((par['ps2'], par['maxrc']))
W = numpy.zeros((par['ps2'], par['maxrc']))
for j in range(len(seg) - 2):
idx = s_idx[seg[j] + range(seg[j + 1])]
cls = dicidx[idx[0]]
D = par['D'][:, :, cls]
S = par['S'][:, cls]
lambdaM = numpy.tile(
(par['c1'] * par['nSig']**2.) / (numpy.sqrt(S) + 0.0001),
par['nlsp'])
for i in range(len(idx)):
#Y = nDCnlX(:,(idx(i)-1)*par['nlsp']+1:idx(i)*par['nlsp'])
Y = nDCnlX[:, (idx[i] - 1) * par['nlsp'] +
range(idx[i] * par['nlsp'])]
b = numpy.transpose(D) * Y
# soft threshold
alpha = numpy.sign(b) * numpy.max(abs(b) - lambdaM / 2., 0.)
# add DC components and aggregation
#X_hat[:,blk_arr[:,idx[i]]] = X_hat[:,blk_arr[:,idx[i]]]+bsxfun(@plus,D*alpha, DC[:,idx[i]])
X_hat[:, blk_arr[:, idx[i]]] = X_hat[:, blk_arr[:, idx[i]]] + (
D * alpha + DC[:, idx[i]])
W[:, blk_arr[:,
idx[i]]] = W[:, blk_arr[:, idx[i]]] + numpy.ones(
par['ps2'], par['nlsp'])
# Reconstruction
im_out = numpy.zeros(h, w)
im_wei = numpy.zeros(h, w)
r = range(par['maxr'])
c = range(par['maxc'])
k = 0
for i in range(par['ps']):
for j in range(par['ps']):
#im_out(r-1+i,c-1+j) = im_out(r-1+i,c-1+j) + reshape( numpy.transpose(X_hat(k,:)), [par['maxr'] par['maxc']])
#im_wei(r-1+i,c-1+j) = im_wei(r-1+i,c-1+j) + reshape( numpy.transpose(W(k,:)), [par['maxr'] par['maxc']])
im_out[r + i, c + j] = im_out[r + i, c + j] + numpy.reshape(
numpy.transpose(X_hat[k, :]), [par['maxr'], par['maxc']])
im_wei[r + i, c + j] = im_wei[r + i, c + j] + numpy.reshape(
numpy.transpose(W[k, :]), [par['maxr'], par['maxc']])
k = k + 1
im_out = im_out / im_wei
# calculate the PSNR and SSIM
PSNR = csnr.csnr(im_out * 255, par['I'] * 255., 0, 0)
SSIM = cal_ssim.cal_ssim(im_out * 255, par['I'] * 255., 0, 0)
print('Iter #d : PSNR = #2.4f, SSIM = #2.4f\n', ite, PSNR, SSIM)
im_out[im_out > 1] = 1
im_out[im_out < 0] = 0
return im_out, par
import numpy
import Parameters_Setting
# set parameters
nSig = 50.
[par, model] = Parameters_Setting.Parameters_Setting(nSig)
# read clean image
I = skimage.io.imread('cameraman.png') / 255.
par['I'] = I
# generate noisy image
#random.seed()
# I doubt that this translates well to python
#par.nim = par.I + par.nSig*randn(size(par.I));
nim = I + par['nSig'] * numpy.random.randn(I.shape[0], I.shape[1])
par['nim'] = nim
PSNR_init = csnr.csnr(nim * 255., I * 255., 0., 0.)
SSIM_init = cal_ssim.cal_ssim(nim * 255., I * 255., 0., 0.)
print('The initial value of PSNR = ' + str(PSNR_init) + ', SSIM = ' +
str(SSIM_init))
# PGPD denoising
[im_out, par] = PGPD_Denoising.PGPD_Denoising(par, model)
# [im_out,par] = PGPD_Denoising_faster(par,model) # faster speed
# calculate the PSNR and SSIM
PSNR_final = csnr.csnr(im_out * 255., I * 255., 0., 0.)
SSIM_final = cal_ssim.cal_ssim(im_out * 255., I * 255., 0., 0.)
print('Cameraman Results: PSNR = ' + str(PSNR_final) + ', SSIM = ' +
str(SSIM_final))
import csnr
import PGPD_Denoising
import numpy
import Parameters_Setting
# set parameters
nSig = 50.
[par, model] = Parameters_Setting.Parameters_Setting( nSig )
# read clean image
I = skimage.io.imread('cameraman.png')/255.
par['I'] = I
# generate noisy image
#random.seed()
# I doubt that this translates well to python
#par.nim = par.I + par.nSig*randn(size(par.I));
nim = I + par['nSig']*numpy.random.randn(I.shape[0], I.shape[1])
par['nim'] = nim
PSNR_init = csnr.csnr(nim*255., I*255., 0., 0.)
SSIM_init = cal_ssim.cal_ssim(nim*255., I*255., 0., 0.)
print('The initial value of PSNR = '+str(PSNR_init)+', SSIM = '+str(SSIM_init))
# PGPD denoising
[im_out, par] = PGPD_Denoising.PGPD_Denoising(par, model)
# [im_out,par] = PGPD_Denoising_faster(par,model) # faster speed
# calculate the PSNR and SSIM
PSNR_final = csnr.csnr(im_out*255., I*255., 0., 0.)
SSIM_final = cal_ssim.cal_ssim(im_out*255., I*255., 0., 0.)
print('Cameraman Results: PSNR = '+str(PSNR_final)+', SSIM = '+str(SSIM_final))
def PGPD_Denoising(par, model):
im_out = par["nim"]
h, w = im_out.shape
# Fill in more parameters
par["nSig0"] = par["nSig"]
par["maxr"] = h - par["ps"] + 1
par["maxc"] = w - par["ps"] + 1
par["maxrc"] = par["maxr"] * par["maxc"]
par["h"] = h
par["w"] = w
r = range(0, par["maxr"], par["step"])
par["r"] = r + range(r[-1] + 1, par["maxr"])
# Unsure if I translated this line correctly
# par['r'] = [r r()+1:par['maxr']]
c = range(0, par["maxc"], par["step"])
par["c"] = c + range(c[-1] + 1, par["maxc"])
par["lenr"] = len(par["r"])
par["lenc"] = len(par["c"])
par["lenrc"] = par["lenr"] * par["lenc"]
par["ps2"] = par["ps"] * par["ps"]
for ite in range(par["IteNum"]):
# iterative regularization
im_out = im_out + par["delta"] * (par["nim"] - im_out)
# estimation of noise variance
if ite == 0:
par["nSig"] = par["nSig0"]
else:
dif = numpy.mean(numpy.mean(numpy.square(par["nim"] - im_out)))
par["nSig"] = numpy.sqrt(abs(par["nSig0"] * par["nSig0"] - dif)) * par["eta"]
# search non-local patch groups
[nDCnlX, blk_arr, DC, par] = CalNonLocal(im_out, par)
# Gaussian dictionary selection by MAP
if (ite) % 2 == 0:
# CHECKPOINT
PYZ = numpy.zeros((model["nmodels"], (DC.shape)[1]))
sigma2I = (par["nSig"] ** 2) * numpy.eye(par["ps2"])
for i in range(model["nmodels"]):
sigma = model["covs"].item()[:, :, i] + sigma2I
R = numpy.linalg.cholesky(sigma)
Q = numpy.transpose(numpy.linalg.lstsq(numpy.transpose(R), nDCnlX)[0])
# print numpy.diag(R).shape, numpy.log(numpy.diag(R)).shape
print numpy.sum(numpy.log(numpy.diag(R))).shape
print (numpy.dot(numpy.transpose(Q), Q) / 2.0).shape
# print Q[0].shape, Q[1].shape
# print type(numpy.dot(Q,Q)/2.), numpy.sum(numpy.log(numpy.diag(R)))
TempPYZ = -numpy.sum(numpy.log(numpy.diag(R))) - numpy.dot(numpy.transpose(Q), Q) / 2.0
print DC.shape, TempPYZ.shape, par["nlsp"]
TempPYZ = TempPYZ.reshape((par["nlsp"], DC.shape[1]))
# print TempPYZ.shape
# print numpy.sum(TempPYZ)
PYZ[i, :] = numpy.sum(TempPYZ)
# find the most likely component for each patch group
tmp, dicidx = max(PYZ)
dicidx = numpy.transpose(dicidx)
idx, s_idx = numpy.sort(dicidx)
idx2 = idx[:-1] - idx[1:]
seq = numpy.nonzero(idx2)
seg = numpy.concatenate([0, seq, len(dicidx)])
seg = seg.reshape(len(seg), 1)
# Weighted Sparse Coding
X_hat = numpy.zeros((par["ps2"], par["maxrc"]))
W = numpy.zeros((par["ps2"], par["maxrc"]))
for j in range(len(seg) - 2):
idx = s_idx[seg[j] + range(seg[j + 1])]
cls = dicidx[idx[0]]
D = par["D"][:, :, cls]
S = par["S"][:, cls]
lambdaM = numpy.tile((par["c1"] * par["nSig"] ** 2.0) / (numpy.sqrt(S) + 0.0001), par["nlsp"])
for i in range(len(idx)):
# Y = nDCnlX(:,(idx(i)-1)*par['nlsp']+1:idx(i)*par['nlsp'])
Y = nDCnlX[:, (idx[i] - 1) * par["nlsp"] + range(idx[i] * par["nlsp"])]
b = numpy.transpose(D) * Y
# soft threshold
alpha = numpy.sign(b) * numpy.max(abs(b) - lambdaM / 2.0, 0.0)
# add DC components and aggregation
# X_hat[:,blk_arr[:,idx[i]]] = X_hat[:,blk_arr[:,idx[i]]]+bsxfun(@plus,D*alpha, DC[:,idx[i]])
X_hat[:, blk_arr[:, idx[i]]] = X_hat[:, blk_arr[:, idx[i]]] + (D * alpha + DC[:, idx[i]])
W[:, blk_arr[:, idx[i]]] = W[:, blk_arr[:, idx[i]]] + numpy.ones(par["ps2"], par["nlsp"])
# Reconstruction
im_out = numpy.zeros(h, w)
im_wei = numpy.zeros(h, w)
r = range(par["maxr"])
c = range(par["maxc"])
k = 0
for i in range(par["ps"]):
for j in range(par["ps"]):
# im_out(r-1+i,c-1+j) = im_out(r-1+i,c-1+j) + reshape( numpy.transpose(X_hat(k,:)), [par['maxr'] par['maxc']])
# im_wei(r-1+i,c-1+j) = im_wei(r-1+i,c-1+j) + reshape( numpy.transpose(W(k,:)), [par['maxr'] par['maxc']])
im_out[r + i, c + j] = im_out[r + i, c + j] + numpy.reshape(
numpy.transpose(X_hat[k, :]), [par["maxr"], par["maxc"]]
)
im_wei[r + i, c + j] = im_wei[r + i, c + j] + numpy.reshape(
numpy.transpose(W[k, :]), [par["maxr"], par["maxc"]]
)
k = k + 1
im_out = im_out / im_wei
# calculate the PSNR and SSIM
PSNR = csnr.csnr(im_out * 255, par["I"] * 255.0, 0, 0)
SSIM = cal_ssim.cal_ssim(im_out * 255, par["I"] * 255.0, 0, 0)
print ("Iter #d : PSNR = #2.4f, SSIM = #2.4f\n", ite, PSNR, SSIM)
im_out[im_out > 1] = 1
im_out[im_out < 0] = 0
return im_out, par
