def wf(y, mask, iter, verbose=False):
m = y.size
n = mask.sum()
mu_max = 0.2
t0 = 330.
y = y**2
lambd2 = y.sum()
x = np.random.rand(*y.shape)
x[~mask] = np.zeros(m - n)
x = x / np.linalg.norm(x, 2)
for i in range(1000):
x = tools.ifft2d(y * tools.fft2d(x))
x[~mask] = np.zeros(m - n)
x = x / np.linalg.norm(x, 2)
x = x * np.sqrt(lambd2)
for t in range(iter):
mu = np.min([1 - np.exp(-t / t0), mu_max])
f = tools.fft2d(x)
x = x - mu * tools.ifft2d((f**2 - y) * f) / lambd2
x[~mask] = np.zeros(m - n)
x = np.real(x)
res = np.real(x)
return res
def oss(y, mask, beta=0.9, verbose=False):
phase = np.random.rand(*y.shape) * 2 * np.pi
x_prev = np.real(tools.ifft2d(y * np.exp(1j * phase)))
sigma1 = np.linspace(y.shape[0], 1. / y.shape[0], 10)
sigma2 = np.linspace(y.shape[1], 1. / y.shape[1], 10)
filtercount = 10
iter = 2000
Rsize = y.shape[0]
X = np.arange(1, iter + 1)
FX = (filtercount + 1 - np.ceil(X * filtercount / iter)) * np.ceil(
iter / (1 * filtercount))
FX = ((FX - np.ceil(iter / filtercount)) *
(2 * Rsize) / np.max(FX)) + (2 * Rsize / 10)
for i in range(iter):
if i == 0 or FX[i - 1] != FX[i]:
sigma = (FX[i], FX[i])
filter = tools.get_gauss2d(y.shape, sigma, normalize=False)
filter = ifftshift(filter)
Fx_curr = y * np.exp(1j * phase)
x_curr = np.real(tools.ifft2d(Fx_curr))
idx = (x_curr < 0) + (mask == False)
x_curr[idx] = x_prev[idx] - beta * x_curr[idx]
Fx_curr = tools.fft2d(x_curr)
idx = (mask == False)
x_curr[idx] = np.real(tools.ifft2d(Fx_curr * filter))[idx]
Fx_curr = tools.fft2d(x_curr)
phase = np.angle(Fx_curr)
x_prev = x_curr.copy()
return x_curr
def prox(self, x, tau=1):
# z = np.zeros(y.shape)
# z[self.mask] = x.flatten()
z = x
zf = tools.fft2d(z)
mag = 1 / (self.alpha + tau) * (self.alpha * self.y + tau * np.abs(zf))
res = mag * np.exp(1j * np.angle(zf))
res = np.real(tools.ifft2d(res))
# return res[self.mask].reshape(x.shape)
return res
def hio(y, mask, iter, beta=0.9, verbose=False):
phase = np.random.rand(*y.shape) * 2 * np.pi
x_prev = np.real(tools.ifft2d(y * np.exp(1j * phase)))
for i in range(iter):
if verbose:
print('Iteration #{:d} '.format(i + 1), end='')
Fx_curr = y * np.exp(1j * phase)
x_curr = np.real(tools.ifft2d(Fx_curr))
idx = ((1 + beta) * x_curr < x_prev) + (mask == False)
# idx = (x_curr < 0) + (mask == False)
x_curr[idx] = x_prev[idx] - beta * x_curr[idx]
Fx_curr = tools.fft2d(x_curr)
phase = np.angle(Fx_curr)
x_prev = x_curr.copy()
# residual = np.power(y - np.abs(new_f), 2).mean()
return x_curr
def prox(self, x, tau=1):
xf = tools.fft2d(x)
mag = 1 / (self.alpha + tau) * (self.alpha * self.y + tau * np.abs(xf))
res = mag * np.exp(1j*np.angle(xf))
res = np.real(tools.ifft2d(res))
return res
def grad(self, x):
f = tools.fft2d(x)
return self.alpha * np.real(tools.ifft2d(f - y * f / np.abs(f)))
def __call__(self, x):
f = tools.fft2d(x)
return np.real(self.alpha * np.sum((y - np.abs(f))**2))
img = np.array(img) / 255.
n, m = img.shape
pad_len_1 = int(img.shape[0] * (np.sqrt(args.samprate) - 1)) // 2
pad_len_2 = int(img.shape[1] * (np.sqrt(args.samprate) - 1)) // 2
n, m = img.shape
img = np.pad(img, ((pad_len_1, pad_len_1), (pad_len_2, pad_len_2)), 'constant', constant_values=((0, 0), (0, 0)))
mask = np.ones(img.shape, dtype=bool) * False
mask[pad_len_1:-pad_len_1, pad_len_2:-pad_len_2] = True
# sd = 0.1 * np.sqrt(n*m)
# noise = np.random.normal(size=img.shape) * sd + 1j * np.random.normal(size=img.shape) * sd
# noise = 0
if args.noise == 'gaussian':
y = np.real(np.abs(tools.fft2d(img))) + np.random.normal(size=img.shape) * args.noiselvl / 255.
elif args.noise == 'poisson':
yy = np.real(np.abs(tools.fft2d(img)))
alpha = args.noiselvl / 255.
intensity_noise = alpha * yy * np.random.normal(size=img.shape)
y = (yy**2 + intensity_noise)
y = y * (y > 0)
y = np.sqrt(y)
elif args.noise == 'rician':
yy = tools.fft2d(img)
y = yy + (np.random.normal(size=img.shape) + 1j * np.random.normal(size=img.shape)) * args.noiselvl / 255. / np.sqrt(2)
y = np.abs(y)
x = algo.hio(y, mask, args.iter, beta=args.beta)
img = Image.open(args.image).convert('L')
img = np.array(img) / 255
pad_len_1 = int(img.shape[0] * (np.sqrt(args.samprate) - 1)) // 2
pad_len_2 = int(img.shape[1] * (np.sqrt(args.samprate) - 1)) // 2
n, m = img.shape
img = np.pad(img, ((pad_len_1, pad_len_1), (pad_len_2, pad_len_2)),
'constant',
constant_values=((0, 0), (0, 0)))
mask = np.ones(img.shape, dtype=bool) * False
mask[pad_len_1:-pad_len_1, pad_len_2:-pad_len_2] = True
if args.noise == 'gaussian':
y = np.real(np.abs(tools.fft2d(img))) + np.random.normal(
size=img.shape) * args.noiselvl / 255.
sigma = args.noiselvl if args.noiselvl > 0 else 1
elif args.noise == 'poisson':
yy = np.real(np.abs(tools.fft2d(img)))
alpha = args.noiselvl / 255.
intensity_noise = alpha * yy * np.random.normal(size=img.shape)
y = (yy**2 + intensity_noise)
y = y * (y > 0)
y = np.sqrt(y)
sigma = np.std(y - yy) * 255.
elif args.noise == 'rician':
yy = tools.fft2d(img)
y = yy + (np.random.normal(size=img.shape) + 1j * np.random.normal(