def registerImagesBetweenThemselves(im1, im2, sliceReference=0):
Is = im1.copy()
Ir = im2.copy()
refImage = Is[sliceReference, :, :]
nbSlices, height, width = Is.shape
for slice in (range(0, nbSlices)):
if slice != sliceReference:
imageToMove = Is[slice, :, :]
shift = getShift(refImage, imageToMove, precision=1000)
if shift[0] > 1:
shift[0] = 0
print('Slice:' + str(slice) + 'Shift:' + str(shift))
offset_image = fourier_shift(np.fft.fftn(imageToMove), shift)
offset_image = np.fft.ifftn(offset_image).real
Is[slice, :, :] = offset_image
refToMove = Ir[slice, :, :]
offset_image = fourier_shift(np.fft.fftn(refToMove), shift)
offset_image = np.fft.ifftn(offset_image).real
Ir[slice, :, :] = offset_image
return Is, Ir
def caculate_TD(d_Ir_xy_roi, I_xy_roi, window, dy, dx):
temp = fourier_shift(np.fft.fftn(d_Ir_xy_roi), (dy, dx)) # shift the reference image of sandpaper
Ir_xy_shift = np.fft.ifftn(temp)
Ir_xy_shift = Ir_xy_shift.real # delta_Ir(x-dx, y-dy)
#Ir_xy_shift = d_Ir_xy_roi
assert np.shape(window) == np.shape(I_xy_roi), " size doesn't match!"
Img_avg_roi = window*I_xy_roi
I_avg_roi = np.sum(window*I_xy_roi)# I_mean
assert np.shape(Ir_xy_shift) == np.shape(window), " size doesn't match!"
T_upper_left =I_avg_roi* np.sum((window*(Ir_xy_shift*Ir_xy_shift)))
T_upper_right1 = np.sum(window*Ir_xy_shift)
T_upper_right2 = np.sum(window*I_xy_roi*Ir_xy_shift)
T_upper_right = T_upper_right1*T_upper_right2
T_bottom_left = np.sum(window*Ir_xy_shift*Ir_xy_shift)
T_bottom_right = (np.sum(window*Ir_xy_shift))**2
T = T_upper_left - T_upper_right
T = T/(T_bottom_left - T_bottom_right)
T = T/Io
#print('calculated T:',T)
D_upper = np.sum((window*I_xy_roi)*Ir_xy_shift)- I_avg_roi*(np.sum(window*Ir_xy_shift))
D_lower = T*(np.sum(window*(Ir_xy_shift*Ir_xy_shift))-(np.sum(window*Ir_xy_shift))**2)
D = D_upper/D_lower
return(T, D)
def test_masked_registration_random_masks_non_equal_sizes():
"""masked_register_translation should be able to register
translations between images that are not the same size even
with random masks."""
# See random number generator for reproducible results
np.random.seed(23)
reference_image = camera()
shift = (-7, 12)
shifted = np.real(
fft.ifft2(fourier_shift(fft.fft2(reference_image), shift)))
# Crop the shifted image
shifted = shifted[64:-64, 64:-64]
# Random masks with 75% of pixels being valid
ref_mask = np.random.choice([True, False],
reference_image.shape,
p=[3 / 4, 1 / 4])
shifted_mask = np.random.choice([True, False],
shifted.shape,
p=[3 / 4, 1 / 4])
measured_shift = masked_register_translation(
reference_image,
shifted,
reference_mask=np.ones_like(ref_mask),
moving_mask=np.ones_like(shifted_mask))
assert_equal(measured_shift, -np.array(shift))
def shift_image(image, shift_yx, interpolation, mode='constant'):
"""
Function to shift an image.
Parameters
----------
image : numpy.ndarray
Input image (2D).
shift_yx : tuple(float, float)
Shift (y, x) to be applied (pix).
interpolation : str
Interpolation type ("spline", "bilinear", or "fft").
Returns
-------
numpy.ndarray
Shifted image.
"""
if interpolation == "spline":
im_center = shift(image, shift_yx, order=5, mode=mode)
elif interpolation == "bilinear":
im_center = shift(image, shift_yx, order=1, mode=mode)
elif interpolation == "fft":
fft_shift = fourier_shift(np.fft.fftn(image), shift_yx)
im_center = np.fft.ifftn(fft_shift).real
return im_center
def subpixel(a, plot=False):
optIntShift, aTemp = closest_pixel(a, plot=plot)
def func(x):
return -(a[::2,...]*np.fft.ifftn(fourier_shift(np.fft.fftn(a[1::2,...]), [0, x[0]])).real).sum() - (a[2::2,...]*np.fft.ifftn(fourier_shift(np.fft.fftn(a[1:-1:2,...]), [0, x[0]])).real).sum()
res = scipy.optimize.minimize(func, x0=[optIntShift], method='Nelder-Mead')
if res['success']:
optimalShift = res['x'][0]
aOut = np.copy(a)
aOut[1::2,...] = np.fft.ifftn(fourier_shift(np.fft.fftn(aOut[1::2,...]), [0, optimalShift])).real
if plot:
upsample_factor=20
shifts = np.arange(optIntShift-1, optIntShift+1, 1./upsample_factor)
corrs = np.zeros(shifts.shape, dtype='float')
for i, shift in enumerate(shifts):
corrs[i] = func([shift])
plt.plot(optimalShift, corrs.min(), 'rx')
plt.text(optimalShift, corrs.min(), 'optimal')
plt.plot(shifts, corrs, 'g.')
plt.xlabel('Horizonal shift')
return optimalShift, aOut
def realign_image(arr, shift, angle=0):
"""
Translate and rotate image via Fourier
Parameters
----------
arr : ndarray
Image array.
shift: float
Mininum and maximum values to rescale data.
angle: float, optional
Mininum and maximum values to rescale data.
Returns
-------
ndarray
Output array.
"""
# if both shifts are integers, do circular shift; otherwise perform Fourier shift.
if np.count_nonzero(np.abs(np.array(shift) - np.round(shift)) < 0.01) == 2:
temp = np.roll(arr, int(shift[0]), axis=0)
temp = np.roll(temp, int(shift[1]), axis=1)
temp = temp.astype('float32')
else:
temp = fourier_shift(np.fft.fftn(arr), shift)
temp = np.fft.ifftn(temp)
temp = np.abs(temp).astype('float32')
return temp
def fourier_imshift(image, xshift, yshift, pad=False):
'''
Shift an image by use of Fourier shift theorem
Parameters:
image : nd array
N x K image
xshift : float
Pixel value by which to shift image in the x direction
yshift : float
Pixel value by which to shift image in the x direction
pad : bool
Should we pad the array before shifting, then truncate?
Otherwise, the image is wrapped.
Returns:
offset : nd array
Shifted image
'''
# Pad ends with zeros
if pad:
padx = np.abs(np.int(xshift)) + 1
pady = np.abs(np.int(yshift)) + 1
pad_vals = ([pady]*2,[padx]*2)
im = np.pad(image,pad_vals,'constant')
else:
padx = 0; pady = 0
im = image
offset = fourier_shift( np.fft.fft2(im), (yshift,xshift) )
offset = np.fft.ifft2(offset).real
offset = offset[pady:pady+image.shape[0], padx:padx+image.shape[1]]
return offset
def shift(img, x, y):
"""
Apply a shift to an image
"""
offset_image = fourier_shift(pyfftw.interfaces.numpy_fft.fftn(img), (x, y))
offset_image = pyfftw.interfaces.numpy_fft.ifftn(offset_image).real
return offset_image
def realign_image(arr, shift, angle=0):
"""
Translate and rotate image via Fourier
Parameters
----------
arr : ndarray
Image array.
shift: float
Mininum and maximum values to rescale data.
angle: float, optional
Mininum and maximum values to rescale data.
Returns
-------
ndarray
Output array.
"""
# if both shifts are integers, do circular shift; otherwise perform Fourier shift.
if np.count_nonzero(np.abs(np.array(shift) - np.round(shift)) < 0.01) == 2:
temp = np.roll(arr, int(shift[0]), axis=0)
temp = np.roll(temp, int(shift[1]), axis=1)
temp = temp.astype('float32')
else:
temp = fourier_shift(np.fft.fftn(arr), shift)
temp = np.fft.ifftn(temp)
temp = np.abs(temp).astype('float32')
return temp
def shift_stack(a, shift):
aOut = np.copy(a)
shiftVector = np.zeros(a.ndim, dtype='float')
shiftVector[1] = shift
aOut[1::2, ...] = np.fft.ifftn(
fourier_shift(np.fft.fftn(aOut[1::2, ...]), shiftVector)).real
return aOut
def setUp(self):
ref_image = ascent()
center = np.array((256, 256))
shifts = np.array(
[
(0.0, 0.0),
(4.3, 2.13),
(1.65, 3.58),
(-2.3, 2.9),
(5.2, -2.1),
(2.7, 2.9),
(5.0, 6.8),
(-9.1, -9.5),
(-9.0, -9.9),
(-6.3, -9.2),
]
)
s = hs.signals.Signal2D(np.zeros((10, 100, 100)))
for i in range(10):
# Apply each sup-pixel shift using FFT and InverseFFT
offset_image = fourier_shift(np.fft.fftn(ref_image), shifts[i])
offset_image = np.fft.ifftn(offset_image).real
# Crop central regions of shifted images to avoid wrap around
s.data[i, ...] = offset_image[center[0] : center[0] + 100, center[1] : center[1] + 100]
self.signal = s
self.shifts = shifts
def register_image(inp):
if 'inRAM_flag' in globals():
inRAM = inRAM_flag
else:
inRAM = True
if 'refIm' in globals():
refIm_ = refIm
if 'crop' not in globals() and 'crop' not in locals():
if inp[1].shape[0]>256:
crop_ = True
refIm_ = inp[0][128:-128,128:-128]
upsample_factor = 10
else:
crop_ = False
refIm_ = inp[0]
upsample_factor = 1
image = inp[1]
else:
crop_ = crop
upsample_factor = 10
if not inRAM:
image_idx = inp[0]
image = inp[1]
regFile = inp[2]
else:
image = inp
#print refIm_.shape,
if crop_==True:
tmpII = image[128:-128,128:-128] - ndimage.gaussian_filter(image[128:-128,128:-128],5)
#shift, _, _ = register_translation(refIm_,image[128:-128,128:-128],upsample_factor=upsample_factor)
shift, _, _ = register_translation(refIm_,tmpII,upsample_factor=upsample_factor)
else:
shift, _, _ = register_translation(refIm_, image, upsample_factor=upsample_factor)
#set this
if np.any(np.abs(shift)>=20):
shift = np.array([0,0])
#print("! WARNING VERY LARGE SHIFTS ! %s" %shift)
if np.sum(np.abs(shift))!=0:
regIm = np.fft.ifftn(fourier_shift(np.fft.fftn(image), shift)).real
else:
regIm = image
#added for reduce memory test
if not inRAM:
regFile[image_idx] = regIm
return shift
else:
return [shift,regIm]
def shift_image(image,
shift_yx,
interpolation,
mode='constant'):
"""
Function to shift an image.
:param images: Input image.
:type images: numpy.ndarray
:param shift_yx: Shift (y, x) to be applied (pixel).
:type shift_yx: (float, float)
:param interpolation: Interpolation type (spline, bilinear, fft)
:type interpolation: str
:return: Shifted image.
:rtype: numpy.ndarray
"""
if interpolation == "spline":
im_center = shift(image, shift_yx, order=5, mode=mode)
elif interpolation == "bilinear":
im_center = shift(image, shift_yx, order=1, mode=mode)
elif interpolation == "fft":
fft_shift = fourier_shift(np.fft.fftn(image), shift_yx)
im_center = np.fft.ifftn(fft_shift).real
return im_center
def shift_im(im, shift):
import numpy as np
from scipy.ndimage import fourier_shift
fim_shift = fourier_shift(np.fft.fftn(im), map(lambda x: -x, shift))
im_shift = np.fft.ifftn(fim_shift)
return im_shift.real
def test_masked_registration_random_masks_non_equal_sizes():
"""masked_register_translation should be able to register
translations between images that are not the same size even
with random masks."""
# See random number generator for reproducible results
np.random.seed(23)
reference_image = camera()
shift = (-7, 12)
shifted = np.real(np.fft.ifft2(fourier_shift(
np.fft.fft2(reference_image), shift)))
# Crop the shifted image
shifted = shifted[64:-64, 64:-64]
# Random masks with 75% of pixels being valid
ref_mask = np.random.choice(
[True, False], reference_image.shape, p=[3 / 4, 1 / 4])
shifted_mask = np.random.choice(
[True, False], shifted.shape, p=[3 / 4, 1 / 4])
measured_shift = masked_register_translation(
reference_image,
shifted,
np.ones_like(ref_mask),
np.ones_like(shifted_mask))
assert_equal(measured_shift, -np.array(shift))
def apply_shifts(im,shift): offset_image = [] for i in range(im.shape[2]): offset_im = fourier_shift(np.fft.fftn(im[:,:,i]), shift) offset_image.append(np.fft.ifftn(offset_im)) offset_image = np.transpose(offset_image,axes=[1,2,0]).astype(np.float32) return offset_image
def phase_correlation_registration(fixed_image,
moving_image,
subpixel=100,
verbose=False,
resample=True):
"""
A simple Phase Correlation based image registration method.
:param verbose: enable print functions
:param subpixel: resampling factor; registration will be perfromed to
1/subpixel accuracy
:param fixed_image: the reference image as MIPLIB Image object
:param moving_image: the moving image as MIPLIB Image object
:return: returns the SimpleITK transform
"""
assert isinstance(fixed_image, Image)
assert isinstance(moving_image, Image)
shift, error, diffphase = register_translation(fixed_image, moving_image,
subpixel)
scaled_shifts = list(
-offset * spacing
for offset, spacing in zip(shift, fixed_image.spacing))
if verbose:
print(("Detected offset (y, x): {}".format(scaled_shifts)))
if resample:
resampled = np.abs(
np.fft.ifftn(fourier_shift(np.fft.fftn(moving_image), shift)).real)
return Image(resampled, fixed_image.spacing)
else:
return scaled_shifts
def setUp(self):
# Starting high-res is a 121x121 image of an Airy disk
hr = np.array(AiryDisk2DKernel(radius=15))
hr = normalize(hr, new_min=0, new_max=1)
# Crop to 120x120 to simplify debugging calculations
hr = hr[0:120, 0:120]
# Arbitrary PSF
psf = np.ones((7, 7)) / 7**2
# Create low-res frames with no motion
camera = ObservationModel(hr,
n=5,
psf=psf,
downsample_factor=10,
translation_range=(0, 0),
rotation_range=(0, 0),
noise_scale=0.001)
self.low_resolution = []
self.translations = []
# Add motion to the lr frames and store the exact amount
for lr in camera.low_res_images:
# There's no significant rotation in the actual data set only translation
tx, ty = np.random.randint(-3, 3, size=(1, 2))[0]
im = fourier_shift(np.fft.fftn(lr), (ty, tx))
im = np.fft.ifftn(im).real
self.low_resolution.append(im)
self.translations.append((tx, ty))
def score(self, ref_im):
"""
Align the reconstruction using phase correlation and calculate the MSE
and SSIM.
Inputs:
- ref_im: A numpy array of shape (N,N) contaning a reference image.
Outputs:
- mse: Mean squared error (float).
- ssim: Structural similarity index (float).
"""
if isinstance(self.res, type(None)):
raise Exception('Result is not yet aviable.')
shift = register_translation(ref_im, self.res)[0]
shifted_res = fourier_shift(np.fft.fft2(self.res), shift)
shifted_res = np.real(np.fft.ifft2(shifted_res))
mse = np.linalg.norm(shifted_res - ref_im)
drange = np.max(shifted_res) - np.min(shifted_res)
ssim = compare_ssim(ref_im, shifted_res, data_range=drange)
return mse, ssim
def setup(self, ndims, image_size, upscale_factor, *args):
if phase_cross_correlation is None:
raise NotImplementedError("phase_cross_correlation unavailable")
shifts = (-2.3, 1.7, 5.4, -3.2)[:ndims]
phantom = img_as_float(
data.binary_blobs(length=image_size, n_dim=ndims))
self.reference_image = np.fft.fftn(phantom)
self.shifted_image = ndi.fourier_shift(self.reference_image, shifts)
def fixImage(image):
for i in range(Sz[3]):
oddLines = image[range(1, Sz[1], 2), :, i]
Shift = (0, medianShift)
shiftedOdd = fourier_shift(np.fft.fftn(oddLines), Shift)
shiftedOdd = np.real(np.fft.ifftn(shiftedOdd))
image[range(1, Sz[1], 2), :, i] = shiftedOdd
return image
def imgsubshift(img,shift):
tmp = img.copy()
fftimg = fft.fft2(tmp)
tmp0 = ndm.fourier_shift(fftimg, shift)
result = fft.ifft2(tmp0).real
return result
def fixPerTimepointFunc(kv):
k, image = kv
shift = (0, shifts[k])
for i in range(Sz[3]):
oddLines = image[range(1, Sz[1], 2), :, i]
shiftedOdd = fourier_shift(np.fft.fftn(oddLines), shift)
shiftedOdd = np.real(np.fft.ifftn(shiftedOdd))
image[range(1, Sz[1], 2), :, i] = shiftedOdd
return image
def test_4d_input_pixel():
phantom = img_as_float(binary_blobs(length=32, n_dim=4))
reference_image = fft.fftn(phantom)
shift = (-2., 1., 5., -3)
shifted_image = fourier_shift(reference_image, shift)
result, error, diffphase = phase_cross_correlation(reference_image,
shifted_image,
space="fourier")
assert_allclose(result, -np.array(shift), atol=0.05)
def test_4d_input_pixel():
phantom = img_as_float(binary_blobs(length=32, n_dim=4))
reference_image = np.fft.fftn(phantom)
shift = (-2., 1., 5., -3)
shifted_image = fourier_shift(reference_image, shift)
result, error, diffphase = register_translation(reference_image,
shifted_image,
space="fourier")
assert_allclose(result, -np.array(shift), atol=0.05)
def apply_shift(res, psi, p, id):
"""Apply shift for one projection."""
# padding to avoid signal wrapping
[nz, n] = psi[id].shape
tmp = np.zeros([nz + nz // 4, n + n // 4], dtype='float32')
tmp[nz // 8:nz + nz // 8, n // 8:n + n // 8] = psi[id]
res0 = np.fft.ifft2(ndimage.fourier_shift(np.fft.fft2(tmp), p))
res[id] = np.real(res0[nz // 8:nz + nz // 8, n // 8:n + n // 8])
return res[id]
def imgsubshift_cupy(img_cupy,shift):
fftimg_cupy = cp.fft.fft2(img_cupy)
fftimg=cp.asnumpy(fftimg_cupy)
tmp0=ndm.fourier_shift(fftimg, shift)
tmp0_cupy=cp.asarray(tmp0)
result_cupy=cp.fft.ifft2(tmp0_cupy).real
return result_cupy
def image_registration(ref_image, off_image, img_to_reg=None, upsample_fac=50):
""" Function to co-register off_image with respect to off_image"""
shift, error, diffphase = phase_cross_correlation(
ref_image, off_image, upsample_factor=upsample_fac)
if img_to_reg is None:
img_to_reg = off_image
reg_image = np.real(
np.fft.ifftn(fourier_shift(np.fft.fftn(img_to_reg), shift)))
return reg_image
def test_correlation():
reference_image = fft.fftn(camera())
shift = (-7, 12)
shifted_image = fourier_shift(reference_image, shift)
# pixel precision
result, error, diffphase = phase_cross_correlation(reference_image,
shifted_image,
space="fourier")
assert_allclose(result[:2], -np.array(shift))
def test_real_input():
reference_image = camera()
subpixel_shift = (-2.4, 1.32)
shifted_image = fourier_shift(np.fft.fftn(reference_image), subpixel_shift)
shifted_image = np.fft.ifftn(shifted_image)
# subpixel precision
result, error, diffphase = register_translation(reference_image,
shifted_image, 100)
assert_allclose(result[:2], -np.array(subpixel_shift), atol=0.05)
def test_subpixel_precision():
reference_image = np.fft.fftn(camera())
subpixel_shift = (-2.4, 1.32)
shifted_image = fourier_shift(reference_image, subpixel_shift)
# subpixel precision
result, error, diffphase = register_translation(reference_image,
shifted_image, 100,
space="fourier")
assert_allclose(result[:2], -np.array(subpixel_shift), atol=0.05)
def test_correlation():
reference_image = np.fft.fftn(camera())
shift = (-7, 12)
shifted_image = fourier_shift(reference_image, shift)
# pixel precision
result, error, diffphase = register_translation(reference_image,
shifted_image,
space="fourier")
assert_allclose(result[:2], -np.array(shift))
def test_4d_input_subpixel():
phantom = img_as_float(binary_blobs(length=32, n_dim=4))
reference_image = fft.fftn(phantom)
subpixel_shift = (-2.3, 1.7, 5.4, -3.2)
shifted_image = fourier_shift(reference_image, subpixel_shift)
result, error, diffphase = register_translation(reference_image,
shifted_image,
10,
space="fourier")
assert_allclose(result, -np.array(subpixel_shift), atol=0.05)
def shift_momentum(self, psik=None, scale=1.0, frac=(0.5, 0.5)):
"""Shifts momentum components of `psi` by a fracion of +/- kL_recoil.
The ground-state solutions of Raman-coupled spinor systems in general
have spinor components with both left- and right-moving momentum
peaks. Providing a manual shift on the momentum-space wavefunction
components better approximates these solutions, i.e. faster convergence
in imaginary time propagation.
Parameters
----------
psik : :obj:`list` of NumPy :obj:`array`, optional.
The momentum-space pseudospinor wavefunction. If `psik` is not
provided, then this function uses the current class attribute
`self.psik`.
scale : :obj:`float`, default=1.0
By default, the function shifts the momentum peaks by a single
unit of recoil momenta `kL_recoil`. `scale` gives the option of
scaling the shift larger or smaller for faster convergence.
frac : :obj:`iterable`, default=(0.5, 0.5)
The fraction of each spinor component's momentum peak to shift in
either direction. frac=(0.5, 0.5) splits into two equal peaks,
while (0.0, 1.0) and (1.0, 0.0) move the entire peak one
direction or the other.
"""
assert self.is_coupling, ("The `is_coupling` option is "
f"{self.is_coupling}. Initialize coupling "
"with `coupling_setup()`.")
if psik is None:
psik = self.psik
shift = scale * self.kL_recoil / self.space['dk'][0]
input_ = ttools.fft_2d(psik, self.space['dr'])
result = [np.zeros_like(pk) for pk in psik]
for i in range(len(psik)):
positive = fourier_shift(input_[i], shift=[0, shift], axis=1)
negative = fourier_shift(input_[i], shift=[0, -shift], axis=1)
result[i] = frac[0] * positive + frac[1] * negative
frac = np.flip(frac)
self.psik = ttools.ifft_2d(result, self.space['dr'])
self.psi = ttools.ifft_2d(self.psik, self.space['dr'])
def test_fourier_shift_real01(self, shape, dtype, dec):
expected = numpy.arange(shape[0] * shape[1], dtype=dtype)
expected.shape = shape
a = fft.rfft(expected, shape[0], 0)
a = fft.fft(a, shape[1], 1)
a = ndimage.fourier_shift(a, [1, 1], shape[0], 0)
a = fft.ifft(a, shape[1], 1)
a = fft.irfft(a, shape[0], 0)
assert_array_almost_equal(a[1:, 1:], expected[:-1, :-1], decimal=dec)
assert_array_almost_equal(a.imag, numpy.zeros(shape), decimal=dec)
def test_size_one_dimension_input():
# take a strip of the input image
reference_image = np.fft.fftn(camera()[:, 15]).reshape((-1, 1))
subpixel_shift = (-2.4, 4)
shifted_image = fourier_shift(reference_image, subpixel_shift)
# subpixel precision
result, error, diffphase = register_translation(reference_image,
shifted_image, 100,
space="fourier")
assert_allclose(result[:2], -np.array((-2.4, 0)), atol=0.05)
def test_3d_input():
phantom = img_as_float(binary_blobs(length=32, n_dim=3))
reference_image = fft.fftn(phantom)
shift = (-2., 1., 5.)
shifted_image = fourier_shift(reference_image, shift)
result, error, diffphase = phase_cross_correlation(reference_image,
shifted_image,
space="fourier")
assert_allclose(result, -np.array(shift), atol=0.05)
# subpixel precision now available for 3-D data
subpixel_shift = (-2.3, 1.7, 5.4)
shifted_image = fourier_shift(reference_image, subpixel_shift)
result, error, diffphase = phase_cross_correlation(reference_image,
shifted_image,
upsample_factor=100,
space="fourier")
assert_allclose(result, -np.array(subpixel_shift), atol=0.05)
def test_3d_input():
phantom = img_as_float(binary_blobs(length=32, n_dim=3))
reference_image = np.fft.fftn(phantom)
shift = (-2., 1., 5.)
shifted_image = fourier_shift(reference_image, shift)
result, error, diffphase = register_translation(reference_image,
shifted_image,
space="fourier")
assert_allclose(result, -np.array(shift), atol=0.05)
# subpixel precision not available for 3-D data
subpixel_shift = (-2.3, 1., 5.)
shifted_image = fourier_shift(reference_image, subpixel_shift)
result, error, diffphase = register_translation(reference_image,
shifted_image,
space="fourier")
assert_allclose(result, -np.array(shift), atol=0.5)
assert_raises(NotImplementedError, register_translation, reference_image,
shifted_image, upsample_factor=100,
space="fourier")
def Register_Image(image,refIm,crop=False):
if crop==True:
shift, _, _ = register_translation(refIm,image[128:-128,128:-128],upsample_factor=10)
else:
shift, _, _ = register_translation(refIm, image, upsample_factor=10)
if np.sum(np.abs(shift))!=0:
regIm = np.fft.ifftn(fourier_shift(np.fft.fftn(image), shift)).real
else:
regIm = image
return [shift,regIm]
def test_masked_registration_vs_register_translation():
"""masked_register_translation should give the same results as
register_translation in the case of trivial masks."""
reference_image = camera()
shift = (-7, 12)
shifted = np.real(np.fft.ifft2(fourier_shift(
np.fft.fft2(reference_image), shift)))
trivial_mask = np.ones_like(reference_image)
nonmasked_result, *_ = register_translation(reference_image, shifted)
masked_result = masked_register_translation(
reference_image, shifted, trivial_mask, overlap_ratio=1 / 10)
assert_equal(nonmasked_result, masked_result)
def fourier_imshift(image, shift):
''' Shift an image by use of Fourier shift theorem
Parameters
----------
image : numpy.ndarray
A 2D (``NxK``) or 3D (``LxNxK``) image.
shift : numpy.ndarray
A 1D or 2D array of shape ``Lx2`` containing pixel values by which
to shift image slices in the X and Y directions.
Returns
-------
offset : numpy.ndarray
Shifted image
'''
ndim = len(image.shape)
if ndim == 2:
shift = np.asanyarray(shift)[:2]
offset_image = fourier_shift(
np.fft.fftn(image),
shift[::-1]
)
offset = np.fft.ifftn(offset_image).real
elif ndim == 3:
nslices = image.shape[0]
shift = np.asanyarray(shift)[:, :2]
if (shift.shape[0] != nslices):
raise ValueError("The number of provided shifts must be equal "
"to the number of slices in the input image.")
offset = np.empty_like(image, dtype=np.float)
for k in range(nslices):
offset[k] = fourier_imshift(image[k], shift[k])
else:
raise ValueError("Input image must be either a 2D or a 3D array.")
return offset
def fourier_imshift(image,xshift,yshift):
''' Shift an image by use of Fourier shift theorem
Parameters:
image : nd array
N x K image
xshift : float
Pixel value by which to shift image in the x direction
yshift : float
Pixel value by which to shift image in the y direction
Returns:
offset : nd array
Shifted image
'''
offset = fourier_shift( np.fft.fftn(image), (-yshift,xshift) )
offset = np.fft.ifftn(offset).real
return offset
def subpixel(a, plot=False):
optIntShift, aTemp = closest_pixel(a, plot=plot)
def func(x):
return -(a[::2,...]*np.fft.ifftn(fourier_shift(np.fft.fftn(a[1::2,...]), [0, x[0]])).real).sum() - (a[2::2,...]*np.fft.ifftn(fourier_shift(np.fft.fftn(a[1:-1:2,...]), [0, x[0]])).real).sum()
res = scipy.optimize.minimize(func, x0=[optIntShift], method='Nelder-Mead')
if res['success']:
optimalShift = res['x'][0]
aOut = np.copy(a)
aOut[1::2,...] = np.fft.ifftn(fourier_shift(np.fft.fftn(aOut[1::2,...]), [0, optimalShift])).real
if plot:
upsample_factor=20
shifts = np.arange(optIntShift-1, optIntShift+1, 1./upsample_factor)
corrs = np.zeros(shifts.shape, dtype='float')
for i, shift in enumerate(shifts):
corrs[i] = func([shift])
plt.plot(optimalShift, corrs.max(), 'rx')
plt.plot(shifts, corrs, 'g.')
return optimalShift, aOut
def setup(self, ndims, image_size, upscale_factor, *args):
shifts = (-2.3, 1.7, 5.4, -3.2)[:ndims]
phantom = img_as_float(binary_blobs(length=image_size, n_dim=ndims))
self.reference_image = np.fft.fftn(phantom)
self.shifted_image = ndi.fourier_shift(self.reference_image, shifts)
def register_dayData(HDF_File,session_ID,inRAM=True,poolSize=4,abs_loc='foo',common_ref=False,show_ref_mean=1):
print('srm',show_ref_mean,type(show_ref_mean))
print(abs_loc)
hdfPath = HDF_File.filename
hdfDir = os.path.split(hdfPath)[0]
global inRAM_flag
inRAM_flag=inRAM
if 'registered_data' not in HDF_File[session_ID].keys():
HDF_File[session_ID].create_group('registered_data')
HDF_File.close()
HDF_File = h5py.File(hdfPath,'a',libver='latest')
print('Creating Registered Data Group')
regData_dir = os.path.join(hdfDir,'regInfo')
if not os.path.exists(regData_dir):
os.mkdir(regData_dir)
for f_idx, file_key in enumerate(HDF_File[session_ID]['raw_data'].keys()):
if f_idx >0:
HDF_File = h5py.File(hdfPath,'a',libver='latest')
if file_key in HDF_File[session_ID]['registered_data'].keys():
del HDF_File[session_ID]['registered_data'][file_key]
print('deleting old version')
st = time.time()
try:
#raw_file = np.array(HDF_File[session_ID]['raw_data'][file_key])
raw_file = HDF_File[session_ID]['raw_data'][file_key]
frames = raw_file.shape[0]
chunkSize = np.max(np.array([x for x in range(1, 11) if frames%x == 0]))
#corrs_ = []
#print frames
#nChunks = np.floor(frames/50)
#raw_file[:50*nChunks].reshape(-50,nChunks)
#for j in raw_file[:frames]:
# corrs_.append(np. corrcoef(i.flatten(),j.flatten())[0,1])
#ord_ = np.argsort(corrs_)
if common_ref and f_idx==0:
"""print 'creating global reference'
#refIm_glob = np.mean(raw_file[ord_[-200:]],axis=0)[:,128:-128]
import Register_Image
############# HERE TRY TO FIND A REFERENCE IMAGE ###############
refss = []
for ix_,as_ in enumerate(HDF_File[session_ID]['raw_data'].keys()):
a_ = np.mean(HDF_File[session_ID]['raw_data'][as_][-500:],axis=0)
refss.append(a_)
plt.figure(ix_)
plt.imshow(a_,interpolation='None',cmap='gray')
plt.title(ix_)
plt.show(block=False)
sel_ref = int(raw_input("selected mean: "))"""
###########
refss = []
im_grads = []
hdf_keys = sorted(HDF_File[session_ID]['raw_data'].keys())
st = time.time()
kix_ = hdf_keys[-2]
temp = np.array(HDF_File[session_ID]['raw_data'][kix_])
print(kix_)
print ('load time:', np.round(time.time() - st,decimals=1))
ixs_sets = []
for ii in range(200):
#kix = np.random.randint(len(hdf_keys))
ixs = np.random.permutation(np.arange(HDF_File[session_ID]['raw_data'][kix_].shape[0]))[:500]
ixs_sets.append(ixs)
a_ = np.mean(temp[np.array(sorted(ixs)),:,:],axis=0)
refss.append(a_)
im_grads.append(np.sum(np.abs(np.gradient(refss[-1]))))
#print time.time() - st
ord_grads = np.argsort(np.array(im_grads))[-10:]
sortImLst = []
print(len(ixs_sets[ord_grads[-1]]))
for iikk,imIx in enumerate(ixs_sets[ord_grads[-1]]):
#print(iikk,)
if iikk==0:
REF__ = refss[ord_grads[-1]]
else:
shift, _, _ = register_translation(REF__[128:-128,128:-128],temp[imIx][128:-128,128:-128],upsample_factor=2)
if np.any(np.abs(shift)>20):
shift = np.array([0,0])
if np.sum(np.abs(shift))!=0:
RegIm = np.fft.ifftn(fourier_shift(np.fft.fftn(temp[imIx]), shift)).real
else:
RegIm = temp[imIx]
sortImLst.append(RegIm)
#REF__ = REF__/((1-1/(iikk+1))) + RegIm/float(iikk+1)
#for ux in ord_grads:
# plt.figure()
# plt.title(ux)
# plt.imshow(refss[ux],interpolation='None',cmap='gray')
# plt.show(block=False)
#sel_ref = int(raw_input("selected mean: "))
#refIm_glob = refss[ord_grads[-1]]
refIm_glob = np.mean(np.array(sortImLst),axis=0)
if show_ref_mean:
plt.figure()
plt.title("NEW")
plt.imshow(refIm_glob,cmap='binary_r')
plt.figure()
plt.title("OLD")
plt.imshow(refss[ord_grads[-1]],cmap='binary_r')
plt.show(block=1)
#np.mean(raw_file[:5000],axis=0)#raw_file[ord_[-100]].astype('float')
#for i in ord_[-99:]:
#image_idx = inp[0]
#image = inp[1]
#regFile = inp[2]
#out=Register_Image.Register_Image(raw_file[i],refIm_glob)
#refIm_glob += out[1].astype('float')
#refIm_glob /= 100.
#plt.imshow(refIm_glob,cmap='gray')
#plt.show()
refIm_glob = refIm_glob[128:-128,128:-128]
refIm_glob = refIm_glob - ndimage.gaussian_filter(refIm_glob,5)
elif not common_ref:
refIm_glob = None
#print 'file open %ss' %(time.time() - st)
st = time.time()
print '\n registering %s \n' %file_key
if not inRAM:
regFile = HDF_File[session_ID]['registered_data'].create_dataset(name=file_key,
shape=raw_file.shape,
chunks=(chunkSize,512,512),
dtype='uint16')
st = time.time()
shifts, tot_shifts = motion_register(raw_file,
regFile,
maxIter=1,
Crop=True,
inRAM=inRAM,
common_ref=common_ref,
refIm_glob=refIm_glob)
regFile.attrs['mean_image'] = np.mean(regFile,axis=0)
else:
#________________________________________________________________
regFile = None
regIms, shifts, tot_shifts = motion_register(raw_file,
regFile,
maxIter=2,
Crop=True,
inRAM=inRAM,
poolSize=16,
common_ref=common_ref,
refIm_glob=None)
print 'Motion Register Duration %ss' %(time.time() - st)
#________________________________________________________________
st = time.time()
regFile = HDF_File[session_ID]['registered_data'].create_dataset(name=file_key,
data=np.round(regIms).astype('uint16'),
chunks=(chunkSize,512,512),
dtype='uint16')
regFile.attrs['mean_image'] = np.mean(regIms.astype('uint16'),axis=0)
regFile.attrs['regInfo'] = regData_dir
try:
regFile.attrs['GRABinfo'] = raw_file.attrs['GRABinfo']
except KeyError:
print "Warning could not find GRABinfo file"
if 'stimattrs' in raw_file.attrs.keys():
regFile.attrs['stimattrs'] = raw_file.attrs['stimattrs']
if common_ref:
regFile.attrs['common_ref'] = True
else:
regFile.attrs['common_ref'] = False
build_registration_log(regFile,abs_loc=regData_dir,tot_shifts=tot_shifts)
HDF_File.close()
print 'file write in ram %ss' %(time.time() - st)
HDF_File = h5py.File(hdfPath,'a',libver='latest')
except IOError:
print '!!!!!!!!!!!!!!!!!!! WARNING !!!!!!!!!!!!!!!! \n %s could not be loaded. Skipping \n !!!!!!!!!!!!!!!!!!! WARNING !!!!!!!!!!!!!!!!' %file_key
HDF_File.close()
#________________________________________________________________
st = time.time()
print 'Write to Disk time %ss:' %(time.time() - st)
HDF_File = h5py.File(hdfPath,'a',libver='latest')
return HDF_File
ystart = int(y*(nr-kerLen))
arr[xstart:xstart+kerLen, ystart:ystart+kerLen] += ker
return arr
##Define or import an image.
image = make_img(num_obj=8, obj_rad=8, nr=512, nc=512)
plt.imshow(image)
## The shift corresponds to the pixel offset relative to the reference image.
shift = (-22.4, 13.32)
## fourier_shift(input, shift) = Multi-dimensional fourier shift filter.
#The array is multiplied with the fourier transform of a shift operation.
#offset_image is a frequency domain shift of reference image.
offset_image = fourier_shift(np.fft.fftn(image), shift) # Compute the N-dimensional DFT of an 'image'.
offset_image = np.fft.ifftn(offset_image) #Computes the N-dimensional inverse discrete Fourier Transform.
print("Known offset (y, x): {}".format(shift))
## pixel precision first
shift, error, diffphase = register_translation(image, offset_image) #Efficient subpixel image
#translation registration by cross-correlation.
fig = plt.figure(figsize=(8, 3))
ax1 = plt.subplot(1, 3, 1, adjustable='box-forced')
ax2 = plt.subplot(1, 3, 2, sharex=ax1, sharey=ax1, adjustable='box-forced')
ax3 = plt.subplot(1, 3, 3)
ax1.imshow(image, cmap='gray')
ax1.set_axis_off()
ax1.set_title('Reference image')
def func(x):
return -(a[::2,...]*np.fft.ifftn(fourier_shift(np.fft.fftn(a[1::2,...]), [0, x[0]])).real).sum() - (a[2::2,...]*np.fft.ifftn(fourier_shift(np.fft.fftn(a[1:-1:2,...]), [0, x[0]])).real).sum()
def frame_shift(array, shift_y, shift_x, imlib='opencv',
interpolation='lanczos4', border_mode='reflect'):
""" Shifts a 2D array by shift_y, shift_x. Boundaries are filled with zeros.
Parameters
----------
array : array_like
Input 2d array.
shift_y, shift_x: float
Shifts in y and x directions.
imlib : {'opencv', 'ndimage-fourier', 'ndimage-interp'}, string optional
Library or method used for performing the image shift.
'ndimage-fourier', does a fourier shift operation and preserves better
the pixel values (therefore the flux and photometry). Interpolation
based shift ('opencv' and 'ndimage-interp') is faster than the fourier
shift. 'opencv' is recommended when speed is critical.
interpolation : str, optional
Only used in case of imlib is set to 'opencv' or 'ndimage-interp'
(Scipy.ndimage), where the images are shifted via interpolation.
For Scipy.ndimage the options are: 'nearneig', bilinear', 'bicuadratic',
'bicubic', 'biquartic' or 'biquintic'. The 'nearneig' interpolation is
the fastest and the 'biquintic' the slowest. The 'nearneig' is the
poorer option for interpolation of noisy astronomical images.
For Opencv the options are: 'nearneig', 'bilinear', 'bicubic' or
'lanczos4'. The 'nearneig' interpolation is the fastest and the
'lanczos4' the slowest and accurate. 'lanczos4' is the default for
Opencv and 'biquartic' for Scipy.ndimage.
border_mode : {'reflect', 'nearest', 'constant', 'mirror', 'wrap'}
Points outside the boundaries of the input are filled accordingly.
With 'reflect', the input is extended by reflecting about the edge of
the last pixel. With 'nearest', the input is extended by replicating the
last pixel. With 'constant', the input is extended by filling all values
beyond the edge with zeros. With 'mirror', the input is extended by
reflecting about the center of the last pixel. With 'wrap', the input is
extended by wrapping around to the opposite edge. Default is 'reflect'.
Returns
-------
array_shifted : array_like
Shifted 2d array.
"""
check_array(array, dim=2)
image = array.copy()
if imlib == 'ndimage-fourier':
shift_val = (shift_y, shift_x)
array_shifted = fourier_shift(np.fft.fftn(image), shift_val)
array_shifted = np.fft.ifftn(array_shifted)
array_shifted = array_shifted.real
elif imlib == 'ndimage-interp':
if interpolation == 'nearneig':
order = 0
elif interpolation == 'bilinear':
order = 1
elif interpolation == 'bicuadratic':
order = 2
elif interpolation == 'bicubic':
order = 3
elif interpolation == 'biquartic' or interpolation == 'lanczos4':
order = 4
elif interpolation == 'biquintic':
order = 5
else:
raise ValueError('Scipy.ndimage interpolation method not '
'recognized')
if border_mode not in ['reflect', 'nearest', 'constant', 'mirror',
'wrap']:
raise ValueError('`border_mode` not recognized')
array_shifted = shift(image, (shift_y, shift_x), order=order,
mode=border_mode)
elif imlib == 'opencv':
if no_opencv:
msg = 'Opencv python bindings cannot be imported. Install opencv or'
msg += ' set imlib to ndimage-fourier or ndimage-interp'
raise RuntimeError(msg)
if interpolation == 'bilinear':
intp = cv2.INTER_LINEAR
elif interpolation == 'bicubic':
intp = cv2.INTER_CUBIC
elif interpolation == 'nearneig':
intp = cv2.INTER_NEAREST
elif interpolation == 'lanczos4':
intp = cv2.INTER_LANCZOS4
else:
raise ValueError('Opencv interpolation method not recognized')
if border_mode == 'mirror':
bormo = cv2.BORDER_REFLECT_101 # gfedcb|abcdefgh|gfedcba
elif border_mode == 'reflect':
bormo = cv2.BORDER_REFLECT # fedcba|abcdefgh|hgfedcb
elif border_mode == 'wrap':
bormo = cv2.BORDER_WRAP # cdefgh|abcdefgh|abcdefg
elif border_mode == 'constant':
bormo = cv2.BORDER_CONSTANT # iiiiii|abcdefgh|iiiiiii
elif border_mode == 'nearest':
bormo = cv2.BORDER_REPLICATE # aaaaaa|abcdefgh|hhhhhhh
else:
raise ValueError('`border_mode` not recognized')
image = np.float32(image)
y, x = image.shape
M = np.float32([[1, 0, shift_x], [0, 1, shift_y]])
array_shifted = cv2.warpAffine(image, M, (x, y), flags=intp,
borderMode=bormo)
else:
raise ValueError('Image transformation library not recognized')
return array_shifted
def func(dframe):
frame1,frame2 = dframe[0], dframe[1]
shift,error,diffphase = register_translation(frame1, frame2, 10)
tframe = fourier_shift(np.fft.fftn(frame2), shift)
tframe = np.fft.ifftn(tframe)
return tframe.real
def patch_register_roi_locs(meanIm1,meanIm2,roiinfo1,im_path,l=32):
""" Function that takes as input two mean images and
roi coordinates on one of the images and returns
the ROI coordinates on a second image
Arguments:
============================
meanIm1: np.array
mean image with known roi coordinates
meanIm2: np.array
mean image to map rois onto
roiinfo1: dict
contains field idxs which is a list of arrays of
roi coordinates on meanIm1
l: int
size of grid on which to motion register patches
"""
xsz,ysz = meanIm1.shape
mask_num = np.zeros([512,512])
for n_,idxp in enumerate(roiinfo1['idxs']):
#if n_ in table[table.columns[1]].tolist():
mask_num[idxp[1],idxp[0]] = n_
nPatch = xsz/float(l)
#First register the whole images in rigid fashion to one another
out = MP.image_registration.Register_Image(meanIm1,meanIm2[128:-128,128:-128],crop=1)
shift_all = out[0]
regIm2 = np.fft.ifftn(fourier_shift(np.fft.fftn(meanIm1), shift_all)).real
#plt.figure(figsize=(12,12))
#plt.imshow(regIm2,cmap='binary_r',interpolation='None')
#pth = os.path.join(im_path,'reg_img'+str(len(os.listdir(im_path)))+'.jpg')
#plt.savefig(pth,dpi=100)
#plt.clf()
rois1 = []
im_pairs = []
roi_pairs = []
roi_locs = cp.deepcopy(roiinfo1['idxs'])
moved_rois = []
shifts = []
for i in range(int(nPatch)):
sty = i*l
for j in range(int(nPatch)):
stx = j*l
patch = meanIm2[stx:stx+l,sty:sty+l]
bigpatch = patch#np.pad(patch,maxShift,mode='median')
shift,_,_ = register_translation(bigpatch,
regIm2[stx:stx+l,sty:sty+l]
,upsample_factor=5.)
#print shift
if np.any(shift>30):
shift = np.array([0,0])
rois_patch = np.unique(mask_num[stx:stx+l,sty:sty+l]).tolist()
for roi in rois_patch:
roi = int(roi)
if roi not in moved_rois:
roi_locs[roi][0] = roi_locs[roi][0] + 1*int(shift[1]) + 1*int(shift_all[1])
roi_locs[roi][1] = roi_locs[roi][1] + 1*int(shift[0]) + 1*int(shift_all[0])
moved_rois = moved_rois + rois_patch
return roi_locs,shift_all
"Efficient subpixel image registration algorithms," Optics Letters 33,
156-158 (2008). :DOI:`10.1364/OL.33.000156`
"""
import numpy as np
import matplotlib.pyplot as plt
from skimage import data
from skimage.feature import register_translation
from skimage.feature.register_translation import _upsampled_dft
from scipy.ndimage import fourier_shift
image = data.camera()
shift = (-22.4, 13.32)
# The shift corresponds to the pixel offset relative to the reference image
offset_image = fourier_shift(np.fft.fftn(image), shift)
offset_image = np.fft.ifftn(offset_image)
print("Known offset (y, x): {}".format(shift))
# pixel precision first
shift, error, diffphase = register_translation(image, offset_image)
fig = plt.figure(figsize=(8, 3))
ax1 = plt.subplot(1, 3, 1)
ax2 = plt.subplot(1, 3, 2, sharex=ax1, sharey=ax1)
ax3 = plt.subplot(1, 3, 3)
ax1.imshow(image, cmap='gray')
ax1.set_axis_off()
ax1.set_title('Reference image')
def shift_stack(a, shift):
aOut = np.copy(a)
shiftVector = np.zeros(a.ndim, dtype='float')
shiftVector[1] = shift
aOut[1::2,...] = np.fft.ifftn(fourier_shift(np.fft.fftn(aOut[1::2,...]), shiftVector)).real
return aOut
