def test_calculate_affine_matrices(angle, scale, xoffset, yoffset):
m = ((models.Scale(scale) & models.Scale(scale)) | models.Rotation2D(angle)
| (models.Shift(xoffset) & models.Shift(yoffset)))
affine = adwcs.calculate_affine_matrices(m, (100, 100))
assert_allclose(affine.offset, (yoffset, xoffset), atol=1e-10)
angle = math.radians(angle)
assert_allclose(affine.matrix,
((scale * math.cos(angle), scale * math.sin(angle)),
(-scale * math.sin(angle), scale * math.cos(angle))),
atol=1e-10)
def align_images_from_wcs(adinput,
adref,
cull_sources=False,
transform=None,
min_sources=1,
search_radius=10,
match_radius=2,
rotate=False,
scale=False,
full_wcs=False,
brute=True,
return_matches=False):
"""
This function takes two images (an input image, and a reference image) and
works out the modifications needed to the WCS of the input images so that
the world coordinates of its OBJCAT sources match the world coordinates of
the OBJCAT sources in the reference image. This is done by modifying the
WCS of the input image and mapping the reference image sources to pixels
in the input image via the reference image WCS (fixed) and the input image
WCS. As such, in the nomenclature of the fitting routines, the pixel
positions of the input image's OBJCAT become the "reference" sources,
while the converted positions of the reference image's OBJCAT are the
"input" sources.
Parameters
----------
adinput: AstroData
input AD whose pixel shift is requested
adref: AstroData
reference AD image
transform: Transform/None
existing transformation (if None, will do brute search)
cull_sources: bool
limit matched sources to "good" (i.e., stellar) objects
min_sources: int
minimum number of sources to use for cross-correlation
search_radius: float
size of search box (in pixels)
match_radius: float
matching radius for objects (in pixels)
rotate: bool
add a rotation to the alignment transform?
scale: bool
add a magnification to the alignment transform?
full_wcs: bool
use the two images' WCSs to reproject the reference image's coordinates
onto the input image's pixel plane, rather than just align the OBJCAT
coordinates?
brute: bool
perform brute (landscape) search first?
return_matches: bool
return a list of matched objects as well as the Transform?
Returns
-------
matches: 2 lists
OBJCAT sources in input and reference that are matched
WCS: new WCS for input image
"""
log = logutils.get_logger(__name__)
if len(adinput) * len(adref) != 1:
log.warning('Can only match single-extension images')
return None
try:
input_objcat = adinput[0].OBJCAT
ref_objcat = adref[0].OBJCAT
except AttributeError:
log.warning('Both input images must have object catalogs')
return None
if len(input_objcat) < min_sources or len(ref_objcat) < min_sources:
log.warning("Too few sources in one or both images. Cannot align.")
return None
largest_dimension = max(*adinput[0].shape, *adref[0].shape)
# Values larger than these result in errors of >1 pixel
mag_threshold = 1. / largest_dimension
rot_threshold = np.degrees(mag_threshold)
# OK, we can proceed
incoords = (input_objcat['X_IMAGE'].data - 1,
input_objcat['Y_IMAGE'].data - 1)
refcoords = (ref_objcat['X_IMAGE'].data - 1,
ref_objcat['Y_IMAGE'].data - 1)
if cull_sources:
good_src1 = gt.clip_sources(adinput)[0]
good_src2 = gt.clip_sources(adref)[0]
if len(good_src1) < min_sources or len(good_src2) < min_sources:
log.warning("Too few sources in culled list, using full set "
"of sources")
else:
incoords = (good_src1["x"] - 1, good_src1["y"] - 1)
refcoords = (good_src2["x"] - 1, good_src2["y"] - 1)
# Set up the initial model
magnification, rotation = 1, 0 # May be overridden later
try:
t = adref[0].wcs.forward_transform | adinput[0].wcs.backward_transform
except AttributeError: # for cases with no defined WCS
t = None
if full_wcs:
log.warning(
"Cannot determine WCS information: setting full_wcs=False")
full_wcs = False
if full_wcs:
refcoords = t(*refcoords)
elif transform is None and t is not None:
transform = t.inverse
if transform is None:
transform = models.Identity(2)
# We always refactor the transform (if provided) in a prescribed way so
# as to ensure it's fittable and not overly weird
affine = adwcs.calculate_affine_matrices(transform, adinput[0].shape)
m_init = models.Shift(affine.offset[1]) & models.Shift(affine.offset[0])
# This is approximate since the affine matrix might have differential
# scaling and a shear
magnification = np.sqrt(abs(np.linalg.det(affine.matrix)))
rotation = np.degrees(
np.arctan2(affine.matrix[1, 0] - affine.matrix[0, 1],
affine.matrix[0, 0] + affine.matrix[1, 1]))
m_init.offset_0.bounds = (m_init.offset_0 - search_radius,
m_init.offset_0 + search_radius)
m_init.offset_1.bounds = (m_init.offset_1 - search_radius,
m_init.offset_1 + search_radius)
m_rotate = Rotate2D(rotation)
if rotate:
m_rotate.angle.bounds = (rotation - 5, rotation + 5)
m_init = m_rotate | m_init
elif abs(rotation) > rot_threshold:
m_rotate.angle.fixed = True
m_init = m_rotate | m_init
log.warning("A rotation of {:.3f} degrees is expected but the "
"rotation is fixed".format(rotation))
m_magnify = Scale2D(magnification)
if scale:
m_magnify.factor.bounds = (magnification - 0.05, magnification + 0.05)
m_init = m_magnify | m_init
elif abs(magnification - 1) > mag_threshold:
m_magnify.factor.fixed = True
m_init = m_magnify | m_init
log.warning("A magnification of {:.4f} is expected but the "
"magnification is fixed".format(magnification))
# Perform the fit
m_final = fit_model(m_init, incoords, refcoords, sigma=10, brute=brute)
if return_matches:
matched = match_sources(m_final(*incoords),
refcoords,
radius=match_radius)
ind2 = np.where(matched >= 0)
ind1 = matched[ind2]
obj_list = [[], []] if len(ind1) < 1 else [
np.array(list(zip(*incoords)))[ind2],
np.array(list(zip(*refcoords)))[ind1]
]
return obj_list, m_final
return m_final
def create_polynomial_transform(transform,
in_coords,
ref_coords,
order=3,
max_iters=5,
match_radius=0.1,
clip=True,
log=None):
"""
This function maps a set of 2D input coordinates to a set of 2D reference
coordiantes using a pair of Polynomial2D object (one for each ordinate),
given an initial transforming model.
Parameters
----------
transform : astropy.models.Model
the initial guess of the transform between coordinate frames
in_coords : 2-tuple of sequences
input coordinates being mapped to reference
ref_coords : 2-tuple of sequences
reference coordinates
order : int
order of polynomial fit in each ordinate
max_iters : int
maximum number of iterations to perform
match_radius : float
matching radius for sources (in units of the reference coords)
clip : bool
sigma-clip sources after matching?
log : logging object
Returns
-------
transform : Model
a model (and its inverse) to map in_coords to ref_coords
matched: ndarray
matched incoord for each refcoord
"""
affine = adwcs.calculate_affine_matrices(transform, shape=(100, 100))
num_params = [
len(models.Polynomial2D(degree=i).parameters) for i in range(order + 1)
]
orig_order = last_order = order
xref, yref = ref_coords
xin, yin = in_coords
if clip:
fit_it = fitting.FittingWithOutlierRemoval(fitting.LinearLSQFitter(),
sigma_clip,
sigma=3)
else:
fit_it = fitting.LinearLSQFitter()
matched = match_sources((xref, yref),
transform(xin, yin),
radius=match_radius)
num_matched = np.sum(matched >= 0)
niter = 0
while True:
# No point trying to compute a more complex model if it will
# be insufficiently constrained
order = min(np.searchsorted(num_params, num_matched, side="right"),
orig_order)
if order < last_order:
log.warning(f"Downgrading fit to order {order} due to "
"limited number of matches.")
elif order > last_order:
log.stdinfo(f"Upgrading fit to order {order} due to "
"increased number of matches.")
xmodel = models.Polynomial2D(degree=order,
c0_0=affine.offset[1],
c1_0=affine.matrix[1, 1],
c0_1=affine.matrix[1, 0])
ymodel = models.Polynomial2D(degree=order,
c0_0=affine.offset[0],
c1_0=affine.matrix[0, 1],
c0_1=affine.matrix[0, 0])
old_num_matched = num_matched
xobj_matched, yobj_matched = [], []
xref_matched, yref_matched = [], []
for i, m in enumerate(matched):
if m >= 0:
xref_matched.append(xref[i])
yref_matched.append(yref[i])
xobj_matched.append(xin[m])
yobj_matched.append(yin[m])
xmodel = fit_it(xmodel, np.array(xobj_matched), np.array(yobj_matched),
xref_matched)
ymodel = fit_it(ymodel, np.array(xobj_matched), np.array(yobj_matched),
yref_matched)
if clip:
xmodel, ymodel = xmodel[0], ymodel[0]
transform = models.Mapping((0, 1, 0, 1)) | (xmodel & ymodel)
matched = match_sources((xref, yref),
transform(xin, yin),
radius=match_radius)
num_matched = np.sum(matched >= 0)
last_order = order
niter += 1
log.debug(f"Iteration {niter}: Matched {num_matched} objects")
if num_matched == old_num_matched or niter > max_iters:
break
xmodel_inv = fit_it(xmodel, np.array(xref_matched), np.array(yref_matched),
xobj_matched)
ymodel_inv = fit_it(ymodel, np.array(xref_matched), np.array(yref_matched),
yobj_matched)
if clip:
xmodel_inv, ymodel_inv = xmodel_inv[0], ymodel_inv[0]
transform.inverse = models.Mapping(
(0, 1, 0, 1)) | (xmodel_inv & ymodel_inv)
return transform, matched
def find_alignment_transform(incoords,
refcoords,
transform=None,
shape=None,
search_radius=10,
match_radius=2,
rotate=False,
scale=False,
brute=True,
sigma=5,
factor=None,
return_matches=False):
"""
This function computes a transform that maps one set of coordinates to
another. By default, only a shift is used, by a rotation and magnification
can also be applied if requested. An initial transform may be supplied
and, if so, its affine approximation will be used as a starting point.
Parameters
----------
incoords: tuple
x-coords and y-coords of objects in input image
refcoords: tuple
x-coords and y-coords of objects in reference image
transform: Transform/None
existing transformation (if None, will do brute search)
shape: 2-tuple/None
shape (standard python order, y-first)
search_radius: float
size of search box (in pixels)
match_radius: float
matching radius for objects (in pixels)
rotate: bool
add a rotation to the alignment transform?
scale: bool
add a magnification to the alignment transform?
brute: bool
perform brute (landscape) search first?
sigma: float
scale-length for source matching
factor: float/None
scaling factor to convert coordinates to pixels in the BruteLandscapeFitter()
return_matches: bool
return a list of matched objects as well as the Transform?
Returns
-------
Model: alignment transform
matches: 2 lists (optional)
OBJCAT sources in input and reference that are matched
"""
log = logutils.get_logger(__name__)
if shape is None:
shape = tuple(max(c) - min(c) for c in incoords)
largest_dimension = max(*shape)
# Values larger than these result in errors of >1 pixel
mag_threshold = 1. / largest_dimension
rot_threshold = np.degrees(mag_threshold)
# Set up the initial model
if transform is None:
transform = models.Identity(2)
# We always refactor the transform (if provided) in a prescribed way so
# as to ensure it's fittable and not overly weird
affine = adwcs.calculate_affine_matrices(transform, shape)
m_init = models.Shift(affine.offset[1]) & models.Shift(affine.offset[0])
# This is approximate since the affine matrix might have differential
# scaling and a shear
magnification = np.sqrt(abs(np.linalg.det(affine.matrix)))
rotation = np.degrees(
np.arctan2(affine.matrix[0, 1] - affine.matrix[1, 0],
affine.matrix[0, 0] + affine.matrix[1, 1]))
m_init.offset_0.bounds = (m_init.offset_0 - search_radius,
m_init.offset_0 + search_radius)
m_init.offset_1.bounds = (m_init.offset_1 - search_radius,
m_init.offset_1 + search_radius)
m_rotate = Rotate2D(rotation)
if rotate:
m_rotate.angle.bounds = (rotation - 5, rotation + 5)
m_init = m_rotate | m_init
elif abs(rotation) > rot_threshold:
m_rotate.angle.fixed = True
m_init = m_rotate | m_init
log.warning("A rotation of {:.3f} degrees is expected but the "
"rotation is fixed".format(rotation))
m_magnify = Scale2D(magnification)
if scale:
m_magnify.factor.bounds = (magnification - 0.05, magnification + 0.05)
m_init = m_magnify | m_init
elif abs(magnification - 1) > mag_threshold:
m_magnify.factor.fixed = True
m_init = m_magnify | m_init
log.warning("A magnification of {:.4f} is expected but the "
"magnification is fixed".format(magnification))
# Tolerance here aims to achieve <0.1 pixel differences in the tests
m_final = fit_model(m_init,
incoords,
refcoords,
sigma=sigma,
scale=factor,
brute=brute,
tolerance=sigma * 1e-5)
if return_matches:
matched = match_sources(m_final(*incoords),
refcoords,
radius=match_radius)
ind2 = np.where(matched >= 0)
ind1 = matched[ind2]
obj_list = [[], []] if len(ind1) < 1 else [
np.array(list(zip(*incoords)))[ind2],
np.array(list(zip(*refcoords)))[ind1]
]
return m_final, obj_list
return m_final
def resampleToCommonFrame(self, adinputs=None, **params):
"""
This primitive applies the transformation encoded in the input images
WCSs to align them with a reference image, in reference image pixel
coordinates. The reference image is taken to be the first image in
the input list if not explicitly provided as a parameter.
By default, the transformation into the reference frame is done via
interpolation. The variance plane, if present, is transformed in
the same way as the science data.
The data quality plane, if present, is handled in a bitwise manner
with each bit of each pixel in the output image being set it it has
>1% influence from that bit of a bad pixel. The transformed masks are
then added back together to generate the transformed DQ plane.
The WCS objects of the output images are updated to reflect the
transformation.
Parameters
----------
suffix : str
suffix to be added to output files
order : int (0-5)
order of interpolation (0=nearest, 1=linear, etc.)
trim_data : bool
trim image to size of reference image?
clean_data : bool
replace bad pixels with a ring median of their values to avoid
ringing if using a high-order interpolation?
conserve : bool
conserve flux when resampling to a different pixel scale?
force_affine : bool
convert the true resampling transformation to an affine
approximation? This speeds up the calculation and has a negligible
effect for instruments lacking significant distortion
reference : str/AstroData/None
reference image for resampling (if not provided, the first image
in the list will be used)
"""
log = self.log
log.debug(gt.log_message("primitive", self.myself(), "starting"))
timestamp_key = self.timestamp_keys[self.myself()]
sfx = params.pop("suffix")
reference = params.pop("reference")
trim_data = params.pop("trim_data")
force_affine = params.pop("force_affine")
# These two parameters are only for GSAOI and will help to define
# the output WCS if there's no reference image
pixel_scale = params.pop("pixel_scale", None)
position_angle = params.pop("pa", None)
# TODO: Can we make it so that we don't need to mosaic detectors
# before doing this? That would mean we only do one interpolation,
# not two, and that's definitely better!
if not all(len(ad) == 1 or ad.instrument() == "GSAOI" for ad in adinputs):
raise OSError("All input images must have only one extension.")
if isinstance(reference, str):
reference = astrodata.open(reference)
elif reference is None and pixel_scale is None:
# Reference image will be the first AD, so we need 2+
if len(adinputs) < 2:
log.warning("No alignment will be performed, since at least "
"two input AstroData objects are required for "
"resampleToCommonFrame")
return adinputs
if reference is None and pixel_scale:
# This must be GSAOI projecting to the requested geometry
ad0 = adinputs[0]
ra, dec = ad0.target_ra(), ad0.target_dec()
# using SkyCoord facilitates formatting the log
center = SkyCoord(ra * u.deg, dec * u.deg)
ra_str = center.ra.to_string(u.hour, precision=3)
dec_str = center.dec.to_string(u.deg, precision=2, alwayssign=True)
log.stdinfo(f"Projecting with center {ra_str} {dec_str}\n"
f"at PA={position_angle} with pixel scale={pixel_scale}")
pixel_scale /= 3600
new_wcs = (models.Scale(-pixel_scale) & models.Scale(pixel_scale) |
models.Rotation2D(position_angle) |
models.Pix2Sky_TAN() |
models.RotateNative2Celestial(ra, dec, 180))
ref_wcs = gWCS([(ad0[0].wcs.input_frame, new_wcs),
(ad0[0].wcs.output_frame, None)])
if trim_data:
log.warning("Setting trim_data=False as required when no "
"reference imagevis provided.")
trim_data = False
else:
if reference is None:
reference = adinputs[0]
else:
log.stdinfo(f"Using {reference.filename} as reference image")
if not trim_data:
log.warning("Setting trim_data=True to trim to size of the "
"reference image.")
trim_data = True
if len(reference) != 1:
raise OSError("Reference image must have only one extension.")
ref_wcs = reference[0].wcs
if trim_data:
params.update({'origin': (0,) * len(reference[0].shape),
'output_shape': reference[0].shape})
# No transform for the reference AD
for ad in adinputs:
transforms = []
if reference is ad:
transforms.append(models.Identity(len(ad[0].shape)))
else:
for ext in ad:
t_align = ext.wcs.forward_transform | ref_wcs.backward_transform
if force_affine:
affine = adwcs.calculate_affine_matrices(t_align, ext.shape)
t_align = models.AffineTransformation2D(matrix=affine.matrix[::-1, ::-1],
translation=affine.offset[::-1])
transforms.append(t_align)
for ext, t_align in zip(ad, transforms):
resampled_frame = copy(ext.wcs.input_frame)
resampled_frame.name = "resampled"
ext.wcs = gWCS([(ext.wcs.input_frame, t_align),
(resampled_frame, ref_wcs.pipeline[0].transform)] +
ref_wcs.pipeline[1:])
adoutputs = self._resample_to_new_frame(adinputs, frame="resampled",
process_objcat=False, **params)
for ad in adoutputs:
try:
trans_data = ad.nddata[0].meta.pop('transform')
except KeyError:
pass
else:
corners = np.array(trans_data['corners'][0])
ncorners = len(corners)
ad.hdr["AREATYPE"] = (f"P{ncorners}",
f"Region with {ncorners} vertices")
for i, corner in enumerate(zip(*corners), start=1):
for axis, value in enumerate(reversed(corner), start=1):
key_name = f"AREA{i}_{axis}"
key_comment = f"Vertex {i}, dimension {axis}"
ad.hdr[key_name] = (value + 1, key_comment)
jfactor = trans_data['jfactors'][0]
ad.hdr["JFACTOR"] = (jfactor, "J-factor in resampling")
ad.update_filename(suffix=sfx, strip=True)
gt.mark_history(ad, primname=self.myself(), keyword=timestamp_key)
return adoutputs
def applyStackedObjectMask(self, adinputs=None, **params):
"""
This primitive takes an image with an OBJMASK and transforms that
OBJMASK onto the pixel planes of the input images, using their WCS
information. If the first image is a stack, this allows us to mask
fainter objects than can be detected in the individual input images.
Parameters
----------
suffix: str
suffix to be added to output files
source: str
name of stream containing single stacked image
order: int (0-5)
order of interpolation
threshold: float
threshold above which an interpolated pixel should be flagged
"""
log = self.log
log.debug(gt.log_message("primitive", self.myself(), "starting"))
source = params["source"]
order = params["order"]
threshold = params["threshold"]
sfx = params["suffix"]
force_affine = True
try:
source_stream = self.streams[source]
except KeyError:
try:
ad_source = astrodata.open(source)
except:
log.warning(f"Cannot find stream or file named {source}. Continuing.")
return adinputs
else:
if len(source_stream) != 1:
log.warning(f"Stream {source} does not contain single "
"AstroData object. Continuing.")
return adinputs
ad_source = source_stream[0]
# There's no reason why we can't handle multiple extensions
if any(len(ad) != len(ad_source) for ad in adinputs):
log.warning("At least one AstroData input has a different number "
"of extensions to the reference. Continuing.")
return adinputs
for ad in adinputs:
for ext, source_ext in zip(ad, ad_source):
if getattr(ext, 'OBJMASK') is not None:
t_align = source_ext.wcs.forward_transform | ext.wcs.backward_transform
if force_affine:
affine = adwcs.calculate_affine_matrices(t_align.inverse, ad[0].shape)
objmask = affine_transform(source_ext.OBJMASK.astype(np.float32),
affine.matrix, affine.offset,
output_shape=ext.shape, order=order,
cval=0)
else:
objmask = transform.Transform(t_align).apply(source_ext.OBJMASK.astype(np.float32),
output_shape=ext.shape, order=order,
cval=0)
ext.OBJMASK = np.where(abs(objmask) > threshold, 1, 0).astype(np.uint8)
# We will deliberately keep the input image's OBJCAT (if it
# exists) since this will be required for aligning the inputs.
ad.update_filename(suffix=sfx, strip=True)
return adinputs
def resampleToCommonFrame(self, adinputs=None, **params):
"""
This primitive applies the transformation encoded in the input images
WCSs to align them with a reference image, in reference image pixel
coordinates. The reference image is taken to be the first image in
the input list.
By default, the transformation into the reference frame is done via
interpolation. The variance plane, if present, is transformed in
the same way as the science data.
The data quality plane, if present, is handled in a bitwise manner
with each bit of each pixel in the output image being set it it has
>1% influence from that bit of a bad pixel. The transformed masks are
then added back together to generate the transformed DQ plane.
The WCS keywords in the headers of the output images are updated
to reflect the transformation.
Parameters
----------
suffix: str
suffix to be added to output files
order: int (0-5)
order of interpolation (0=nearest, 1=linear, etc.)
trim_data: bool
trim image to size of reference image?
clean_data: bool
replace bad pixels with a ring median of their values to avoid
ringing if using a high-order interpolation?
"""
log = self.log
log.debug(gt.log_message("primitive", self.myself(), "starting"))
timestamp_key = self.timestamp_keys[self.myself()]
sfx = params.pop("suffix")
force_affine = True
if len(adinputs) < 2:
log.warning("No alignment will be performed, since at least two "
"input AstroData objects are required for "
"resampleToCommonFrame")
return adinputs
# TODO: Can we make it so that we don't need to mosaic detectors
# before doing this? That would mean we only do one interpolation,
# not two, and that's definitely better!
if not all(len(ad) == 1 for ad in adinputs):
raise OSError("All input images must have only one extension.")
ad_ref = adinputs[0]
ndim = len(ad_ref[0].shape)
# No transform for the reference AD
for i_ad, ad in enumerate(adinputs):
if i_ad == 0:
ref_wcs = ad[0].wcs
t_align = models.Identity(ndim)
else:
t_align = ad[0].wcs.forward_transform | ref_wcs.backward_transform
if force_affine:
affine = adwcs.calculate_affine_matrices(t_align, ad[0].shape)
t_align = models.AffineTransformation2D(matrix=affine.matrix[::-1, ::-1],
translation=affine.offset[::-1])
resampled_frame = copy(ad[0].wcs.input_frame)
resampled_frame.name = "resampled"
ad[0].wcs = gWCS([(ad[0].wcs.input_frame, t_align),
(resampled_frame, ref_wcs.pipeline[0].transform)] +
ref_wcs.pipeline[1:])
adoutputs = self._resample_to_new_frame(adinputs, frame="resampled",
process_objcat=False, **params)
for ad in adoutputs:
try:
trans_data = ad.nddata[0].meta.pop('transform')
except KeyError:
pass
else:
corners = np.array(trans_data['corners'][0])
ncorners = len(corners)
ad.hdr["AREATYPE"] = (f"P{ncorners}",
f"Region with {ncorners} vertices")
for i, corner in enumerate(zip(*corners), start=1):
for axis, value in enumerate(reversed(corner), start=1):
key_name = f"AREA{i}_{axis}"
key_comment = f"Vertex {i}, dimension {axis}"
ad.hdr[key_name] = (value + 1, key_comment)
jfactor = trans_data['jfactors'][0]
ad.hdr["JFACTOR"] = (jfactor, "J-factor in resampling")
ad.update_filename(suffix=sfx, strip=True)
gt.mark_history(ad, primname=self.myself(), keyword=timestamp_key)
return adoutputs