Tutorial

A simple usage example

Suppose we have two images of about the same portion of the sky, and we would like to transform one of them to fit on top of the other. Suppose we do not have WCS information, but we are confident that we could do it by eye, by matching some obvious asterisms on the two images.

In this particular use case, astroalign can be of great help to automatize the process.

After we load our images into numpy arrays, we simple choose one to be the source image to be transformed, and the other to be the target.

The usage for this simple most common case would be as follows:

>>> import astroalign as aa
>>> registered_image, footprint = aa.register(source, target)

registered_image is now a transformed (numpy array) image of source that will match pixel to pixel to target.

footprint is a boolean numpy array, True for masked pixels with no information.

Note

Warning

Flux may not be conserved after the transformation.

Note

If your image requires special care see Examples.

Images with RGB channels

Astroalign can work with color images provided the channel index be the last axis in the array. Adding the channel dimension in the last axis of the array is the default behavior for pillow and scikit-image. The transformation is found on the mean average of all the channels. PNG images with RGBA channels work similarly.

Example:

from PIL import Image
import astroalign as aa
source = Image.open("source.jpg")
target = Image.open("target.jpg")
registered, footprint = aa.register(source, target)
# Convert back to pillow image if necessary:
registered = Image.fromarray(registered.astype("unit8"))

Pillow may require array to be unsigned 8-bit integer format.

Mask Fill Value

If you need to mask the aligned image with a special value over the region where transformation had no pixel information, you can use the footprint mask to do so:

>>> registered_image, footprint = aa.register(source, target)
>>> registered_image[footprint] = -99999.99

Or you can pass the value to the fill_value argument:

>>> registered_image, footprint = aa.register(source, target, fill_value=-99999.99)

Both will yield the same result.

Finding the transformation

In some cases it may be necessary to inspect first the transformation parameters before applying it, or we may be interested only in a star to star correspondence between the images. For those cases, we can use find_transform:

>>> transf, (source_list, target_list) = aa.find_transform(source, target)

The inputs source and target can be either numpy arrays of the image pixels, or any iterable of (x, y) pairs, corresponding to star positions.

Having an iterable of (x, y) pairs is especially useful in situations where source detection requires special care. In situations like that, source detection can be done separately and the resulting catalogs fed to find_transform.

find_transform returns a scikit-image SimilarityTransform object that encapsulates the matrix transformation, and the transformation parameters. It will also return a tuple with two lists of star positions of source and its corresponding ordered star postions on the target image.

The transformation parameters can be found in transf.rotation, transf.translation, transf.scale and the transformation matrix in transf.params.

If the transformation is satisfactory, we can apply it to the image with apply_transform. Continuing our example:

>>> if transf.rotation > MIN_ROT:
...     registered_image = aa.apply_transform(transf, source, target)

If you know the star-to-star correspondence

Note

estimate_transform from scikit-image is imported into astroalign as a convenience.

If for any reason you know which star corresponds to which other, you can call estimate_transform.

Let us suppose we know the correspondence:

  • (127.03, 85.98) in source –> (175.13, 111.36) in target

  • (23.11, 31.87) in source –> (0.58, 119.04) in target

  • (98.84, 142.99) in source –> (181.55, 206.49) in target

  • (150.93, 85.02) in source –> (205.60, 91.89) in target

  • (137.99, 12.88) in source –> (134.61, 7.94) in target

Then we can estimate the transform:

>>> src = np.array([(127.03, 85.98), (23.11, 31.87), (98.84, 142.99),
...                 (150.93, 85.02), (137.99, 12.88)])
>>> dst = np.array([(175.13, 111.36), (0.58, 119.04), (181.55, 206.49),
...                 (205.60, 91.89), (134.61, 7.94)])
>>> tform = aa.estimate_transform('affine', src, dst)

And apply it to an image with apply_transform or to a set of points with matrix_transform.

Applying a transformation to a set of points

Note

matrix_transform from scikit-image is imported into astroalign as a convenience.

To apply a known transform to a set of points, we use matrix_transform. Following the example in the previous section:

>>> dst_calc = aa.matrix_transform(src, tform.params)

dst_calc should be a 5 by 2 array similar to the dst array.

Objects with data and mask property

If your image is stored in objects with data and mask properties, such as ccdproc’s CCDData or astropy’s NDData or a NumPy masked array you can use them as input for register, find_transform and apply_transform.

In general in these cases it is convenient to transform their masks along with the data and to add the footprint onto the mask.

Astroalign provides this functionality with the propagate_mask argument to register and apply_transform.

For example:

>>> from astropy.nddata import NDData
>>> nd = NDData([[0, 1], [2, 3]], [[True, False], [False, False]])

and we want to apply a clockwise 90 degree rotation:

>>> import numpy as np
>>> from skimage.transform import SimilarityTransform
>>> transf = SimilarityTransform(rotation=np.pi/2., translation=(1, 0))

Then we can call astroalign as usual, but with the propagate_mask set to True:

>>> aligned_image, footprint = aa.apply_transform(transf, nd, nd, propagate_mask=True)

This will transform nd.data and nd.mask simultaneously and add the footprint mask from the transformation onto nd.mask:

>>> aligned_image
array([[2., 0.],
   [3., 1.]])
>>> footprint
array([[False,  True],
   [False, False]])

Creating a new object of the same input type is now easier:

>>> new_nd = NDData(aligned_image, mask=footprint)

The same will apply for CCDData objects and NumPy masked arrays.


See Module API for the API specification.