plenoptic.process.shrink

plenoptic.process.shrink(x, factor)

Shrink a signal by a factor.

We do this in the frequency domain: cropping out the center of the Fourier transform of x, putting it in a new tensor, and taking the IFFT.

Parameters:

• x (Tensor) – The signal for expansion.

• factor (int) – Factor by which to resize image. Must be larger than 1 and factor / x.shape[-2:] must give integer values.

Return type:

Tensor

Returns:

expanded – The expanded signal.

Raises:

• ValueError – If factor is less than or equal to 1.

• ValueError – If the height or width of x divided by factor is not an integer.

See also

expand

The inverse operation.

blur_downsample

An alternative downsampling operation.

Examples

>>> import plenoptic as po
>>> import torch
>>> img = po.data.einstein()
>>> img.shape
torch.Size([1, 1, 256, 256])
>>> shrunk = po.process.shrink(img, factor=2)
>>> shrunk.shape
torch.Size([1, 1, 128, 128])
>>> po.plot.imshow([img, shrunk], title=["original", "shrunk"])
<PyrFigure...>

(png, hires.png, pdf)

Note the horizontal/vertical lines in the shrunk version of the image. These are the result of aliasing. To avoid these, use blur_downsample:

>>> po.plot.imshow(
...     [img, shrunk, po.process.blur_downsample(img)],
...     title=["original", "shrunk", "blurred"],
... )
<PyrFigure...>

(png, hires.png, pdf)

You can invert shrink using expand, but the inversion is not perfect; shrinking discards information that can not be recovered:

>>> expand_after_shrink = po.process.expand(
...     po.process.shrink(img, factor=2), factor=2
... )
>>> torch.allclose(img, expand_after_shrink, atol=1e-2)
False
>>> po.plot.imshow(
...     [img, expand_after_shrink],
...     title=["original", "expand after shrink"],
... )
<PyrFigure...>

(png, hires.png, pdf)

Even in the opposite order, i.e., shrinking an expanded image, the inversion is not perfect. In this example with pixel values between 0 and 1, there are differences on the order of 1e-3:

>>> shrink_after_expand = po.process.shrink(
...     po.process.expand(img, factor=2), factor=2
... )
>>> torch.allclose(img, shrink_after_expand, atol=1e-2)
True
>>> torch.allclose(img, shrink_after_expand, atol=1e-3)
False
>>> po.plot.imshow(
...     [img, shrink_after_expand, img - shrink_after_expand],
...     title=["original", "shrink after expand", "difference"],
... )
<PyrFigure...>

(png, hires.png, pdf)