MotionBlurGenerator#

class deepinv.physics.generator.MotionBlurGenerator(psf_size: tuple, num_channels: int = 1, rng: Generator | None = None, device: str = 'cpu', dtype: type = torch.float32, l: float = 0.3, sigma: float = 0.25, n_steps: int = 1000)[source]#

Bases: PSFGenerator

Random motion blur generator.

See https://arxiv.org/pdf/1406.7444.pdf for more details.

A blur trajectory is generated by sampling both its x- and y-coordinates independently from a Gaussian Process with a Matérn 3/2 covariance function.

\[f_x(t), f_y(t) \sim \mathcal{GP}(0, k(t, t'))\]

where \(k\) is defined as

\[k(t, s) = \sigma^2 \left( 1 + \frac{\sqrt{5} |t -s|}{l} + \frac{5 (t-s)^2}{3 l^2} \right) \exp \left(-\frac{\sqrt{5} |t-s|}{l}\right)\]
Parameters:

  • psf_size (tuple) – the shape of the generated PSF in 2D, should be (kernel_size, kernel_size)

  • num_channels (int) – number of images channels. Defaults to 1.

  • l (float) – the length scale of the trajectory, defaults to 0.3

  • sigma (float) – the standard deviation of the Gaussian Process, defaults to 0.25

  • n_steps (int) – the number of points in the trajectory, defaults to 1000

Examples:

>>> from deepinv.physics.generator import MotionBlurGenerator
>>> generator = MotionBlurGenerator((5, 5), num_channels=1)
>>> blur = generator.step()  # dict_keys(['filter'])
>>> print(blur['filter'].shape)
torch.Size([1, 1, 5, 5])
f_matern(batch_size: int = 1, sigma: float | None = None, l: float | None = None)[source]#

Generates the trajectory.

Parameters:

  • batch_size (int) – batch_size.

  • sigma (float) – the standard deviation of the Gaussian Process.

  • l (float) – the length scale of the trajectory.

Returns:

the trajectory of shape (batch_size, n_steps)

matern_kernel(diff, sigma: float | None = None, l: float | None = None)[source]#

Compute the Matérn 3/2 covariance.

Parameters:

  • diff (torch.Tensor) – the difference t - s

  • sigma (float) – the standard deviation of the Gaussian Process

  • l (float) – the length scale of the trajectory

step(batch_size: int = 1, sigma: float | None = None, l: float | None = None, seed: int | None = None)[source]#

Generate a random motion blur PSF with parameters \(\sigma\) and \(l\)

Parameters:

  • batch_size (int) – batch_size.

  • sigma (float) – the standard deviation of the Gaussian Process

  • l (float) – the length scale of the trajectory

  • seed (int) – the seed for the random number generator.

Returns:

dictionary with key ‘filter’: the generated PSF of shape (batch_size, 1, psf_size[0], psf_size[1])

Examples using MotionBlurGenerator:#

A tour of blur operators

A tour of blur operators

Imaging inverse problems with adversarial networks

Imaging inverse problems with adversarial networks