SpaceVaryingBlur#

class deepinv.physics.SpaceVaryingBlur(filters=None, multipliers=None, padding='valid', device='cpu', **kwargs)[source]#

Bases: LinearPhysics

Space varying blur via product-convolution.

This linear operator performs

\[y = \sum_{k=1}^K h_k \star (w_k \odot x)\]

where \(\star\) is a convolution, \(\odot\) is a Hadamard product, \(w_k\) are multipliers \(h_k\) are filters.

Parameters:

  • filters (torch.Tensor) – Filters \(h_k\). Tensor of size (B, C, K, h, w) where B is the batch size, C the number of channels, K the number of filters, h and w the filter height and width which should be smaller or equal than the image \(x\) height and width respectively.

  • multipliers (torch.Tensor) – Multipliers \(w_k\). Tensor of size (B, C, K, H, W) where B is the batch size, C the number of channels, K the number of multipliers, H and W the image \(x\) height and width.

  • padding (str) – options = 'valid', 'circular', 'replicate', 'reflect'. If padding = 'valid' the blurred output is smaller than the image (no padding), otherwise the blurred output has the same size as the image.

  • device (str, torch.device) – cpu or cuda

Examples:

We show how to instantiate a spatially varying blur operator.

>>> from deepinv.physics.generator import DiffractionBlurGenerator, ProductConvolutionBlurGenerator
>>> from deepinv.physics.blur import SpaceVaryingBlur
>>> from deepinv.utils.plotting import plot
>>> psf_size = 32
>>> img_size = (256, 256)
>>> delta = 16
>>> psf_generator = DiffractionBlurGenerator((psf_size, psf_size))
>>> pc_generator = ProductConvolutionBlurGenerator(psf_generator=psf_generator, img_size=img_size)
>>> params_pc = pc_generator.step(1)
>>> physics = SpaceVaryingBlur(**params_pc)
>>> dirac_comb = torch.zeros(img_size).unsqueeze(0).unsqueeze(0)
>>> dirac_comb[0,0,::delta,::delta] = 1
>>> psf_grid = physics(dirac_comb)
>>> plot(psf_grid, titles="Space varying impulse responses")
A(x, filters=None, multipliers=None, padding=None, **kwargs)[source]#

Applies the space varying blur operator to the input image.

It can receive new parameters \(w_k\), \(h_k\) and padding to be used in the forward operator, and stored as the current parameters.

Parameters:

  • filters (torch.Tensor) – Filters \(h_k\). Tensor of size (b, c, K, h, w). \(b \in \{1, B\}\) and \(c \in \{1, C\}\), \(h\leq H\) and \(w\leq W\).

  • multipliers (torch.Tensor) – Multipliers \(w_k\). Tensor of size (b, c, K, H, W). \(b \in \{1, B\}\) and \(c \in \{1, C\}\)

  • padding – options = 'valid', 'circular', 'replicate', 'reflect'. If padding = 'valid' the blurred output is smaller than the image (no padding), otherwise the blurred output has the same size as the image.

  • device (str) – cpu or cuda

A_adjoint(y, filters=None, multipliers=None, padding=None, **kwargs)[source]#

Applies the adjoint operator.

It can receive new parameters \(w_k\), \(h_k\) and padding to be used in the forward operator, and stored as the current parameters.

Parameters:

  • h (torch.Tensor) – Filters \(h_k\). Tensor of size (b, c, K, h, w). \(b \in \{1, B\}\) and \(c \in \{1, C\}\), \(h\leq H\) and \(w\leq W\).

  • w (torch.Tensor) – Multipliers \(w_k\). Tensor of size (b, c, K, H, W). \(b \in \{1, B\}\) and \(c \in \{1, C\}\)

  • padding – options = 'valid', 'circular', 'replicate', 'reflect'. If padding = 'valid' the blurred output is smaller than the image (no padding), otherwise the blurred output has the same size as the image.

  • device (str) – cpu or cuda

update_parameters(filters=None, multipliers=None, padding=None, **kwargs)[source]#

Updates the current parameters.

Parameters:

  • filters (torch.Tensor) – Filters \(h_k\). Tensor of size (b, c, K, h, w). \(b \in \{1, B\}\) and \(c \in \{1, C\}\), \(h\leq H\) and \(w\leq W\).

  • multipliers (torch.Tensor) – Multipliers \(w_k\). Tensor of size (b, c, K, H, W). \(b \in \{1, B\}\) and \(c \in \{1, C\}\)

  • padding (str) – options = 'valid', 'circular', 'replicate', 'reflect'.

Examples using SpaceVaryingBlur:#

Tour of blur operators

Tour of blur operators