BlurFFT#

class deepinv.physics.BlurFFT(img_size, filter=None, device='cpu', **kwargs)[source]#

Bases: DecomposablePhysics

FFT-based blur operator.

It performs the operation

\[y = w*x\]

where \(*\) denotes convolution and \(w\) is a filter.

Blur operator based on torch.fft operations, which assumes a circular padding of the input, and allows for the singular value decomposition via deepinv.Physics.DecomposablePhysics and has fast pseudo-inverse and prox operators.

Parameters:

img_size (tuple) – Input image size in the form (C, H, W).
filter (torch.Tensor) – torch.Tensor of size (1, c, h, w) containing the blur filter with h<=H, w<=W and c=1 or c=C e.g., deepinv.physics.blur.gaussian_blur().
device (torch.device, str) – Device on which the physics’ buffers will be created. If a buffer is updated via physics.update_parameters(), if not None, it will be automatically casted to the device of the replaced buffer, else, use the device of the provided value. To change the device of all buffers, please use physics.to(device).

Examples:

BlurFFT operator with a basic averaging filter applied to a 16x16 black image with a single white pixel in the center:

>>> from deepinv.physics import BlurFFT
>>> x = torch.zeros((1, 1, 16, 16)) # Define black image of size 16x16
>>> x[:, :, 8, 8] = 1 # Define one white pixel in the middle
>>> filter = torch.ones((1, 1, 2, 2)) / 4 # Basic 2x2 filter
>>> physics = BlurFFT(filter=filter, img_size=(1, 16, 16))
>>> y = physics(x)
>>> y[y<1e-5] = 0.
>>> y[:, :, 7:10, 7:10] # Display the center of the blurred image
tensor([[[[0.2500, 0.2500, 0.0000],
          [0.2500, 0.2500, 0.0000],
          [0.0000, 0.0000, 0.0000]]]])

static get_filter_parameters(img_size, filter, device='cpu')[source]#

Create filter parameters for BlurFFT operator.

Parameters:

img_size (tuple[int]) – size of the input image (C, H, W).
filter (torch.Tensor) – Filter to be applied to the input image.
device (torch.device, str) – Device where the filter tensor will be created.