BilateralFilter#

class deepinv.models.BilateralFilter(device='cpu')[source]#

Bases: Denoiser

Bilateral filter.

The bilateral filter as it was introduced in Tomasi and Manduchi[1] where each output pixel is a normalized weighted average of neighboring pixels in a spatial window. The weights factor into a spatial kernel and a range kernel:

\[\hat{x}_i = (1 / W_i) * \sum_{j \in \omega(i)} k_d(i - j) k_r(x_i - x_j) x_j\]

with Gaussian spatial kernel \(k_d\) and Gaussian range kernel \(k_d\). The spatial standard deviation \(\sigma_d\) controls how fast the kernel decays with distance, while the range standard deviation \(\sigma_r\) controls how strongly intensity differences are penalized.

Example:

>>> import torch
>>> import deepinv as dinv
>>> x = dinv.utils.load_example("butterfly.png")
>>> rng = torch.Generator().manual_seed(0)
>>> physics = dinv.physics.Denoising(dinv.physics.GaussianNoise(rng=rng))
>>> y = physics(x)
>>> model = dinv.models.BilateralFilter()
>>> x_hat = model(y,sigma_d=1.1,sigma_r=0.3,window_size=9)
>>> dinv.metric.PSNR()(x_hat, x) > 26
tensor([True])

References:

forward(x, sigma=None, *, sigma_d=1, sigma_r=1, window_size=5, **kwargs)[source]#

Apply a bilateral filter to the input x.

Parameters:

  • x (torch.Tensor) – Input tensor of shape (B, C, H, W) or (C, H, W).

  • sigma_d (float) – Spatial standard deviation (spatial domain). Larger values yield larger effective receptive fields.

  • sigma_r (float) – Range standard deviation (intensity domain). Larger values yield stronger smoothing across edges.

  • window_size (int) – Size of the (square) spatial window. Must be an odd integer.

Returns:

Filtered image.

Return type:

Tensor