BlindRLIteration

class deepinv.optim.optim_iterators.BlindRLIteration(k_prior=None, normalize_kernel=True, use_fft=False, eps=1e-8, **kwargs)

Bases: OptimIterator

Iterator for Blind Richardson-Lucy deconvolution.

This iterator performs one step to estimate the next kernel, and one step to estimate the next image.

The current iterate is stored as X["est"] = (x, k). The kernel update assumes 2D circular convolution and a spatially invariant kernel shared by all image channels.

Parameters:

• k_prior (deepinv.optim.Prior, None) – optional kernel prior. Default: None.

• normalize_kernel (bool) – whether to normalize the kernel to unit sum. Default: True.

• use_fft (bool) – whether to use the FFT implementations for convolutions. Default: False.

• eps (float) – numerical stability constant used for divisions. Default: 1e-8.

forward(X, cur_data_fidelity, cur_prior, cur_params, y, physics, *args, **kwargs)

Single Blind Richardson-Lucy iteration.

Parameters:

• X (dict[str, tuple[torch.Tensor, torch.Tensor] | torch.Tensor]) – Current iterate with X["est"] = (x, k).

• cur_data_fidelity (deepinv.optim.DataFidelity) – Data fidelity term.

• cur_prior (deepinv.optim.Prior) – Image prior.

• cur_params (dict) – Parameters containing x_steps, k_steps, lambda_reg_x, lambda_reg_k, g_param and g_param_kernel.

• y (torch.Tensor) – Blurry observation of shape (B, C, H, W).

• physics (deepinv.physics.Physics) – Blur physics updated in-place with the current kernel for the image update.

Returns:

Dictionary {"est": (x, k), "cost": F, "it": it} containing the updated image, kernel, cost, and iteration number.

Return type:

dict[str, tuple[torch.Tensor, torch.Tensor] | torch.Tensor]