DScCP

class deepinv.models.DScCP(depth=20, n_channels_per_layer=64, pretrained='download', pretrained_2d_isotropic=False, device=None, dim=2)

Bases: Denoiser

DScCP denoiser network.

The network architecture is based on the paper from Le et al.[1]. and has an unrolled architecture based on the fast Chambolle-Pock algorithm using strong convexity. DScCP stands for Deep Strongly Convex Chambolle Pock.

The pretrained weights are trained with the default parameters of the network, i.e. depth=20 layers, n_channels_per_layer=64 channels. They can be downloaded via setting pretrained='download'.

Parameters:

• depth (int) – depth i.e. number of convolutional layers.

• n_channels_per_layer (int) – number of channels per convolutional layer.

• pretrained (str, None) – pretrained='download' to download pretrained weights, or path to local weights file. When building a 3D network, it is possible to initialize with 2D pretrained weights by using pretrained='download_2d', which provides a good starting point for fine-tuning.

• pretrained_2d_isotropic (bool) – when loading 2D pretrained weights into a 3D network, whether to initialize the 3D kernels isotropically. By default the weights are loaded axially, i.e., by initializing the central slice of the 3D kernels with the 2D weights.

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

• dim (str, int) – Whether to build 2D or 3D network (if str, can be “2”, “2d”, “3D”, etc.)

---

References:

[1]

Hoang Trieu Vy Le, Audrey Repetti, and Nelly Pustelnik. Unfolded proximal neural networks for robust image gaussian denoising. IEEE Transactions on Image Processing, 2024.

forward(x, sigma=0.03)

Run the denoiser on noisy image.

Parameters:

• x (torch.Tensor) – noisy image.

• sigma (float) – noise level.