DScCP

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

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) – ‘download’ to download pretrained weights, or path to local weights file.

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

---

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.