Deep Reconstruction Models

The simplest method for reconstructing an image from measurements is to pass it through a feedforward model architecture that is conditioned on the acquisition physics, that is \(\inversef{y}{A}\). We offer a range of architectures for general and specific problems.

Artifact Removal

The simplest reconstruction architecture first maps the measurements to the image domain via a non-learned mapping, and then applys a denoiser network to the obtain the final reconstruction.

The deepinv.models.ArtifactRemoval class converts a denoiser deepinv.models.Denoiser or other image-to-image network \(\phi\) into a reconstruction network deepinv.models.Reconstructor \(R\) by doing

• Adjoint: \(\inversef{y}{A}=\phi(A^{\top}y)\) with mode='adjoint'.

  This option is generally to linear operators \(A\).

• Pseudoinverse: \(\inversef{y}{A}=\phi(A^{\dagger}y)\) with mode='pinv'.

• Direct: \(\inversef{y}{A}=\phi(y)\) with mode='direct'.

  This option serves as a wrapper to obtain a Reconstructor, and can be used to adapt a generic denoiser or image-to-image network into one that is specific to an inverse problem.

General reconstruction models

We provide the following list of reconstruction models trained on multiple various physics and datasets to provide robustness to different problems.

See Description of weights for more information on pretrained denoisers.

Table 11 Multiphysics reconstruction models

Model	Type	Tensor Size (C, H, W)	Pretrained Weights	Noise level aware
deepinv.models.RAM	CNN-UNet	C=1, 2, 3; H,W>8	C=1, 2, 3	Yes

Specific reconstruction models

We also provide some architectures for specific inverse problems.

Table 12 Specific architectures

Model	Description
deepinv.models.PanNet	PanNet model for pansharpening.