Introduction#

Reconstruction algorithms define an inversion function \(\hat{x}=\inversef{y}{A}\) which recovers a signal \(x\) from measurements \(y\) given an operator \(A\).

x_hat = model(y, physics)

Defining your own reconstructor#

All reconstruction algorithms inherit from the deepinv.models.Reconstructor base class, take as input measurements y and forward operator physics, and output a reconstruction x_hat.

To use your own reconstructor with DeepInverse, simply define the forward method to follow this pattern.

Summary#

Below we provide a summary of existing reconstruction methods, and a qualitative description of their reconstruction performance and speed.

For the models that require training, you can do this using the trainer and loss functions.

Table 6 Reconstruction methods#
Family of methods	Description	Requires Training	Iterative	Sampling
Deep Reconstruction Models	Deep model architectures for reconstruction.	No if pretrained, yes otherwise	No	No
Plug-and-Play (PnP)	Leverages pretrained denoisers as priors within an optimisation algorithm.	No	Yes	No
Unfolded Networks	Constructs a trainable architecture by unrolling a PnP algorithm.	Yes	Only `DEQ`	No
Diffusion	Leverages pretrained denoisers within a ODE/SDE.	No	Yes	Yes
Non-learned priors	Solves an optimization problem with hand-crafted priors.	No	Yes	No
Markov Chain Monte Carlo	Leverages pretrained denoisers as priors within an optimisation algorithm.	No	Yes	Yes
Generative Adversarial Networks and Deep Image Prior	Uses a generator network to model the set of possible images.	No	Yes	Depends
Multi-physics models	Models trained on multiple various physics and datasets for robustness to different problems.	No	No	No

Note

Some algorithms might be better at reconstructing images with good perceptual quality (e.g. diffusion methods) whereas other methods are better at reconstructing images with low distortion (close to the ground truth).

Introduction#

Defining your own reconstructor#

Summary#

This Page