Blind Inverse Problems#

Following the notation of the library, here we consider measurements of the form \(y = \noise{\forw{x, \theta}}\), where \(\theta\) represents unknown physics parameters. Noise parameters associated to \(\noise{\cdot}\) may also be unknown. In this section, we consider two classes of problems:

  • Calibration problems: Estimate the unknown parameters \(\theta\) given paired signal and measurement data \((x,y)\)

  • Blind inverse problems: Jointly estimate the signal \(x\) and \(\theta\) parameters (and other noise parameters) from the measurements \(y\). Some methods directly estimate the signal without explicitly estimating the parameters.

Calibration problems#

If paired measurement and signal data is available at inference time, physics parameters can be estimated using optimization methods. See the example Calibrating physics operators for more details.

Physics parameters estimation#

If only measurement data is available \(\theta\) at inference time, we can estimate the parameters from the observed data, and then use any non-blind reconstructor to recover the image. The library provides the following parameter estimation models/algorithms:

Table 24 Identification models#

Model/Algorithm

Tensor Size (C, H, W)

Pretrained Weights

Physics

Parameters estimated

Examples

KernelIdentificationNetwork

C=3; H,W>8

RGB

SpaceVaryingBlur

filters, multipliers

blind deblurring.

ESPIRiT

C=2; H,W>64

(non-learned)

MultiCoilMRI

coil_maps

MRI coil map estimation.