spyrit.core.noise.PoissonApproxGaussSameNoise

class spyrit.core.noise.PoissonApproxGaussSameNoise(meas_op: Linear | LinearSplit | HadamSplit, alpha: float)[source]

Bases: NoNoise

Simulates measurements corrupted by identical Gaussian-approximated Poisson noise.

To accelerate the computation, we consider a Gaussian approximation to the Poisson distribution. Contrary to PoissonApproxGauss, all measurements in a batch are corrupted with the same noise sample, i.e. the noise depends only on the measurement number (and order).

Assuming incoming images \(x\) in the range [-1;1], measurements are first simulated for images in the range [0; \(\alpha\)]: \(y = \frac{\alpha}{2} P(1+x)\). Then, Gaussian noise is added: \(y + \sqrt{y} \cdot \mathcal{G}(\mu=0,\sigma^2=1)\).

The class is constructed from a measurement operator \(P\) and an image intensity \(\alpha\) that controls the noise level.

Warning

Assumes that the incoming images \(x\) are in the range [-1;1]

Args:

meas_op: Measurement operator \(H\) (see the meas submodule)

alpha (float): Image intensity (in photoelectrons)

Example 1: Using a Linear measurement operator

>>> H = torch.rand([400,32*32])
>>> meas_op = Linear(H)
>>> noise_op = PoissonApproxGaussSameNoise(meas_op, 10.0)

Example 2: Using a HadamSplit operator

>>> Perm = torch.rand([32*32,32*32])
>>> meas_op = HadamSplit(H, Perm, 32, 32)
>>> noise_op = PoissonApproxGaussSameNoise(meas_op, 200.0)

Methods

`forward`(x)	Simulates measurements corrupted by Poisson noise
`reindex`(x[, axis, inverse_permutation])	Sorts a tensor along a specified axis using the indices tensor.