Core Concepts
A novel Moment-Aggregation (MA) framework that can reconstruct imaging signals by considering all candidate parameters of the uncertain forward model simultaneously during the training of a neural network prior.
Abstract
The content discusses the problem of inverse imaging, where the goal is to reconstruct an original image from its compressed measurements. This problem is often ill-posed, as there can be multiple interchangeably consistent solutions. The reconstruction process typically relies heavily on the parameters of the imaging (forward) model, which may not be fully known or may undergo calibration drifts. These uncertainties in the forward model create substantial challenges, as inaccurate reconstructions usually occur when the postulated parameters of the forward model do not fully match the actual ones.
To address this issue, the authors propose a novel Moment-Aggregation (MA) framework that is compatible with the popular solution of using a neural network prior. Specifically, the MA framework can reconstruct the signal by considering all candidate parameters of the forward model simultaneously during the update of the neural network. The authors provide a theoretical analysis of the MA framework, demonstrating its convergence properties and similar complexity to reconstruction under the known forward model parameters.
The authors conduct proof-of-concept experiments on two applications: compressive sensing and phase retrieval. The results show that the proposed MA framework achieves performance comparable to the forward model with the known precise parameter in reconstruction across various datasets, including MNIST, X-ray, Glas, and MoNuseg, with a PSNR gap of 0.17 to 1.94 compared to the upper bound.
Stats
The measurement y is obtained by applying a forward model with a known parameter A(x0; θ*) to the ground truth data x0, with y = A(x0; θ*) + η, where η is the measurement noise.
The set of candidate forward model parameters is Θ = {θ1, ..., θnc}, where nc is the total number of candidates.
The objective function is F(x; θi) = 1/2 || y - A(x; θi) ||_2^2, where θi ∈ Θ.
Quotes
"Inverse imaging problems (IIPs) aim to reconstruct a sought-after image x0 ∈ Rn from its measurements y ∈ Rm, where m is often much smaller than n and the observation is typically contaminated by some sort of observation noise η."
"It is worth mentioning that one key issue of IIPs is that the quality of signal reconstruction can be severely declined if the designed and implemented parameters of forward models do not match."