1: \begin{abstract}

2: Iterative neural networks (INN) are rapidly gaining attention 

3: for solving inverse problems in imaging, image processing, and computer vision.

4: INNs combine regression NNs and an iterative model-based image reconstruction (MBIR) algorithm, 

5: often leading to both good generalization capability and 

6: outperforming reconstruction quality over existing MBIR optimization models.

7: This paper proposes the first fast and convergent INN architecture, \emph{Momentum-Net},

8: by generalizing a block-wise MBIR algorithm that uses momentum and majorizers with regression NNs.

9: For fast MBIR, Momentum-Net uses \emph{momentum} terms in extrapolation modules,

10: and noniterative MBIR modules at each iteration by using \emph{majorizers},

11: where each iteration of Momentum-Net consists of three core modules:

12: image refining, extrapolation, and MBIR.

13: Momentum-Net guarantees convergence to a fixed-point

14: for general differentiable (non)convex MBIR functions (or data-fit terms) and convex feasible sets, 

15: under two asymptomatic conditions.

16: To consider data-fit variations across training and testing samples, 

17: we also propose a regularization parameter selection scheme 

18: based on the \dquotes{spectral spread} of majorization matrices.

19: Numerical experiments for light-field photography using a focal stack and sparse-view computational tomography 

20: demonstrate that, given identical regression NN architectures,

21: Momentum-Net significantly improves MBIR speed and accuracy over several existing INNs;

22: it significantly improves reconstruction quality 

23: compared to a state-of-the-art MBIR method in each application.  

24: \end{abstract}

25: