Current browse context:
eess.IV
Change to browse by:
References & Citations
Electrical Engineering and Systems Science > Image and Video Processing
Title: Stochastic Primal-Dual Deep Unrolling
(Submitted on 19 Oct 2021 (v1), last revised 15 Feb 2022 (this version, v4))
Abstract: We propose a new type of efficient deep-unrolling networks for solving imaging inverse problems. Conventional deep-unrolling methods require full forward operator and its adjoint across each layer, and hence can be significantly more expensive computationally as compared with other end-to-end methods that are based on post-processing of model-based reconstructions, especially for 3D image reconstruction tasks. We develop a stochastic (ordered-subsets) variant of the classical learned primal-dual (LPD), which is a state-of-the-art unrolling network for tomographic image reconstruction. The proposed learned stochastic primal-dual (LSPD) network only uses subsets of the forward and adjoint operators and offers considerable computational efficiency. We provide theoretical analysis of a special case of our LSPD framework, suggesting that it has the potential to achieve image reconstruction quality competitive with the full-batch LPD while requiring only a fraction of the computation. The numerical results for two different X-ray computed tomography (CT) imaging tasks (namely, low-dose and sparse-view CT) corroborate this theoretical finding, demonstrating the promise of LSPD networks for large-scale imaging problems.
Submission history
From: Junqi Tang [view email][v1] Tue, 19 Oct 2021 16:46:03 GMT (933kb,D)
[v2] Fri, 19 Nov 2021 14:54:10 GMT (1190kb,D)
[v3] Tue, 25 Jan 2022 16:42:13 GMT (4066kb,D)
[v4] Tue, 15 Feb 2022 11:10:56 GMT (4063kb,D)
Link back to: arXiv, form interface, contact.